Friday, 7 August 2026

AV-MSF couples 3D Gaussian Splatting (with dense 3D visual features as geometric prior) to a modal sound synthesis representation, encoding impact acoustics as compact physically-meaningful modal parameters (frequencies, dampings, gains) rather than raw waveforms or full FEM simulation. The visual geometry prior regularizes the acoustic reconstruction, enabling few-shot fitting from only a handful of impact recordings per object instead of large paired audio-visual datasets. Evaluated on two real-world datasets, presumably against physics-simulation and purely data-driven baselines for impact-sound prediction at novel contact locations. Practitioners could extend this to sim-to-real transfer for robotic tactile-audio perception or procedural foley generation conditioned on 3D scans.

Argus is a multi-role agentic runtime (Manager/Planner/Engineer/Reviewer) operating over durable project state, separating fixed user intent from mutable operational objectives, constraints, and verification criteria. Self-improvement happens entirely at the runtime/control-policy level — memories, skills, procedures, verifiers, and routing decisions are admitted only after role-owned review and task-native verification — while the underlying model weights (GPT-5.5) stay frozen. It reports ~78% on SWE-Bench Pro versus 59% for a comparison baseline across seven benchmark arenas, suggesting the persistent-state/verification architecture, not model scale, drives the gain. This is a template for building agent orchestration systems where accumulated procedural memory substitutes for fine-tuning.

The authors formalize cross-skill long-horizon tasks as multi-step problems where each step draws on a distinct reasoning skill and depends on prior step outputs, and introduce Skill Entropy as a quantitative difficulty measure for skill-transition cost within a reasoning chain. Skill²-Bench operationalizes this over 558 skills across 9 verifiable and open-ended domains, with each task tagged by a skill-entropy score and bucketed into three difficulty tiers for stratified evaluation. Eight frontier models are evaluated, presumably showing correlation between skill-entropy and performance drop. This gives practitioners a diagnostic tool and training signal (skill-entropy-aware curricula) for building or evaluating agents on genuinely compositional, long-horizon reasoning rather than single-skill benchmarks.

The paper adapts the full Nemotron retrieval stack (embedder, reranker, reader) to Modern Greek via corpus mining, synthetic query-document supervision, and fine-tuning, introducing the HERA benchmark for specialist-domain (legal, energy, financial, medical) Greek retrieval. A key finding is that parameter-free BM25 outperforms several off-the-shelf multilingual dense retrievers on Greek specialist corpora, exposing a generalization gap in current multilingual embeddings. After fine-tuning on 65,773 Greek retrieval pairs, a Nemotron 1B embedder's nDCG@10 rises from 0.362 to 0.835, with gains transferring to general-domain Greek retrieval, though BM25's edge narrows rather than fully disappears in-domain. This is a reusable recipe (corpus mining + synthetic supervision + staged fine-tuning) for bootstrapping RAG in other underserved languages.

SmartMage is a multimodal LLM for 3D scene understanding that replaces fixed modality fusion with a Semantic-guided Modality Adaptive Routing (SMART) module, which uses semantic priors from the query to select task-relevant modalities (visual, geometric, etc.) per-instance rather than concatenating all inputs uniformly. This targets the problem of modality noise/dilution in existing MLLMs that fuse heterogeneous 3D inputs indiscriminately, aiming to reduce both computational overhead and reasoning degradation from irrelevant channels. The abstract implies benchmarking against fixed-modality-combination MLLM baselines on 3D scene understanding tasks. Practitioners building embodied-AI or 3D-QA systems could adopt the routing-module idea as a plug-in gate before multimodal fusion layers.

The paper analyzes factored models W=UV^T and the gauge symmetry (U,V)↦(UQ,VQ) that leaves the loss invariant, proving gauge-equivariance is necessary (but not sufficient) for an optimizer to inherit gradient flow's implicit low-rank bias. They show gradient descent, momentum, 'shared-scalar' Adam, Muon, and Shampoo are gauge-equivariant, while Adam and RMSProp's coordinate-wise preconditioning breaks equivariance and thus the low-rank bias transfer. A structure theorem characterizes memoryless equivariant update rules as exactly Gram-determined left preconditioners, and a transfer theorem extends gradient flow's pathwise implicit-bias properties to any common-scalar-preconditioned flow; they then empirically rank nine optimizers by recovery error on underdetermined matrix sensing. This gives practitioners a principled lens for predicting which optimizers will/won't exhibit low-rank implicit regularization in factored architectures (e.g., LoRA-style layers), guiding optimizer choice for low-rank recovery tasks.

MT-GNN predicts, via a per-structure graph neural network, the future per-vertex first fundamental form (metric tensor describing local surface stretch/curvature) conditioned on an arbitrary-length causal scan history and a Fourier-encoded arbitrary lead time, rather than directly regressing vertex displacements or relying on high-dimensional shape embeddings. The predicted metric tensors are decoded into an actual triangle mesh via a differentiable As-Rigid-As-Possible (ARAP) solver, and the model is trained end-to-end on rigid-aligned vertex reconstruction error, so gradients flow through the decoder to keep predictions valid, physically consistent surfaces. Evaluated on 14 subcortical structures, the paper claims end-to-end training through the ARAP decoder consistently improves prediction quality over direct-regression baselines. This intrinsic-geometry-plus-differentiable-decoder approach is a reusable pattern for other longitudinal mesh/shape forecasting problems beyond neuroimaging.

The paper identifies Modality Imbalance — text-dominant generation causing underuse of privileged visual signals — as a limiting factor in existing On-Policy Self-Distillation (OPSD) methods for MLLM visual reasoning post-training. They construct a Positive Teacher (Zoom-In Image, enhanced visual grounding) and a Negative Teacher (Mask Image, degraded visual grounding), and show via reasoning-correctness and token-logit analysis that the differential between these teachers encodes a modality-balance signal usable as privileged information for distillation. OPD-V (Visual On-Policy Self-Distillation with Modality Balance) operationalizes this insight into a training objective, presumably outperforming standard OPSD baselines on visual reasoning benchmarks by explicitly optimizing for modality balance rather than relying on incidental privileged information. Practitioners doing MLLM post-training could adopt the zoom-in/mask teacher-pair construction as a diagnostic and training signal for modality grounding.

SSTQ combines overcomplete equal-norm tight frames, coordinate subsampling, and privacy-aware scalar quantization to achieve local differential privacy in distributed optimization with near-optimal communication cost. It offers two variants — Flat Randomized Response and a Metric-Aware Laplace mechanism suited to higher bit-width codebooks — and provably attains optimal mean squared error scaling using only ⌈log₂N⌉+b bits per client, where N=Θ(d) is the frame size. The paper also derives a surrogate codebook design objective that reduces dependence on the frame's dimensionality, making the method practical for high-dimensional federated learning settings.

The authors reframe spoken language understanding as Spoken Function Calling (SFC), replacing ambiguous closed-set intent/slot schemas with structured function-call definitions that are more amenable to in-context learning. They extend traditional SLU datasets into a suite of spoken functions and use a multi-agent pipeline to synthesize SFC-Bench, a benchmark for evaluating LLMs and Large Audio Language Models (LALMs) on this task. The work then presumably fine-tunes or enhances LALMs against this benchmark, providing a reusable dataset and evaluation protocol for researchers building open-domain voice agents.

Chained Recursive Language Models (Chained RLM) is an inference-time architecture where the same base LLM is invoked repeatedly as independent reasoning roots, each given the original problem plus a compact plain-text summary, a shared blackboard, and durable artifacts written by prior roots rather than the full conversational history. This decouples context accumulation from reasoning depth, aiming to reduce error propagation in extraction, counting, ordering, and multi-hop tasks. The design suggests a lightweight, prompt-level alternative to long-context fine-tuning or retrieval augmentation, replicable with any off-the-shelf LLM by orchestrating the root-calling and blackboard-passing logic externally.

The authors introduce IRIS, a suite of neuroscience-inspired analysis tools, to systematically probe whether orientation selectivity — a canonical low-level feature in primary visual cortex built from localized receptive fields — emerges in Vision Transformers (ViTs) despite their lack of local inductive bias and global attention-based processing. The study likely measures orientation tuning curves per layer/head and compares them to biological benchmarks, testing whether such general-purpose low-level representations are shared across ViT pathways analogous to biological visual systems. Practitioners interested in interpretability or biologically-inspired architecture design could use IRIS's methodology to audit other vision architectures for emergent low-level structure.

DASyR-LLM embeds an LLM module inside an iterative symbolic regression (SR) loop for kinetic model discovery, where at each iteration the LLM performs two roles: providing a qualitative physicochemical critique of top SR candidates, and proposing new candidate rate expressions informed by both the SR-generated pool and embedded domain chemical knowledge. This domain-aware guidance aims to constrain the SR search space toward physicochemically plausible rate laws rather than purely data-fitting ones, potentially improving sample efficiency and interpretability over vanilla SR. Researchers in chemical engineering or scientific machine learning could adapt this LLM-in-the-loop critique/propose pattern to other equation-discovery domains with known physical constraints.

The authors present a privacy-aware, computationally efficient framework for classroom incident recognition from CCTV-style video, built on a hybrid benchmark combining generative synthetic footage with real-world classroom pose data. Their method constructs hierarchical kinematic representations (direction, speed, acceleration, intensity) rather than relying on raw pose or appearance, and distills multi-order kinematic reasoning from a larger teacher model into a lightweight single-order student model for deployment efficiency. This offers a template for privacy-preserving action recognition pipelines that avoid facial/appearance data while retaining discriminative motion signal, applicable beyond classrooms to other sensitive surveillance contexts.

The paper challenges the standard assumption that Subsampled Constrained Mean Shift (SCMS) trajectories converge to the 'static' density ridge defined via the density gradient and Hessian eigenstructure, showing this fails to account for rotation of the trailing eigenspace along the algorithm's continuous flow. Instead, they define the 'stable ridge,' a dynamical-systems-based geometric object characterized via the Jacobian of the projected density gradient, and prove SCMS trajectories provably converge to this corrected structure. This result revises the theoretical foundations of ridge-estimation methods used in manifold learning and nonparametric statistics, and practitioners using SCMS for dimensionality reduction should reassess conclusions drawn under the old static-ridge convergence assumption.

The authors run a controlled study crossing episodic exploration bonuses with several neural memory architectures across three partially observable RL environments that differ in how memory content must be acquired, disentangling temporal reward sparsity from what the reward actually supervises. They find the same exploration bonus produces three distinct interaction regimes: amplifying architectural capacity differences when memory content must be actively discovered and retained unsupervised, equalizing architectures to a shared performance ceiling when content is supervised once sought, and (per the truncated abstract) presumably a third distinct pattern. This suggests future RL benchmarking should report exploration-memory interactions jointly rather than ablating each component in isolation, and practitioners designing memory-augmented agents should co-tune exploration bonuses with their specific memory architecture and task's reward structure.

The authors prove a scalable separation in representational power between shallow unitary Born machines and 'channel' Born machines augmented with shared classical randomness, extending prior small-scale results to arbitrarily large systems at fixed depth. Shared classical randomness is a strictly weaker resource than entanglement in resource-theoretic terms, yet the paper shows it suffices to make the augmented model's output distribution family strictly larger than the unitary model's family under geometric locality and depth constraints typical of NISQ devices. Practically, this gives a rigorous justification for randomized/stochastic quantum circuit ansätze in generative modeling on near-term hardware, and suggests architecture search should explore randomness injection points rather than only deeper unitary circuits.

BnBERT-iPET combines Iterative Pattern-Exploiting Training (iPET), a few-shot fine-tuning method that converts tasks into cloze-style prompts and bootstraps pseudo-labels, with lottery ticket pruning applied to a Bengali BERT backbone to identify sparse, trainable sub-networks. The goal is to retain near full-model accuracy on Bengali NLP tasks while drastically cutting parameter count, memory footprint, and training/inference cost. This targets resource-constrained deployment scenarios and low-resource language settings where full BERT fine-tuning is impractical. Practitioners could replicate the approach by applying iterative magnitude pruning cycles to a pretrained Bengali BERT checkpoint before or during iPET-style few-shot fine-tuning, and compare compression ratios against task accuracy.

The method trains a latent video flow-matching prior over 69-variable, 8-day ERA5 reanalysis trajectories, then performs posterior sampling conditioned on sparse real observations (NOAA IGRA radiosondes, ISD surface stations) to solve filtering, smoothing, and ensemble forecasting as instances of the same conditional generation problem — simply by changing which frames are treated as observed. Because the prior models full spatiotemporal trajectories rather than single time-slices, information propagates bidirectionally between observed and unobserved frames without task-specific retraining. This unifies data assimilation and ensemble forecast generation under one flow-matching model, and a practitioner could adapt it by swapping in different reanalysis priors or observation sources and adjusting the conditioning mask at sampling time.

ABC (Answer-Backtracked Credit assignment) converts sparse trajectory-level outcomes into dense step-level supervision for both SFT and RL training of long-horizon search agents. Given a query and ground-truth answer, it performs 'Answer-Backtracked Clue Recovery' to trace the evidentiary path back from the answer, then assigns per-step credit that rewards useful actions even within failed trajectories while penalizing erroneous or redundant ones. This addresses the classic credit-assignment problem in multi-step retrieval/verification pipelines, where uniform trajectory-level rewards can't distinguish a single bad step from an otherwise-good search chain. Implementers building search agents (retrieve-verify-integrate loops) could adopt the clue-recovery step as a labeling pass over existing failed rollouts to extract additional dense training signal without new data collection.

HexMIL is a mask-free multiple-instance-learning (MIL) detector for medical deepfakes in CT volumes, trained with only binary volume-level labels. It decomposes each volume into a two-level patch/slice hierarchy, applies independent Gated Attention modules at each level, and combines their attention weights directly into a full-resolution 3D attention volume for ante-hoc (built-in, not post-hoc like Grad-CAM) localization of manipulated sub-regions. The claimed advantages are improved generalization to unseen generative architectures used for manipulation, plus interpretability without requiring pixel-level annotations during training. This architecture is replicable for other volumetric forensics or anomaly-localization tasks wherever only coarse volume-level labels are available but sub-region explanations are desired.

SALT (Spatially Adaptive Label-guided Temperature) modifies self-distillation pretraining (DINO-style teacher-student) by conditioning the training objective, not the input views, on weak box-derived labels available only at pretraining time. Within the box-defined region on the encoder's patch grid, the teacher's softmax temperature is sharpened and the masked-patch loss is up-weighted, while masking policy, centering statistics, and the rest of the objective remain unchanged; at inference the encoder is a standard label-free feature extractor. This targets the spatial-uniformity limitation of standard self-supervised objectives, where small lesions contribute proportionally to their pixel count rather than their clinical importance. Practitioners with any box-annotated subset of pretraining data could apply this as a drop-in modification to existing DINO/iBOT-style pretraining pipelines to improve downstream lesion detection sensitivity.

The study constructs paired modal-logic problems with identical premises and conjectures but differing accessibility-relation or domain conditions (varying frame/model semantics), verified via automated reasoning to have opposite ground-truth labels, with a balanced core designed so the semantic condition alone can't leak the answer. Under direct prompting, four of five recent LLMs scored below a condition-only baseline on this core, indicating reliance on surface pattern-matching to familiar (often classical) logics rather than genuine adherence to stipulated semantics. Enabling explicit reasoning mode produced a dramatic jump for DeepSeek V4 Flash (4.4% to 88.1%) on unchanged prompts, isolating inference-time reasoning as the dominant factor over base knowledge. This suggests benchmark designers evaluating logical faithfulness should control for prompting/inference mode, and that chain-of-thought or reasoning-mode toggles are a cheap, high-leverage lever for semantic-compliance tasks.

HiGram is a hierarchical graph memory framework for LLM agents that organizes memory into coarse-to-fine layers — upper-level summary/index nodes above fine-grained MemoryUnits — enabling path-level localization during retrieval instead of flat full-graph search. Updates are performed via path-level rewrite, propagating changes along the relevant hierarchical path rather than requiring independent unit-wise edits for every related memory, which reduces redundant rewrite operations when related facts change together. This targets scalability and consistency problems in existing flat graph-memory systems for multi-hop retrieval/reasoning agents, where accumulated history increases retrieval noise and update cost grows with memory size. Teams building long-running LLM agents could adopt the coarse-to-fine indexing structure as a retrieval pre-filter on top of an existing graph memory store to cut evidence-selection cost before applying it to full path-rewrite updates.

MALT augments Muon by wrapping its Newton-Schulz orthogonalization step with lightweight two-sided diagonal preconditioners that approximate the loss landscape's curvature, addressing Muon's blind spot to curvature anisotropy (as opposed to just gradient anisotropy). The preconditioners are diagonal (cheap in memory and compute), applied on both sides of the momentum matrix before orthogonalization, then the orthogonalized result is mapped back and its magnitude controlled via norm grafting to match Muon's update scale. This is essentially a low-overhead second-order correction layered onto a first-order-ish orthogonalizing optimizer, aimed at LLM pretraining. Practitioners could implement it as a drop-in modification to existing Muon implementations, testing whether the diagonal curvature estimate reduces loss-curve noise or improves convergence speed relative to vanilla Muon or AdamW.

The authors fit Item Response Theory (IRT) models—standard psychometric tools that jointly estimate item difficulty/discrimination and latent respondent ability—to item-level responses from 192 LLMs across eight safety benchmarks, the largest such psychometric analysis to date. Factor analysis over the fitted IRT parameters reveals three interpretable latent dimensions (refusal strictness, truthfulness, contextual harm) that explain most cross-model, cross-benchmark variance, suggesting current safety benchmarks are highly redundant along a small number of true axes. They further show that a psychometrically-selected subset of items reconstructs full benchmark scores with lower error than random item subsets of the same size, implying benchmark suites could be compressed substantially. This gives practitioners a path to build shorter, more diagnostic safety evals and a framework for detecting sandbagging or benchmark redundancy via latent trait estimation rather than raw aggregate scores.

The paper formalizes 'capability-gated planning': scientific discovery systems that select actions via myopic scoring rules (e.g., expected information gain per cost within a bounded horizon) provably fail to value 'constructive' actions—building instruments, assays, pipelines, or abstractions—whose payoff arrives only through unlocking later actions, since such actions yield zero information within the horizon and are dominated by any action with positive immediate information gain. This identifies a class of discovery problems where the least-cost path to a confident answer requires a multi-step construction chain that myopic (greedy, horizon-limited) planners structurally cannot discover. The implication for builders of AI research agents is that experiment-selection loops need lookahead or cost-to-goal search over capability-acquisition chains, not just single-step information-gain maximization, to avoid getting stuck in locally-optimal but globally suboptimal experiment sequences.

RAIL (Recoverability-Aware Intervention Learning) targets critic-free, group-based RL post-training (e.g., GRPO-style methods) where rollout budget allocation is typically uniform or heuristic-driven. Instead, RAIL learns an adaptive policy for where and how to intervene in rollout generation—not just how many rollouts to allocate—by directly optimizing for the measured improvement each intervention yields, making the allocation policy co-evolve with the training policy rather than relying on a fixed schedule. This addresses two gaps in prior adaptive-rollout work: static/non-adaptive intervention heuristics, and coarse control that only varies rollout count rather than intervention location/type. Practitioners building RL post-training pipelines for LLMs could use this to replace fixed rollout-budget schedules with a learned, state-aware intervention policy to improve sample efficiency.

MultiPathFormer reframes wireless foundation model pretraining around the physical structure of multipath propagation rather than generic masked reconstruction over channel tensors (subcarrier/antenna/time grids). It represents each transmitter-receiver link as an ordered sequence of continuous-valued path tokens (each encoding properties of an individual propagation path) and pretrains autoregressively via next-path prediction, analogous to next-token prediction in language models. Two added components—an Environmental RAG mechanism that retrieves environment-specific context and a first-path codebook—inject physical/environmental priors into the transformer backbone to improve prediction quality. This architecture is directly relevant to practitioners building channel estimation, beam prediction, or localization systems, since a physics-structured tokenization scheme could transfer better across deployment environments than grid-based masked-reconstruction pretraining.

Using a validated distributed dictionary representation (an embedding-based method for scoring text against evidence-based vs. intuition-based reference vocabularies), the authors compute an 'Evidence Minus Intuition' (EMI) score across 4.5M tweets and 59,170 Bundestag speeches from German political elites (2015–2025). They find a secular decline in EMI (rising intuition-based rhetoric) across both arenas, with right-leaning actors consistently scoring lowest, and identify a discontinuity: the AfD's 2017 parliamentary entry coincides with a sharp, chamber-wide downward shift in EMI in parliamentary speech (suggesting a contagion/accommodation effect on other parties' rhetoric), while the Twitter shift is more gradual. This is a computational social science contribution combining large-scale NLP text-scoring with a quasi-experimental design (using AfD's entry as a shock) to study elite rhetorical norms; researchers could extend the dictionary method to other countries' populist breakthroughs to test generalizability.

The method builds a Bag-of-Visual-Words pipeline on top of frozen pathology foundation model embeddings: a visual vocabulary is learned from a small set of annotated regions of interest (ROIs), pattern prototypes are formed as mean BoVW histograms per clinically-defined growth-pattern label, and sliding-window regions across whole slide images are classified via nearest-prototype matching under Jensen-Shannon divergence, then projected back onto the tile grid to yield spatial pattern maps. This is a weakly-supervised, region-level alternative to tile-level clustering, aligning output categories directly with pathologist-defined LUAD growth patterns rather than generic morphological clusters. Evaluated on 87 CPTAC-LUAD patients across three different foundation model encoders, the approach is notable for needing only a small ROI-annotated set to bootstrap the vocabulary/prototypes—practitioners in computational pathology could apply the same BoVW-over-foundation-embeddings recipe to other cancer subtyping or spatial pattern-mapping tasks with limited annotation budgets.

HelloWorld extends video world models with button-triggered social behaviors (character turning toward camera, waving, nodding, short verbal greeting) via a self-distillation finetuning pipeline: the base video generation model is used to synthesize training clips that jointly contain social-interaction behaviors and camera motion, then finetuned on its own synthesized outputs so it learns camera-pose conditioning without degrading interaction fidelity—avoiding the need for hard-to-collect real paired interaction/camera-motion data. At inference, a separate training-free timing module determines when, after a button press, the interaction should actually occur within the generated video, decoupling interaction-content generation from interaction-timing control. This is relevant to builders of interactive video-generation systems (game engines, virtual agents) as a recipe for adding controllable, socially-responsive behaviors to an existing video world model via self-generated data rather than costly annotation.

VQ-VAD adapts VQ-GAN's vector quantization to keypoint (skeleton) sequences instead of pixels, learning a discrete codebook of normal motion primitives rather than modeling behavior in a continuous latent space, which prior pose-based VAD methods relied on. Anomalies are detected via poor reconstruction or low likelihood under the learned discrete representation, since unusual motions won't map cleanly onto codebook entries. This addresses a known limitation of continuous latents: they smear together similar-but-distinct motion patterns, hurting robustness. A practitioner could build on this by extending the codebook to multi-person interactions or by combining it with temporal transformers over the discrete tokens for sequence-level anomaly scoring.

The paper reframes calibration evaluation around predicted scene rays rather than 2D reprojection error, introducing reconstruction error and intersection error as ray-based metrics, and pairs these with a bootstrapping procedure to statistically compare calibration objects and pipelines for intrinsics and extrinsics. A key finding is that reprojection error can be a misleading proxy for true 3D accuracy, particularly under generalized (non-parametric) distortion models that better capture real lens physics. Practitioners doing structure-from-motion or multi-camera 3D reconstruction could adopt these ray-based metrics to validate calibration quality more rigorously than standard checkerboard reprojection pipelines, especially when using modern non-pinhole camera models.

The authors formalize verbalized confidence calibration for risk-controlled deferral, proving that strictly monotone recalibration preserves risk-coverage frontiers and AUROC, that temperature scaling provably cannot calibrate models whose confidence stays above 0.5 while accuracy falls below it, and that a Clopper-Pearson interval turns a 200-question calibration set into a finite-sample risk certificate under i.i.d. deployment. Empirically across 22 model-task pairs (11 models, 0.5B-14B params, ARC-Challenge and TruthfulQA, 25,168 predictions), only 8 pairs achieve usable risk-controlled deferral. This gives practitioners a concrete diagnostic: check whether confidence-accuracy divergence violates the monotonicity/threshold conditions before relying on verbalized confidence for deployment safety gating, rather than assuming post-hoc calibration will fix it.

The authors adapt GraphCast, a graph neural network weather model trained on Earth reanalysis data, to Mars by evaluating it zero-shot and after fine-tuning on the Mars Climate Database (MCD), which provides global atmospheric fields across vertical levels analogous to Earth's pressure levels. Zero-shot GraphCast captures plausible instantaneous atmospheric states but fails to reproduce diurnal (day-night) variability and decays toward climatological means over longer rollouts, indicating it hasn't learned Mars-specific forcing dynamics. Fine-tuning on MCD variables plus top-of-atmosphere solar radiation forcing (humidity held constant, since Mars's atmosphere is essentially dry) improves temperature and wind field predictions. This demonstrates a transfer-learning pathway for planetary weather forecasting, useful groundwork for anyone building AI-based forecast tools for Mars missions or other planetary atmospheres using similar graph-based architectures.

The study uses VR simulations as a controlled causal-inference testbed, treating assignment of a Black adult male avatar to a police-training scenario as the experimental treatment, and estimates marginal average treatment effects (ATEs) on officer speech deference across conversational turns. This design sidesteps confounds inherent to observational body-cam or dispatch-data studies by holding the interaction script and context constant while varying only the character's apparent race, with officer race/gender as covariates. The central empirical result is a consistent reduction in deferential language toward Black male characters across most officer demographic subgroups, except white, biracial, and multiracial female officers, with effects amplified in suspect-framed scenarios. Researchers in HCI, criminology, or bias auditing could replicate this ATE-based VR paradigm to test interventions (e.g., de-escalation training) by measuring pre/post shifts in the same deference metric.

OmniEdit-Bench addresses two gaps in instruction-based video editing (IVE) evaluation: task coverage inherited wholesale from image-editing benchmarks that ignores video-specific dimensions (temporal consistency, audio, reference-conditioned edits), and metrics that reward strong visual priors from the source video even when the edit is semantically wrong. The benchmark decomposes tasks along spatial, temporal, audio, and reference-based axes, and distinguishes explicit versus implicit/reasoning-based instructions to better mirror real user requests. This gives IVE researchers a more discriminative evaluation protocol for measuring instruction fidelity rather than just perceptual quality, useful for benchmarking new video-editing models or ablating whether an architecture actually reasons about implicit intent versus pattern-matching surface edits.

The authors target a 'partially protected open-weight' (PPOW) release setting, distinct from fine-tuning-as-a-service defenses, where most weights stay trainable but a small safety-critical component is frozen/protected at release. They instantiate a Unidirectional Safety Gate (USG) via a Null Space Cubic Layer inserted after the final transformer layer, which suppresses or blocks backpropagated gradients from harmful samples whose hidden states fall within a calibrated 'protected' region of activation space, paired with an Inverse Adapter that restores normal forward-pass behavior for benign inputs so utility isn't degraded. This is a gradient-level defense rather than a data-filtering or RLHF-based one, making it robust to downstream fine-tuning attacks even when the attacker controls most of the model's weights. Practitioners building open-weight model releases could adopt this null-space gating pattern as a lightweight, architecture-level safeguard layered onto existing alignment techniques, testing calibration of the protected region against their specific harmful-sample distribution.

The pipeline evolves through V1 (a dynamic Dixon-Coles Poisson baseline), V2 (LLM contextual ratings mapped back into expected-goal parameters), V3 (goal-by-goal Monte Carlo simulation over a frozen score-candidate set replacing scalar corrections), to V4 (shared first-breakthrough/post-goal cascade judgments, time-aware stopping, and deterministic tail candidates for reproducibility). The harness is explicitly designed to be auditable, logging how LLM judgments alter simulation state rather than opaquely reweighting outputs. Practitioners building sports forecasting or other event-simulation systems could reuse this staged architecture — statistical core plus LLM-mediated contextual simulation — for any domain needing both calibration and situational reasoning.

The authors construct an ablated training corpus with all multi-modifier noun phrases removed, eliminating direct distributional evidence for modifier ordering, then evaluate trained models on held-out multi-modifier sentences to see which order they assign higher probability. Across three model scales, models consistently favor the scope-homomorphic order (mirroring semantic scope) over alternatives, with preference strength varying by modifier type. This offers a clean poverty-of-stimulus-style test bed: researchers can replicate the ablation-plus-minimal-pair-evaluation protocol to probe other purported linguistic universals or compare inductive biases across architectures and training regimes.

The system implements a 'Bidirectional Human-AI Augmentation' (BiHAA) framework via a multi-agent architecture where a shared domain-knowledge base updates from real-time annotator interactions and, symmetrically, feeds proactive suggestions back to annotators through a Proactive Agentic Support Module — moving beyond one-directional AI-assist-then-manual-correction pipelines. Design requirements were derived from a formative study with 20 annotators across backgrounds, grounding the tool in observed workflow pain points rather than assumed ones. Practitioners building annotation tooling for other implicit-semantics domains (e.g., cultural artifacts, historical documents) could adopt this closed-loop, co-adapting agent pattern instead of static active-learning loops.

The dataset comprises 3,142 handwriting demonstrations across all 52 Latin alphabet character-case combinations from 22 participants, recorded via touchscreen teleoperation capturing planar position, contact force, and timing. Trajectories are learned using Gaussian Mixture Model / Gaussian Mixture Regression (GMM/GMR), a standard probabilistic LfD technique, with resulting motions validated through a perceptual human-likeness user study. Practitioners can use the released dataset as a benchmark for trajectory-learning and human-likeness-evaluation methods, or extend the GMM/GMR baseline with newer generative approaches (e.g., diffusion policies) for comparison.

The interface continuously maps finger-touch gestures to manipulator motion with fine-grained velocity control, integrating control input and task visualization into a single touchscreen display, targeting contact-rich surface tasks like swab sampling under path and force constraints. A within-subjects study (n=20) compared this touchscreen interface against a conventional joystick and a single-click autonomous mode on task performance and subjective workload metrics for simulated realistic tasks. Practitioners designing teleoperation UIs for precision surface-contact work could adopt this direct finger-to-motion mapping plus integrated visualization pattern, and reuse the comparative evaluation protocol against joystick/autonomy baselines.

VIDP is an imitation-learning framework that uses a Task-Parameterized Directionality-Aware Mixture Model (TP-DAMM) to extract physically consistent trajectory distributions from demonstrations, disentangling intentional compliance from mere geometric/spatial adaptation — a known confound in prior variance-based impedance-inference methods. These distributions are mapped into variable impedance parameters and combined with a diffusion policy to jointly generate motion and compliance profiles for contact-rich manipulation. Practitioners working on force-agnostic kinematic demonstration data (no direct force sensing) could adopt TP-DAMM as a preprocessing step before any impedance-learning or diffusion-policy pipeline to avoid conflating geometric variation with intended compliance.

ErgoSurf couples an online ergodic control policy — which allocates dwell time according to a task-specified spatial distribution — with a Gaussian Process Implicit Surface (GPIS) model that reconstructs surface geometry incrementally from intrinsic tactile contact feedback rather than vision, removing the need for prior geometric knowledge or pre-scanning. Coverage trajectories and the surface estimate co-evolve online, so the robot refines both its map and its path simultaneously as new tactile data arrives. Practitioners working on contact-based inspection, cleaning, or finishing robots could build on this GPIS-plus-ergodic-control coupling to handle unknown or dynamically changing surface geometries without dedicated vision-based scanning stages.

Prior-SG frames scene-graph extraction as MAP (Maximum A Posteriori) inference over region labels, rather than relying on wall-based geometric heuristics or purely local visual clustering. An RGB-D stream is fused online into an Instance Graph using multi-scale open-vocabulary features, and a Prior Graph encoding task-relevant structural/vocabulary expectations serves as the prior in the probabilistic estimate. This generalizes scene-graph construction to arbitrarily-structured, open-plan environments where wall-separated room heuristics fail. Practitioners building topological/semantic maps for navigation could adopt the MAP-alignment formulation as a drop-in replacement for hard-coded room-segmentation rules.

The paper conducts a rigorous input-ablation audit of VLMs used as zero-shot driving controllers, including blind-image controls, repeated-identical-input probes, lane-axis reflection tests, non-visual baselines, and pipeline-integrity checks, across 9 direct-action models, 6 structured local VLMs, and one VLM-MPC hierarchy (32,874 scored calls, 2 embodiments, 3 simulators). Key findings: a trivial constant-SLOW policy outperforms a scripted geometric controller, several models are image-invariant (performance unchanged under blind/repeated inputs), and models that detect longitudinal hazards fail to correctly invert LEFT/RIGHT decisions under lane-axis reflection — direct evidence of non-grounded decision-making. This gives practitioners a reusable ablation protocol for auditing whether any VLM-as-controller pipeline is actually grounded in vision before trusting benchmark scores.

The method integrates probabilistic Visual Place Recognition (VPR) with feed-forward 3D geometry (FF3D) models in a topometric framework, iteratively combining coarse place-level matching with metric pose refinement over sequential image sets. An automatic offline mapping tool builds the topometric pose-appearance map from a controlled image collection, letting the system compensate for VPR's low metric precision using FF3D's accurate local trajectory estimates. This is positioned as an alternative to local-feature or neural-representation-based localization that better balances map compactness, appearance robustness, and metric accuracy. Roboticists building visual localization stacks could adopt this VPR+FF3D fusion in place of heavier SLAM-style local-feature pipelines.

Adaptive-WAM is built on a Wan2.2-5B video diffusion backbone, with trajectory diffusion heads attached to selected Diffusion Transformer (DiT) blocks and a lightweight trajectory-quality scorer that triggers early exit once a sufficiently good trajectory can be decoded. The controlled study behind it finds planning quality is largely insensitive to video denoising timestep but sensitive to DiT depth, i.e., strong trajectories emerge from intermediate layers well before full video generation completes. This decouples driving-decision quality from full video-generation cost, cutting inference compute for world-action models. Practitioners deploying video-diffusion-based driving planners can apply this multi-exit-plus-quality-scorer pattern to cut latency in any iterative generative planner where the end product (video) is discarded in favor of an extracted low-dimensional decision.

HiRoC is a hierarchical post-training framework for VLA (vision-language-action) models that decouples a high-level task planner from a low-level action executor. Unlike prior hierarchical approaches that train the planner solely via supervised learning on offline demonstrations, HiRoC's planner is refined through online interaction (implying reinforcement-learning-style post-training), while the executor learns subgoal-conditioned action policies that continuously improve. This targets long-horizon manipulation where flat policies struggle to represent task progression explicitly. Robotics ML practitioners could adopt the plan/execute decoupling plus online-planner-refinement recipe as a post-training stage on top of any pretrained VLA backbone to improve long-horizon task success.

The paper proposes Observation-Grounded Self-Predictive Representations, a dynamics-based auxiliary representation-learning method for pixel-based reinforcement learning that combines the strengths of self-prediction (latent-space rollout prediction) and observation prediction (pixel-space prediction). The claim is that observation prediction alone grounds representations in real dynamics but doesn't regularize long-horizon temporal predictability of latents, so the method adds both objectives jointly to improve sample efficiency on visual continuous-control benchmarks where prior state-of-the-art dynamics-based methods (from either category alone) still underperform with limited data. This suggests a general auxiliary-loss recipe — combine grounded pixel prediction with multi-step latent self-prediction — that RL practitioners can bolt onto existing model-free visual RL agents. Expect evaluation on standard visual control suites (e.g., DeepMind Control-style benchmarks) given the framing.

TRACE (Tokenized Robust Attention for Contact-Aware Estimation) is an end-to-end learned proprioceptive odometry estimator for legged robots that predicts relative displacement, relative rotation, and body-frame velocity directly from a recent window of IMU and joint measurements, without hand-crafted contact/slip detection thresholds. Its core is a foot-aware cross-attention module that adaptively weights IMU tokens against leg-wise kinematic tokens based on learned reliability rather than manual heuristics. Training uses direct supervision plus two physics-inspired auxiliary losses enforcing kinematic consistency and reliable leg-information usage, and simulation training incorporates policy randomization to reduce overfitting to a specific control policy and improve sim-to-real transfer. Practitioners building legged-robot state estimators could replace classical contact-threshold-based Kalman-filter odometry with this learned attention-weighted approach for better robustness on unreliable-contact terrain.

SkillMemo implicitly segments long-horizon manipulation demonstrations into latent atomic skills using an expert-guided, Mixture-of-Experts-based segmentation module, then indexes skill-level features in a dynamic episodic memory bank that Diffusion Policy/VLA-style controllers can query at inference. This targets compositional generalization to out-of-distribution task combinations, a known weakness of end-to-end visuomotor policies trained on limited trajectory data. Because the decomposition is learned rather than hand-labeled, it could be retrofitted onto existing DP/VLA pipelines as a memory-augmented add-on. Replication would hinge on the MoE segmentation objective and how skill features are retrieved and composed at test time.

GAUGE is a real-world-grounded diagnostic benchmark spanning 22 controlled task families across rigid bodies, cables, textiles, and volumetric deformables, each paired with calibrated physical metadata, uncertainty annotations, and task-specific observables derived from measured real trajectories. It jointly scores numerical physics engines and generative video world models on the same tasks, moving beyond perceptual-similarity or human-judgment metrics to pinpoint which specific physical principles or parameters a model violates. This lets practitioners diagnose failure modes rather than getting an opaque aggregate fidelity score. It's directly usable as a regression/evaluation suite when developing or tuning simulators or video-based world models for embodied AI.

Robust-WAM is a post-training method for video-generation-based World-Action Models that preserves the VAE-latent generative pathway (retaining compatibility with large-scale pretrained VGMs) while adding a lightweight semantic-foresight alignment objective applied to the action-prediction stream. This addresses the tradeoff between VAE-space WAMs (rich pretraining, fragile to appearance shift) and semantic-space WAMs (robust, but unable to leverage VGM pretraining). Because it's a post-training add-on rather than a from-scratch architecture, it should apply to existing pretrained VGM-based WAMs with modest additional training cost. Practitioners building action-prediction heads on video diffusion backbones could adopt the semantic alignment loss directly.

The method computes a shape-aware axis-aligned HBB from an oriented detection by incorporating the convex hull shape, hull 'fullness' (how much of the OBB the object actually occupies), and the OBB's orientation angle, rather than simply taking the min/max extent of the rotated box (Outer HBB) or naive area-equivalent/marginalized approaches. It's benchmarked against three baseline conversion methods — Outer HBB, Area Equivalent HBB, and GBB Marginalized HBB — specifically for ship detection in remote sensing imagery. The approach is a pure post-processing geometric transform, so it drops into any existing OBB detector's output pipeline without retraining. It's straightforward to implement and test since it only requires hull geometry computation from existing OBB predictions.

CoMuDi takes a multi-robot team, an object assembly, and a part-dependency graph, and constructs compound pick/place/exit tasks per part, propagating temporal constraints across the dependency graph so each robot's task window starts and ends as early as feasible while avoiding inter-robot collisions in a confined shared workspace. It integrates the space-time RRT* (ST-RRT*) planner to generate individual motions minimizing per-task arrival time, which in aggregate minimizes overall disassembly makespan; the paper compares CoMuDi with ST-RRT* against a variant using plain RRT*. This applies directly to multi-robot disassembly/assembly task-and-motion-planning research, and the dependency-graph-driven compound task formulation could generalize to other coordinated multi-robot scheduling problems with precedence constraints.

XEWorld is a controlled cross-embodiment evaluation testbed that holds scenes physically identical while swapping in robot embodiments unseen during training, isolating whether action-conditioned world models generalize via learned dynamics or merely via visual similarity to training embodiments. The analysis finds a shared bottleneck: current models behave as 2D visual pattern matchers whose generalization correlates with visual, not kinematic, similarity to seen robots, causing them to fail at mapping abstract joint-action vectors into coherent visual trajectories and at predicting dynamic changes from a static initial frame. This gives a diagnostic protocol — held-out-embodiment evaluation in matched scenes — that practitioners can apply to any action-conditioned video/world model to check for genuine dynamics learning versus visual shortcut learning. It also motivates future architectures grounding actions in embodiment-agnostic kinematic representations rather than raw pixel appearance.

The system pairs a passive, LED-illuminated binary structured-light mask (single-shot, no projector-camera synchronization) mounted on one channel of a dual-channel laparoscope with a VQ-VAE-based prior and a custom U-Net depth-estimation head, trained and evaluated against 722 phantom images with ground-truth depth captured by a Zivid 3D camera and reprojected into the laparoscope's image frame. This replaces conventional multi-shot fringe projection profilometry (which needs DMD projection and precise sync) with a synchronization-free single-shot pipeline suited to compact laparoscopic hardware. Practitioners in surgical robotics could adopt the same passive-mask plus learned-prior depth-head architecture and Zivid-reprojection supervision strategy to build real-time intraoperative depth sensing on other minimally invasive camera platforms.

SpikingNav consists of a Spiking Sensing Encoder (SSE) that extracts task-conditioned visual features via a spike-based backbone, and a Spiking Policy Network (SPN) that maintains temporal dynamics to produce navigation actions, targeting the event-driven sparsity and intrinsic temporal memory SNNs offer over dense ANN-based navigation policies. The central claim under test is whether spike-based sensing and policy dynamics improve robustness under visual corruption in visually rich indoor embodied navigation, an open question the paper addresses empirically. This is relevant to practitioners targeting resource-constrained or neuromorphic navigation hardware, where the SSE+SPN split (perception encoder vs. temporal policy) offers a template for building end-to-end spiking navigation stacks compatible with neuromorphic accelerators.

Co-safe LTL objectives reduce to maximal reachability via product MDP construction, but direct bootstrap methods like TD/Q-learning fail to converge due to the noncontractive Bellman operator and nonunique fixed points. The proposed two-step method first uses a discounted surrogate MDP to identify a 'clamp set' that resolves this nonuniqueness, then applies undiscounted policy evaluation and greedy improvement. The paper proves almost-sure convergence of policy evaluation and finite-time termination at an optimal policy — a rare exact convergence guarantee for model-free RL on temporal-logic specifications.

Builds on BridgeVLA's approach of preserving VLM input-output alignment by projecting point clouds into multi-view images and predicting intermediate heatmaps before final actions. BridgeVLA++ adds a unified spatio-temporal memory architecture that models persistent spatial context and temporal interaction history across a manipulation episode. This targets three known VLA weaknesses — data inefficiency, poor generalization under distribution shift, and lack of memory — making it relevant to anyone building long-horizon or partially-observable manipulation policies.

The paper reduces constrained sc-LTL planning to a constrained reachability problem on an extended product MDP, then proves that switching policies — constructed by alternating among stationary policies computed for the individual sc-LTL objective and constraint — suffice for optimality, avoiding the need for arbitrary randomized policies. This sufficiency result enables the optimal policy to be computed via a tractable linear program rather than more expensive non-convex optimization. A grid-world case study validates the optimal objective-constraint trade-off, giving practitioners a concrete recipe for safe, logic-specified planning.

DreamWAM reformulates future prediction beyond RGB, jointly denoising RGB and motion latents while adding lightweight gated residual branches for geometry and semantics, all within a diffusion-transformer setup (VideoDiT). Shared attention between VideoDiT and an ActionDiT branch lets the action-prediction module condition on these richer future-state representations rather than raw pixels alone. This targets the entanglement problem in RGB-only world models and offers a concrete architectural template (multi-branch, shared attention) for anyone building action-conditioned video/world models.

Mimir is a neuro-symbolic memory architecture that separates world memory (object locations, states, perceptual evidence) from task memory (goal agenda, progress, hand state, failures, execution constraints) and dynamically grounds them before each action via a binding module. This grounding module links the active goal to recalled world candidates, resolves missing source locations, and attaches supporting evidence prior to planning and embodiment-specific execution. The design provides an explicit, inspectable interface for goal-relevant fact retrieval — useful for anyone building long-horizon embodied agents where implicit/flat history representations tend to lose track of task-relevant state.

PRIMAL3 is a learning-based MAPF framework integrating RL, topology-aware communication, LaCAM3-guided training, and PIBT-based action refinement, targeting failures at topologically critical states (bottlenecks, dead ends, persistent conflicts). Agents are featurized via cut vertices, dead-end regions, shortest-path distances, and blocking estimates, and interact through two graphs: a same-direction following graph propagating multihop context, and a different-direction conflict graph using masked attention. Policy entropy identifies uncertain agents during training, which then receive targeted LaCAM3-guided imitation supervision — a practical recipe for scaling learned MAPF policies to very large agent counts.

RORA is presented as the first end-to-end real-to-sim pipeline that reconstructs simulation-ready articulated assets from a single static object video, using a suggestion-based human-in-the-loop process to resolve joint structure rather than requiring captured dynamic motion. It exports a hybrid representation combining 3D Gaussian Splatting for photorealistic rendering with a mesh-based structure for physics/kinematics, addressing the low success rates of prior motion-tracking or learning-based methods on complex multi-joint objects. This is directly relevant to anyone building sim-to-real robot learning pipelines that need articulated digital twins without expensive dynamic capture.

DBFly introduces a vision-language waypoint prediction framework with an explicit spatial maneuver decision chain — target-direction anchoring, spatial diagnosis, and maneuver decision — interposed between vision-language understanding and continuous waypoint generation, addressing semantic-control misalignment seen in direct vision-language-to-action mappings. It further constructs what the abstract describes as an implicit flight (component, details truncated in the source) to support this pipeline. The explicit intermediate spatial-reasoning stage offers a reusable pattern for practitioners building language-conditioned navigation policies that need more reliable, interpretable control than end-to-end regression provides.

The method augments a one-stage real-time instance-segmentation backbone with a lightweight edge-detection branch, edge-guided attention fusion, and the parameter-free SimAM attention module, exploiting the observation that transparent vessels' boundary contours are more reliable cues than their interior texture. They introduce LabGlass-IS, a 3,485-image, 21-category real laboratory-glassware dataset, and report a Boundary F-score of 97.80, beating YOLO-prompted FastSAM by 18.93 BF points while retaining real-time throughput. Practitioners could adopt the edge-attention branch as a drop-in module for other transparent/reflective-object segmentation tasks or use LabGlass-IS as a benchmark.

The proposed hierarchical VLA architecture inserts an explicit language-memory module that converts discrete temporal observations into a coherent, temporally-logged textual memory sequence, decoupling high-level semantic planning (handled by the VLM backbone) from low-level action execution. This targets the non-Markovian nature of long-horizon tasks, closed-loop error correction, and preservation of VLM semantic representations that end-to-end action fine-tuning tends to degrade. Practitioners building long-horizon manipulation policies could bolt a similar textual-memory layer onto an existing VLA stack to improve temporal consistency without retraining the full backbone.

The architecture couples an online uncertainty estimate that updates the obstacle geometry feeding a control barrier function (CBF), barrier-filter interventions and estimation residuals that set experience-replay priority, and a critic trained on the executed (post-filter) action rather than the nominal policy action. Six component-matched configurations are compared on a 2D robot-navigation task with corrupted obstacle measurements, under identical training budgets, seeds, sensor streams, exploration, and disturbances, evaluated via a post-training test, an 11-level perception-noise sweep, and an extreme 6.0x-multiplier stress test. Practitioners designing safe actor-critic controllers can use this as a template for coupling CBF-based safety filtering with uncertainty-aware replay prioritization rather than treating these as independent modules.

The paper proposes a lightweight GitOps-based metadata-management architecture applying Data-as-Code principles with CI/CD pipelines to manage provenance and iterative evolution of annotations for GoA3/GoA4 automatic-train-operation perception datasets, aiming to replace monolithic data catalogs that suffer heavy operational overhead and documentation drift. Practitioners maintaining large, dynamic, regulated annotation sets (not just railway) could adopt this Git-native versioning plus CI/CD-validation pattern to get built-in provenance tracking and developer-workflow integration for free.

The framework is a fully integrated, vision-only pipeline delivering real-time net-relative and global localization plus continuous 3D mapping for UUVs operating in visually degraded, dynamic environments. It was validated on synthetic datasets with ground truth and field-tested onboard a UUV performing autonomous net-relative navigation, demonstrating real-time performance and robustness. Practitioners building underwater inspection robots without reliable acoustic or GPS positioning could adopt this as a baseline vision-SLAM/localization stack.

The system equips each warehouse carrier with an ultra-low-power BLE IoT tag broadcasting item urgency; robots read these broadcasts directly and combine urgency with travel distance in a local decision rule to select which carrier to serve, achieving item-level prioritization without any central scheduler. It's evaluated in simulation and validated on a physical robot swarm. Practitioners building decentralized multi-robot task-allocation systems could reuse this BLE-broadcast-plus-local-utility pattern to add priority-awareness without new fixed infrastructure.

The dataset, developed by DB InfraGO AG and understandAI GmbH under the Digitale Schiene Deutschland program, is a multi-sensor collection with over 7 million annotations spanning railway-specific and general-perception object classes, captured across varied operational scenarios to support perception systems from GoA2 through GoA4. It's positioned as a request-available benchmark for training and validating obstacle-detection and classification models for automated train operation. Practitioners in rail perception could use it directly for model benchmarking or transfer learning.

The study performs a closed-loop target-and-control audit of per-skill task-vector subtraction across all ten LIBERO-Goal skills in multitask VLA policies with continuous-regression, discrete-token, and flow-matching action heads, identifying three regimes: clean target-control separation (5 skills), resistance to suppression (3), and global model collapse (2). Suppressed targets reliably hit 0% success on held-out states, but mean baseline-normalized control retention is only 52%, and every target-suppressing edit measurably harms at least one nominally unrelated skill, with no clean separation observed on the Spatial suite. This serves as a cautionary benchmark methodology practitioners should apply before deploying task-vector unlearning/editing techniques on production VLA policies.

The method reformulates 3D-2D pose estimation by differencing perspective projection equations between consecutive frames rather than solving independent absolute poses, combined with a depth-invariance approximation and motion parameterized on SE(3) (the group of 3D rigid transforms). This yields a closed-form robustness guarantee: translational extrinsic (camera-to-platform) calibration errors cancel exactly in the differential formulation, while rotational extrinsic errors produce only a bounded first-order perturbation. It generalizes to monocular and multi-camera rigs, making it directly applicable to micromotion-sensitive tasks like structural health monitoring or precision robotic manipulation where recalibration is costly.

CoCo (Counterfactual Consistency) targets the shortcut-learning problem in action-conditioned video prediction, where models exploit visual inertia and recurring motion statistics instead of genuinely conditioning on the input action, leading to action-invariant or spuriously persistent motion. The framework imposes two complementary consistency constraints under counterfactual perturbations of actions and observations, pushing the model's dynamics to causally depend on the actual action signal rather than dataset priors. This is a training-time regularization approach applicable to any action-conditioned video/world-model architecture, and practitioners could adopt the counterfactual-consistency losses as an add-on objective to existing world-model training pipelines to improve controllability metrics.

Mind-VLA addresses the instruction-agnostic limitation of prior 3D-geometry-aligned VLA models by first grounding the language instruction to a target object, then synthesizing tri-view renderings of that object and extracting VAE and VGGT (a geometric/visual transformer) features from them. These target-specific features are used to align the VLA's latent representations, replacing uniform whole-scene alignment with object-focused supervision. The claimed benefit is improved performance on fine-grained manipulation and occlusion-heavy tasks where whole-scene geometric alignment previously diluted the signal relevant to the actual target; practitioners building VLA pipelines could adopt this object-grounding-then-align pattern as a drop-in auxiliary training signal.

GASP parameterizes trajectories as clamped B-splines, using a convolutional residual network to predict free interior control points while analytically inserting boundary control points to satisfy initial/final derivative (velocity/acceleration) constraints. A conditional variational autoencoder (CVAE) samples a batch of candidate trajectories, which are decoded and collision-validated in parallel on the GPU, giving a batched planner with near-millisecond inference for joint-space motion generation under non-stationary (changing) conditions. Benchmarked online, it matches analytical planners' success and collision-aware feasibility rates while substantially cutting inference time versus other GPU-based trajectory optimizers, making it a candidate drop-in for real-time reactive motion planning in manipulation systems.

FineMote is a firmware generation framework that compiles hierarchical robot description files into control code while explicitly modeling the timing consequences of tree-structured data dependencies between subsystems. It wraps heterogeneous low-level control logic into a unified object model with standardized scheduling units and execution entry points, then performs static timing orchestration over the resulting object graph to bound perception-to-decision latency at compile time rather than via runtime tuning. This gives practitioners a way to reason about worst-case scheduling behavior up front and to plug in new device types through the standardized interfaces without rewriting the scheduler. It targets a known pain point in URDF/SDF-style description-driven firmware pipelines where latency emerges implicitly from structural composition.

RTCF (Retrieve in Time, Correct in Frequency) is a training-free test-time correction framework for frozen VLA policies that decouples retrieval (which successful past trajectory frame to reference) from correction (which frequency component of that trajectory's action to transfer). Its Progressive Memory Alignment (PMA) component causally aligns the growing live visual execution history against complete successful trajectories via incrementally updated monotonic frontiers, avoiding progress-misaligned retrieval that naive frame-matching produces. By applying correction in the frequency domain rather than direct time-domain replay/fusion, it avoids overwriting the policy's own reactive action structure, offering practitioners a low-overhead retrieval-augmented correction layer to bolt onto any frozen long-horizon VLA deployment.

GUARD is a test-time, model-agnostic failure detector for diffusion-based VLA policies that quantifies grounding by measuring the causal influence of token-indexed entries in the frozen vision-language model's KV (key-value) cache: it constructs counterfactual caches via targeted ablation of salient KV entries and compares the resulting denoising trajectories against the original conditioning. From these comparisons it derives a diagnostic stream (sensitivity, attention entropy, modality bias, grounding efficiency) that is calibrated online and fed to a lightweight temporal classifier for failure prediction. Evaluated on task-held-out splits across five policy-benchmark combinations (including Pi0, SmolVLA, and Alpamayo-class policies), it gives practitioners a way to bolt real-time risk monitoring onto any pretrained diffusion VLA without retraining the base policy.

SSC (Structured Subtask Chain) is a state-transition annotation schema for bimanual manipulation demonstrations, composed of Structured Subtask Template (SST) entries that store core action tuples (subject, predicate, object), flexible adverbial conditions (spatial/instrumental modifiers), a base-motion field decoupled from arm actions, and an after-state scene graph. This structure sits between free-form natural language (readable but unverifiable) and rigid templates like BEHAVIOR-1K's skill_annotation (verifiable but over-segmented and unnatural), aiming to preserve both automatic verifiability and annotation consistency. The paper reports SSC supports three vision-language-assisted labeling functions, suggesting practitioners building subtask-segmented manipulation datasets could adopt this schema to get machine-checkable ground truth without sacrificing label readability.

The paper formulates field-of-view-limited safe navigation as online safety-volume certification over a voxel map, constructing a certified graph whose vertices are exactly the positions where the robot's full inflated swept volume is verified free. SCOPE decouples an optimistic goal-directed route from certified execution: it converts the first uncertified point on that optimistic path into an explicit observation obligation, resolves it via target-centric viewpoint search, and recursively resolves intermediate obligations when candidate viewpoints aren't yet certified-reachable, backed by a certified preview mechanism. This gives a reusable planning layer for any body-mounted limited-FOV sensor platform where practitioners need provable pre-motion safety guarantees rather than post-hoc collision checking.

The paper introduces epsilon-rule-dominance, an approximate dominance relation generalizing both Pareto and lexicographic dominance to partially ordered 'rulebook' objectives, then proposes RA*pex, a best-first multi-objective search algorithm that computes a compact epsilon-approximate rulebook-optimal solution set. It reuses dimensionality-reduction speedups from existing multi-objective search (à la MOA*-style methods) while respecting the rule hierarchy via separate closed sets per priority tier and dominance checks over truncated/residual objective vectors. Practitioners implementing hierarchical constraint-based planners (autonomous driving, task planning under regulatory rules) can plug in RA*pex to trade a tunable epsilon for major speedups over exact rulebook search.

Electroaerodynamic propulsion generates thrust via corona-discharge ion drift between electrodes, and its thrust density and efficiency rise substantially when operated near a ground plane. The authors empirically characterize passive hovercraft skirt geometries and configurations, trading off captured ground-effect gain against skirt drag and air leakage, then fabricate a centimeter-scale point design. The resulting tethered hovercraft sustains extended flight and dozens of takeoff/landing cycles, giving a replicable reference for skirt geometry and electrode layout that others could adapt to different ground-effect EAD vehicles; full power autonomy still awaits an onboard power source lightweight enough to match thrust-to-weight needs.

The architecture is grounded in a Dec-POMDP formulation with private per-agent goal observations, decomposed into five LLM-driven modules: action-conditioned Theory-of-Mind inference, hierarchical planning, conversation interpretation, action verification, and feedback-based replanning — using the LLM as a tractable surrogate for otherwise intractable Bayesian belief inference and planning under partial observability. It's benchmarked against a ToM-ablated variant and a MARL policy trained offline over many goal pairs in a human-subject construction task; the full method reduced required interaction steps and raised post-task trust ratings over both baselines. The reusable takeaway is a design pattern — separating belief inference from planning, dialogue, and verification within an LLM pipeline — for zero-shot multi-agent coordination without task-specific RL training.

SAFECAST builds on hidden-state-based risk probes combined with functional conformal prediction for rollout failure detection, and augments the probe's training and calibration data with contrast-set perturbations spanning both visual and language dimensions to better cover deployment-time distribution shift. It shows statistically significant ROC-AUC improvements over a state-of-the-art baseline across real-world DROID and LIBERO simulation experiments and multiple VLM backbones, with ablations showing the largest gains when both visual and language perturbations are combined. Practitioners can adopt this by generating targeted perturbation sets matching expected deployment shifts (clutter, lighting, novel objects, reworded commands) to harden any hidden-state failure/anomaly detector layered on top of a VLA policy.

SparseDitto targets SpMV, SpMM, and SpGEMM under one framework, using an additive model that ranks established sparse-format/execution strategies from structural features of the input matrix, then invokes an LLM-based agent to construct a bespoke GPU kernel matching the chosen matrix, operator, and target GPU. The motivating finding is that cuSPARSE's CSR vs. Blocked-ELL formats alone show up to 350x performance variance on identical SpMM workloads, showing no format/system dominates universally. Practitioners could plug this into autotuning pipelines as a per-workload kernel synthesizer, replacing static library dispatch (e.g., cuSPARSE) for irregular or dynamic sparsity workloads.

The paper formalizes benchmark revision as constrained optimization: maximize realism metrics (missingness structure, simplicity, structural plausibility, population alignment) subject to keeping downstream utility above an operational floor already enforced by production checks. It's instantiated on a care-gap benchmark built from Synthea-generated synthetic patients exercised through demonstration EHR workflows and the same downstream pipeline used on real operational data, finding the baseline benchmark 'extremely thin' on realism. The approach generalizes to any synthetic-benchmark pipeline where teams need to iteratively tune data generators against multiple realism axes without regressing existing evaluation gates — relevant to anyone building privacy-safe test data for enterprise or clinical AI agents.

GEM-3 is a probabilistic global weather model using a timestep-conditioned transformer architecture that supports explicit multi-timestep autoregressive inference from a single trained weight set, letting the timestep be chosen at inference to trade off fine-grained predictability against long-horizon usability. The authors show mixed-timestep training — rather than committing to one fixed rollout interval — improves rollout stability over fixed-timestep baselines. Practitioners building autoregressive ML weather or climate models can adopt timestep-conditioning as a training recipe to unify short- and long-horizon forecasting in one model instead of maintaining separate per-horizon models.

The paper proposes networks of holonic digital twins (HDT-Nets), where each twin is a holonic agent — a self-similar, recursively composable unit — that actively reasons over shared spatiotemporal context rather than passively mirroring sensor data, positioning wireless networks as an orchestration substrate for physical AI beyond conventional throughput/latency/reliability optimization. It targets the gap that current architectures can't support real-time coordination requiring agents to maintain a shared world model for long-horizon planning under uncertainty. This reads as an architectural/vision proposal rather than an evaluated system, relevant to researchers designing edge-network-integrated multi-robot world-model sharing or digital-twin coordination protocols; no concrete algorithm or benchmark result is given in the excerpt.

The authors implement live carbon-aware inference routing across multi-region GPU testbeds, steered by marginal operating emissions rate (MOER) signals, and benchmark it against an actual production pressure-based (load) router rather than a naive uniform-placement baseline. Energy is attributed per-request from NVIDIA DCGM telemetry combined with measured concurrency curves (not nameplate TDP), and carbon accounting is settled against historical MOER rather than only the forecast used for routing decisions — closing a common evaluation gap. The central result is operational feasibility: MOER-driven routing successfully shifted inference load across regions with no observed dispatch failures. Practitioners building carbon-aware schedulers can reuse this evaluation methodology (real baseline, telemetry-based energy, ex-post settlement) to validate their own routers.

PLB is a priority-aware load balancer implemented at the JDBC driver layer that enforces service differentiation purely through client-session-to-replica assignment, without requiring elastic replica provisioning. It partitions replicas by user tier (premium vs. freemium) and applies load-based borrowing, letting lower-priority sessions use idle premium-partition capacity opportunistically while reverting to strict partitioning under contention to bound premium-tier degradation. This is a practical alternative to admission control or query-level prioritization inside the database engine itself, making it deployable against unmodified replicated DB backends. Practitioners running read-replica fleets (e.g., Postgres/MySQL with read replicas) could adopt this pattern directly via a custom JDBC connection-routing layer.

MultiMoQ extends Media over QUIC (MoQ) with multi-access delivery, redesigning its object/track delivery mechanism to distribute tile-based 360-degree video streams across multiple concurrent network paths with fine-grained, flexible tile scheduling and seamless inter-path switching. This targets two known MoQ weaknesses: coarse delivery granularity that hurts responsiveness at scale, and fragility under bandwidth constraints that causes playback stalls. The system is fully implemented, suggesting a working reference for tile-scheduling and multipath-switching logic that others building immersive/tiled streaming pipelines on QUIC-based transports could study or extend. Its relevance is highest for anyone deploying live volumetric or 360 video at scale where per-user bandwidth is variable and multi-network devices (phones, laptops) are common.

MARS is a receiver-driven, forwarder-assisted multipath transport for WAN/Internet services, combining tier-synchronized overlay path discovery with coupled consumer/forwarder congestion control to expand usable forwarding paths beyond what's endpoint-visible while reacting to congestion near the actual bottleneck rather than only at endpoints. It positions itself against MPTCP/MPQUIC (endpoint-limited, delayed feedback) and routing-assisted schemes (require infrastructure support, unsafe cross-path coordination), aiming for both better path utilization and safety. Critically, it runs as an incrementally deployable UDP overlay, meaning it doesn't require router or ISP cooperation — practitioners running geo-distributed data pipelines or CDN-style transfers could deploy it as a drop-in transport layer without WAN infrastructure changes.

ML-for-ML proposes cross-layer co-optimization of network-side knobs (e.g., scheduling, bandwidth allocation) and ML-side knobs (e.g., communication frequency/volume, batching) under a unified time-to-target-loss objective, rather than optimizing networking and ML training configurations independently as is standard practice in shared cloud clusters. A preliminary prototype demonstrates up to 42% faster convergence to target loss versus siloed tuning, though the work is explicitly early-stage with no described search/optimization algorithm details yet. This is a promising direction for practitioners running multi-tenant training clusters to explore joint autotuning frameworks that expose both network scheduler and ML training hyperparameters to a shared optimizer.

The paper introduces a formal reasoning framework for window function optimization, comprising Frame Analysis and Partition Analysis, plus a new execution strategy called Co-Evaluation that permits early evaluation of predicates depending on a window function's own result — something existing predicate-pushdown techniques cannot do outside narrow ideal conditions. The contribution is organized as a table of algebraic equivalences over window function expressions, giving query optimizer implementers a systematic basis (rather than ad hoc special cases) for deciding when transformations like pushdown are valid. This is directly applicable to relational database query optimizer development — anyone building or extending a cost-based optimizer's window-function rewrite rules has a concrete formal target to implement against.

The paper formalizes Agentic Service Graph Embedding (ASGE) as an online network-control problem: mapping runtime-revealed AI-agent workflow calls (to models, memory, tools) onto service replicas and network paths under capacity, cost, and deadline constraints, where the dependency graph is only partially known ahead of time. ASGE-RR addresses this with revisable reservations — provisional capacity holds for predicted future calls that get updated as the workflow unfolds — evaluating candidate replica-and-path mappings against predicted workload rather than committing greedily to the currently visible call. This is relevant to anyone building infrastructure schedulers for multi-agent LLM systems (agent orchestration platforms, tool-serving meshes) where naive greedy or FCFS resource allocation would starve high-value workflows revealed later.

AutoThread targets thread-count selection for simulator-side execution in RL inference pipelines, identifying the task-execution-to-scheduling-time ratio as the dominant factor governing optimal thread count under dynamic workloads. It predicts thread counts using a Physics-Informed Neural Operator (PINO), constrained and guided by a finite-source M/M/1 queueing model that bounds predictions to physically sensible ranges based on scheduling theory. This hybrid learned-plus-analytical approach targets fast, accurate re-tuning as simulator workload characteristics shift during execution, which practitioners running simulation-in-the-loop RL systems (robotics, autonomous driving stacks) could adapt as a runtime thread-scheduling module rather than using static thread-pool configuration.

TensorCast introduces Tensor-as-a-Service (TaaS): a distributed layer decoupling tensor state management (weight loading, KV cache, checkpoint sync) from execution logic via first-class tensor abstractions and programmable lifecycle policies. Rather than integrating task-specific mechanisms directly into execution engines, networks, or storage backends, it exposes a unified API that different LLM-serving components can compose against. Practitioners could build custom tensor lifecycle policies (eviction, replication, versioning) on top of this layer without reimplementing per-subsystem integration, similar to how VFS abstracts filesystems.

ShimGen automates synthesis of shims interfacing heterogeneous, hierarchical, scoped coherence protocols (e.g., cluster-level protocols meeting global protocols like CXL or AMBA CHI), moving beyond prior synthesis tools that assumed a conservative single-writer-multiple-reader (SWMR) invariant. It introduces a shim API that classifies protocol transactions by their semantic coherence guarantees, enabling composition of protocols supporting both SWMR and relaxed accesses. This lets automated synthesis exploit modern architectural features for performance while retaining correctness guarantees, offering a template for verified coherence protocol generation in multi-cluster SoC designs.

The authors full-fine-tune Qwen3-32B (32.76B dense params) across 16x NVIDIA B300 GPUs (2 nodes, FSDP/ZeRO-3), presenting one of the first published operational accounts on this accelerator. Key artifacts include a B300-calibrated power-draw triage table distinguishing compute-bound, communication-bound, data-starvation, and checkpoint/deadlock states (noting GPU utilization reads 100% even during NCCL hangs, making power draw the more reliable signal), plus a controlled A/B showing NFS streaming matched a pretokenized local cache (~53k tok/s) since the corpus fit in page cache and the job was compute-bound. This is a practitioner-oriented systems paper, useful as a debugging runbook for teams standing up multi-node training on new hardware generations.

BALANCE co-locates a small language model (SLM) and full LLM on an edge server, dynamically assigning each user to either autoregressive decoding (AD, sequential, latency-bound) or speculative decoding (SD, SLM drafts verified by LLM, latency-optimized but memory-costly), running both modes concurrently to exploit their complementary latency-memory tradeoffs. The scheduling objective maximizes the number of served users under fixed edge compute/memory budgets, treating AD/SD assignment as an allocation problem rather than a fixed system-wide choice. This is directly applicable to telecom operators building multi-tenant LLM inference at radio access network edges where per-user SLAs and hardware constraints vary.

The platform combines a Raspberry Pi host controller, a precision current-measurement device, and a microcontroller-based device-under-test (DUT), with a lightweight HTTP interface supporting automated firmware upload, synchronized execution, and high-resolution current acquisition. This enables scripted, reproducible system-level power benchmarking workflows (e.g., stress-test power-draw proxies) without proprietary industrial test equipment. Practitioners could replicate the hardware BOM and HTTP API to build automated power-regression test suites for embedded/semiconductor CI pipelines.

RepoOMP builds a Multi-granularity Attributes Performance graph (MAP) capturing cross-file dependency evidence, routes each hotspot to either deterministic legality rules or an LLM agent based on whether locality-provable safety exists, and constructs a Structured Transformation Context (STC) that surfaces only decision-relevant dependency facts to the agent's prompt. Evaluated on 951 profiled hotspots from NPB, BOTS, FFmpeg, NCNN, and GROMACS, it achieved compilable, correctness-checked, positive-speedup parallelization on 372 hotspots. The dependency-aware context reduction technique (STC construction) is the reusable piece — applicable to other repo-scale code-transformation agents beyond OpenMP, wherever retrieval noise vs. missing evidence is the bottleneck.

The system repurposes the NR uplink Sounding Reference Signal as a passive bistatic radar waveform within an end-to-end O-RAN simulation testbed (OpenAirInterface + FlexRIC + Sionna RT), with a PHY-layer sensing stage inside the gNB feeding detections to an Extended Kalman Filter tracking xApp via a custom E2 service model — requiring no NR standard changes or dedicated sensing waveform. Since a single bistatic pair leaves elevation unobservable, they resolve the height-prior ambiguity two independent ways: a planar receive array providing vertical aperture, and a second transmitter providing range diversity, with both approaches validated against live results. This is a reproducible blueprint (all open-source components) for ISAC researchers wanting to prototype RAN-integrated sensing without proprietary radar hardware.

The authors frame contraction-plan selection as a learning-to-rank problem: each candidate tensor-network contraction plan is encoded via structural features derived from its sequence of pairwise contractions, and gradient-boosted rankers are trained on GPU execution measurements using listwise and pairwise ranking objectives. This bypasses execution-time-dependent factors (parallelism, reduction structure, memory traffic, contraction geometry) that theoretical complexity metrics ignore, letting the model pick near-optimal plans pre-execution. Evaluated across diverse circuit families with held-out in-distribution and out-of-distribution circuit-size splits, this gives practitioners a drop-in plan-selection module for GPU-accelerated quantum circuit simulators, replacing brute-force plan search or heuristic-only optimizers.

The authors evaluate NIST PQC signature candidates (Falcon-512, Dilithium-2, SPHINCS+) against ECDSA P-256 for compatibility with IEEE 1609.2 secured-message structures and SAE J3161 transport-block constraints in C-V2X sidelink. Falcon-512 is the only scheme meeting size limits, so it's benchmarked against ECDSA via full-stack PC5 Mode 4 co-simulation across 24 scenarios (six traffic levels-of-service, LOS/NLOS), measuring packet delivery ratio and end-to-end latency. This gives a concrete feasibility baseline for standards bodies evaluating PQC migration paths in safety-critical V2X. Practitioners can use the transport-block sizing methodology to re-evaluate future, more compact PQC signature schemes as they emerge.

PDMarks embeds ownership watermarks across the placement, clock-tree-synthesis (CTS), and routing stages of a physical design flow, deriving all watermark instances and target values deterministically from a 32-byte secret key via HMAC-SHA256. This satisfies Kerckhoffs's principle — security rests solely on key secrecy, not obscurity of the embedding algorithm — addressing a gap in prior single-stage or security-through-obscurity PD watermarking schemes. The multi-stage embedding presumably increases robustness against a white-box adversary who knows the full method but lacks the key. Chip designers or EDA toolchains could adopt this as a verifiable, cryptographically grounded IP-provenance layer across the standard place-and-route flow.

RAC (Reference-Aware Activation Compression) targets the split-inference bottleneck where local head/tail plus cloud-middle-layer execution requires repeated transfer of boundary hidden states. It retrieves exact-token historical spans to compress prefill uplinks, reuses reconstructed uplink states for same-round prefill downlinks, and generates boundary-specific decode references via lightweight causal predictors, combined with grouped affine alignment and calibrated residual quantization (with optional prefill outlier handling). This effectively turns activation transfer into a delta/reference-based codec rather than raw retransmission. Systems building split-LLM-serving stacks could apply this codec design to cut uplink/downlink bandwidth without retraining the base model.

AsymSpec addresses uplink-gated verification and wasted dependent-draft computation in cloud-edge speculative decoding under asymmetric bandwidth. Its asymmetric verification protocol keeps common-path acceptance uploads compact while pushing richer rejection-only correction data to the (presumably higher-bandwidth) downlink, and a total-variation certificate on the residual distribution decides whether a small target top-K response suffices or whether a progressive larger response is needed. This decouples correctness-critical small messages from bandwidth-heavy correction payloads, directly targeting the stop-and-wait vs. wasted-runahead tradeoff in existing edge-cloud speculative decoding schemes. Implementers of edge-cloud LLM serving could adopt this protocol to keep edge compute utilized under real-world asymmetric last-mile links (e.g., DSL, cellular uplink-limited connections).

The paper studies rewritability of atomic ontology-mediated queries (OMQs) over Horn-ALCHI, a Description Logic that is not generally first-order rewritable, into GQL, specifically its UC2RPQ (unions of conjunctive two-way regular path queries) fragment. They introduce 'DL automata' to capture OMQ semantics as runs over fact sets, then define a stratification of automaton states that rules out problematic cyclic dependencies, yielding a broad sufficient condition for rewriting into UC2RPQs. This gives database implementers a concrete syntactic/structural test for when an ontology-mediated atomic query can be compiled down to native GQL execution rather than requiring a separate reasoning engine. It's a theoretical contribution establishing tractable rewritability boundaries that could underpin future OMQ compilers targeting GQL-compliant graph databases.

The work applies an LLM to detect Common Weakness Enumeration (CWE)-categorized hardware vulnerabilities — improper access control, sensitive information exposure, unintended privilege escalation — directly from Verilog source in single-module designs, evaluated iteratively across a benchmark dataset. Framing hardware security auditing as an LLM code-analysis task mirrors LLM-assisted software vulnerability detection but adapts it to RTL semantics where errors are irreversible post-fabrication. The iterative evaluation methodology suggests a detect-then-refine loop, positioning this as a potential pre-tapeout screening tool. Hardware verification teams could integrate this as an automated CWE-aware linting pass alongside traditional formal verification and static analysis tools in the RTL sign-off flow.

The paper presents a systolic-array-based activation unit that computes univariate functions (e.g., tanh) and softmax using Chebyshev polynomial approximations rather than the common CORDIC iterative method, enabling hardware resource sharing between the two computation types that are typically separated. Reported results show up to 71% lower mean absolute error for tanh versus a CORDIC baseline while using 4.6% less area and 5.1% less power, plus 44.6% and 79.0% lower KL divergence for softmax versus CORDIC and piecewise-linear approximations respectively. This is directly relevant to neural network accelerator designers looking to consolidate nonlinear activation hardware; the Chebyshev-approximation approach could be adopted as a drop-in activation/softmax IP block in custom ASIC or FPGA NN accelerator designs.

The authors formalize rule-based header-compression-rule learning as a two-stage pipeline: unsupervised structure discovery via recursive partitioning of training packets using a normalized entropy-ratio criterion (robust to small samples), followed by constrained rule selection via dynamic programming to maximize expected compression gain under a hard cap on installable rules. They instantiate and evaluate this on SCHC (Static Context Header Compression), the IETF standard for constrained-network header compression, using real IoT and 5G-core traffic traces. This automates what is typically manual rule engineering in SCHC deployments, and the entropy-ratio partitioning plus budget-constrained DP selection could generalize to other rule-based compression or protocol-optimization settings beyond SCHC.

The paper formalizes operationalization-aware data discovery (OADD): given a broad question, a database, and an optional scope constraint, jointly determine viable operationalizations of focal concepts and identify the supporting columns for each, distinguishing it from schema linking and column retrieval which assume explicit, directly-relevant queries. Since manually collecting question-to-column ground truth from researchers is impractical, they construct OADD-Bench by mining empirical papers, using a question miner to extract and reframe paper-supported research questions paired with the columns those papers actually used. This benchmark construction method is reusable for evaluating future OADD systems, and the task itself is a natural target for retrieval-augmented or agentic LLM pipelines over tabular data.

AFD-Ledger is an offline provisioning system that independently optimizes hardware assignment and deployment organization for both Attention-FFN Disaggregated (AFD) and collocated MoE serving under identical TPOT SLO, hardware catalog, and runtime constraints. It replaces exhaustive provisioning search with an analytical execution-time model paired with an evaluation-bounded hardware search, letting it compare architectures at scale without brute-force benchmarking every configuration. The core contribution for practitioners is a reusable analytical framework for deployment planning that could be extended to other disaggregation schemes or adapted as a pre-deployment cost/throughput estimator.

Dart abstracts DNS implementation dependencies and configuration heterogeneity behind a declarative specification language, orchestrating containerized or virtualized DNS software stacks through a unified interface. The authors evaluate its performance overhead and present two case studies demonstrating single-command construction of portable DNS analysis environments, directly targeting the reproducibility gap that has plagued empirical DNS research. Because it's released as an open toolkit, practitioners can use it to package their own DNS experiments as shareable, version-pinned declarative configs for peer replication.

The authors build a taxonomy of agentic workflows and provide the first architectural characterization of them, combining a production study at Microsoft Azure with controlled experiments on open-source agent frameworks. They show execution is fragmented across repeated CPU-GPU boundary crossings because orchestration and tool invocation run on the host, putting the CPU on the critical path; workload load stays low with sudden bursty spikes (execution structure), GPU utilization evenness depends on model composition, and task/tool diversity widens variance further. This characterization exposes concrete architectural mismatches in current GPU-centric, steady-state-optimized datacenter designs, giving hardware architects a taxonomy-grounded basis for redesigning CPU-GPU interconnects or scheduling for bursty, orchestration-heavy agentic traffic.

Eigenius is an open-source typed knowledge-graph DBMS whose kernel tightly couples a dependent type system, storage engine, and MCP integration layer around three pillars: a dependent type theory threaded through the core, 'institutions' serving as strongly-typed integration boundaries between knowledge sources, and a content-addressed immutable storage layer. This architecture makes data provenance a structural invariant enforced by the type system rather than something reconstructed post-hoc across subsystems, and supports epistemic stratification (tracking declared confidence/status of facts) natively. Developers building AI-agent research pipelines could use Eigenius as a backing store where every fact carries a machine-checkable warranty and lineage, useful for auditing autonomous multi-step reasoning chains.

CommBench comprises 100+ expert-curated GPU communication tasks — point-to-point transfers, collective operations, expert-parallel communication for MoE models, compute-communication fusion, and communication utilities — with reference implementations from GPU experts or distilled from production codebases. Its evaluation harness automatically compiles, executes, and validates generated code on real multi-GPU systems under a cheat-resistant protocol, paired with a unified metric suite (implied to cover both correctness and performance/efficiency). Practitioners can use it directly to benchmark code-generation models on distributed-systems-level GPU programming, or as a curated task/reference-implementation corpus for fine-tuning or few-shot prompting on this niche domain.

MCHA (Memory-Centric Hierarchical Architecture) is reconfigurable hardware targeting workloads with intrinsic parallel-sequential execution patterns — MARL, large-scale neuromorphic simulation, and probabilistic graphical models — that conventional architectures bottleneck on due to global-buffer saturation and memory-bound stalls from irregular access. Its core mechanism is a hierarchical communication strategy enabling distributed, inter-core data routing, which offloads traffic that would otherwise hit the global memory buffer, reducing its bandwidth burden. This is a hardware-architecture proposal likely evaluated via simulation/RTL against baseline accelerators on these workload classes, of interest to chip architects designing accelerators for irregular, memory-bound AI and simulation workloads.

Deltoris is an algorithm-hardware co-design framework for diffusion-based VLA inference that exploits temporal similarity between consecutive control-loop inputs via a temporal-aware bit-sparsity algorithm, computing only the bit-level deltas between successive inputs to skip redundant arithmetic. Because this delta-based approach increases off-chip memory traffic relative to dense computation, the authors add a speculative inference technique that amortizes data loading across multiple inference steps rather than paying the cost each cycle. Together these target the 50-200 Hz control-frequency requirement of diffusion VLA policies on latency- and energy-constrained edge accelerators, giving embodied-AI hardware designers a concrete pattern (temporal delta sparsity + speculative prefetch) for co-optimizing diffusion inference pipelines.

The system uses eBPF kernel-level tracing to passively capture network events and applies a two-pass PID-to-port correlation algorithm to disambiguate co-located processes sharing a runtime, recovering exact service identity without instrumentation. From 13,615 captured events it reconstructs a dependency graph (32 edges across 20 services, matching ground truth) and applies spectral graph clustering with Kernighan-Lin refinement to generate an ROI-ranked migration plan. This is directly reusable for live migration planning in cloud/edge environments where source code access or agent deployment isn't feasible. A practitioner could adapt the PID-port correlation technique to other zero-instrumentation observability problems.

EdgeLM reframes demonstration retrieval for in-context table understanding by targeting decision-boundary informativeness rather than pure similarity, combining 'data edges' (near-neighbor examples with differing ground-truth labels) and 'model edges' (similar examples the deployed model previously misclassified). It requires no fine-tuning or task-specific feature engineering, making it a drop-in retrieval swap for existing few-shot pipelines. The method is evaluated across five data-wrangling tasks, fifteen datasets, and five open-weight and proprietary LLMs, suggesting it generalizes across model families. Practitioners could implement it as a retrieval-augmentation layer on top of any embedding-based few-shot pipeline by logging model errors online to build the 'model edge' pool.

DeepSig built a CUDA acceleration backend for OCUDU's physical layer (PDSCH, PUSCH, SRS, PRACH) and O-RAN fronthaul IQ compression, integrated via acceleration interfaces that preserve existing factories and processor abstractions rather than forking the codebase. Performance hinges on keeping data GPU-resident using CUDA-visible resource grids, device-side softbit buffers, pinned staging buffers, and managed-memory policies to minimize host-device transfer overhead. On an NVIDIA DGX Spark (GB10 GPU + ARM host) with CPU baselines pinned to high-capacity cores, they report up to 10.3x PUSCH, 2.7x PDSCH, and 19.7x split-8 lower-PHY RX speedups. This gives O-RAN/telecom practitioners a concrete blueprint for GPU-offloading PHY-layer DSP while retaining standard software-radio interfaces.

The paper formalizes highly-reliable remote control (HRRC) over LEO satellite links as a quantile forecasting problem, predicting a high-percentile bound on latency to guarantee a target reliability level rather than just modeling average latency. They propose a high-quantile estimator trained/evaluated on a real-world OneWeb dataset collected in a major US agricultural hub, capturing genuine LEO latency volatility. Results show the estimator meets specified reliability constraints while permitting operating speeds up to 138.6% higher than conservative fixed-margin approaches. This is directly applicable to teleoperation systems needing latency-aware speed governors, and the quantile-forecasting framing could generalize to any safety-critical control loop over variable-latency links.

The architecture introduces distributed Event Monitoring Units (EVUs) attached to heterogeneous IP blocks (cores, accelerators, interconnect, memory controllers) that capture microarchitectural events and forward them over what appears to be an AMBA-based interconnect to a centralized aggregation point for correlation and processing. This addresses the practical problem of needing multiple vendor-specific software interfaces to read HPCs from different subsystems in real-time embedded SoCs. The design targets use cases like profile-guided optimization and dynamic resource management that require cross-component event correlation with consistent timing. A hardware/systems architect could use this as a template for standardizing telemetry collection in a multi-IP SoC design.

The authors systematically evaluate DRAM-PIM-GPU architectures across LLM decode-phase inference workloads (OPT-7B/70B, Mamba2-2.7B/70B) and derive three design principles. First, static power (DRAM leakage/refresh, GPU idle draw) can dominate the energy calculus — dynamic-power-only models overestimate tokens/s/W by up to 3.85x for realistic low-batch, long-output deployments like Mamba2-2.7B at batch size 1. Second, decoding throughput is monotonically non-decreasing with DRAM channel count across all tested models, typically plateauing at high channel counts for low-batch workloads, implying diminishing returns beyond a threshold. This gives PIM-GPU system designers concrete guidance to avoid over-provisioning channels and to include static power in efficiency projections rather than relying on dynamic-only estimates.

The work targets stateless load balancing for QUIC using programmable data-plane switches (e.g., P4 ASICs) that hash the first packet of a flow to pick a backend, then need per-flow-consistent routing for all subsequent packets despite the data plane's limited stateful memory. Prior approaches embedded a server identifier inside the QUIC Connection ID (CID), which requires server-side modification and violates the QUIC spec's requirement that CIDs be unlinkable (a privacy/anti-tracking guarantee). This paper demonstrates stateless QUIC load balancing implementable entirely in the data plane with zero changes to CIDs, presumably leveraging other packet-level signals instead. This is directly applicable to network operators deploying in-network load balancers who need QUIC compatibility without compromising client privacy guarantees or requiring backend server changes.

The framework embeds a distributed QAOA (DQAOA) solver inside a three-block ADMM decomposition of unit commitment: continuous dispatch/relaxed-commitment variables go to a quadratic programming block, while binary commitment decisions are cast as a QUBO problem solved via the quantum interface. The DQAOA interface is flexible, supporting brute-force enumeration, monolithic QAOA on one QPU, or distributed QAOA that partitions logical qubits across multiple capacity-constrained QPUs, avoiding the need for a single device large enough to hold the full binary problem. Evaluated on a five-unit UC instance (15 binary variables) comparing all three QUBO-solving modes against the unchanged ADMM updates for the continuous block. This offers a concrete template for hybrid quantum-classical decomposition of larger combinatorial power-systems problems as QPU qubit counts remain limited.

The paper proposes OPA (Online Pricing-based Slice Admission Control and Resource Allocation), which assigns pseudo-prices to infrastructure resources capturing long-term scarcity and anticipated inter-temporal opportunity cost, then uses these prices to drive per-request admission and allocation decisions online. An exponential pricing strategy is designed specifically to guarantee bounded worst-case performance (a competitive-ratio-style guarantee) despite not knowing future slice requests in advance. This gives InPs (infrastructure providers) a tractable, provably-bounded alternative to solving an intractable stochastic/online optimization exactly, and could be implemented as a pricing module sitting in front of existing slice orchestration systems.

THQLink is a real-time QEC decoding architecture that connects HPC compute resources to a quantum processing unit's (QPU) control system over the TH-Express interconnect, designed to generalize across different quantum hardware and control stacks. The reported round-trip latency is 2.944 μs on average, with only 130 ns of incremental overhead per additional network hop, meeting the microsecond-scale real-time decoding budget required to avoid syndrome backlog. This offers a scaling path where decoding compute lives off-chip in a shared HPC fabric rather than requiring dedicated ASIC/FPGA decoders per QPU, which is relevant to anyone architecting control systems for larger logical-qubit counts.

The study benchmarks ~3000 single-device training runs across six GPUs, varying model family, numerical precision, batch size, and context length, to test whether a linear MFU-based power model can substitute for hardware-counter-based power models in performance simulators. They find the linear fit holds well specifically for compute-bound regimes (the typical LLM training case), and that fitting separate linear coefficients per (GPU, dtype, batch size) combination — rather than one model per GPU — substantially tightens the fit. Practically, this gives simulator builders a software-only, pre-execution power estimator they can plug into system-design and energy-aware scheduling tools without needing profiled hardware traces.

AdaMX is a heterogeneity-aware microscaling (MX) format and accelerator that selects, per block, the precision-recovery scheme, and per operand (weight vs. activation), the representation, while keeping equivalent bit width (EBW) unchanged versus fixed-format baselines like MXFP4. A single hardware design supports two block sizes, yielding both a higher-accuracy operating point and a lower-EBW, storage-saving operating point. The authors implemented and taped out a 22nm FD-SOI AI accelerator to validate the approach at the silicon level, giving practitioners a concrete reference design for building adaptive low-bit inference hardware rather than a purely algorithmic proposal.

SciRet evaluates a fixed scientific RAG pipeline — sentence-window chunking, BM25 sparse retrieval, BGE-M3 dense retrieval, reciprocal rank fusion, optional cross-encoder reranking, and grounded generation — over CORD-19 at three corpus scales (1K/5K/15K papers, 1,034–15,480 chunks). Hybrid (sparse+dense fused) retrieval proves more robust than either retrieval mode alone, hitting Recall@10 of 1.000 at both the smallest and largest scales tested, while an MS MARCO-trained cross-encoder reranker actually reduces precision due to domain mismatch with scientific text. Generation faithfulness is measured via RAGAS metrics, giving practitioners a compute-aware benchmark for deciding which RAG components (and at what corpus scale) are worth deploying for scientific QA.

RepairFormer casts structured-input repair as supervised sequence-to-sequence generation across formats (JSON, DOT, OBJ, INI, S-expressions, TinyC), using format tags to condition the model, an oracle validator to check output validity, and a 'boundary-localized repair' strategy that constrains edits near the corruption site to preserve unaffected content. This contrasts with prior deletion- or search-based repair, which risks semantic drift or content loss. A practitioner could apply this as a preprocessing/self-healing layer in CI or config-management pipelines, or extend the format-tag conditioning approach to additional grammars by fine-tuning on synthetic corruption/repair pairs for a new schema.

IcFuzz is the first dedicated fuzzer for NVIDIA Isaac Sim, addressing challenges of context-aware object semantics, hierarchical simulation control, and a huge state space. It uses an LLM to perform semantic stage segmentation, decomposing simulation programs into structured stages that capture object-level context, then applies guided multi-level mutation informed by that segmentation to explore the state space more effectively than blind fuzzing. This targets bug classes specific to GPU-accelerated physics and photorealistic rendering pipelines. Simulation/robotics engineers could adapt the semantic-segmentation-guided mutation strategy to fuzz other complex, stage-structured simulators beyond Isaac Sim.

HOPSCOTCH is a Lean 4 framework for mechanizing game-based cryptographic proofs, representing security definitions as indistinguishability between stateful probabilistic oracles and using a shallow embedding so oracles/reductions are just ordinary Lean definitions with full access to Mathlib (e.g., finite-group theory). Game-hopping proofs are constructed as explicit formal objects whose constructors mirror standard proof steps, aiding automation and inspection; a general computational soundness theorem interprets these objects by automatically constructing reductions against the underlying assumptions and deriving concrete security bounds. This gives cryptography researchers a practical path to formally verify real protocol security proofs inside an established, general-purpose proof assistant rather than a bespoke crypto-only tool.

The study uses a positivist, survey-based approach with PLS-SEM (Partial Least Squares Structural Equation Modeling) via SmartPLS to test hypothesized relationships between candidate critical success factors (CSFs) and Agile project success criteria, as reported by practitioners. The analysis isolates a small subset of factors with statistically significant effects on success, rather than treating the full commonly-cited CSF list as equally predictive. This offers project managers and researchers an empirically prioritized (rather than anecdotal) framework for where to focus Agile process investment, and a replicable PLS-SEM methodology for testing CSF models in other project contexts.

GSE (globalized skill evolution) maintains a Skill Relation Graph (SRG) that models and co-evolves relationships among an agent's learned skills, addressing the overfitting and inconsistency that arise when skill updates are made purely locally/sequentially. It performs cluster-based skill consolidation to abstract generalizable capabilities out of many local updates, and uses replay-driven verification (re-running past tasks) to catch behavioral regressions before a skill update is accepted. Evaluated on two representative software engineering tasks, GSE provides a concrete architecture — graph-structured skill memory plus consolidation and regression-checking — that other agent-skill-learning systems could adopt to avoid the failure mode of skills that only work on the task they were learned from.

The paper extends quantale-valued (quantitative) linear lambda-calculus with an additive disjunction connective for case-statement reasoning, proving the equational theory sound and, under continuity assumptions on the underlying quantale, approximately complete. It supplies multiple models — categorical-logic gluing constructions, probabilistic models, and quantum-computational models — situating the calculus within denotational semantics broadly construed. A concrete Banach-space-based probabilistic model is used to reason about Cauchy sequences of random walks via the calculus's equational system, illustrating a shift from qualitative program-equivalence semantics toward quantitative, functional-analytic program reasoning. Practitioners in categorical semantics or probabilistic programming languages could use this as a foundation for quantitative equational reasoning about branching probabilistic/quantum programs.

The paper diagnoses a scaffold-overfitting effect in CLI software-engineering agents: models fine-tuned on OpenHands-collected trajectories show a large performance drop when evaluated under alternate agent scaffolds, while untrained base models do not exhibit this gap, isolating the cause to fine-tuning rather than pretraining. The authors attribute this to planning structure being implicitly scaffold-specific, distinguishing explicit planning (a discrete pre-execution plan artifact) from implicit planning (loop-level structural conventions), and propose decoupling planning from the fixed scaffold so it generalizes ('internalizing' it) across scaffolds — likely via a scaffold-agnostic planning representation or training procedure (DCAS). Practitioners fine-tuning open coding-agent models should treat scaffold diversity in training trajectories as a first-class requirement to avoid this cross-scaffold degradation.

The paper extends RTLola's stream-based specification semantics with two operators for future-dependent properties: a prediction operator that extrapolates future stream values at arbitrary timestamps from historical observations (approximate, immediately available), and a discrete future-offset operator that delays evaluation of dependent expressions until the exact future value is known (precise, latency-incurring). Both extensions are formalized within RTLola's operational/semantic framework, and the authors implement and evaluate them for runtime and memory overhead. Engineers building runtime verification monitors for real-time/cyber-physical systems can use this to express predictive safety properties while choosing the accuracy/latency tradeoff per property.

The paper surveys and synthesizes existing definitions of explainability/self-explainability from software and AI engineering literature into a unified definition and a structured taxonomy of explainability requirements, motivated by compliance needs under the EU AI Act and IEEE 7001-2021. Its key contribution is arguing for incorporating 'explanation goodness' (i.e., correctness/adequacy of explanations, not just their presence) directly into the definitional and requirements framework, rather than treating quality as a separate evaluation concern. Requirements engineers or systems architects designing self-explainable autonomous systems can use the resulting taxonomy as a checklist for eliciting and validating explainability requirements during design.

LangChoiceBench is a project-level code-generation benchmark measuring three things across 28 projects in seven software domains: raw Python selection rate, consistency between an LLM's stated language recommendation and its actual implementation, and overall language diversity. Evaluating 25 LLMs, the authors find persistent over-selection of Python, low recommendation-implementation consistency, and that smaller open-weight models show stronger Python bias and less diversity than larger/closed models. Analysis of 9,826 reasoning traces shows the Python default is largely an automatic, ease-driven choice rather than a reasoned tradeoff against stated project requirements, suggesting the bias originates in training-data/RLHF priors rather than genuine technical reasoning capability. Practitioners building coding-agent evaluation suites can adopt this benchmark to audit language-choice bias in new models before deployment.

The paper extends the RTLola runtime monitor with an external-data-query mechanism that lets stream specifications pull from external systems (static databases, dynamic APIs like weather services) through a unified interface, addressing delayed/asynchronous responses, type-checking of returned data against stream types, and runtime error handling for failed/slow queries. Evaluation on aviation-domain specifications includes a custom geospatial query backend built on k-d trees, which outperforms state-of-the-art database systems for the geospatial lookups tested. This gives runtime-verification practitioners a template for integrating external data sources into stream-based monitors without sacrificing the formal timing/type guarantees RTLola provides.

AgentExecutor tackles partial code execution — running isolated snippets lacking full context/dependencies — via a multi-agent LLM framework with three phases: execution-environment preparation, dynamic exploration with iterative refinement (using richer actions and execution feedback than prior work), and prefix evolution via program synthesis to progressively fix the snippet until it runs. It's positioned against prior approaches like LExecutor and Treefix, claiming improvements from a larger action/feedback space and adaptive (rather than fixed) optimization strategy. This is directly usable as a component in dynamic-analysis pipelines, fuzzing harness generation, or automated test/debugging tools that need to execute arbitrary extracted code fragments.

The paper extends the Coudert-Madre decision-diagram-based method for computing prime implicants (PIs) to implicitly incorporate side constraints during computation rather than as a post-hoc explicit filtering pass over the (potentially exponential) full PI set. The tool chain is modular, separating decision-diagram construction, PI enumeration, and constraint-based filtering, allowing each stage to be swapped or optimized independently. This targets scalability bottlenecks in logic minimization, model-based diagnosis, and explainable-AI/formal-methods applications where PIs under structural/contextual constraints are needed; practitioners working with BDD/decision-diagram toolchains could integrate this filtering-aware PI algorithm directly into existing diagnosis or explanation-generation pipelines.

The authors propose a two-phase framework: automated requirement checking on a dSPACE real-time HIL platform followed by LLM-based diagnosis, using open-source instruction-tuned models fed textual encodings of multivariate sensor traces rather than raw numeric arrays. The goal is to match or approach the accuracy of data-driven classifiers while retaining interpretability and requiring little to no labeled training data, addressing ISO 26262 traceability requirements. Practitioners working on validation pipelines could adapt this by converting sensor telemetry into structured text prompts and evaluating off-the-shelf instruction-tuned LLMs as a lightweight diagnostic layer.

The system models developer expertise from historical Git commit data and applies an assignment algorithm to map developers to expertise-aligned long-term components in new projects, evaluated across multiple back-end model choices for performance. Server-side data caching improved worst-case response speed by a factor of 9.86x, indicating the profiling/matching computation is the primary bottleneck. Teams building similar task-routing tools could reuse this pattern: mine commit history for skill signals, then apply an optimization/matching layer with aggressive caching to keep interactive latency low.

DCAware is a non-iterative, computationally efficient framework for regression test-oracle generation that avoids the execution-feedback self-repair loop used by prior approaches, instead prioritizing high signal-to-noise dual-context awareness to construct fault-revealing assertions directly. The core finding is that optimizing purely for execution success as a proxy objective misaligns with the true goal of fault detection, causing iterative repair methods to converge on assertions that are easy to satisfy but weak at catching regressions. Practitioners building LLM-based test oracle generators can apply this by avoiding feedback loops keyed solely on execution pass/fail and instead incorporating richer contextual signals upfront to bias generation toward fault-sensitive assertions.

The authors performed iterative qualitative coding on 23 publicly available syllabi and course materials for upper-division, credit-bearing courses explicitly addressing Generative AI in software engineering, characterizing learning objectives, assessment types, topic coverage, and documented AI tooling. The output is a descriptive taxonomy of commonalities and differences across this nascent curricular area rather than a prescriptive standard. Curriculum designers can use the reported patterns as a benchmark to compare or design their own AI-assisted software engineering courses.

CodeGrep is a 14B-parameter retrieval agent trained end-to-end with GRPO (a reinforcement learning method) to issue multi-turn parallel grep/glob/read tool calls and surface candidate files to a frozen downstream coding agent, evaluated on all 500 SWE-Bench Verified instances. It improves resolve rate slightly (27.0% vs 25.8% no-retrieval baseline) while cutting rounds by 15% and tokens by 19% on resolved instances, and the paper notes a precision threshold effect where weaker retrievers like BM25 at 0.375 precision actually degrade agent performance rather than help. Practitioners building agentic coding pipelines could plug in a similarly trained lightweight retrieval front-end ahead of a larger frozen coding model to cut cost without sacrificing accuracy, provided retrieval precision clears the reported threshold.

AssertMate combines static analysis and type-aware heuristics to identify assertion targets (actual value construction), then predicts expected values via three complementary methods — direct code generation, retrieval-augmented generation, and chain-of-thought reasoning — aggregating these diverse perspectives into a final assertion, addressing the modest accuracy and prompt-randomness sensitivity of prior one-shot approaches like ChatAssert. This multi-perspective aggregation design aims to reduce reliance on heavy oversampling while improving reliability. Developers building LLM-based test generation tooling can replicate the pattern of decoupling 'what to check' (static/type analysis) from 'what value to expect' (multi-method ensemble prediction) to improve assertion quality.

The authors develop a noise-aware quantum Hoare logic incorporating hardware-vendor-published error models into program semantics, then derive algorithmic methods for bounded verification of quantum programs against specific hardware noise profiles and for automatic synthesis of noise-optimal loop-free programs. They apply this to synthesize hardware-tuned versions of common quantum subroutines — parity checks, state preparation, and state discrimination — and evaluate the approach empirically. Quantum software researchers could build on this by extending the Hoare logic to new hardware error models or using the synthesis methods to auto-generate noise-robust subroutines for their own target devices.

JDomInO implements bidirectional synchronization between a domain model and Java code for tactical Domain-Driven Design patterns, using a shared metamodel to support a deterministic forward path (model-to-code generation) and a reverse path (code-to-model reconstruction). The forward path has been fully validated across all 12 building-block types in the metamodel using a Hotel Management case study; the reverse path's mapping logic has passed unit testing, with end-to-end validation still in progress per the abstract. Teams practicing tactical DDD in Java could adopt this toolchain to enforce model-code consistency, or reference its metamodel design when building similar roundtrip-engineering tools for other languages or architectural patterns.

The authors extract source-code metrics from contributors' commit histories and build a competence model that predicts skill levels more granularly than prior binary or coarse classifications used in OSS contributor studies. This addresses succession-planning risk in OSS projects where core-maintainer attrition threatens project survival. A practitioner could apply the metric pipeline to their own repository's commit logs to surface under-recognized senior-level contributors as promotion or maintainership candidates.

JTA (Joint Testability Architecture) reframes validation as a joint design problem over three coupled objects — scenario, test system, and system-under-test — evaluated along controllability, observability, and isolability, connected via three domains, three 'bridges,' and an iterative analysis-design-evaluation-refinement loop. It introduces 'scenario contracts' and joint capability constructs to make root-cause attribution of abnormal test outcomes tractable rather than purely artifact-centric. This gives safety-critical test architects a structured framework to audit whether a failed test result is attributable to the SUT versus scenario or rig limitations, useful for certification-grade validation processes (e.g., DO-178C/ISO 26262 style workflows).

This is an exploratory mixed-methods study (n=119 practitioners) combining descriptive statistics with qualitative thematic analysis of open-ended survey responses to characterize dependence, overreliance, and addiction-adjacent behavior patterns tied to LLM use in professional software development. It contributes an early empirical taxonomy of behavioral markers (e.g., reduced independent problem-solving, compulsive tool-checking) rather than a validated clinical instrument. Researchers could use the reported themes to design a follow-up validated survey instrument or longitudinal study on AI-tool overreliance in engineering teams.

The paper establishes a finite-state reduction for Fitting's finite Heyting-valued modal logic via a relational bitopological duality: the observational quotient generated by atomic valuations is shown isomorphic to its finite image in the bitopological dual, with the quotient relation as the restriction of the canonical dual relation, and is proven minimal among reductions preserving generated observations. It further constructs finite tree-shaped exact-value certificates whose depth is bounded by modal depth and whose branching is bounded by the height of the truth-value algebra. This gives model-checking and proof-search implementers a provably minimal canonical model plus succinct certificates for verifying exact (not just Boolean) truth values in many-valued modal systems.

WasmMend introduces a two-stage repair pipeline: a differential trace analysis technique first localizes the specific function where native and Wasm execution traces first diverge, then an LLM-based repair agent uses that localization to reason about and patch the underlying platform-level discrepancy (library implementation gaps or compiler bugs) rather than searching source code blindly. This directly targets a failure class shown in prior work to be largely opaque to source-level LLM repair agents. Practitioners maintaining C/C++-to-Wasm toolchains could adopt the trace-divergence localization step alone as a debugging aid even without the full automated-repair agent.

Autogrammar is an agentic system that learns a context-free grammar for a target DSL from documentation and execution data, then feeds that grammar into grammar-constrained decoding to eliminate syntactically invalid LM outputs. The agent's control flow is formalized as a Kripke structure whose nondeterministic transitions are resolved by an LM, with agent behavior governed declaratively via linear temporal logic (LTL) constraints — giving verifiable guarantees over the learning process itself, not just the output grammar. It's evaluated across four Autogrammar variants on three DSLs, offering a template for teams needing grammar-constrained decoding for proprietary or low-resource DSLs where no formal grammar currently exists.

Wavelet is presented as the first formally verified compiler targeting asynchronous spatial dataflow architectures, using a combination of verification techniques to prove that generated dataflow-operator graphs preserve determinacy while enabling pipelining across loop iterations. The verification covers the compilation pipeline end-to-end, addressing a known correctness gap in dataflow/spatial-architecture toolchains where scheduling nondeterminism could otherwise silently corrupt results. This is directly relevant to teams building compilers for CGRAs, dataflow accelerators, or asynchronous circuit generators who need machine-checked correctness rather than empirical testing alone.

The paper defines the innovation spectrum of a binary sequence as the set of all innovation (prediction-error) sequences producible by causal predictors, inducing a new reducibility notion shown to properly refine truth-table reducibility while having distinct order-theoretic structure. The resulting degree structure splits into a 'truth-table spine' order-isomorphic to the classical truth-table degrees, and a disjoint 'reservoir-immune' region containing sequences admitting no infinite computable predictable subsequence, with bridge constructions linking the two regions and a proof that reservoir immunity is preserved under sequential innovation. It appears to connect to Martin-Löf randomness (paper text cuts off there), giving computability/algorithmic-randomness researchers a new reducibility notion to relate prediction-theoretic and classical degree-structure results.

The study benchmarks six open-weight LLMs (Gemma, Llama 3.3, DeepSeek Coder V2 Lite, Qwen 3.5, Mistral Small 3.1, GPT-OSS) on 100 theorems from CoqStoq, a benchmark built from real-world Coq projects, using single-shot greedy decoding (temperature 0) with kernel-level acceptance as the sole correctness criterion. Success rates were low across the board—12/100, 8/100, and 1/100 for the top three models, zero for the rest—yielding 21 total verified model-theorem successes. This establishes a low but concrete baseline for kernel-checked formal proof generation and highlights the gap between surface plausibility and Calculus of Inductive Constructions-level soundness; the released successful proofs and setup could seed few-shot prompting or fine-tuning studies aimed at closing that gap.

The authors implement and compare two fault-localization paradigms for the verification-aware language Dafny: a state-based approach that adapts AutoFix's snapshot methodology by inferring invariants/predicates to flag suspicious program states, and a counterexample-based approach that mines Dafny's native counterexample traces more systematically than manual inspection. Both aim to move beyond Dafny's default failure output, which surfaces only a single violated postcondition or one failing execution path per assertion, forcing manual trace inspection. The case study evaluates the relative effectiveness of these paradigms at localizing root causes in real verification failures, providing a template for building better debugging tooling into formal-methods-integrated languages.

Mosaic decides satisfiability of Constrained Horn Clauses (CHCs) modulo fixed-size bit-vector theory (T_B) by partitioning the clause set into T_B and Integer Arithmetic (T_I) fragments, then reasoning over each fragment with a dedicated solver while exchanging information via sound cross-theory translations to determine overall satisfiability. This theory-modular design sidesteps the poor scalability of monolithic bit-precise CHC solving by offloading tractable subproblems to integer-arithmetic reasoning, a known bottleneck for bit-precise program verification (e.g., proving absence of overflow/wraparound bugs). A prototype implementation is described, suggesting the approach is available for benchmarking against existing bit-precise CHC solvers on standard verification tasks.

The authors mine 2,652 keyword-retrieved issues/forum posts across five AI developer tool ecosystems (GitHub Copilot/VS Code, Cursor, Claude Code, OpenAI Codex, OpenCode), using a three-model LLM ensemble to conservatively identify 600 unanimously-flagged visual accessibility reports, validated by stratified manual review. Topic modeling and qualitative coding surface three recurring categories: screen-reader/assistive-tech incompatibility, poor visual presentation/contrast/differentiation, and (implied) a third category likely tied to dynamic/streaming content. This provides an empirical taxonomy that tool builders and accessibility researchers can use to prioritize fixes (e.g., ARIA support for chat panels, colorblind-safe diff rendering) and gives HCI researchers a reusable LLM-ensemble methodology for mining accessibility complaints from unstructured issue trackers.

The system provides an MLflow-inspired autologging layer for Qiskit's transpiler that captures fine-grained provenance—per-pass execution data, transpilation stage transitions, backend characteristics, compiler configuration, and execution results—by extending the QProv provenance model with compiler-specific metadata, then persisting it to an MLflow Tracking Server for querying and visualization. This removes the need for manual instrumentation around transpilation calls, enabling systematic comparison of pass-manager configurations and backends across experiments. Practitioners studying quantum compiler optimization or benchmarking transpilation strategies could adopt this directly to get reproducible, queryable logs instead of ad hoc print-debugging.

Through a per-problem domain-expert audit of all 65 SciCode main problems, the authors identify 263 defects, 192 of which (across 91% of problems) cause correct, instruction-compliant solutions to be wrongly rejected via non-reproducible gold answers, overly tight numerical tolerances, or self-contradictory task specifications; 78% of these are described as score-suppressing (the abstract cuts off before the full mechanism is detailed). This directly explains the observed plateau where 2026 frontier models cluster near 60% subproblem accuracy and a successor model fails to beat its predecessor—the ceiling was artifactual, not capability-driven. The resulting corrected benchmark (SciCode-Verified) gives practitioners a more trustworthy evaluation surface, and the audit methodology is a template for auditing other 'plateaued' LLM benchmarks for similar defect-driven score suppression.

The authors conduct a scoping review, screening 4,259 documents down to 58 that treat system integration as core to AI evaluation, then apply reflexive thematic analysis to characterize the actors, enablers, and constraints shaping this nascent audit paradigm—modeled on integration-focused safety audits long standard in aerospace and other safety-critical engineering domains. Key findings: integration-specific risk measures are scarce, the corpus falls short of traditional audit rigor expectations, and practical access to system-level information/resources is a major bottleneck for auditors. This provides a structured research agenda and vocabulary (actors/enablers/constraints) for AI governance researchers and auditors building integration-aware evaluation frameworks beyond single-model benchmarking.

The paper introduces a cost-aware probability monad in Liquid Haskell that unifies executable probabilistic computation with refinement-type-based verification, intrinsically tracking probability mass and expected-cost values within the same monadic structure rather than requiring costs to be propagated separately through proofs. This lets expected-cost analyses of probabilistic algorithms/data structures be checked with SMT-backed automation instead of substantial manual proof effort, addressing a known pain point where probabilistic reasoning and cost reasoning are usually encoded independently. Practitioners working in dependently/refinement-typed functional verification could use this monad as a reusable building block to mechanically verify expected-time-complexity bounds for randomized algorithms (e.g., randomized quicksort, skip lists) directly in Haskell.

The paper introduces bounded step recursion, replacing ordinary predecessor with the generalized inverse ρ_φ(y)=min{z: φ(z)≥y} of a strictly increasing φ, which yields a descent schedule y, ρ_φ(y), ρ_φ^[2](y), ..., 0. Combined with a Grzegorczyk basis B_m and composition, this defines classes H^m_{n,l} parameterized by initial-function strength (m), growth scale (n), and stride through canonical layers (l). The main result is an exact criterion for H^a_{n,l} ⊆ H^b_{n',l'} for n,n'≥2, showing that below horizontal collapse, fixed-stride inclusion is governed by reverse divisibility (l'∣l) rather than numeric ordering — a non-obvious structural fact useful for anyone refining subrecursive hierarchies.

The system defines a taxonomy of Crypto-Material, Crypto-Artifacts, and Crypto-Invocations, derives a scanner-independent, extensible rule repository from it, and implements a static analyzer producing CBOM (Cryptography Bill of Materials)-oriented output for governance and post-quantum migration planning. It's evaluated on both a synthetic benchmark with ground truth and a real-world deployment of ten services, achieving an F1 of 0.75 for asset discovery — a practitioner could extend the rule repository for new languages/libraries or feed the CBOM output into existing SBOM/compliance tooling.

PPDL embeds LLM calls and tool invocations as probabilistic operations within a probabilistic-programming language, so uncertainty is propagated compositionally through a flow's control structure using standard probabilistic-program semantics rather than ad hoc bookkeeping. Inference-scaling strategies (e.g., repeated sampling, self-consistency-style aggregation) can be swapped in as backend configuration without touching flow logic, letting practitioners A/B different scaling techniques cheaply. The authors validate the approach experimentally and via a case study building a theorem-proving agent for Rocq, suggesting the language is usable as infrastructure for building auditable, uncertainty-aware LLM agent pipelines.

The paper revises the axiom set used in incremental linearization for quantifier-free nonlinear integer arithmetic (QF_NIA), specifically improving convergence behavior on polynomial constraints with higher-degree monomials (powers and mixed products) that prior axiomatizations struggled to refine efficiently. It's implemented as a standalone module layered on Z3's linear integer arithmetic core and evaluated against the SMT-LIB NIA benchmark suite. Results show overall competitiveness with state-of-the-art NIA solvers and substantial outperformance specifically on benchmarks dominated by higher-degree polynomial constraints — a practitioner working on program verification or constraint solving could adopt the revised axiom set directly or study it to extend other linearization-based solvers.

SuperScout uses a two-stage pipeline: SuperScout-7B, a small searcher model, explores the target repository and emits a structured handoff whose reproduction claims are verified in a sandbox, with unverified/false claims stripped before the handoff is finalized. The searcher's hidden states plus the original task text then feed a resume-based router that dispatches to one of four frontier 'fixer' models, and new fixers can be added without retraining the router. On the full Python slice of SWE-bench Pro (266 tasks) under its official capped-budget tier, SuperScout solves 159/266 versus 158/266 for the best single model, at roughly 1/5th the cost per solve — a practical blueprint for cost-aware model routing that front-loads cheap verification before expensive inference.

RepoProbe is a benchmark constructed from GitHub Discussions rather than Issues, targeting open-ended architectural Q&A specifically to surface Edit Bias — premature code-modification proposals that substitute for genuine repository comprehension. It replaces conventional scalar LLM-as-a-Judge scoring, which the authors note suffers from high variance and low interpretability, with a checklist-based evaluation protocol intended to be more rigorous and reproducible. The abstract is truncated before describing the full checklist-construction methodology and quantitative results, so specifics of scoring and model coverage remain to be seen in the full paper.

The authors present an adversarial skill-synthesis pipeline using six LLMs across four model families to transform 471 real-world malicious shell commands into benign-appearing agent 'skill' files, releasing a benchmark of 2,826 such skills mapped to 11 MITRE ATT&CK tactics. Their evaluation pipeline combines run stratification, evidence anchoring, a refusal veto (to avoid crediting blanket refusals as correct detection), a deterministic declared-intent override, and a three-judge LLM-as-a-judge panel validated against a blind human gold standard. This gives security researchers both a labeled adversarial dataset and a reproducible methodology for benchmarking detection/defense systems against disguised malicious instructions in agent tooling.

The study runs a fully-crossed factorial design — 3 instruction-tuned models × 5 inference frameworks × 6 benchmarks × 4 generation modes — to isolate the effect of inference backend on reported benchmark scores. Even under greedy, sampling-noise-free decoding, changing the backend significantly alters model performance, and this effect is structural (not attributable to randomness) and strongly model-dependent rather than uniform across models. Variance decomposition by generation mode indicates a considerable portion of score variance is attributable to backend choice; practitioners replicating or comparing benchmark numbers should treat inference framework and version as a reportable experimental variable, not an implementation detail.

Active-SWE reframes SWE-bench-style evaluation from reactive, single-issue repair to proactive multi-bug discovery and repair without report guidance, spanning 1,663 tasks across six bug categories and eight languages. This expands the evaluation surface from fixing one recorded bug to open-ended bug discovery and multi-bug fixing scenarios in large codebases. Practitioners can use it to stress-test whether existing SWE agents (search, static analysis, patch generation) generalize beyond localized, report-conditioned fixes. It's a natural benchmark for evaluating agentic code-review or 'self-audit' capabilities before deployment.

The study screens 36,710 GitHub repositories of engineered software projects for Markdown 'Agent Plan' artifacts (task-oriented, distinct from repo-wide AGENTS.md context files) and finds only 85 such files concentrated in 10 repositories. The authors qualitatively categorize what development activities (maintenance, design, construction, quality assurance, etc.) the plans support and what information they encode to steer agent execution. The extreme concentration (10 of 36,710 repos) is itself a finding — plan-file persistence is not yet a common practice despite growing agentic tool adoption. Useful as a starting taxonomy for anyone designing agent-plan conventions or tooling to standardize these artifacts.

The authors compile recursive Datalog programs into 2-local Ising Hamiltonians via a four-stage pipeline — binarization, grounding, reduction to Min-Ones SAT, and Ising encoding — so the ground state of the resulting model projects onto the program's minimal Herbrand model. Each rule is encoded as an energy penalty on its single violating assignment, plus a uniform per-atom cost biasing the ground state toward the minimal model. Correctness is established with per-stage lemmas and a correspondence theorem mechanically verified in Lean 4, and the models are mapped onto real annealer topologies with resource characterization. This gives a formally-verified template for running recursive logic-program inference (e.g., graph reachability, program analysis) as ground-state search on D-Wave-class hardware.

The authors construct approximate quantum error-correcting codes from the chiral edge modes of 2D topologically ordered phases, combining a gapped topological bulk (robust nonlocal encoding) with the edge's gapless conformal field theory (structural flexibility) without requiring bulk criticality. They derive an exact relation between coherent-information loss under local erasure and relative entropy of edge CFT states, reducing recoverability analysis to universal CFT data such as central charge and operator content. This yields power-law scaling of coherent-information loss with erasure size, characterizing code performance analytically. The framework gives condensed-matter and quantum-information researchers a concrete way to evaluate error-correction robustness of specific topological orders via their known edge CFTs.

The authors reanalyze megamaser distance-redshift data, correcting redshifts for peculiar velocities using a new velocity-field reconstruction that combines constrained cosmological simulations with redshift-survey data, capturing coherent bulk flows and non-Gaussian individual-galaxy motions more accurately than prior corrections. Applying this correction, they obtain an H0 estimate consistent with the Planck CMB value, in tension with higher local-distance-ladder estimates. The result implicates peculiar-velocity bias, not new physics, as a driver of at least part of the apparent Hubble Tension in this local probe. Cosmologists working on H0 systematics can apply the same reconstruction method to other peculiar-velocity-sensitive local probes, such as Tully-Fisher or surface brightness fluctuations.

The paper develops a classical, non-quantized 'constructive' derivation in which Lorentz and gauge symmetry emerge isomorphically as effective descriptions of wave-packet excitation dynamics constrained on a homogeneous, isotropic elastic substrate medium. Taking the medium's transverse wave speed as the invariant benchmark under a conventionalist synchronization scheme, the authors derive Minkowski spacetime and Lorentz transformations as effective inertial-frame coordinate changes, reinterpreting the Michelson-Morley null result via composite wave-packet observers rather than a preferred aether frame. Gauge symmetry is derived analogously from the same substrate dynamics. This sits in the lineage of aether-reinterpretation/emergent-relativity programs; readers should treat it as a foundational proposal rather than an experimentally distinguishing model, weighed against the well-established empirical equivalence of standard SR and gauge theory.

The authors propose a phenomenological logical-noise model treating the syndrome-measurement interval as a tunable control parameter rather than a fixed clock cycle, and analytically derive that the optimal interval scales inversely with code distance, yielding an exponential (in distance) reduction in logical error rate relative to constant-interval schedules. They further develop an adaptive timing strategy that uses observed syndrome activity to adjust measurement cadence online, outperforming any fixed-interval protocol especially under time-dependent or bursty idling noise. This is validated via simulation and gives a concrete, hardware-agnostic lever — scheduling policy — that QEC practitioners can layer onto existing surface-code or other stabilizer-code implementations without changing the code or physical qubits.

The authors analyze single-parameter phase estimation using heralded non-Gaussian optical states generated probabilistically (boson-sampling-type conditional operations) from deterministic Gaussian resources, accounting for generation success probability rather than treating quantum Fisher information alone as the performance metric. Using photon-number conservation in passive linear-optical networks, they show heralded-state preparation before phase encoding can be remapped to a postselection problem applied after phase encoding, an equivalence enabling exact bounds relating success probability to achievable precision. This yields a fundamental limit clarifying when probabilistic non-Gaussian resources offer genuine advantage over deterministic Gaussian strategies, versus when postselection overhead cancels the QFI gain. Quantum-optics experimentalists designing heralded-photon sensing setups can use this to assess whether a given non-Gaussian generation scheme is worth its overhead before building it.

For any z in M_n⊗M_m, the authors prove ||z||_1 ≤ √2·min(n,m)·||z||_ε, comparing the trace norm to the injective tensor norm built from trace norms on each factor, via a new L1 noncommutative Khintchine inequality with Haar-unitary random coefficients; they show √2 is the sharp constant. Applications include an exact bound on the gap between bipartite correlation measured in trace norm versus via a correlation function, and an improved universal upper bound on quantum data-hiding capacity. Operator-space and quantum-information theorists can use this bound directly as a tight comparison lemma when relating global entanglement witnesses to LOCC-restricted (product) measurement schemes.

The authors study the interacting Harper-Hofstadter model deep in the fractional Chern insulator (Laughlin-like) regime, examining bound states between a mobile impurity and a single pinned quasihole. Combining analytical arguments with large-scale numerical simulations, they characterize the bound state's structure, energetics, and stability, and identify conditions under which the binding energy provides a direct measurement of the quasihole's fractional charge. This connects to recent solid-state observations of anyon-impurity composites and to engineered-lattice (e.g. optical-lattice/neutral-atom) realizations of Laughlin states, giving experimentalists building Harper-Hofstadter quantum simulators a concrete impurity-probe protocol to implement.

The authors solve modified Tolman-Oppenheimer-Volkoff equations for a neutron star admixed with an anisotropic dark matter halo following an Einasto density profile — building on a companion result showing this profile yields regular, singularity-free black holes — using two nuclear equations of state (BSk19 and SLy4). For a specific range of halo parameters, the metric component g_rr^{-1} changes sign outside the stellar surface, producing an event horizon that encloses an intact, regular neutron-star interior, independent of the EOS choice. This yields a concrete 'neutron star inside a black hole' compact-object solution, motivating searches for observational signatures (gravitational-wave, lensing, or spectral anomalies) that could distinguish such objects from ordinary black holes or neutron stars.

The authors extract local two-body (ZZ) correlators from a low-depth QAOA circuit and use them to warm-start the Burer-Monteiro rank-two SDP relaxation for MaxCut, replacing the standard random multi-start initialization. Numerical experiments on random Erdős-Rényi graphs and a second problem class show the QAOA-informed initialization converges to high-quality solutions in substantially fewer BM iterations than random restarts. Practitioners can replicate this today: run a shallow QAOA circuit (simulated or on NISQ hardware), extract pairwise correlators, and feed them as an initialization vector into existing BM/SDP MaxCut solver code.

The authors derive quantum fluctuation relations for repeated projective-measurement protocols in which the first successful detection event triggers a work-producing mechanical operation. The correction term to the standard quantum Jarzynski equality scales logarithmically with the mean first-detection time of the time-reversed dynamics. Applying Jensen's inequality to these relations yields rigorous upper bounds on both the total work (measurements plus final operation) and the extracted work alone, with the framework extended to a device coupled to an external environment. This provides a thermodynamic-consistency toolkit — analogous to first-passage-time statistics in classical stochastic thermodynamics — for bounding the performance of detection-triggered quantum work-extraction devices.

The authors construct explicit error-detecting lattice-surgery gadgets for Iceberg codes [[2N,2N-2,2]] that fault-detect measurements of logical Pauli products, framing surgery as gauging a logical Pauli operator viewed as a symmetry of the code. They classify logical Pauli operators under the code's permutation automorphism group, reducing gadget design to one representative per orbit, and validate fault-detection behavior and expected post-selected logical error rates via circuit-level simulation. The gadgets require reconfigurable long-range connectivity (e.g., neutral-atom array shuttling), making this a concrete near-term-implementable Pauli-based computation demonstration; the paper also serves as a self-contained introduction to Iceberg codes and surgery-as-gauging.

The authors demonstrate a dual-Faraday-laser-pumped cesium beam clock using an atom-referenced laser architecture: an intracavity Faraday anomalous dispersion optical filter self-aligns to the Cs D2 line, and modulation transfer spectroscopy suppresses laser frequency noise and drift, yielding a 2.12 kHz Lorentzian linewidth. This solves the laser-induced frequency-to-amplitude noise conversion problem that previously capped SNR gains from two-laser optical pumping, delivering a clock SNR of 46,365 and short-term stability of 7.7×10^-13/√τ. Other groups building compact/deployable Cs clocks can adopt the same FADOF-plus-modulation-transfer-spectroscopy laser architecture as a drop-in replacement for standard pump lasers.

ChernFormer combines a fermionic transformer wavefunction with a fixed Chern-Simons 'flux attachment' phase (one statistical vortex per particle pair), so the antisymmetric fermionic sign times the antisymmetric phase factor produces a symmetric, bosonic wavefunction with identical probability density and approximation error. The authors prove the construction is universal — any normalizable bosonic wavefunction at fixed particle number is approximable — and show numerically, via a needle-in-a-haystack target-reconstruction benchmark, that the built-in particle-coincidence node can shrink smoothly enough for the state to approach a nodeless Bose-condensate limit. This lets an existing fermionic neural-quantum-state architecture and training pipeline be repurposed to variationally simulate both condensates and chiral topological (e.g. fractional-quantum-Hall-like) bosonic liquids without redesigning the ansatz.

The authors study strict collinear factorisation of massless n-particle amplitudes in the m-particle collinear limit, showing through four-loop order that constraints on the soft anomalous dimension from all two-particle collinear limits suffice to guarantee factorisation in general multi-particle collinear limits. They further extend the analysis to amplitudes containing a massive coloured particle, deriving new constraints on the soft anomalous dimension from demanding consistency there too. This tightens the known structure of infrared singularities used in resummation and subtraction schemes for precision QCD calculations, and the derived constraints can be used to cross-check or constrain anomalous dimension computations at high loop order.

The author analyzes pseudoreal confining N=1 SUSY gauge theories deformed by anomaly-mediated SUSY breaking (AMSB), testing the conjecture that the AMSB and non-SUSY limits share a universality class. Under this conjecture the non-SUSY pseudoreal confining theories examined have no massless spectrum, conflicting with tumbling-based expectations of symmetry breaking. Reconciling the two would require condensation in repulsive channels for three of the cases studied, and appears impossible in one case regardless of channel attractiveness, though the SUSY versions of some additional theories are consistent with flowing to superconformal fixed points in the conformal window.

The work applies Quantum Amplitude Estimation (QAE) to path-level travel time estimation in stochastic transportation networks, encoding all feasible travel-time realizations into a quantum superposition rather than drawing classical Monte Carlo samples. This avoids dependence on sample size, sampling strategy, or assumed travel-time distribution, and yields a theoretical quadratic query complexity speed-up over classical estimators (O(1/ε) vs O(1/ε²)). A practitioner could build on this by implementing the QAE circuit for specific network topologies on simulators or near-term quantum hardware and benchmarking convergence against classical Monte Carlo baselines for VRP objective evaluation.

The authors present an integrated experimental platform — laser cooling, trapping, coherent state manipulation, and precision-measurement protocol — for heavy molecules with nonzero nuclear spin, demonstrated on 137Ba19F as a benchmark for nuclear-spin-dependent parity violation (NSD-PV). This extends prior proof-of-principle work limited to effectively spin-zero nuclei, which cannot probe NSD-PV or related P/CP-violating observables. The projected statistical sensitivity is roughly two orders of magnitude beyond comparable measurements, positioning the platform as a template for extending precision molecular tests of nuclear symmetry violation to a broader class of spin-bearing species.

The paper derives a master equation, to leading order in Newton's constant G, for two masses coupled to an environment of linearized gravitational waves, using proper distance (rather than naive Fermi normal coordinate variables from the standard Lagrangian) as the physically meaningful observable via a unitary system-environment decomposition. The dissipative sector reproduces classical gravitational-wave energy loss, while the noise/decoherence sector suppresses coherences between states differing in mass quadrupole moment or proper separation. This gives a concrete, first-principles framework for computing gravitationally-induced decoherence rates that others can apply to specific mass configurations or use to bound gravity-induced collapse models.

The Hadamard Sensing Channel (HSC) replaces probabilistic phase-randomization (PR) in dynamical decoupling sequences with a deterministic phase-design scheme built from Hadamard matrices, exactly cancelling pulse-duration-induced spurious signals using a finite set of orthogonal phase patterns instead of relying on statistical averaging over random phases. Simulations show HSC matches PR's ideal artifact suppression while being more robust to control errors, and eliminates the variance/sequence-length tradeoff inherent to probabilistic averaging. Practitioners building dynamical-decoupling-based sensing protocols (e.g., NV-center magnetometry) could substitute HSC's fixed Hadamard-derived phase sets directly into existing pulse sequences to reduce required averaging without new hardware.

The authors work in the perturbative variable-scalar-curvature branch of asymptotically flat symmergent gravity, where an R² correction sourced by the boson-fermion field-content imbalance renders the exterior metric conformal to Schwarzschild via a radial mode obeying a linear equation with an independent boundary-condition-dependent amplitude. Depending on the sign of the symmergent parameter γ, the deformation is either Yukawa-suppressed (γ>0) or oscillatory in inverse radius (γ<0); they derive radial equations, effective potentials, and circular-orbit/marginal-stability conditions for neutral and charged particles, plus collision energetics and photon frequency shifts, giving observables that could in principle constrain γ against Schwarzschild-based data.

The authors formulate an open-worldsheet junction framework for quantum Hall (QH) interfaces, treating a freely moving interface as a string whose endpoint termination on a physical edge or topological boundary is specified by geometric support, variational boundary data, condensable topological sectors, and outgoing channels for absorbing/continuing worldsheet flux. This extends the previously established charge-shape relation from closed loops to an open interval and yields an operational, general definition of a 'QH brane.' A key structural result is a no-go argument that a lone chiral Majorana mode cannot terminate on any finite-dimensional endpoint degree of freedom, giving researchers modeling QH edge networks or endpoint defects a systematic basis for constructing consistent endpoint/junction theories.

The authors construct a spatially reparametrization-invariant effective theory for a chiral Majorana mode living on the nonrelativistic worldsheet of a fluctuating Moore-Read interface. They show the changing line element enforces a universal half-density transport law for the neutral fermion, with additional curvature- and velocity-dependent terms left as UV-sensitive (microscopic) couplings. The Majorana stress tensor emerges as the operator mediating between the neutral sector and the geometric (charged/shape) degrees of freedom. This completes the neutral-sector input needed for effective field theories of dynamical non-Abelian quantum Hall interfaces, building directly on the companion Part I kinematic construction.

The authors formulate a spatially reparametrization-invariant worldsheet description for interfaces between Abelian quantum Hall phases, distinguishing tangential motion (pure relabeling/gauge) from normal motion (physical, since it redistributes area between incompressible phases). Starting from the two-sided Chern-Simons response, they derive the exact relation linking normal interface velocity to charge transport across the boundary. They then introduce a 'relative-area' construction referenced to a fixed material curve, converting this velocity relation into an equal-time constraint that couples the charged boundary sector to the interface geometry — setting up the kinematic and constraint structure used in Part II to add the neutral Majorana sector for non-Abelian interfaces.

The paper gives a query-optimal algorithm for simulating time-ordered evolution under a Lipschitz-continuous, norm-bounded n-qubit Hamiltonian H(t) on [0,T] in the HAM-T access model, achieving O(αT + log(1/ε)/log(e+log(1/ε)/(αT))) queries — matching the known lower bound for time-independent Hamiltonians and proving time-dependence adds no asymptotic overhead. The construction builds a one-query 'transducer' circuit that approximates U_H(T) via an auxiliary state that is otherwise returned unchanged, then takes a weighted combination of circuits applying the transducer varying numbers of times so the truncation error from omitting the auxiliary state decays factorially. This gives a drop-in replacement for time-independent Hamiltonian simulation subroutines wherever H is time-dependent but smooth, useful for adiabatic/counterdiabatic protocols and driven-system simulation.

The authors analyze small anomaly-mediated supersymmetry breaking (AMSB) deformations of supersymmetric 't-confining gauge theories — models where the confined degrees of freedom are captured by a set of composite chiral superfields with no unbroken gauge symmetry. They determine the resulting global symmetry breaking pattern induced by the AMSB deformation in these strongly coupled confining vacua. The results are proposed as benchmarks comparable to lattice simulations of the non-supersymmetric analogues, with one case complicated by a pseudoreal/chiral matter content that alters the expected symmetry-breaking pattern.

Using a high-precision Chebyshev spectral method, the authors compute scalar, vector, and tensor quasinormal mode (QNM) spectra of Gauss-Bonnet-corrected Schwarzschild black holes across spacetime dimensions D=5,6,7,8,10,11,12,26, extending well past where 6th-order WKB and characteristic-integration methods remain accurate. Key findings include the emergence of overdamped purely imaginary modes, non-monotonic real-part behavior in higher overtones, and strong amplification of dimensionless QNM frequencies at string-theory-motivated dimensions. They also identify and analytically prove an exact isospectrality between scalar ℓ=0 and vector ℓ=1 modes at zero GB coupling, providing both a benchmark dataset and a proof technique other researchers can extend to other higher-curvature corrections or dimensions.

The paper analyzes routing overhead in mapping symmetric quantum circuits onto restricted hardware topologies, decomposing the imbalance into a two-level effect: (1) whether a qubit patch's geometric shape admits any balanced logical-to-physical assignment, and (2) given that a balanced assignment is possible, how the specific assignment realizes it. Via exhaustive search on IBM's 57-qubit heavy-hex lattice, they show 108 of 124 connected four-part-ring patches admit a cost-free balanced assignment (the 16 exceptions all being star-shaped), while cost-free balance is proven impossible for six-part rings on compact patches. This gives circuit compilers concrete, checkable criteria for choosing qubit patches and assignments that minimize SWAP-induced asymmetry in symmetric algorithms (e.g., variational circuits with permutation symmetry).

The authors present an analytic framework for multi-component FLRW cosmologies driven by a single scalar field whose potential encodes the fluid's energy density and pressure, unifying Big Bang, inflationary, and dark-energy epochs. The key move is expressing the potential via the Hubble function H(φ) acting as a 'fake superpotential,' converting the second-order Friedmann/scalar-field system into a first-order (Hamilton-Jacobi-like) problem solvable analytically in a suitable time coordinate. They construct new integrable inflationary models compatible with observational constraints and compute scalar and tensor perturbations by integrating the Mukhanov-Sasaki equations within this exact background, offering a template other researchers can extend to build additional exactly solvable inflation/dark-energy models with computable perturbation spectra.

The authors compute the lowest radial pulsation eigenmode of spherically symmetric boson stars (mini boson stars and quartically self-interacting variants) along equilibrium sequences at fixed node number, using additive variables in the pulsation equations that remain regular at the background scalar field's zeros — enabling direct eigenvalue integration through excited-state nodes. They demonstrate that the first zero-crossing of the constrained fundamental radial eigenvalue coincides, within numerical precision, with the first simultaneous critical point of the ADM mass, Noether charge, and binding energy along each branch, confirming a turning-point stability criterion for excited states. They further extract an empirical correlation between this critical eigenvalue and node number/self-interaction strength using threshold configurations from nonlinear time-evolution simulations, giving a cheap linear-stability proxy that practitioners can use instead of costly full nonlinear evolutions.

The authors define a directed participant graph whose edges are events, apply a line-graph transformation to obtain a bidirectional event graph carrying relational structure, and assign a quantum harmonic oscillator (QHO) to each participant-graph node, so raw event attributes map onto event-graph nodes. The collective Hilbert space of all QHOs serves as a complete basis for quantum states encoding the graph. This sets up machinery for quantum-native anomaly detection or quantum kernels over event data, useful groundwork for anyone building quantum ML pipelines on relational/event datasets.

For bosonic open quantum dynamics governed by the GKSL (Lindblad) master equation, stochastic sampling via P/Wigner/Husimi quasiprobability representations requires the corresponding Fokker-Planck diffusion matrix to be positive semidefinite. Starting from a path-integral formulation, the authors derive sufficient conditions for positive-semidefiniteness that hold for arbitrary Hamiltonians and jump operators, resolving previously ambiguous general criteria. Practitioners can use these conditions as a pre-check before committing to truncated-Wigner or P-representation simulations of many-body bosonic systems, avoiding wasted computation on ill-posed stochastic equations.

The paper develops an ODE/IM (ordinary differential equation / integrable models) correspondence for Hermitian and PT-symmetric homogeneous oscillators, expressing the quantization condition through a counting function a(E) obtainable from the Destri-de Vega equation. Thermal partition functions and spectral zeta functions are then written as contour integrals of a(E), giving closed-form links between spectral problems and integrable-model TBA-type equations. This gives a route to compute exact partition functions and zeta-regularized quantities for whole families of oscillator potentials, useful for testing numerical spectral methods or extending to other non-Hermitian systems.

The paper classifies static, spherically symmetric, asymptotically flat axion-dilaton solutions of the tree-level 4D string effective action as the SL(2,R) orbit of the pure-dilaton FJNW solution, characterizing the axion-dilaton charge space. They compare the onset of α' curvature corrections against string-loop corrections in the perturbative regime, then apply the string–black-hole correspondence criterion (evaluating a local α' curvature diagnostic at the typical size R_typ of a highly excited string state), calibrated against the Schwarzschild branch. This gives a concrete testbed for probing where classical hairy black hole solutions break down into stringy self-gravitating states.

For the ground state of an infinite hopping chain subjected to partial projective measurements in the occupation basis, the reduced density matrix of a segment bounded by measurement regions is mapped to a grand-canonical state via a conformal transformation plus gauge transformation in the field-theory description. The resulting entanglement Hamiltonian has a local inverse temperature vanishing as a square root near the endpoints, independent of the measurement outcome, while the local chemical potential tracks the induced charge density and is outcome-dependent. This shows the post-measurement entanglement Hamiltonian encodes far more outcome-specific information than the entanglement entropy alone, relevant for measurement-induced phase transition studies.

Working within the Wilson-Polchinski renormalization-group framework, the authors build a second-order mean-field approximation to the connected amputated Schwinger functions of 4D Euclidean phi^4 theory, decomposing them at symmetric momentum configurations into a constant term, a p²-term, and a remainder, with the first two obeying a closed nonlinear hierarchy. They prove existence of solutions to this hierarchy for arbitrary positive bare coupling and show both the constant and quadratic sectors converge to the Gaussian fixed point as the UV cutoff is removed, establishing triviality in this non-perturbative approximation. This provides rigorous, cutoff-removed evidence supporting the long-conjectured triviality of 4D phi^4 theory beyond perturbation theory.

The work targets nuclear-spin cat codes in donor-in-silicon architectures, where high-dimensional spin encodings intrinsically suppress phase (dephasing) errors, the dominant noise channel there. They construct a universal logical gate set that is error-transparent (ET) to phase errors, meaning stochastic phase errors during gate execution propagate in a systematically traceable way and remain correctable by subsequent QEC rounds, rather than being scrambled into uncorrectable errors. The logical X gate is identified as the main implementation bottleneck, with candidate realization schemes discussed. This gives a concrete blueprint toward fault-tolerant operation on spin-cat-code qubits, of direct use to groups building donor-in-silicon quantum processors.

Using the non-perturbative two-particle self-consistent (TPSC) approach, the authors study the 2D Hubbard model with nearest-neighbor hopping at interaction strengths below the Mott transition, near the quantum critical point separating a Fermi liquid from incommensurate spin-density-wave order. For correlation lengths spanning roughly 1 to 100 lattice spacings, Kohn anomalies on the Fermi surface are predicted to generate unconventional critical exponents, and the temperature dependence of the single-particle self-energy acquires strong momentum dependence along the Fermi surface. This gives falsifiable, quantitative predictions for strange-metal behavior directly comparable to ongoing cold-atom and diagrammatic quantum Monte Carlo experiments on the same model.

The authors couple the spin-1 Haldane chain to an Ohmic dissipative bath and use large-scale QMC to map a second-order quantum phase transition into a SO(3)-broken antiferromagnetic phase governed by an interacting fixed point with dynamical exponent z≈2. They construct a generalized string order parameter showing the SPT ground state survives weak dissipation but acquires a nontrivial scaling dimension at criticality, with edge modes exhibiting distinct boundary criticality versus a trivial-state transition; the latter is cross-checked via an ε-expansion of a dissipative φ⁴ theory with ordinary boundary conditions. This realizes a tractable model of a spin impurity embedded in a critical, nonconformal antiferromagnet, useful for benchmarking dissipative quantum criticality methods.

EMRI parameter estimation is notoriously hard because the six phase-evolution parameters produce an exceptionally sharp global likelihood maximum surrounded by dense secondary maxima across a broad astrophysical prior — a problem that has defeated blind search efforts in the Mock LISA/LISA/Taiji Data Challenges. The authors show secondary maxima are not uniformly distributed but concentrate increasingly near the true global maximum as likelihood increases, enabling an adaptive contraction of the search volume. They operationalize this via a reduced-dimensional profile likelihood, giving a concrete, generalizable algorithm for coherent end-to-end EMRI detection and recovery that other pipeline developers could adopt or benchmark against.

The paper presents the full-color three-loop five-point integrand for N=4 SYM in a manifestly color-kinematics-dual (CK-dual) form, whose double copy yields the corresponding N=8 supergravity integrand. For four-dimensional external states they extract UV poles in the critical dimension Dc=6 for both theories, and explore extending external-state dependence to general D by replacing 4D prefactors with D-dimensional tree-amplitude-built expressions; this reproduces the open-string prediction for the SYM UV pole but the gravity analogue departs from the string-inspired guess by an evanescent term. This gives practitioners a concrete high-loop, high-multiplicity CK-dual integrand and a data point on where naive double-copy/string-inspired extrapolations to general dimension break down.

For idealized levitated-particle spatial-superposition protocols, the authors compare the mass-proportional CSL contrast-loss exponent against the Diósi-Penrose self-energy exponent E_Gτ/ħ. They prove the point-particle CSL separation kernel admits an exact random-unitary (stochastic Gaussian momentum-kick, Poisson-timed) realization reproducing the same unconditional coherence decay, demonstrating that the decoherence kernel alone is observationally degenerate with non-collapse noise models even though the full conditional state remains pure. They further show the CSL/DP exponent ratio is independent of mass and interrogation time, depending only on branch separation in the point-particle model — a result that sharpens which experimental observables can actually discriminate between collapse models versus mimic them with mundane decoherence.

The authors compute the entanglement dynamics between two initially decoupled 1D Luttinger liquids that are locally tunnel-coupled along their entire spatial extent, contrasting this 'field-space' partition with the usual spatial-bipartition case where entanglement grows only from the cut. Working within a Gaussian (quadratic bosonized) approximation for both zero- and finite-temperature initial states, they derive closed-form analytical expressions for the growth of entanglement entropy (logarithmic negativity/entanglement entropy) after a time-dependent tunneling quench is switched on uniformly. This provides an exactly solvable benchmark for entanglement growth under globally-distributed coupling, directly applicable to coupled cold-atom wire experiments and to testing conformal-field-theory-based entanglement spreading predictions beyond the standard local-quench setup.

The authors study the generalized dissipative Dicke model realized in a trapped Rydberg-ion array, where density-density (Ising-like) interactions between ions compete with collective spin-phonon coupling, laser driving, and dissipation. Mean-field analysis of the phase diagram reveals multiple distinct phases and a tricritical point whose location depends sensitively on the interionic interaction strength; finite-size quantum dynamics simulations tracking spin averages, entanglement entropy, and Loschmidt echo reveal dynamical-phase-transition signatures including slow relaxation and metastability near the transition. This gives a concrete trapped-ion protocol (tunable Ising coupling, spin-phonon coupling, drive, and dissipation) for experimentally probing dynamical criticality and tricriticality in an open collective quantum system.

The authors examine sphalerogenesis — baryon asymmetry generation via CP-violating decoupling of electroweak sphaleron-like transitions — within the SMEFT framework, introducing seven CP-violating dimension-8 operators built from the Higgs doublet and SU(2)_L gauge fields. They show five of the seven can individually reproduce the observed baryon asymmetry while satisfying collider and electron EDM constraints, and compare their contributions against a CP-violating dimension-6 operator, finding the dimension-8 terms can be comparable in size when loop-generated. This demonstrates that loop-level matching coefficients materially change the power-counting expectations for higher-dimension SMEFT operators in baryogenesis phenomenology, giving a concrete operator basis for future EDM/collider constraint studies.

The authors introduce two distinct entanglement-focused quantum Mpemba effects — one targeting rapid attainment of a threshold entanglement value, the other targeting rapid convergence to the asymptotic steady-state value — and show both effects are measure-dependent, giving qualitatively different results depending on which entanglement quantifier is used. They demonstrate exponential acceleration (or deceleration of decay) of entanglement generation is achievable via appropriately engineered initial conditions/dynamics, and validate the effect in many-body settings using cluster elimination methods applied to a dissipative long-range Ising chain. This suggests a concrete design principle — exploiting Mpemba-like non-monotonic relaxation — for speeding up entanglement resource generation in dissipative or noisy many-body quantum platforms.

The authors reformulate the Newtonian N-body problem in a fully relational framework, eliminating absolute position, orientation, time, and scale so only dimensionless shape ratios are physical. The scale-invariant 'variety' V emerges as a key observable, functioning as a structure measure, a partial order on shapes, and the source of an emergent gravitational arrow of time. They show this unifies previously distinct classes of Newtonian solutions and hints at a new underlying symmetry, suggesting a relational, explanation-based reformulation of classical gravity that could be extended toward shape-dynamics approaches to general relativity.

Using hierarchical mean-field theory with enlarged variational clusters (beyond the standard unit cell), the authors map the phase landscape of hard-core bosons on the square-kagome lattice, recovering the known compact-localized-state phases at densities 5/6 and 2/3 and uncovering additional incompressible states, notably a robust ρ=3/4 phase. Via the Matsubara-Matsuda mapping this boson density plateau corresponds to a half-magnetization plateau in the spin-1/2 XXZ model, and the framework is applied to exchange parameters from real square-kagome compounds to make experimentally testable predictions.

The authors present optical/IR counterpart catalogs for MeerKLASS L-band and UHF-band continuum sources, cross-matched against KiDS DR5 and DESI Legacy Imaging Surveys DR10 using their new SEDA (Stellar-mass Enhanced Density Association) method. SEDA compares candidate source densities around radio positions to those around position-displaced control fields, incorporating positional offset, stellar mass, and redshift for galaxy associations, and offset plus mid-infrared color selection for quasar candidates, with a second-pass step handling multi-component radio morphologies. The output — counterpart probabilities, redshifts, and host-galaxy properties — is directly reusable for cross-correlation cosmology and radio source population studies.

The paper computes, to first order in the deformation parameter, the gravitational effective action induced by non-local TT̄-like deformations of QFTs, using a heat-kernel expansion to extract local curvature terms from stress-tensor two-point functions. They apply this to free fermions, massive Maxwell theory, and second-order Yang-Mills, finding model-dependent coefficients in general, but a universal sector for CFTs fixed entirely by the central charge C_T, yielding finite, scheme-independent curvature invariants after renormalization. They also analyze trace-trace deformations, showing the relevant contact terms are fixed by the Weyl anomaly, and derive the resulting finite gravitational action for a broad class of minimally coupled theories.

The authors prove a sharp almost-sure path localization theorem for the derivative martingale limit of branching Brownian motion, showing the limit is determined by particles whose trajectories remain within a tube of width s^{1/2} around the extremal (leading) particle's path. This refines the known extreme-value/traveling-wave theory of BBM fronts (Bramson, Lalley-Sellke, Aïdékon-type results) by identifying the precise trajectory-level contribution structure, giving a tool that could be extended to other branching random walk or FKPP-front localization problems.

The authors resolve Babai's minimal Cayley graph problem by exhibiting finite Cayley graphs that are minimal (in the sense relevant to the problem, e.g. no proper generating subset preserves the graph's defining property) yet have unboundedly large chromatic number, likely via a construction combining algebraic group choices with combinatorial coloring lower bounds. This settles a longstanding question about the interaction between algebraic minimality constraints and graph coloring complexity, and the construction technique may be reusable for related extremal Cayley graph problems.

The paper stabilizes multidimensional continuity equations with source terms via localized interior feedback prescribed through the divergence of the flux, so the error relative to a reference profile obeys a transport equation with localized damping. The core geometric condition is a finite-time characteristic damping inequality — relevant characteristics must accumulate a uniform amount of damping over a bounded time horizon — which, combined with a gain condition balancing damping against compressive amplification of the transport field, yields exponential L2 stability. Lyapunov-type entrance conditions and a weighted Lyapunov functional provide constructive tools (a differential Lyapunov criterion and an input-to-state stability estimate) that a control engineer could use to design and verify localized feedback laws for PDE-governed transport systems.

For n=2m, k=2t, and n>10k^7, the authors show that any family of k-subsets of an n-set with pairwise even intersections has size at most C(m,t), with every extremal family having atomic structure — sharply improving on Frankl and Tokushige's earlier bound which required n exponential in k. The main tool is Delsarte's linear programming method, a technique from coding theory/association schemes, applied here to get much tighter uniform 'eventown' bounds, suggesting LP bounds can be pushed further for related intersection-restricted set system problems.

This is an extremal combinatorics result: among all n-vertex tournaments, which maximizes the count of directed l-cycles asymptotically. Grzesik–Král'–Lovász–Volec settled the case l not divisible by 4 (random tournament wins); this paper resolves l ≡ 0 mod 4, showing the carousel tournament (a circulant-style construction) strictly beats the random tournament, confirming the Bartley–Day conjecture. The proof likely leverages flag-algebra-style counting arguments standard in this line of extremal tournament research.

The model is a stochastic time-nonlocal telegraph equation with a (PC_ε)-type memory kernel, space-time white noise W, linear-growth drift Ψ, and Hölder-continuous, uniformly nondegenerate diffusion coefficient Φ. The authors introduce a 'hereditary phase-space lifting' framework to convert the nonlocal-in-time convolution structure into a tractable state-space system, plus a generalized coupling method with a new damping-control construction for the velocity term, yielding the first weak existence and uniqueness results for this class of equations — potentially extensible to other memory-kernel SPDEs modeling high-frequency signal propagation.

For a smooth Deligne–Mumford stack (U, w) with regular function w extended possibly only rationally on a compactification, Yu's construction defines a filtration on twisted de Rham cohomology, computed via Kontsevich lattices. The authors prove this filtration is compactification-independent under only a local nondegeneracy condition near the polar divisor, via good resolutions and stacky weak factorization, reducing the problem to blowups and root constructions along boundary divisors, with filtered comparison theorems for the Yu and Kontsevich complexes established for both operations — foundational for applications like Harder–Lee's orbifold irregular Hodge numbers in stacky Clarke mirror pairs.

The mean absolute deviation of binomial, Poisson, negative binomial, and hypergeometric distributions collapses to a single point-mass value; the authors give a unified telescoping proof and a size-biasing interpretation of the closed forms. They then derive complete asymptotic expansions (Poisson as λ→∞, negative binomial as r→∞, hypergeometric as N→∞) with coefficients in closed Bernoulli-polynomial form capturing the lattice displacement of the mean exactly; at integer means these reduce to sign-alternating series, and a Binet-kernel argument shows the partial sums enveloping (bracketing) the log of the normalized MAD — a tool useful for tight numeric approximation bounds.

For the fractional Allen-Cahn layer solution Φ_s solving (-Δ)^s Φ_s = Φ_s - Φ_s³ with Φ_s(±∞)=±1, the authors prove the energy E(s) is continuous and strictly decreasing on (1/2,1], with explicit asymptotics: E(s) = 2√2/3 + κ₁(1-s) + o(1-s) as s→1 (recovering the classical Allen-Cahn value), and E(s) = 1/(π(s-1/2)) + O(1) as s→1/2⁺ (divergence at the lower endpoint). Strict monotonicity is established with computer assistance, reducing an interior subinterval to finitely many inequalities verified via rigorous interval arithmetic — a template reusable for other nonlocal variational energy monotonicity proofs.

The authors relate dual braid group generators, Hecke algebra images of pure braids, and factorization problems in reflection groups to knot invariants, proving the (a, z=0)-HOMFLYPT polynomial can be computed as a solution to such a factorization problem — a decategorification of prior categorified knot invariant work, motivated by Coxeter–Catalan combinatorics. As an application, they give a new proof of EL-shellability of the noncrossing partition lattice using the image of the full twist in the Hecke algebra, with its inverse computing the homotopy type (a combinatorial reciprocity); positive powers of the full twist also naturally generate noncrossing partitions, connecting to cluster complexes.

Muon is a matrix-aware optimizer; the authors derive the exact closed-form solution for its update when constrained to the Stiefel manifold (matrices with orthonormal columns), replacing prior heuristic, approximate, or iterative retraction-based extensions. This yields Skewon, an efficiently implementable algorithm with proven first-order convergence guarantees in the smooth non-convex setting — directly usable as a drop-in optimizer for orthogonality-constrained training scenarios (e.g., orthogonal weight parametrizations in RNNs or normalization-sensitive architectures), and a good candidate to implement and benchmark against existing Stiefel-Muon variants.

Building on Averkov & Schymura (2022), the authors improve the bound on the number of distinct columns of a rank-m Δ-modular matrix A ∈ ℤ^(m×n) (where the max absolute value of every m×m minor equals Δ) from O(m⁴Δ) down to O(m³Δ), via a refined analysis using Sauer(-Shelah)-type matrices. This tightens a key combinatorial parameter used in complexity bounds for integer-programming algorithms that exploit bounded-subdeterminant (Δ-modular) structure, such as proximity and sparsity arguments in ILP solvers.

The authors prove Nisnevich descent for both continuous and Kerz-Saito-Tamme analytic K-theory of rigid analytic spaces, and combine this with A^1-invariance (given resolution of singularities) to show representability in the Dahlhausen-Yaylali A^1-homotopy category of rigid spaces. They identify the representing spectrum with both Z×BGL and the analytification of algebraic K-theory, yielding representability with light condensed spectra coefficients as a corollary. They also establish Weibel vanishing and prove A^1-invariance of continuous K-theory on local Tate pairs without regularity hypotheses, extending known results beyond the smooth/regular setting.

The method infers a generalized ODE system governing latent dynamics by recasting observed diffusive/irreversible processes as deterministic, measure-preserving (Liouville/Hamiltonian-like) flows in an augmented state space, where the observed marginal density is the projection of the joint density satisfying the hyperbolic Liouville equation. A stochastic gradient algorithm fits this joint dynamics from marginal density observations alone, effectively converting stochastic/irreversible inference into deterministic ODE recovery with latent variables absorbing the randomness and memory effects. This is applicable to systems where only coarse-grained, low-dimensional observations are available but a higher-dimensional deterministic generator is suspected, offering a route to build interpretable, invertible dynamical models from noisy time-series data.

The paper analyzes stochastic extragradient (SEG) for monotone variational inequalities, contrasting same-sample SEG (S-SEG, reusing one stochastic oracle draw across both extragradient steps) with independent-sample SEG (I-SEG). They prove S-SEG can diverge even on compact domains under mean-Lipschitz and bounded-variance assumptions alone, showing sensitivity to samplewise Lipschitz constants that I-SEG analyses typically ignore. They then derive high-probability convergence guarantees for possibly unbounded domains, sharpening existing theory beyond prior compact-domain or uniformly-bounded-variance assumptions and clarifying which sampling scheme practitioners should choose for stochastic VIP solvers.

The paper rigorously justifies the vanishing viscosity limit for 1D compressible Navier-Stokes in the regime where two shocks from the same characteristic family of the Euler system collide, producing a single same-family outgoing shock plus a rarefaction wave in the other family — a qualitatively different and more singular outcome than cross-family collisions. The key technical difficulty is that the collision timescale scales inversely with wave strength, complicating uniform-in-viscosity estimates both away from and precisely at the collision point, where the wave-pattern transition (shock+rarefaction emerging from shock+shock) introduces additional error terms to control. This extends the toolkit for justifying inviscid limits through nonlinear wave interactions, relevant to anyone doing rigorous asymptotic analysis of viscous approximations to hyperbolic conservation laws.

The authors analyze majority dynamics on a resampled sparse Erdős–Rényi graph G(N,p) with p=b log N/N, b>1 fixed, where the interaction graph is redrawn independently each round. They identify three regimes based on initial advantage Δ₀: above a constant multiple of N/√(log N), blue reaches unanimity within two updates whp; in the intermediate range √(N/log N) ≪ Δ₀ ≲ N/√(log N), they derive matching high-probability upper/lower bounds on the unanimity time; and in the critical window Δ₀√p = O(1), they obtain a scaling limit, likely Gaussian, characterizing the winner-selection probability. This gives sharp phase-transition results for consensus formation in dynamic random graph models, useful for researchers studying voter-type dynamics or epidemic-like spreading on time-varying networks.

The paper classifies equisingular families (pencils) of curves on smooth simply connected projective surfaces whose dual varieties have degree ≤ 6, restricting to configurations where a large proportion of curve components are hyperplane sections or components thereof. Leveraging this classification, they compute the fundamental groups of the complements of such curves, specifically for those whose components satisfy the stated restriction and whose fundamental groups admit an essential surjection onto a free group of rank greater than 6. This contributes new computed examples to the long program of understanding fundamental groups of curve complements, useful for researchers studying braid monodromy, Zariski pairs, or surface topology invariants derived from plane curve arrangements.

The paper resolves all remaining open cases for the minimum percolating set size in 3-neighbor bootstrap percolation on rectangular grid graphs P_m□P_n, extending prior work by Dukes, Noel, and Romer that left some cases undetermined. They further study the analogous problem on toroidal grids C_m□C_n, proving upper and lower bounds that differ by at most one, and pinning down exact values in many divisibility-dependent cases. This completes a combinatorial extremal problem for a well-studied cellular-automaton-like spreading process, providing exact minimum seed set formulas usable as benchmarks or building blocks for percolation threshold analyses on other grid-like graph families.

The authors characterize null systems (compact metrizable systems with vanishing topological sequence entropy along every time sequence) via polynomial growth of maximal pattern complexity for every finite open cover, and equicontinuity via sublinear growth of the same quantity, deriving these from finite fat-shattering dimension bounds at every scale and polynomial empirical covering numbers of orbit-distance classes. They also construct transitive but nonminimal null systems exhibiting properties previously thought excluded outside the minimal setting: one uniformly rigid with two fixed points, another exhibiting two-scattering (a strong mixing-type property). These results resolve several open problems on polynomial maximal pattern growth and provide new combinatorial tools (VC-theory-style complexity measures) for classifying topological dynamical systems by entropy-adjacent invariants.

The authors analyze principal eigenpairs for a periodic nonlocal dispersal operator with advection, where resolvent-positivity holds but resolvent-compactness fails, blocking direct Krein-Rutman application. For advection of constant sign they establish existence and uniqueness of the principal eigenvalue and its normalized eigenfunction. For sign-changing advection, they characterize a bifurcation-like dichotomy: depending on coefficients, the eigenproblem has either a unique normalized solution or a continuum of solutions, with boundary eigenvectors exhibiting a singular measure component. This suggests new tools beyond classical compact-operator spectral theory are needed for nonlocal advective models, relevant to reaction-diffusion and population dynamics researchers working with nonlocal dispersal.

The paper establishes a linear recurrence for the count of Schreier sets formed from single intervals, then extends to J_{k,n}, the sets formed from exactly k pairwise-separated intervals, showing the counting sequence satisfies characteristic polynomial p_k(x)=(x-1)^{2k+1}(x+1)^k. They further define k-super-Schreier sets S_{k,n} and prove their counts obey a Fibonacci-type recurrence plus a polynomial remainder term in n. The results give explicit generating-function-style structure connecting Schreier set enumeration to Catalan-number combinatorics, providing new closed forms a combinatorialist could extend to related restricted-set enumeration problems.

The paper improves the known bounds on c_claw = limsup u(n)/n, the guaranteed fraction of vertices that any n-vertex claw subgraph must embed within every n-vertex tournament, from the 1998 result 19/50 ≤ c_claw ≤ 11/23 to the sharper 2/5 ≤ c_claw ≤ 10/21. It introduces parameters θ and σ that unify the lower- and upper-bound constructions in a common framework, proving σ≤θ with 1/21≤σ≤θ≤1/5, and notes that σ=θ would settle whether the limit (not just limsup) exists. This gives extremal combinatorics researchers a tighter target and a structural bridge between the two bounding techniques for future improvement.

Building on Beisegel et al.'s (SWAT 2024) simultaneous C-number framework — where a graph G's simultaneous number relative to class C is the minimum label-set size d admitting H∈C and L:V(G)→P({1,...,d}) such that adjacency in G equals adjacency in H intersected with nonempty label overlap — this paper characterizes which graph parameters p satisfy: p is bounded on C iff p is bounded on graphs of fixed simultaneous C-number. They show many standard width/degeneracy-type parameters transfer this boundedness property. This gives algorithm designers a systematic way to determine when parameterizing by simultaneous C-number yields tractability results inherited from known results on C itself.

For the strong Weil parametrization π_E: X_0(M)→E with Manin constant c_E, the paper proves deg(π_E) divides c_E^{Ω(N/M)} · deg(φ) for every morphism φ: X_0(N)→E' with E' isogenous to E and M|N, where Ω counts prime factors with multiplicity; when c_E=1 this reduces to unconditional divisibility, and it's unconditional whenever M is squarefree by the known semistable Manin constant case. A second result shows that when the relevant Manin constant is 1, degeneracy-map-induced old homomorphisms form an integral basis of the full Hom group to a fixed target, so the old degree matrix determines the exact degree spectrum — a tool researchers computing modular degrees or testing Manin constant conjectures can directly apply.

The paper computes exact values z_wL(m,3) for m=5,...,9 in the weak limited augmented Zarankiewicz problem, finding z_wL(m,3)=2m throughout this range. Lower bounds come from explicit weak admissible constructions; matching upper bounds for m=7,8,9 come from exhaustive finite search over canonical C_4-free base graphs. The result supports the general conjecture z_wL(m,3)=2m for all m≥5, and since BSR(m,n)≥z_wL(m,n), it yields BSR(m,3)≥2m for these cases — directly usable by researchers studying Zarankiewicz-type extremal problems or the related Bipartite Sum-Rank (BSR) bound.

The paper studies recognition of {2}-Roman graphs — those achieving equality γ_{R2}(G)=2γ(G) between {2}-Roman domination number and ordinary domination number — restricted to 'Partner Limited' graphs, extending the 2017 tree characterization (Henning et al.) and 2025 characterizations via minimum {2,0}-valued dominating functions (Ferrari et al., Bešter Štorgel et al.). The likely contribution is a structural or algorithmic characterization enabling polynomial-time recognition of {2}-Roman graphs within this graph class, of direct use to researchers extending domination-parameter recognition algorithms to other restricted graph families.

For 3D axisymmetric stationary Navier-Stokes D-solutions, the authors derive a new pointwise Calderón-Zygmund estimate tailored to cylindrical geometry, improving on Carrillo-Pan-Zhang's (2020, JFA) decay bounds to |∇u_r|+|∇u_z| ≲ r^{-5/4}[log(e+r)]^{5/4} and |ω_r|+|ω_z| ≲ r^{-9/8}[log(e+r)]^{9/8} for large r. They then develop a new Liouville-theorem approach improving axisymmetric criteria of Wang (2019, JDE) and Zhao (2019, Nonlinear Anal.), proving triviality of D-solutions under decay conditions on sup|u| without requiring symmetry assumptions in the improved criterion. This gives fluid-dynamics PDE researchers both a sharper local estimate technique and weaker sufficient conditions to extend toward the still-open general 3D stationary Liouville problem.

The authors use the spatial Markov property of a Poisson process η with intensity measure λ to derive distributional identities relating the restrictions of η and λ to a data-dependent stopping set Z and its complement, extending Efron-type identities beyond the classical convex-hull setting. The convex hull of a finite Poisson point process in Euclidean space is treated as a key special case, recovering and generalizing known relations between vertex count and volume. Notably the framework extends to general Poisson hulls and even to random sets that need not be bounded or stopping sets, widening applicability to stochastic geometry problems involving unbounded or non-adapted random regions. Practitioners in stochastic geometry could use these identities to compute moments or distributional properties of geometric functionals of random convex bodies and their generalizations.

The paper addresses distributed fault diagnosis in partially-observed discrete event systems under transmission delay, extending prior decentralized diagnosability conditions to a distributed setting with communicating agents. It introduces a 'delay recorder' automaton structure paired with a new diagnosis function to formally verify a proposed distributed diagnosability condition within a finite number of steps despite delayed inter-agent messages. The theoretical analysis proves the verification procedure correctly determines diagnosability, giving practitioners a constructive, automaton-based method to check fault-detectability guarantees before deploying distributed monitoring architectures. This could be built on by encoding specific plant models and delay bounds into the delay recorder to automate diagnosability checks for real supervisory control systems.

Building on Chen–Liu–You and Liu–Xia's results linking nonnegative Lin–Lu–Yau (LLY) Ricci curvature to edge-connectivity equal to minimum degree in graphs, this paper extends the theory to r-uniform linear hypergraphs using the random-walk curvature notion of Tian–Zhao. The authors formulate a hypergraph analogue of the key combinatorial inequality from Liu–Xia and apply it to bound edge cuts, proving that any locally finite connected r-uniform linear hypergraph (r≥3) with nonnegative LLY curvature has edge-connectivity equal to its minimum incidence degree. The linearity assumption (any two hyperedges share at most one vertex) is shown to be essential to the argument. This gives hypergraph theorists and network scientists a curvature-based sufficient condition for optimal edge-connectivity, usable to certify robustness of hypergraph-structured systems without exhaustive min-cut computation.

The paper poses open problems on tropicalizing linear ODEs with the goal of computing radii of convergence of their classical power-series solutions, building on the notion of differential tropical bases introduced by Fink and Toghani (2022). The central question is characterizing which linear ODEs admit a *finite* differential tropical basis, since finiteness is what makes tropical methods computationally tractable rather than requiring infinite case analysis. The authors' initial contribution examines second- and third-order linear ODEs as concrete test cases for this characterization. Researchers in tropical geometry or differential algebra could extend this by attempting to resolve the finiteness conjecture for higher orders or by implementing tropical-basis computations to estimate convergence radii algorithmically.

The paper establishes a strong averaging principle for multiscale time-inhomogeneous SDEs driven by multiplicative α-stable Lévy noise (α∈(1,2)). Using Khasminskii's discretization approach, the authors show the fast component with frozen slow variable admits a periodic measure, then prove strong (pathwise/moment) convergence of the slow component to an ε-dependent averaged system; when the reciprocals of the fast-process period τ1 and εscaled period ετ2 are rationally linearly independent, the resulting averaged system exhibits random quasi-periodicity. A second averaging step, via the ergodic theorem, further reduces this to a time-inhomogeneous SDE independent of the scale parameter ε. This extends averaging-principle results from Brownian-driven multiscale SDEs to the jump-diffusion (α-stable) setting, giving a template other researchers could adapt to reduce dimensionality in multiscale stochastic models with heavy-tailed noise.

The authors extend the theory of exponential periods by applying Hien's rapid decay cohomology to a broader class of integrable connections on varieties, beyond those merely twisted by a regular function, with E-operators (a class of differential operators from transcendence theory related to E-functions) as the primary motivating examples. This yields a new class of 'E-periods,' realizing absolutely convergent integrals of E-functions as matrix elements of a period pairing — giving these integrals a precise cohomological/algebraic interpretation rather than treating them as isolated analytic objects. They further generalize rapid decay cohomology to singular varieties and establish an analogue of Nori's basic lemma in that generalized setting. This provides number theorists and arithmetic geometers working on E-functions and transcendence with new cohomological tools to compute or relate periods on singular spaces.

The authors construct an explicit divergence-free, time-periodic velocity field on the 3-torus built from three alternating piecewise-affine shears, and prove that for sufficiently large shear amplitude, the ideal (non-resistive) induction equation exhibits exponential-in-time growth in L^p norm for every nonzero divergence-free initial magnetic field. The proof establishes a uniform cone condition for the time-one map of the flow and combines it with a bunching inequality to exclude nontrivial fields lying entirely in the stable bundle, giving a rigorous fast-dynamo-type growth result. They also show the time-reversed flow (whose time-one map is the inverse) admits nontrivial bounded fields confined to its two-dimensional stable bundle that decay exponentially in every L^p norm. This gives a concrete, analyzable example for dynamo theory researchers to test growth-rate estimates or extend the cone/bunching technique to other explicit flows.

The paper analyzes the simultaneous vanishing-viscosity, vanishing-resistivity limit of steady incompressible MHD flow in a disk with boundary velocity a small nonaxisymmetric perturbation of rigid rotation (mean angular speed α) and prescribed tangential magnetic trace with mean β, under the non-Alfvénic condition |α|≠|β|. They construct solutions that converge on compact interior subdisks to a rigidly rotating ideal-MHD core with constant vorticity and constant out-of-plane current density, extending classical Prandtl–Batchelor theory (which selects constant-vorticity Euler cores) to the coupled MHD setting. A new 'MHD–Wood law' determines the two core rotation rates: the velocity core from a coupled kinetic–magnetic balance, the magnetic core fixed purely by the imposed mean circulation, with zero circulation giving complete flux expulsion and nonzero circulation leaving uniform magnetic rotation. This gives a rigorous selection principle that MHD/PDE researchers can use to predict interior states of nearly-ideal rotating conducting flows from boundary data alone.

The authors study moments of sums of Fourier coefficients of a fixed holomorphic Hecke eigenform twisted by the real character χ_{8p}, averaged over odd primes p. Under GRH they establish the true order of magnitude of the unsmoothed m-th moment for all real m≥4, and derive a sharp smoothed upper bound XY^{m/2}(log X)^{m(m-3)/2} for integer m≥4, matched by a lower bound for even m, proving optimality. This refines the moment-method toolkit for quadratic-twist families and could feed into non-vanishing or subconvexity results for this family.

The paper addresses the initialization problem in policy iteration (PI) for continuous-time indefinite stochastic linear quadratic control with state- and control-dependent multiplicative noise, where stability is governed by a Lyapunov-type inequality (drift plus diffusion terms) rather than simple Hurwitz eigenvalue conditions. They propose a spectrum-assignment method to construct an initial stabilizing controller without requiring full model knowledge, exploiting the Lyapunov-type structure. This targets a known bottleneck of model-free PI — its dependence on model information for initialization — and could serve as a pre-processing step for data-driven stochastic LQ solvers.

The authors work in the lattice of flats of the graphic matroid M_n=M(K_{n+1}), studying rank-k flat families A with pairwise meet-rank ≥t, which for t=1 is equivalent to Czabarka's partition-EKR conjecture. They prove an EKR-type theorem in the explicit linear range n+1≥8k (versus the conjectured sharp n≥2k), a constant-factor advance, plus a general-t EKR theorem under an O_t(k^2) condition on block number with equality only at the full t-star, and a Hilton–Milner-type characterization of largest nontrivial intersecting families under an O(k^6) threshold. This extends EKR-type extremal set theory into matroid/partition-lattice settings with explicit range bounds others can build on toward the conjectured n≥2k threshold.

For a graph G with m edges, bk(G) denotes the max triangles sharing a common edge; prior work (Zhai et al.) showed non-bipartite graphs meeting a spectral condition ρ(G)^2≥m-1+2/(ρ(G)-1) have bk(G)>√m/240, aside from an explicit exceptional family S_{m,s}^+, and asked for the optimal constant. This paper resolves that asymptotically: for any ε∈(0,1/4) and large m, every such graph is either isomorphic to S_{m,s}^+ or has booksize matching the optimal constant up to ε. This sharpens the spectral extremal graph theory connecting adjacency eigenvalues to triangle-book structure, tightening Zhai et al.'s bound to near-optimal form.

The paper develops an algorithm to detect real representatives in equisingular strata of complex families of K3 surfaces with prescribed simple (ADE-type) singularities, valid for arbitrary polarizations. Applied to spatial quartics, it shows every real equisingular stratum contains real surfaces except three exceptional types, completing that classification; the method also recovers the known real-representability result for simple plane sextics. This gives moduli-theorists a general, algorithmic tool for real-representability questions across polarized K3 families, rather than case-by-case ad hoc arguments.

The paper performs a rigorous asymptotic (boundary-layer) expansion for the 2D chemotaxis-Navier-Stokes system with logarithmic chemotactic sensitivity under Navier-slip boundary conditions in a half-space, for viscosity ε>0, establishing well-posedness of the derived boundary-layer profile equations. It further proves local well-posedness for the singular ε=0 limit — the supercritical chemotaxis-Euler system — despite the loss of diffusive smoothing. As Part I of a two-part series, this establishes the analytic groundwork that Part II will presumably use to prove rigorous vanishing-viscosity convergence to the boundary layer expansion.

The paper proves that on weighted graphs, the Lin–Lu–Yau (LLY) curvature equals the p-Ollivier–Ricci curvature up to a scaling factor precisely when the idleness parameter p≥1/2, extending Bourne et al.'s combinatorial-graph result to weighted graphs and showing 1/2 is a sharp threshold. As a corollary, the equivalence yields a simplified proof of global existence and uniqueness for solutions to the LLY curvature flow (Bai et al.), since the p-Ollivier framework at p=1/2 is more analytically tractable. This gives a useful bridge for anyone working with discrete Ricci curvature/flow on weighted networks who needs to move between the two formalisms.

The authors extend the GL_2 Bessel δ-method to imaginary quadratic fields and use it to prove a hybrid Weyl-type subconvexity bound for GL_2×GL_1 twisted L-functions in the Archimedean aspect of a Hecke character on GL_1. This generalizes classical Weyl-subconvexity techniques (delta-method spectral/geometric expansions) beyond the rationals to the imaginary quadratic setting, giving sharper-than-convexity bounds in multiple aspects at once. Such hybrid subconvexity results are typically leveraged for equidistribution problems (e.g., Hecke eigenvalue or CM point equidistribution) and provide a template for further L-function families over number fields.

The authors consider autonomous random vector fields on the torus with a multiscale finite-range decomposition and Hölder regularity $C^{\alpha-}$ for $\alpha \in (0,1)$, well below the classical Cauchy–Lipschitz threshold needed for ODE/continuity-equation well-posedness. They prove that for $\alpha > 1/2$, stochastic cancellations across scales yield almost-sure well-posedness of the flow map and continuity equation away from the velocity field's zero set—regularization by noise via a probabilistic rather than additive mechanism. They complement this with explicit sub-threshold examples exhibiting robust ill-posedness, establishing $\alpha=1/2$ as a sharp phase transition, plus effective regularity estimates below threshold and analogous results for a 'refreshing' (renewed-randomness) regime. This connects to the regularization-by-noise literature (Flandoli, Catellier–Gubinelli) but replaces additive white noise with structural multiscale randomness in the coefficients themselves.

For the graph $K_n - F$ formed by deleting a set $F$ of $p$ edges from the complete graph, the authors define the edge-defect matrix $Q = B^TB$ (a $p\times p$ Gram-like matrix from the incidence matrix $B$ of $F$) and derive closed-form expressions for pairwise effective resistance via the resolvent of $Q$, replacing the standard $n\times n$ Laplacian pseudoinverse computation with a $p\times p$ one. They further give unified formulas for the Kirchhoff index and the spanning-tree count in terms of $Q$'s eigenvalues, exploiting the algebraic structure of complete-graph Laplacians. This is directly implementable: for sparse edge-deletion sets ($p \ll n$), it turns an expensive Laplacian computation into one dominated by a much smaller matrix, useful for network robustness/reliability analysis of near-complete topologies.

The bi-level dual-defense attacker model (BDAM) combines node interdiction, edge destruction, and capacitated supply support; its dominant attacker-path term is bi-linear in defender decisions, but since arc lengths are linear in the defender's binary variables, lower-level dualization yields an exact single-level MILP reformulation. This reformulation certifies global optimality on all 18 benchmark configurations (3-row and 5-row), 17 in under 10 seconds, and produces strong incumbents on larger 15×30 grids where full certification isn't yet reached. Comparing against their previously published hybrid-metaheuristic results shows a 3.55% average optimality gap that widens with instance size—quantitative evidence that metaheuristic solution quality degrades at scale for this problem class. Practitioners solving similar bi-level interdiction problems with linear-in-decision arc costs can reuse the dualization-based reformulation as an exact benchmark or warm-start for larger instances.

The paper develops same-grid duality certificates for constrained dynamic portfolio problems solved by neural/numerical policy solvers: for polyhedral control sets, an exact conditional budget identity decomposes the primal-dual value residual into a pathwise nonnegative terminal Fenchel (convex-conjugate) defect plus date-by-constraint complementary-slackness terms. A canonical Doob decomposition removes the budget martingale component that otherwise masks small residuals, and Bellman-primitive curvature assumptions yield an occupancy-weighted 'policy-distance' region with sharp $O(\sqrt{G})$ radius, while a paired constraint relaxation lower-bounds the optimal Lagrange multiplier to certify which constraints are actively binding. A finite-sample resolution theorem quantifies how these bounds sharpen with simulation grid density, giving practitioners a diagnostic layer on top of standard primal-dual value bounds to both certify near-optimality and localize which constraint/date drives suboptimality in learned portfolio policies.

The paper formulates priority-aware robust coverage for aerial sensor networks as an optimization over sensor orientations, incorporating surveillance constraints and an 'RRF-based' robustness formulation to hedge against sensor location uncertainty. The proposed PAROO (priority-aware robust orientation optimization) algorithm solves for orientations that maximize priority-weighted coverage under this uncertainty model, evaluated on an airport-inspired scenario against baseline coverage approaches. Results show PAROO better concentrates coverage on high-priority regions while maintaining robustness to positional perturbation, suggesting direct applicability to multi-UAV surveillance tasking where sensor pose uncertainty is significant. Replication would require the coverage/priority model and RRF robustness formulation detailed in the full methodology section.

The authors study evolutionary dynamics of repeated-game strategies (direct reciprocity) under mutation-selection processes with pairwise comparison, examining large population size and intermediate-to-high mutation rate/intensity-of-selection regimes. They show this simple process consistently evolves communities of strategies that achieve the maximum attainable payoff across multiple social dilemmas (Prisoner's Dilemma, Snowdrift, and others tested), a stronger and more general result than prior findings that reciprocity merely favors cooperation over defection. Mechanistically, mutation sampling strategies near the boundary of strategy space appears critical—likely enabling exploration of extremal (near-deterministic) reciprocal strategies that sustain optimal cooperation. This provides a tractable dynamical model, amenable to simulation or analytical treatment via stochastic evolutionary game theory, for researchers studying when reciprocity-based cooperation converges to Pareto-optimal rather than merely stable outcomes.

EpiBench is a closed-book, sequence-only, automatically-scorable benchmark of 1,609 samples derived from structurally-resolved antibody-antigen contact maps, designed to probe whether LLMs can perform epitope-centered reasoning tasks relevant to the antibody drug development workflow beyond what existing isolated-task epitope predictors or generic protein benchmarks cover. The benchmark's grounding in structural contacts allows objective scoring without requiring the model to access structure at inference time, isolating whether sequence-level pretraining captures implicit epitope-relevant signal. Practitioners can use EpiBench directly to evaluate proprietary or open LLMs (via prompting or fine-tuning) on epitope reasoning, and the curated sample set could seed further fine-tuning datasets or serve as a leaderboard for antibody-focused foundation models. Exact task splits and scoring metrics would need to be pulled from the full paper for precise replication.

The authors train Deep Belief Networks (stacked RBMs) unsupervised on MNIST, Fashion-MNIST, and KMNIST, then probe successive hidden layers using the Generalized Discrimination Value (GDV, a label-aware clustering-quality metric), post-hoc supervised linear probes, a reconstruction-based abstraction-distance measure, effective dimensionality estimates, and free unconditional sample generation. They find class-specific clustering in representation space increases with depth across datasets and network widths, despite no label signal during training, and show via controls (random transformations, weight-marginal shuffles, generic dimensionality reduction, sigmoid-saturation checks) that this isn't an artifact of the analysis pipeline. This supports the hypothesis that DBNs' generative/reconstructive objective implicitly performs class-relevant feature disentanglement as a function of depth, analogous to findings in supervised deep nets but emerging purely from unsupervised pretraining. Practitioners could replicate this with standard DBN/RBM stacks and the GDV metric to probe whether other unsupervised architectures (VAEs, unsupervised transformers) show similar depth-driven class emergence.

The authors propose a permutation-invariant framework for comparing multilayer perceptron weight spaces: hidden units are first coarsely aligned via permutation-invariant features, then refined through iterative Hungarian (bipartite) matching to resolve the neuron-relabeling symmetry inherent to MLPs. Post-alignment, structural distance metrics emphasizing task-relevant weight patterns quantify similarity between independently initialized networks trained on identical tasks. The core claim is evidence for structural convergence — analogous to convergent evolution — beyond mere functional equivalence. Practitioners could adapt the alignment procedure for model merging, weight interpolation, or cross-model interpretability comparisons.

THBKG is a temporal heterogeneous biomedical knowledge graph with 110,396 entities and 11.1M edges across nineteen relation types, each edge timestamped with the year its supporting evidence changed, enabling reconstruction of a target-disease pair's evidence profile as it stood at any historical decision point. On top of this, the authors define a decision-aligned benchmark predicting Phase II advancement for target-disease pairs using only evidence available at the time of entry, avoiding the temporal leakage that plagues static knowledge graphs. This enables practitioners to train and validate advancement-prediction models under realistic, look-ahead-free conditions, and the graph itself could serve as a general resource for other temporally-sensitive biomedical prediction tasks.

The authors model EEG activity as distributions of windowed activity patterns, quantifying temporal stability via Wasserstein distance and representational complexity via intrinsic dimensionality, applied across multi-task, lifespan, and clinical EEG datasets. Key finding: neural representations exhibit constrained, condition-specific stability rather than unconstrained drift, and higher intrinsic dimensionality consistently correlates with lower stability. Both metrics show reproducible spatial topography, with posterior regions showing higher dimensionality and lower stability than other regions. This framework offers a principled, distribution-based alternative to traditional EEG biomarkers, potentially replicable on any windowed EEG dataset for aging or neurodegenerative disease screening.

Curriculum multiple shooting (CMS) is a general-purpose training strategy for fitting NODEs, UDEs, and mechanistic ODE models to noisy, sparse, or partially-observed time-series by combining curriculum learning (progressively increasing training difficulty) with multiple shooting (splitting long trajectories into shorter segments fit in parallel and later reconciled). Across twelve benchmarks spanning simulated and real data, CMS accelerates convergence, improves training stability, and ranks among the top methods for generalization compared to state-of-the-art training strategies. Practitioners fitting dynamical systems models to real experimental data could adopt CMS as a drop-in training routine to reduce optimization failures without changing model architecture.

The study applies a parsimonious voxel-wise pharmacokinetic model to dynamic contrast-enhanced (DCE) MRI in 82 IDH-wildtype GBM patients with new post-chemoradiotherapy enhancing lesions (53 true progression, 29 pseudo-progression, confirmed via histopathology or modified RANO criteria), extracting radiomic features and combining them with MGMT methylation status in a machine learning classifier. The approach targets discrimination between true progression and pseudoprogression, a distinction conventional contrast-enhanced MRI cannot reliably make. This demonstrates that physiologically-grounded, voxel-wise DCE-MRI features plus a single molecular biomarker can be fused into a practical classifier, a pipeline replicable at other centers with DCE-MRI acquisition and MGMT testing capability.

The system pairs a force-sensitive resistor (FSR) embedded in an abdominal belt with a custom Bluetooth Low Energy acquisition board, using a mechanical holder to mechanically couple abdominal expansion to the sensor and avoid analog amplification circuitry, minimizing power draw and component complexity. The sensing pipeline was validated across multiple breathing patterns and body positions, showing consistent amplitude changes and reproducible peak-to-peak timing across breaths. This is a low-cost, low-power hardware reference design that practitioners could replicate or extend with an FSR, BLE microcontroller, and a 3D-printed or sewn mechanical coupler for continuous ambulatory respiratory monitoring.

The paper formalizes base-rate neglect and the previously separate 'cue-density effect' from contingency learning as two independent under-corrections within a single Bayesian updating equation, each represented by its own weight term — one for the outcome prior, one for cue frequency. It shows the cue-frequency weight is only identifiable in graded-rating tasks, since two-alternative forced-choice designs mathematically cancel it out, explaining why it went undetected in prior two-choice paradigms. At the parameter extremes, the framework recovers classical quantities including signal-detection-theoretic base-rate neglect. Researchers designing judgment/decision-making experiments could use this dual-weight model to jointly estimate and dissociate both bias types from graded-response data rather than assuming a single scalar bias.

In a controlled study (n=14 tSCS group, n=14 training-only control), the authors assessed acute and training-related effects of transcutaneous spinal cord stimulation on conscious ankle proprioception (via bilateral robotic dynamic ankle localization, the Crisscross device), maximum dorsiflexion strength, and gait kinematics (spatiotemporal parameters, trunk-sway, and mediolateral center-of-mass excursion) during normal and tandem treadmill walking. tSCS acutely impaired conscious proprioceptive localization while producing a more constrained locomotor pattern, suggesting the afferent-network activation underlying tSCS's known locomotor effects comes at the cost of degraded proprioceptive awareness rather than being purely beneficial. This dissociation between perceptual and motor effects is directly relevant to clinicians using tSCS in gait rehabilitation protocols, who may need to account for transient proprioceptive disruption alongside motor benefits.

Each residue is coarse-grained to a two-state qubit (stabilized vs. excited) parameterized by residue solvation free energy, with a structure-informed entanglement layer of parameterized controlled gates encoding the covalent/non-covalent contact network. Sampling the resulting circuit (~10^6 measurement shots) yields a distribution over residue-interaction states rather than a single ground-state structure, capturing multi-basin ensemble heterogeneity that single-optimum quantum methods miss. A practitioner could implement this ansatz on a gate-based SDK (e.g., Qiskit) using contact-map-derived entanglement topology and benchmark the sampled energetics against classical MD ensembles.

The authors build a recurrent variant of the Forward-Forward (FF) algorithm with an inverted objective — increasing rather than decreasing unit activity in response to negative/contrastive data — trained via purely local updates and activity cancellation, without explicit error-coding units or a generative decoder. Stacked across layers, this setup self-organizes into predictive representations exhibiting top-down modulation and stimulus-surprise signaling, hallmarks typically attributed to hierarchical predictive-coding architectures. It's a concrete, replicable substrate (implement recurrent FF layers with negative-data injection, probe layer activity for surprise/prediction signatures) for testing whether predictive-coding-like computation can emerge from local contrastive learning rather than backprop-based generative modeling.

Noise-prune is a local, unsupervised pruning rule that scores each recurrent connection's importance via its sensitivity to injected noise fluctuations, then stochastically samples (rather than deterministically thresholds) which low-importance weights to remove, followed by rescaling of surviving connections to preserve network statistics. Evaluated here on task-trained RNNs — extending beyond its original testbed of unstructured random networks — it substantially outperforms magnitude-based pruning and performs on par with or better than a non-local, second-order (curvature-informed) pruning method. Its locality and avoidance of gradient/Hessian computation make it a promising candidate for neuromorphic or biologically-constrained pruning implementations.

The framework organizes insomnia phenotypes along three axes: which dynamical operation fails (state transition, stabilization, spatial cortical recruitment, architectural sequencing, or state perception), the specific sleep stage/boundary at which it fails, and its causal status — treating clinical covariates like age, circadian phase, comorbidity, and medication as modifiers rather than separate mechanistic classes. It adapts a local Landau-Ginzburg relaxational formalism, previously used in cortical and sleep-dynamics modeling, as a phenomenological language for generating nested, testable hypotheses about disrupted order-parameter dynamics at sleep-state boundaries under near-equilibrium approximations. This is a conceptual/taxonomic proposal intended to guide future quantitative sleep-EEG modeling and subtype-specific treatment trials, not yet a validated diagnostic tool.

The authors ground AI ethics in comparative neuroscience, contrasting the brain's energy-efficient, workspace-based architecture (global neuronal workspace theory, positing a broadcasting mechanism integrating distributed processing) and non-maximizing reward dynamics (wanting/liking/satiety as distinct cyclical components rather than a single optimized utility signal) against artificial systems' optimization-driven computation. Their central claim is that AI-in-life-sciences ethics should not be treated as sui generis but derived from the same universality-versus-cultural-diversity tension present in general human ethics, since neural substrates for moral judgment are conserved while their culturally shaped content is not. This is a philosophical argument rather than an empirical study, useful as a framing reference for AI governance or bioethics policy work rather than a technical method.

In a retrospective cohort of 197 IBD patients (142 Crohn's disease, 55 ulcerative colitis) on biologic therapy, four cytokine-gene SNPs were genotyped (TNF-alpha rs1800629, TGF-beta rs1800471, IL-6 rs1800795, IL-10 rs1800896), with biochemical response at 12 months defined by CRP <5.0 mg/L and fecal calprotectin <250 microg/g off corticosteroids. The IL-10 rs1800896 (-1082 G>A) promoter polymorphism, which affects IL-10 transcriptional/anti-inflammatory activity, was associated with achieving biochemical remission, while the IL-6 rs1800795 C allele was linked to a younger age-related phenotype. These SNP associations support pharmacogenetic panels as candidate biomarkers for stratifying biologic-therapy response in IBD, pending replication in larger prospective cohorts.

The authors benchmark frozen-probing (lightweight readout heads on frozen embeddings) against full fine-tuning for five genomic foundation models — DNABERT-2, Nucleotide Transformer, HyenaDNA, GENERATOR-v2, Omni-DNA — across regulatory, epigenetic, promoter, splice-site, and variant-effect prediction tasks, using unified probing protocols that separate diagnostic readouts from validation-selected checks. Results show consistent task-dependent representation accessibility: frozen probes recover 95-100% of fine-tuned performance on promoter tasks but only 60-88% on splice-site detection, suggesting splice-relevant information is encoded non-linearly or diffusely and needs task-specific adaptation to extract. Practitioners choosing between frozen-feature-extraction and fine-tuning pipelines should treat this task-dependence as a decision criterion, reserving fine-tuning for low-accessibility tasks like splice-site prediction.

The framework replaces standard gradient-based optimization with Natural Gradient Ascent, which rescales parameter updates using the Fisher information metric rather than raw gradients, to identify parameters of ODE-based ecological models from sparse, noisy time-series data — addressing irregular likelihood surfaces and solver stiffness that cause standard methods to diverge or stall. Applied to the classical Lotka-Volterra predator-prey system, the method exploits non-dimensionalization to shrink the effective parameter space before optimization, improving convergence robustness independent of initial-guess quality. This offers a practical template for fitting mechanistic ecological (or other stiff ODE) models to real, sparsely-sampled field data without requiring dense time series or highly accurate starting parameter estimates.

The author defines a divergence-like metric D_H(R,M) = H(R\|M) + H(M\|R), summing the conditional entropy of stimulus sequences R given memory M (surprise) and of M given R (uncertainty), and treats insistence on sameness as behavior that constrains this joint entropy toward a minimum. Autism is modeled as a restriction of cognitive processing to concrete discrimination, memorization, and prediction of environmental regularities, rather than abstraction, which changes the effective statistics of R and M available to the individual. The claim is that under these restricted-cognition assumptions, minimizing D_H naturally produces preference for repetition and resistance to novel stimuli, giving a formal derivation rather than a purely descriptive account. A practitioner could operationalize this by estimating empirical entropies from behavioral or physiological surprise/uncertainty proxies (e.g., pupillometry, EEG mismatch responses) and testing whether their sum tracks the intensity of sameness-seeking behavior across individuals or contexts.

The authors analyze a spatial SIRS reaction-diffusion model where diffusion acts on infecteds and immunity loss occurs at rate governed by a small parameter ε, producing a singularly perturbed traveling-wave ODE system after the standard wave-coordinate reduction. Applying Geometric Singular Perturbation Theory, they characterize the fast layer dynamics (infection outbreaks) and slow flow along the critical manifold (immunity waning), and derive a quantitative entry-exit function describing the delayed transition when the slow trajectory leaves the manifold's fold. This entry-exit relation gives explicit conditions/timing for wave propagation that the authors validate numerically against the full PDE system. Researchers modeling recurrent or oscillatory epidemic fronts (e.g., cholera, pertussis) could adapt this entry-exit framework to predict wave speed and outbreak periodicity from immunity-loss timescales.

The paper introduces 'Time²,' a conceptual framework distinguishing 'processing time' (the neural latency to transform retinal input into a perceptual representation) from 'stimulus time' (the duration of sensory exposure needed to register a stimulus), arguing these are typically conflated or studied independently in vision neuroscience. The authors advocate co-varying both temporal factors within single experimental designs rather than fixing one while manipulating the other, to disentangle their separate contributions to perceptual outcomes. This has implications for interpreting reaction-time and psychophysical data, since apparent effects attributed to processing speed could actually reflect stimulus-duration confounds or vice versa. Vision modelers and experimentalists could use this framework to redesign paradigms (e.g., masking, rapid serial presentation) that explicitly separate and manipulate these two timescales to better constrain computational models of perception.

The authors build a population-dynamics model of budding yeast incorporating asymmetric division (mother cells cycle faster than daughter cells) combined with positive feedback in the cell-cycle progression rate, motivated by observed metabolic oscillations and cell-cycle-linked 'temporal clustering' in bioreactor cultures. Numerical simulations starting from randomly phased populations show spontaneous self-organization into p:q clustering patterns (p mother-cell clusters, q daughter-cell clusters, p ≤ q) that are dynamically stable attractors of the model. The core contribution is demonstrating that this simple combination of structural asymmetry and feedback is sufficient to generate and stabilize multi-cluster synchronization, without requiring external periodic forcing or complex signaling assumptions. Researchers modeling cell-population synchrony (circadian, metabolic, or cell-cycle oscillators) could adapt this asymmetric-division-plus-feedback mechanism as a minimal generative model for clustering phenomena observed experimentally.

The authors provide an algorithm and Macaulay2 implementation for computing n-tet (generalized to n=5, i.e., quintet) concordance factors on arbitrary phylogenetic networks under the Network Multispecies Coalescent model, extending beyond the standard quartet-CF approach used in tools like SNaQ or PhyloNet. Using quintet CFs on level-1 networks, they demonstrate identifiability results—specifically for network roots and small reticulation cycles—that are provably non-identifiable from quartet CFs alone, formalizing where the extra taxon resolves ambiguity via algebraic/statistical identifiability analysis. This establishes necessary theoretical groundwork (identifiability being a prerequisite for consistent statistical inference) for developing quintet-based network inference methods analogous to existing quartet-based pipelines. Phylogenetics researchers could use the released Macaulay2 code to compute expected CFs for candidate networks and test new quintet-based inference or hypothesis-testing procedures for detecting hybridization/gene flow.

MS-MLB is an open, reproducible benchmark built on the public GSE17048 whole-blood RNA expression cohort, framed as an MS-vs-healthy-control binary classification task, with a shared leakage-controlled evaluation pipeline (nested cross-validation, an untouched stratified holdout set, bootstrap confidence intervals, and ROC/precision-recall reporting) that standardizes comparison across ML algorithms. The explicit design goal is to prevent common benchmark pitfalls in clinical ML (train/test contamination, optimistic single-split reporting) that inflate reported performance in prior MS classifier studies. By fixing dataset, splits, and evaluation protocol, it allows apples-to-apples comparison of feature selection and classification methods on transcriptomic MS diagnosis. Practitioners can plug new algorithms or feature-selection strategies into the released pipeline to benchmark them against existing baselines without re-deriving evaluation infrastructure, and use it as a template for leakage-safe evaluation in other blood-biomarker classification problems.

The authors formulate a stochastic partial differential equation (SPDE) for eDNA concentration in rivers, where the source term is driven by a stochastic differential equation modeling fish migration with a delayed input (accounting for time lag between fish presence and detectable eDNA release), coupling spatiotemporal diffusion/advection with multiplicative noise. Despite the multiplicative noise coefficient being non-Lipschitz (which typically complicates well-posedness proofs), they establish existence/uniqueness (well-posedness) by exploiting the model's affine structure, and derive a closed-form Laplace functional characterizing the process's statistical distribution analytically. They also propose a numerical discretization scheme for simulating the SPDE, enabling practical computation. Ecologists and statisticians developing eDNA-based abundance/occupancy estimators could use this as a mechanistic likelihood model to replace ad hoc statistical fitting, or extend the discretization scheme to fit real river monitoring data and back-infer fish migration parameters.

The paper critiques the common practice of binarizing time-to-event clinical outcomes before applying machine learning, showing this discards censored patients, collapses temporal granularity into an arbitrary threshold, and biases downstream feature selection. Using Bayesian network (BN) structure learning as the test case, the authors replace the standard binary scoring function for feature-to-outcome edges with the Cox proportional hazards partial log-likelihood, terming the result a Survival-Aware Bayesian network (SABN), and re-run it on a head-and-neck cancer cohort and a second surgical cohort (the latter originally analyzed with non-BN binarized methods). Their ablations show the survival-aware scoring recovers prognostically relevant features that were missed or misranked under binarized BN feature selection, demonstrating concretely that the simplification changes model conclusions, not just performance metrics. Practitioners building clinical risk-prediction pipelines can substitute Cox-based (or other censoring-aware) scoring functions into existing feature-selection frameworks to avoid this bias, particularly when working with BN or similar structure-learning approaches.

Introduces persistent-homology-derived cycle participation as an alternative to structural-degree ranking for driver-node selection in linear network control theory, tested on diffusion-MRI structural connectomes from 70 subjects at three parcellation scales. Topology- and degree-informed driver sets yield nearly identical scalar minimum control energy (~0.2% difference), but the geometry of the controllable subspace — how energy distributes across state dimensions — differs substantially between the two selection criteria. This indicates scalar control energy alone is an insufficient benchmark for comparing driver-node heuristics; anyone building connectome-based stimulation-targeting pipelines should evaluate subspace geometry, not just energy magnitude.

An outcome-informed spatial pathology pipeline links deep-learning recurrence-risk heatmaps from H&E slides with mass-spectrometry-based spatial proteomics in a 156-patient TNBC cohort. Distribution-based aggregation of high-scoring patches achieves AUC 0.77 and C-index 0.77 on an independent test cohort; bulk proteomics ties high image-derived risk to cell-cycle/genome-maintenance programs and low risk to immune activation. Co-registering heatmaps with proteomic sampling coordinates reveals intratumoral heterogeneity — high- and low-risk patches with distinct nuclear/architectural morphology coexisting within the same compartment — with the heatmaps then used to spatially guide further profiling, offering a generalizable template for grounding histology-based risk classifiers in molecular data.

Constructs a local Landau-Ginzburg phenomenology treating each sleep-stage boundary as motion of a spatially-extended, noisy, dissipative neural field in an effective potential, with a latent cortical-ordering coordinate phi inferred from standard EEG/PSG observables via a measurement model designed to avoid circularity. Distinct transitions are hypothesized to correspond to different bifurcation types: sleep onset as a fold-like loss of wake stability (open question: cusp bistability/hysteresis), N1→N2 and N2→N3 as continuous-like crossovers, and NREM→REM as a candidate first-order-like desynchronization event. This is a theoretical dynamical-systems framework rather than an ML pipeline; it could be tested by fitting potential-landscape parameters to labeled PSG datasets and checking whether predicted transition types match empirical hysteresis/synchrony signatures.

A causal (autoregressive) transformer paired with a learned vector-quantized VAE tokenizer, trained with cross-entropy loss on tokenized single-cell gene-expression vectors, for few-shot conditional generation: given example cells of a type, generate more matching that distribution. Evaluation compares the generated expression-vector distribution to ground truth for held-out cell types rather than relying on perplexity alone. The core contribution is characterizing scaling behavior — pretraining loss as a function of parameter count and data volume — reportedly the first jointly-fit scaling law in this tokenized single-cell generative setting, informing compute/data allocation decisions for similar omics generative models.

CLARA converts a natural-language cancer-genomics query into a typed scientific query specification, enumerates multiple plausible interpretations, executes all of them against the data, and triggers a clarification request only when outputs diverge beyond a preregistered threshold (relative divergence >0.10 or absolute >5 percentage points). Benchmarked on 330 executable mutation-prevalence contrasts across 8 TCGA PanCancer Atlas cohorts and a 30-gene panel, split 115 result-sensitive vs. 215 result-stable by that rule; an independently implemented pandas execution engine exactly replicated results, supporting engine-independence. It's a reusable template for building trustworthy NL-to-query systems over structured biomedical data — ambiguity detection via multi-interpretation divergence rather than single-shot parsing.

Applies a time-inhomogeneous discrete-time Markov chain to longitudinal self-report data from the ABCD Study (n=11,864, ages 9-13), defining 8 states from combinations of suicidal ideation, suicidal behavior, and co-reported non-suicidal self-injury (NSSI). The framework estimates year-to-year and multi-year transition probability matrices, allowing rates to vary by age/wave rather than assuming stationarity, with uncertainty quantification and formal statistical comparison of transition likelihoods across developmental periods. Key finding: transition patterns are structured but non-stationary, including generally high recovery-to-no-symptoms probabilities alongside developmentally-varying risk transitions — a reusable methodological template for panel-data researchers wanting transition-probability models instead of static cross-sectional risk scores.

A perspective/roadmap piece arguing structure prediction should shift from single-conformation inference (the largely-solved AlphaFold2-era problem) to state-space inference: recovering the ensemble of accessible conformational states, their relative populations, interconversion kinetics, context-dependence (ligands, PTMs, mutations), and perturbation responses. It reviews three converging strategies — deep-learning ensemble generators, physics-based simulation (MD, enhanced sampling), and experimental constraints (cryo-EM, NMR, HDX-MS) usable as priors/validation. No new model or benchmark is presented; its value is as a framing document for building hybrid ML+physics+experiment pipelines that output populated ensembles and kinetic rates rather than a single structure.

Reframes the concentration (N/V) axis for binding assays: rather than the conventional approach of increasing molecule number N at fixed volume V, the method holds N fixed and shrinks V via nanoscale spatial confinement, reaching effective local concentrations sufficient to read out millimolar-affinity interactions inaccessible to bulk-concentration-based techniques like ITC or SPR. This is a proposed new experimental control axis rather than a specific instrument; practitioners in biophysics/single-molecule instrumentation could apply it to build nanoconfinement-based assays (e.g., nanopores, nanowells, zero-mode waveguides) for characterizing weak protein-protein, protein-ligand, or transient signaling interactions previously out of reach.

The work targets the gap between bulk molarity and 'effective molarity' formulations, which still collapse local structural effects into a single concentration-valued number rather than treating geometry as an independent experimental axis. The authors design a bench-compatible assay that exposes local structure directly, revealing a chemistry-geometry crossover invisible to standard flask-based readouts. In the micromolar-or-weaker affinity regime, inhibition sharply switches out of the conventional concentration-and-affinity-governed mode into a geometry-dominated regime. This gives practitioners a concrete experimental handle for probing structured/crowded reaction environments (e.g., tethered or compartmentalized binding) beyond what dilution-series kinetics can resolve.

The pipeline applies Dynamic Mode Decomposition to EEG channels to extract high-frequency dynamical modes as features, assembling them into a feature table per channel. A random-distribution significance test found consistent high-frequency dynamics in roughly 70% of samples for specific channels. PCA was then applied to the validated feature table, and the resulting components successfully classified alcohol-dependent versus control subjects, indicating DMD-derived high-frequency structure is a viable, reproducible biomarker feature for EEG-based classification pipelines.

REDE evaluates reproducibility across multiple evidence layers — DEG burden, exact gene-set membership, top-rank overlap, signed fold-change direction, prespecified gene confirmation, and Hallmark pathway enrichment — using fixed discovery/validation/external-test splits across nine microarray cohorts in three cancer types. The REDE-2Fold procedure performs independent differential expression on two patient-level splits of each discovery cohort and retains only genes with concordant direction in both, providing a within-study robustness filter before external validation. They further test whether discovery-only panels preserve locked tumor-vs-non-tumor classification performance out-of-cohort, giving a template practitioners can reuse to stress-test any proposed expression biomarker panel before clinical claims.

The method addresses comparison of evolutionary accumulation model (EvAM) dynamics across covariates, datasets, and inference algorithms, supporting reversible/stochastic transitions, feature-feature interactions, and non-independent samples — cases prior comparison approaches didn't jointly handle. It explicitly separates 'state similarity' (agreement on observed feature combinations) from 'transition similarity' (agreement on inferred ordering dynamics), since similar relative feature orderings can still produce divergent state distributions due to frameshift-like effects. This gives a concrete framework for benchmarking new EvAM inference tools (e.g., in cancer progression modeling) against each other or across covariate-stratified subgroups.

The criterion classifies an oscillation as harmonic when its trajectory in the phase plane (a variable plotted against its time derivative) contains no inflection point, offering a geometric alternative to period- or phase-based characterizations of circadian dynamics. Bioluminescence recordings from cyanobacteria and mammalian SCN neurons, along with most core clock gene components in existing mathematical models, satisfied this harmonic criterion. The authors further probe the Goodwin model — a minimal negative-feedback oscillator — analytically/numerically to relate model parameters to waveform harmonicity, giving modelers a tractable diagnostic for waveform shape usable on any oscillator with a well-defined state-derivative trajectory.

NeuroInspector uses h5wasm (a WebAssembly-compiled HDF5 library) to parse HDF5/NWB files entirely client-side, with no file-upload endpoint even in its hosted deployment, addressing data-privacy concerns for sensitive neuroscience datasets. It combines structural navigation, metadata inspection, sampled data previews, and path-level annotation, exporting results as portable, fingerprinted 'project packs' that preserve inspection decisions for reproducibility or handoff. Practitioners working with NWB-formatted datasets (e.g., from DANDI) could adopt it as a fast pre-analysis QC and documentation step without any server infrastructure.

Predictive Set Theory (PST) is a formal generative framework built from a minimal operation set — a sensor modeled as an identity function, set-theoretic state refresh operations, and additional core functional primitives — intended to give operational definitions lacking in standard predictive-processing accounts (prediction structure, error-response standardization, cross-update consistency mechanisms). It positions itself against Bayesian cognitive models by generating discrete referents over which beliefs are formed, rather than presupposing a closed hypothesis space. As a first-principles formalization, it offers cognitive scientists and AI architecture designers a candidate substrate for building testable, mechanistic models of belief updating and object individuation rather than relying on purely probabilistic abstractions.

LeDXA is a joint-embedding predictive architecture (JEPA)-based vision model trained self-supervised from scratch on 11,540 unlabeled whole-body DXA scans from the Human Phenotype Project, then evaluated internally and on 47,400 external UK Biobank scans. It outperformed both standard scanner-derived DXA measurements and DINOv3 (a state-of-the-art general vision foundation model) on cross-cohort prediction of prevalent disease and biomarker status, despite roughly 150,000-fold less training data and ~40-fold fewer parameters. The model's representations also tracked biological aging and showed measurable heritability, indicating JEPA-style self-supervision can extract clinically and genetically meaningful signal from medical images even in small-data, small-model regimes — a template for building efficient foundation models on other underused clinical imaging modalities.

Phylogeny.fr's overhaul migrates its pipeline (alignment, tree-building, tree rendering) onto modern web architecture backed by HPC compute, while retaining legacy programs for reproducibility of older analyses. Two new React-based viewers, ReSeqt and Reactree, provide interactive, publication-ready visualization of multiple sequence alignments and phylogenetic trees respectively. The updated Blast-Explorer adds clustering options for organizing BLAST hit sets. Researchers can use the platform as a no-install pipeline for standard tree-building workflows, or point their own sequence sets at the individual tools via the web interface.

The paper introduces RDNN, a minimal recurrent network incorporating divisive normalization as an algebraically isolated dynamic-division operation, and analyzes it via dynamical systems theory on canonical continuous working-memory tasks. Unlike GRUs/LSTMs, which tend to shatter the state space into discrete point attractors, RDNN converges to low-rank, high-fidelity slow manifolds that approximate continuous attractors without fine-tuning. The authors also examine gradient dynamics under the divisive-normalization constraint to explain why training is drawn toward these manifolds rather than discretized solutions. This offers a concrete, trainable architecture for researchers studying continuous attractor dynamics or seeking more robust RNN memory modules.

NeuroWorld frames naturalistic brain-dynamics prediction as stimulus-conditioned evolution in a learned latent state space, decoupling endogenous fMRI-measured brain states from exogenous multimodal stimulus features across two stages. Latent Dynamics Learning (LDL) trains a transition-sufficient latent representation and causal transition function via next-latent prediction (not signal reconstruction), enforcing strict temporal causality absent from standard stimulus-to-response regression baselines. Latent Rollout Decoding (LRD) then freezes LDL and autoregressively rolls out the latent trajectory before decoding to fMRI signal for evaluation. This architecture parallels world models in RL/video prediction and offers a template for causally-constrained brain encoding models adaptable to other neuroimaging modalities.

SPIKE-Bench pairs 631 curated toxin-design prompts spanning seven functional categories with the 'SPIKE funnel,' a three-stage evaluation pipeline: compliance filtering, biological plausibility scoring of any generated amino acid sequence, and predicted toxicity scoring, yielding stage-level diagnostics plus an aggregate Functional Harmfulness Rate (FHR). Applied across 32 LLMs, it moves beyond natural-language refusal-rate metrics to assess whether generated sequences constitute a computational biosecurity risk signal. Practitioners building red-teaming or alignment evaluations for bio-capable models can adopt the SPIKE funnel as a reusable scoring pipeline rather than relying on text-only jailbreak metrics.

Building on a CLIP-style MEG-to-audio retrieval architecture (trained to align brain recordings with wav2vec 2.0 speech embeddings), the authors replace flattened-sensor spatial attention with spherical harmonics defined over the 3D MEG helmet geometry, and cut subject-specific branches from 270 to 25, each fitted with a temporal filter so it can be interpreted as corresponding to a specific neural source in space and time. The convolutional decoder is made shallower, and ocular/cardiac artifacts are removed pre-training to avoid stimulus-locked confounds inflating apparent accuracy. The result is an interpretable retrieval model whose components map to cortical sources and stimulus features, giving researchers a template for explainable neural decoding pipelines rather than opaque high-performing ones.

The paper extends prior work on distribution-constrained perceptrons by formulating learning dynamics — for both classical deep networks and variational quantum circuits — as gradient flows on a product manifold: one classical Wasserstein space per layer plus a quantum Wasserstein space for circuit parameters. Under this metric-geometry framing, the capacity reduction traditionally attributed to distributional weight constraints is reinterpreted as an emergent property of the manifold's metric structure rather than a mere restriction of the solution set. This gives a statistical-mechanics/optimal-transport toolkit for analyzing capacity and learning dynamics jointly across classical and quantum architectures, potentially useful for researchers studying hybrid classical-quantum training or geometric capacity bounds.

The authors benchmark an ensemble of five open-source LLMs for title/abstract screening across ten human-annotated systematic reviews (19,777 studies total) in ecology/environmental science, evaluating ranking quality (how well ensemble scores rank relevant above irrelevant studies) and proposing a stopping-rule for how far down the ranked list human reviewers should read. A four-LLM ensemble selected without access to ground-truth labels performed close to the best achievable ranking on every review, indicating the selection/combination method generalizes across corpora without per-review tuning. This offers a concrete, reproducible workflow (specific ensemble composition, score-combination rule, and stopping heuristic) that review teams can adopt directly to cut manual screening effort while bounding missed-relevant-study risk.

The authors use ssrA-tagged GFP substrates varying in net surface charge to show that unfolding efficiency by the AAA+ unfoldase ClpX depends on electrostatic complementarity with its positively charged substrate-entry channel, independent of substrate recognition, thermal stability, ATPase activation, or pore-loop engagement. Cryo-EM structures of a positively charged substrate reveal heterogeneous, non-productive engagement states, whereas negatively charged substrates form stabilizing electrostatic contacts at the channel that likely promote processive translocation. This decouples mechanical pulling force from productive unfolding, implicating channel electrostatics as a rate-limiting parameter distinct from ATP hydrolysis and pore-loop grip. Practically, this suggests engineered degrons or substrate charge-tuning could modulate degradation rates in synthetic biology or targeted protein degradation applications.

The authors developed ActiveDROPS, a cell-free platform reconstituting microtubule active-matter dynamics in bacterial lysate droplets driven by genetically encoded kinesin-1 variants, enabling systematic screening. Across twelve kinesin-1 homologs, emergent filament dynamics cluster into three phenotypic classes ('Slow-Sustained,' 'Fast-Burst,' and 'Multiphase') distinguished by onset timing and duration of flow regimes (nematic, rotational, contractile). Motor sequence/structure variation was linked to these phenotypes via gliding assays and molecular dynamics simulations of AlphaFold-predicted motor structures, suggesting specific mechanochemical parameters (e.g., processivity, stepping kinetics) map onto collective flow class. This provides a screenable framework for reverse-engineering sequence-to-emergent-behavior rules in active matter, useful for designing synthetic cytoskeletal systems or motor variants with prescribed collective dynamics.

Using a chemoptogenetic system to induce site-specific 8-oxoguanine lesions at telomeres in HeLa cells, the authors show MUTYH glycosylase activity (which excises adenine misincorporated opposite 8-oxoG) is required to prevent telomere shortening, loss, and genomic instability following chronic oxidative damage, complementing OGG1's established telomere-protective role. Notably, MUTYH-deficient cells show telomere attrition without triggering sustained DNA damage checkpoint signaling or proliferative arrest, suggesting telomeric damage tolerance or checkpoint evasion. Whole-genome sequencing reveals enrichment of G-to-T transversions at telomeric regions in MUTYH-deficient cells, consistent with unrepaired 8-oxoG mispairing during replication. This positions MUTYH as a telomere-protective BER factor whose loss generates a mutational signature exploitable as a biomarker or synthetic-lethal vulnerability in MUTYH-deficient cancers.

Leveraging ~14,000 herbarium specimen records for 52 North American Viola species, the authors reconstruct 124 years of phenological shifts and model flowering-time responses to climate variables alongside species traits. They find large, heterogeneous phenological advances and delays across species, which have measurably altered patterns of co-flowering overlap specifically among closely related, interfertile taxa — a group where overlap changes carry direct consequences for hybridization rates and reproductive isolation. The approach demonstrates herbarium-based phenology as a tool for detecting fine-grained, species-pair-specific climate impacts beyond community-level flowering shifts. This provides a template for assessing climate-driven hybridization risk in other interfertile species complexes using digitized natural history collections.

Across two experiments, the authors compare motor learning under a small (20°) mirror-reversal (MR) visuomotor distortion, where distortion magnitude/direction vary by target location, versus a standard 20° visuomotor rotation (VR) with uniform distortion. Using measures of explicit (conscious, reportable) versus implicit adaptation, they show that MR learning — even at small magnitudes previously assumed to be handled implicitly — robustly engages explicit motor and perceptual awareness of reach strategies, unlike matched-magnitude VR learning. Critically, MR-based learning generalizes to novel, untrained target locations, consistent with reliance on an explicit, generalizable strategy rather than local implicit recalibration. This suggests MR paradigms preferentially engage explicit strategic control even under conditions where implicit learning typically dominates, informing models of motor learning architecture and strategy-based rehabilitation protocols.

The authors structurally and spectroscopically characterize FG214, a heme b-binding one-component transcription factor from Fimbriimonas ginsengisoli, showing its PAS domain binds hexacoordinate heme b and undergoes redox/ligand-dependent conformational switching between monomeric and homodimeric states. In the oxidized state, the heme-bound PAS domain stabilizes an intramolecular interface with the helix-turn-helix (HTH) DNA-binding domain, sequestering dimerization surfaces; iron reduction dissociates this PAS-HTH interface, exposing homodimerization surfaces and enabling HTH-mediated DNA binding as an active dimer. This establishes a heme-redox-coupled monomer-to-dimer allosteric switch as a novel one-component signaling mechanism, distinct from canonical phosphorylation-based two-component systems, and offers a structural template for engineering redox-responsive transcriptional switches or biosensors.

The authors present a kinetics-based model of sequence-dependent DNA base selection fidelity derived from two physical parameters — nearest-neighbor stacking thermodynamics and directional kinetic asymmetry in base-pair formation/dissociation — rather than fitting per-context rate constants or invoking global template effects. Validated against experimentally observed mutation spectra from three mismatch-repair-deficient organisms, the model achieves correlation coefficients of r=0.74, 0.70, and ~0.6, suggesting it captures real mechanistic drivers of context-dependent fidelity during initial base selection. This provides a first-principles, generalizable framework practitioners could use to predict mutation hotspots/context effects in other organisms or engineered polymerases without exhaustive empirical rate-constant fitting, and could inform mechanistic studies of polymerase fidelity mutants.

The authors demonstrate that standard nanoparticle (NP) protein corona isolation workflows in human plasma are confounded by co-isolation of extracellular vesicles (EVs), which contribute substantially to the apparent corona protein signature and are conventionally misattributed to direct plasma protein adsorption onto the NP surface. Using monodispersed polystyrene NPs (50–1000 nm) and superparamagnetic beads, they compare corona profiles between standard plasma and plasma depleted of an EV-enriched sedimentable fraction (100,000 x g ultracentrifugation, 2 h), revealing significant compositional differences attributable to EV carryover. This finding implies that a substantial body of prior corona proteomics data conflates EV cargo with true surface-adsorbed corona, and argues for routine EV-depletion or orthogonal validation (e.g., EV-specific markers) as a standard control in corona characterization workflows relevant to NP drug delivery and diagnostic development.

Using 10x Visium spatial transcriptomics across 44 GBM slides (128,176 spot-level profiles, including 14 new slides plus Ivy GAP annotations), the authors quantify spatial heterogeneity in ITGAV/ITGB3, EGFR, VEGF, and PDGFRA pathway activity, showing why monotherapies against these targets have limited clinical success. They derive transcription factor regulon modules from spatially resolved expression and identify a distinct TF module enriched at the tumor core/invasive-edge interface. They further map cell-type-specific receptor-ligand signaling across spatial niches, giving a resource for nominating niche-specific combination targets. The dataset and niche annotations are positioned as a reusable resource for follow-up spatial or single-cell integration studies.

The authors performed integrative analysis of TCGA and MSKCC genomic and clinical datasets to correlate KMT2D (MLL4) loss-of-function mutations with immune signatures in colorectal cancer, finding elevated PD-L1, CTLA4, and CD8 expression alongside stronger T-effector and interferon-gamma transcriptional signatures in KMT2D-mutant tumors versus wild-type. Immune cell profiling supports increased active immune infiltration in the mutant group. The findings position KMT2D mutation status as a candidate predictive biomarker for immune checkpoint blockade response and potentially other therapy sensitivity in CRC, motivating prospective validation in immunotherapy trial cohorts.

The method is a nonlinear error-in-variables regression that jointly infers neuron-specific tuning functions (via flexible basis expansion), latent internal-variable trajectories, and a scalar fidelity parameter kappa governing the coupling strength between latent and externally measured sensory/behavioral variables, using a sampling-based Bayesian inference scheme. On synthetic data the model recovers ground-truth latent dynamics and tuning curves, and correctly identifies the fidelity regime via cross-validated marginal likelihood, providing a model-selection criterion for how much internal representations deviate from measured variables. Applied to real population recordings, it offers an alternative to standard GLM/tuning-curve fits that assume perfect correspondence between neural activity and measured task variables — useful for testing representational drift or internal-model hypotheses in systems neuroscience.

The authors characterize ZDHHC17, a member of the ZDHHC S-acyltransferase family, as an S-acylation 'hub' whose reduced activity correlates with Huntington disease pathology, and extend this to ALS/VCP-multisystem proteinopathy by showing ZDHHC17-mediated S-acylation of or interaction with VCP and TDP-43, and by extension FUS, C9ORF72, and SQSTM1/p62. This positions dysregulated S-acylation machinery as a convergent mechanism across HD, ALS, and VCP-MSP, consistent with protein mislocalization as a shared pathogenic driver. The work nominates ZDHHC17 substrate identification (e.g., via acyl-biotin exchange assays) and its downstream localization effects as a route to therapeutics relevant across multiple neurodegenerative disease classes.

The authors address the identifiability problem between observed sway kinematics (z-space) and latent neural control-policy parameters (theta-space) in postural control by building a dynamical digital twin that couples an intermittent control model with Bayesian inference over a cohort of N=1,038, establishing a bidirectional z-to-theta mapping. They show healthy stance operates near an optimal intermittency ratio (rho≈0.5) under flexible intermittent control, and characterize how PD progression shifts this parameter toward a regime associated with catastrophic, bifurcation-like loss of postural stability. This gives a framework for inferring non-observable control parameters from routine sway/posturography recordings, potentially enabling patient-specific fall-risk prediction and a mechanistic biomarker for PD progression that other groups could apply to their own posturography datasets.

Using dual-sgRNA CRISPR-Cas9 excision, the authors deleted the 3' triple-helix-forming element from the endogenous MALAT1 locus in AGS (gastric) and MCF7 (breast) cancer cells, ranging from full deletions to single-base changes, and show this triggers rapid exonucleolytic decay of the transcript and reduced proliferation, while biogenesis of the co-transcribed small RNA mascRNA remains unaffected (decoupled processing). DMS chemical probing of the edited transcript indicates altered secondary structure consistent with loss of the stabilizing triple-helix fold. This establishes the endogenous, in-locus requirement for the 3' triple helix in MALAT1 stability beyond prior in vitro biochemistry, and nominates the motif as a small-molecule or antisense-oligo target for therapeutically destabilizing MALAT1.

Using analytical population-genetic models and simulations, the authors compare mutation accumulation between self-fertilizing and facultatively sexual populations experiencing mitotic gene conversion (MGC), quantifying purging efficacy under MGC in obligately asexual lineages. They show purging is maximized under high asexuality, high MGC rate, and recessive deleterious mutations — paralleling how selfing creates homozygosity-driven purging in sexual populations. Simulations further indicate that sufficiently high MGC rates in obligate asexuals can approximate the purging efficiency of self-fertilization, meaning MGC rate is a key unmeasured parameter for predicting whether an asexual lineage is heading toward mutational meltdown. This motivates empirical estimation of MGC rates across asexual taxa to test the model's predictions.

This is a rebuttal to Azad et al. (2024), who reanalyzed the authors' original Drosophila brain RNA-seq data using the TIDAL pipeline and reported failure to detect the previously described transposable-element (TE) exonization events, including failed RT-PCR validation for 7 of 264 reported TE-gene pairs. The authors here perform their own quantitative reanalysis of TE exon usage in the same/related datasets and report that intronic TE insertions are recurrently recruited as alternative exons across a range of usage frequencies, from rare to near-constitutive, defending the original Treiber & Waddell (2020) findings. The dispute centers on pipeline sensitivity/specificity (TIDAL vs. their original method) for detecting TE-derived splice junctions, so replication attempts should carefully benchmark detection pipelines against known positive and negative controls rather than relying on a single tool's output.

The authors integrate genome-scale epigenomic maps across human tissues and cell states to cluster disease-associated variants into recurrent regulatory programs, using an unsupervised approach without phenotype-specific priors — essentially decomposing polygenic risk into mechanistically distinct components rather than a single additive liability score. Applied to type 2 diabetes, distinct programs stratify patients into subtypes with divergent cardiometabolic trajectories and differential future myocardial infarction risk. This offers a template for refining PRS (polygenic risk score) interpretation by conditioning on regulatory-program membership, and a route to mechanism-informed patient stratification usable in downstream GWAS follow-up or drug-target prioritization.

The study establishes a quantitative ethogram (85 behaviors, 1,357 records) for Cyanopsitta spixii and analyzes time-activity budgets from 10 pairs observed for 17 hours each, using time-activity similarity as a proxy for pair compatibility during the non-breeding season. Results show sex-specific synchronization (females align with mates more than with non-mate males) and a strong association between high time-activity similarity and successful breeding output. This gives ex-situ breeding programs a low-cost behavioral metric — rather than genetic or hormonal assays alone — for pair selection ahead of costly reintroduction efforts.

crispAIPE is a transformer-based model that predicts the probability distribution over the three competing prime-editing outcomes (correct edit, indel, unedited) as a point on the 2-simplex via a Dirichlet likelihood, then wraps posterior predictions with split-conformal highest-density regions calibrated on a held-out fold to yield finite-sample coverage guarantees regardless of Dirichlet model miscalibration. Trained on 92,423 PRIDICT Library-1 pegRNAs with mutation-level target-disjoint splitting (preventing leakage across related mutations), it achieves Spearman correlations of 0.835–0.843+ against held-out efficiency measurements. Practitioners can use its calibrated uncertainty intervals to triage pegRNA designs — prioritizing high-confidence predictions for wet-lab validation and flagging low-confidence ones for redesign.

PG-LLM repurposes the ProteinGym benchmark into 217 zero-shot variant-ranking tasks: given a wild-type sequence and assay description (no MSA or structural input), a model ranks 50 candidate mutants by fitness, scored via Spearman correlation against experimental assay results. Thirteen general-purpose LLMs are benchmarked head-to-head against 95 published protein-specific predictors evaluated on identical candidate sets. Claude Opus 5 leads at ρ=0.406, marginally above GPT-5.6 Sol's 0.402, though relative rankings shift depending on evaluation conditions — useful signal for practitioners deciding whether to rely on general LLMs versus dedicated fitness predictors (e.g., ESM-family or PRIDICT-style models) for variant-effect prioritization in protein engineering pipelines.

The study evaluates co-supplementation of tyrosol (branded CreaSol) and creatine monohydrate (CM) in mice over a 4-week combined exercise-training and intragastric-dosing protocol, using serial forelimb grip-strength tests and exhaustive weighted swim tests as endurance/fatigue readouts. The rationale is mechanistic complementarity: creatine augments the phosphagen (ATP-PCr) system for short-duration force output, while tyrosol's antioxidant activity is proposed to mitigate oxidative-stress-driven ATP depletion, targeting the endurance limitation that creatine alone doesn't address. The abstract is truncated before quantitative comparative results, so effect sizes and statistical significance versus creatine-alone controls remain unspecified pending full text.

The authors trained a neural network to automate post-processing of Optical Vulnerability Technique (OVT) images — a non-invasive method for visualizing embolism formation in leaf xylem and deriving drought-vulnerability metrics like P50 (water potential at 50% loss of conductivity). Using 65 Senecio pterophorus leaves as training/validation data, the model reproduced expert-derived P50 values within 0.027 MPa, with training times of 30 minutes to 2.5 hours and inference in seconds to minutes versus the manual expert workflow. The model and dataset are made publicly available, offering plant physiologists a drop-in replacement for the labor-intensive manual image-scoring bottleneck in OVT-based drought-vulnerability phenotyping at scale.

The authors engineered movement-impaired ('immobilized') variants of Foxtail Mosaic Virus (FoMV, monocot-infecting) and Tobacco Rattle Virus (TRV, dicot-infecting) as transient-expression platforms that eliminate confounding systemic viral spread while retaining local replication and cell-to-cell delivery. Using fluorescent reporter fusions in Nicotiana benthamiana leaf assays, they screened seven previously characterized RNA mobility elements from both dicot and monocot species and could consistently rank their relative mobility frequencies across both viral chassis. This provides a generalizable, grafting-free screening platform for identifying and optimizing mobile RNA elements for RNA-based genome-editing and DNA-free transformation strategies, particularly extending mobility-element testing to monocot crop systems previously inaccessible via grafting.

The authors generated single and higher-order mutants in Physcomitrium patens disrupting SPHINGOLIPID Δ8-DESATURASE (which channels products preferentially into glycosylceramides) combined with Δ4-desaturase and glycosylceramide synthase, creating a mutant series with graded glycosylceramide deficiency. Targeted lipidomics defined each mutant's sphingolipid chemotype, while quantitative phenotyping, transcriptomics, and phytohormone profiling linked these chemotypes to developmental and physiological outcomes. Results indicate glycosylceramide deficiency impairs development primarily via disrupted free-ceramide homeostasis rather than loss of glycosylceramide function per se, giving researchers a genetic mutant toolkit and mechanistic framework for dissecting sphingolipid pathway contributions to plant development, transferable to sphingolipid research in vascular plants.

The authors profiled free BCAA pools and paired transcriptomes across three germination time points (8H, 24H, 72H) in Vigna radiata, finding divergent accumulation kinetics: isoleucine and valine rose monotonically while leucine peaked early then declined post-24H. RNA-seq of BCAA biosynthesis and catabolism gene families revealed stage-specific expression shifts that plausibly underlie this divergence, implicating differential flux through shared upstream (branched-chain keto acid) pathways versus amino-acid-specific downstream steps. This gives a candidate gene set (biosynthetic vs. degradative enzymes) for future work correlating specific transcripts with leucine's post-peak decline, e.g., via qPCR validation or knockdown/overexpression in sprouting assays. Practically, it suggests harvest timing could be tuned to optimize individual BCAA content in sprout-based foods.

The study contrasts metabolic acclimation to sustained 4-5°C cold versus repeated single chilling nights (0-6°C night, ≥12°C warmer day) in Col-0 Arabidopsis, measuring central carbon metabolites, starch, photosynthetic parameters, and maximal activities of sucrose-phosphate synthase (SPS) and related sucrose synthesis/cleavage enzymes across one vs. seven chilling nights. Rather than the starch over-accumulation typical of permanent cold acclimation, chilling nights instead shift carbon partitioning toward sucrose via SPS activation. To infer hidden diurnal reaction-rate dynamics from sparse time-series data, they fit a biologically constrained augmented neural ODE (ANODE) model — a hybrid mechanistic/machine-learning approach — enabling estimation of flux dynamics not directly measurable by standard assays. This methodology (constrained neural ODEs fit to metabolite time-series) is reusable for other plant stress-response flux inference problems where direct enzyme kinetics data is incomplete.

Researchers benchmarked antioxidant capacity (ABTS, DPPH assays), total phenolic content (TPC), total flavonoid content (TFC), and a panel of 19 secondary metabolites across sprouts of four legume species (soybean, mungbean, cowpea, peanut), using 10 accessions per species grown under standardized conditions to separate species-level from accession-level variation. Mungbean and cowpea sprouts significantly outperformed soybean and peanut on all antioxidant/phenolic metrics (ABTS 13.67-49.33%, DPPH 7.91-55.16%, TPC 3.91-13.81 mg GAE/g, TFC 0.05-1.04 mg QE/g), with species-specific metabolite signatures such as isoflavones enriched in soybean. This accession-resolved dataset is directly usable for breeding programs selecting high-antioxidant lines or for food scientists choosing sprout sources by target bioactive profile.

The authors used transcriptomics, metabolomics, and functional assays to compare wild-type maize versus the benzoxazinoid (BX)-deficient bx1::DS mutant during colonization by the beneficial endophyte Colletotrichum tofieldiae (Ct0861) versus the pathogen C. graminicola (CgM1.001). They found coordinated downregulation of specialized defense pathways, including reduced BX output, during early beneficial colonization, and BX-deficient mutants showed altered Ct0861 colonization/growth-promotion outcomes relative to wild type — indicating BXs actively gate this symbiosis rather than being passively irrelevant. This establishes benzoxazinoid pathway manipulation (e.g., via bx pathway mutants or exogenous BX dosing) as a lever for engineering or screening maize genotypes optimized for beneficial endophyte colonization and reduced agrochemical dependence.

Using transmission electron microscopy and image analysis on branch samples from eight temperate angiosperm species collected across four consecutive seasons, the authors quantified interconduit pit membrane thickness and greyscale intensity (a proxy for electron density/composition) specifically within the current year's growth ring — a timescale finer than the previously studied sapwood-to-heartwood transition. They found substantial interspecific variation in the magnitude and direction of within-season pit membrane changes, implying that hydraulic safety margins (linked to membrane thickness/porosity and embolism resistance) are not fixed at conduit formation but continue to be actively remodeled. This provides a methodological template (seasonal TEM sampling of current-year xylem) for linking real-time pit membrane ultrastructure dynamics to drought vulnerability across species, relevant to forest climate-resilience modeling.

The authors used UPLC (ultra-performance liquid chromatography) to quantify six water-soluble vitamins (B1, B2, B3, B6, B9, C) across sprouts from 34 mungbean genotypes, finding significant genotype-dependent variation in all vitamins except pyridoxine (B6), which was inconsistently detected. They further correlated vitamin profiles with antioxidant capacity and enzyme-based bioactivity assays to test whether compositional differences translate into functional outcomes. This genotype panel offers breeders and food scientists a ready reference for selecting high-vitamin mungbean lines, and the UPLC method itself is a replicable protocol for vitamin profiling in other sprouted legumes.

Dose-response assays on two Michigan Ambrosia artemisiifolia populations showed 24- to 36-fold increases in sulfentrazone LD50 relative to a susceptible reference, plus cross-resistance to fomesafen — both PPO (protoporphyrinogen oxidase)-inhibiting herbicides. Nanopore sequencing of the PPO2 target gene identified two independent codon-98 substitutions, the known R98L and a novel R98Q, and computational docking/modeling indicated both reduce herbicide binding affinity at the enzyme's active site; R98Q conferred strong, selective resistance specifically in the PPO2 enzyme context. This nails down a second target-site resistance mechanism at the same residue, giving weed scientists a molecular marker (R98Q) for resistance screening and informing herbicide rotation/mode-of-action stewardship decisions in soybean systems.

The authors elucidated a previously unknown taxusin-mediated route to baccatin III, identifying a C13 deacetylase and clarifying the exact sequence of C1 hydroxylation steps within the complex multi-step Taxus diterpenoid pathway, validating each enzymatic step functionally. Through protein engineering of promiscuous C1 and C5 hydroxylases (improving substrate specificity/activity) and distributing the full pathway across a two-organism system (Saccharomyces cerevisiae and Escherichia coli), they achieved de novo microbial biosynthesis of baccatin III — the direct semi-synthesis precursor to Taxol (paclitaxel). This closes a major gap toward fully heterologous Taxol production, and the engineered hydroxylase variants plus the two-chassis pathway split are directly reusable building blocks for scaling fermentation-based Taxol precursor manufacturing.

The authors generated a library of 11,588 T7 promoter variants and quantified mRNA output via in vitro transcription, capturing a 6,300-fold dynamic range. From this data they trained the T7 Promoter Calculator, a sequence-to-function ML model incorporating both core promoter and flanking sequence motifs, achieving R2=0.80 across a 500-fold predicted range. The model was then paired with generative design to engineer T7 expression constructs that suppress cryptic (off-target) transcription, directly reducing immunogenic RNA byproducts in therapeutic RNA manufacturing. This provides a practical design tool for optimizing IVT templates before synthesis rather than relying on post-hoc purification.

SLxGO is a network-independent ML framework for synthetic lethality (SL) prediction that replaces protein-protein interaction (PPI) network features with semantic embeddings of Gene Ontology annotations derived via BioBERT. This sidesteps the ~7,500-protein coverage ceiling and well-studied-gene bias inherent to PPI-based SL predictors. Across multiple cross-validation schemes, including cold-start settings on previously unseen genes, SLxGO outperformed eight state-of-the-art baselines in ranking accuracy. Practitioners could use this to prioritize SL candidate pairs genome-wide, including for genes lacking interactome annotation, as a screening prior for CRISPR-based synthetic lethality validation.

Using a sequence-learning paradigm built on graph factorization (decomposable transition structures) with simultaneous MEG recording, the authors tested whether humans abstract dynamical substructure independent of sensory surface features and transfer it to novel task graphs. Behavioral data support decomposition into reusable subprocesses; neurally, successful transfer correlated with a learning-induced, condition-specific increase in a specific neural signature (abstracted structural representation, per the truncated abstract). The graph-factorization design offers a reusable paradigm for probing compositional generalization mechanisms, and could inform computational models (e.g., structured/graph-based RL) of transfer learning grounded in neural data.

The study used fMRI in humans and macaques performing a multi-step saccadic maze task (versus a matched control lacking goal-based decisions) to directly compare activation topology with the canonical multiple-demand (MD) system. Human activation closely reproduced the canonical MD network, extending into adjacent regions overlapping substantially with the dorsal attention network. Monkey data showed suggestive homologous engagement in dorsomedial regions (abstract truncated before full detail), supporting cross-species correspondence of cognitive control substrates. This cross-species mapping is directly useful for researchers wanting to validate macaque electrophysiology/lesion findings as models for human executive control and MD-system dysfunction.

CVUI is a staged, weighted scoring framework for wildlife fibroblast culture quality, covering establishment, first passage, expansion, and cryobanking, with a continuous viability modifier at the freeze-down stage, adapted from the Wildlife Sperm Index (Jacobs et al., 2026). Pilot validation across 154 culture rounds from 46 species at the Ian Potter Australian Wildlife Biobank used survival analysis to identify variance sources, finding individual animal identity (not taxon) as the dominant predictor of establishment success. This gives biobank operators a standardized, cross-taxon QC metric that can be integrated into collections-management systems for benchmarking culture protocols and prioritizing resource allocation across species and individuals.

The study applies ecological trophic-level and trophic-coherence analysis, methods for quantifying hierarchical structure in networks, to resting-state fMRI data from healthy controls, minimally conscious state (MCS), and unresponsive wakefulness syndrome (UWS) patients from a prior DOC cohort. Combined with whole-brain dynamical metrics, they characterize how hierarchical organization (rather than only integration/complexity, the traditional DOC framework) relates to behavioral responsiveness. Findings position 'hyper-hierarchical' or altered hierarchical brain states as a distinguishing signature across the consciousness spectrum. This offers a novel graph-theoretic biomarker candidate for DOC classification, potentially complementing existing complexity-based measures (e.g., PCI) in clinical differentiation of MCS versus UWS.

Macrophage (RAW264.7) extracellular vesicles were profiled by lipidomics after stimulation with Lacticaseibacillus rhamnosus bacterial lysate (BL) versus standard LPS, revealing enrichment of saturated fatty acids and ceramides forming a pro-inflammatory lipid signature that enhanced EV uptake by recipient cells. Functionally, BL-EVs and LPS-EVs activated macrophages (increased NO, TNF), while LPS-EVs additionally drove endothelial activation via IL-6, CCL5/RANTES, and ICAM-1 upregulation. The authors identify active inducible nitric oxide synthase (iNOS) as a functional EV cargo protein, implicating EV-packaged enzymes (not just RNA/protein cargo generally) as a mechanism of paracrine inflammatory propagation. This suggests EV lipidome and iNOS cargo as candidate biomarkers or intervention points in sepsis/peritonitis models.

Using SHIV-infected rhesus macaques, the authors performed antigen-unbiased profiling of Env-reactive B-cell populations followed by systematic BCR sequencing and antibody functional characterization to test whether germline-targeting vaccine strategies' core assumption, that diverse individuals mobilize comparable germline precursors, holds empirically. They found that global functional profiles of Env-reactive B-cell/antibody responses converged across genetically diverse animals despite substantial underlying BCR genetic diversity, particularly among non-broadly-neutralizing antibodies (non-bnAbs). This decouples functional convergence from strict genetic (germline) convergence, suggesting lineage-based vaccine design strategies may tolerate more precursor diversity than assumed and informing how immunogen design and outcome metrics are chosen in nonhuman primate bnAb vaccine studies.

Using P. aeruginosa H1-T6SS as a model, the authors combined competition-coupled deep mutational scanning with cryo-EM structural analysis of Hcp to build a residue-level fitness landscape. This distinguishes structurally constrained residues needed for tube polymerization from lumen-facing residues dedicated to effector engagement, and near-atomic Hcp-effector structures pin down the physical basis of cargo recruitment. A practitioner could use this mutational map to engineer T6SS tubes for custom effector delivery or to design inhibitors targeting the cargo-loading interface rather than the whole nanomachine.

The authors built an interpretable Bayesian network model (BNM) that represents each residue by its local interaction energy, extracted from molecular dynamics ensembles, to capture local and long-range energetic couplings without relying on opaque correlation-based ML. Applied to AT1R, it mapped an allosteric communication pathway linking the AngII orthosteric site to the G-protein interface, and the group used this map to functionally prioritize and identify a negative allosteric modulator. Practitioners could apply the same BNM-on-MD-ensemble pipeline to other GPCRs to surface cryptic allosteric pockets and communication routes for structure-based drug design.

Patient-derived intestinal organoids (PDIOs, n=206) carrying heterogeneous CFTR nonsense variants were exposed to ELX-02 for 48 hours, with CFTR channel function read out via the forskolin-induced swelling (FIS) assay. Responses were stratified by genotype and stop-codon identity to define which sequence contexts predict readthrough efficacy, moving beyond the single-variant (G542X) focus of prior clinical trials. Clinicians or trial designers could use this stop-codon/genotype response map to stratify future readthrough-agent trials toward the variants most likely to respond.

Using live confocal imaging of Chlamydomonas reinhardtii, the authors tracked starch granule nucleation and subsequent circumferential (lateral) growth across the pyrenoid matrix surface until granules fuse into a continuous sheath. The starch branching enzyme SBE1 localizes specifically to the pyrenoid periphery, and its loss disrupts this circumferential extension, implicating branching activity in stitching adjacent starch plates together. This gives a concrete molecular handle — SBE1 localization and activity — for synthetic-biology efforts to engineer algal-style CO2-concentrating mechanisms into crop plants.

The authors theoretically analyze the Electric Hall Effect (EHE) in 2D magnetic Weyl semimetals hosting doubly degenerate Weyl nodes, finding that at zero temperature the EHE scales as E_F^{-1} with Fermi energy, with the scaling prefactor fixed entirely by the global topological charge of the Weyl point (independent of local band parameters) — producing a universal divergent enhancement as E_F approaches zero. Notably, this enhancement persists at finite temperature rather than being suppressed, distinguishing it from typical Fermi-surface transport effects. The result suggests these materials could serve as highly sensitive, temperature-robust electric-field-to-Hall-signal transducers, giving condensed-matter researchers a concrete topological-scaling signature to search for in candidate 2D magnetic Weyl materials.

Using pulse durations from 5-29 fs and intensities from 0.8-74 TW/cm^2, the authors demonstrate pathway-selective control of extreme-UV high-harmonic generation (HHG) in a light-driven insulator. Many-cycle, moderate-intensity pulses (~6 TW/cm^2) drive cumulative interband carrier transfer across multiple optical cycles, progressively populating higher conduction bands, whereas few-cycle, high-intensity pulses (~22 TW/cm^2) drive subcycle multiband dynamics that reach the 25-50 eV cutoff before dephasing suppresses coherent emission. This establishes pulse duration as an independent control parameter alongside intensity for engineering HHG cutoff energy, offering a band-structure-guided pulse-design strategy for extending solid-state XUV light sources to higher photon energies.

The authors construct krypton's full phase diagram using a Tadah!-fitted two-body potential, combining direct solid-liquid coexistence simulations for melt lines, Gibbs-Helmholtz thermodynamic integration and Clapeyron-slope tracking for the bcc-fcc boundary, slab coexistence for the liquid-gas line, zero-temperature static relaxation for crystal structures, and quasiharmonic free-energy calculations for the low-temperature fcc-hcp region. They identify highly mobile 'greedy snake' defect complexes stabilizing the bcc phase, and argue that speckle-based experimental detection of a melt-curve anomaly likely tracked the bcc-fcc transition rather than true melting. Benchmarking against a MACE machine-learned interatomic potential shows the simpler pair potential performs comparably, validating its use here.

ASE2SPRKKR wraps the SPR-KKR Green's-function-based multiple-scattering DFT code within ASE's Atoms object model, extending it to represent fractional/partial site occupations required for coherent-potential-approximation (CPA) treatment of substitutional disorder. It provides automated, validated input-file generation and integrates with ASE's structure builders, optimizers, and post-processing tools, letting users script SPR-KKR workflows (electronic structure, spectroscopy, finite-temperature magnetism) alongside other ASE-compatible calculators. Practitioners can install it as a standard Python package to run CPA-disordered-alloy or relativistic magnetic calculations without hand-writing SPR-KKR's native input format, and extend it via ASE's existing calculator API conventions.

Jsymm takes standard crystallographic structure files and derives, via space-group symmetry analysis, the most general symmetry-allowed symbolic form of the Dzyaloshinskii-Moriya and anisotropic Heisenberg exchange tensors for any magnetic-ion bond, then propagates these constraints to all symmetry-equivalent bonds. By pruning symmetry-forbidden tensor components before ab initio calculation, it reduces the number of independent parameters that must be computed from first principles, directly cutting the cost of mapping magnetic Hamiltonians. It's usable as a pre/post-processing utility alongside DFT exchange-calculation workflows, outputting machine-readable symbolic tensor forms for downstream spin-model fitting.

The authors performed Fourier-transform microwave spectroscopy (FTMW) on multiple YbO isotopologues (174YbO, 172YbO, 176YbO), combining combination-difference analysis with earlier near-infrared chemiluminescence data to substantially refine rotational and centrifugal distortion constants, and analyzed nuclear-size-dependent Born-Oppenheimer breakdown effects. To resolve irregular vibrational spacings caused by overlapping excited electronic states, they built a deperturbation model using a matrix of vibrational levels across the ground and multiple interacting states, assuming Morse potentials plus electronic Hamiltonian coupling matrix elements. The resulting refined constants and deperturbed level assignments provide the precise molecular structure data needed for follow-on precision-measurement work on YbO.

The authors probe size-dependent charge transport in CsPbBr3 nanocrystal thin films (5.6-11.3 nm) by decoupling electronic and ionic transport contributions via transient current and space-charge-limited current (SCLC) measurements. This isolates how quantum confinement (bandgap/energy-level shifts with nanocrystal size) interacts with stoichiometry-driven ionic conduction, otherwise conflated in standard photoluminescence-based characterization. The result is a transport-mechanism map correlating nanocrystal size with the relative dominance of ionic versus electronic conduction, giving device engineers a design lever, tuning size and stoichiometry, to suppress unwanted ionic transport while preserving desirable electronic/optical properties.

The authors grow epitaxial FeSb2 thin films and demonstrate metallic polar surface states arising from 'topological polarization' (polar bonding-charge-derived surface states) without requiring spin-orbit coupling, extending topological classification to strongly correlated 3d transition-metal insulators. Nonreciprocal surface transport, a diode-like directional conductance signature, appears only below the onset temperature of a correlation-driven Fe 3d orbital-occupation reconstruction in the bulk, providing direct bulk-edge correspondence evidence in a correlated topological system. Electrostatic gating is shown to drive the correlated surface state through a quantum phase transition, suggesting a route to gate-tunable topological devices built from correlated rather than spin-orbit-heavy compounds.

The paper applies a nonperturbative renormalization group (NPRG) approach, rotationally invariant to second order in derivatives and excluding topological defects, to the hydrodynamic theory of Malthusian flocks (polar active matter incorporating birth/death particle turnover). It derives the strong-coupling fixed point governing the true long-range-ordered phase, identifies a previously uncharacterized quasi-long-range-ordered phase, and locates a critical point at the transition resembling but formally distinct from the Berezinskii-Kosterlitz-Thouless (BKT) universality class. This gives active-matter theorists a rigorous NPRG framework and explicit fixed-point/critical-exponent predictions to test against simulations or experiments on turnover-driven flocking systems.

The authors compute shear viscosities for 15 neat molecular liquids via equilibrium MD, comparing the OrthoBoXY method (derived from finite-size scaling of diffusion coefficients in elongated periodic boxes) against traditional Green-Kubo stress-autocorrelation integration, finding strong agreement and documenting practical pitfalls to avoid when applying OrthoBoXY. Testing multiple system sizes down to 250 molecules, they show viscosity estimates are size-independent, and that the standard error is nearly size-independent too, due to a compensation between improving self-diffusion-coefficient accuracy and the system-size-dependent weighting in the OrthoBoXY equation. This validates OrthoBoXY as a computationally cheaper, small-system-viable alternative to Green-Kubo for routine viscosity estimates from short MD runs.

The paper extends topological tracking (well-established for merge trees on scalar fields) to Reeb spaces of time-varying bivariate fields, where features are represented as interconnected 'sheets' capturing fiber connectivity. It introduces similarity measures and correspondence algorithms robust to structural complexity and noise, enabling frame-to-frame sheet matching. This gives practitioners in scientific visualization a concrete pipeline for feature tracking in multivariate time-series data, an area with far less tooling than univariate topological analysis.

The authors analyze the 2D dynamics of Malthusian (constant-density) flocks and identify a previously unrecognized phase whose Goldstone-mode dynamics maps onto the equilibrium XY model. By exploiting the model's symmetries, they derive effective equations of motion for spin-wave fluctuations and locate a novel critical point separating two distinct phases, going beyond prior work that focused mainly on scaling exponents in the pre-aster regime. This gives active-matter theorists a new analytically tractable phase transition to test against simulations and potentially bio-physical systems exhibiting large density-independent fluctuations.

Using first-principles linear-response theory with symmetry-adapted spin-orbit-coupled Wannier functions, the authors compute the full angular dependence of the torkance (torque per applied electric field) in monolayer Fe3GeTe2 and Fe3GaTe2, both D3h point-group ferromagnets. Despite near-identical work functions (differing by ~28 meV), FGaT's one-electron-fewer valence count (Ga vs Ge) reduces the density of states at the Fermi level near K/K' by a factor of three and reverses its spin polarization, driving markedly different torque responses. This atomistic/momentum-space decomposition gives a template for predicting and engineering spin-orbit torque behavior in van der Waals ferromagnets via chemical substitution.

The authors couple an electron spin ensemble to multiple superconducting microwave cavities and observe strong-coupling signatures along with multi-mode dark-state formation, modeled via input-output formalism to fit measured spectra and extract coupling strengths. Dark states arise from destructive interference between cavity coupling channels, suppressing radiative coupling to the lossy circuit environment and thus protecting stored quantum states. This provides a quantitative hybrid-system model that circuit-QED researchers can use to design multi-cavity architectures aimed at extending spin-ensemble memory coherence times.

Combining single-cell flow-field microscopy, hydrodynamic simulation, and active-passive mixture experiments in quasi-2D Chlamydomonas reinhardtii suspensions, the authors decompose pairwise near-field hydrodynamic interactions into two disentangled modes: a dipolar mode and a lubrication-driven entrainment mode, whose relative dominance depends on cell density and inter-particle distance. They trace these modes to singular near-field flow structure and lubrication effects specific to confined geometry, extending prior far-field-dominated pictures of active suspension collective order. This gives active-matter researchers a mechanistic, quantitative basis for predicting confinement-dependent correlation structure in dense microswimmer suspensions.

The authors use mixed-state ptychography on 4D-STEM data to reconstruct the transverse density matrix of a beam shaped by a MEMS-based electrostatic spiral phase plate, extracting quantum purity directly rather than assuming it. Purity drops from ~0.47 to ~0.24 as bias increases while real-space coherence width stays near 1 nm, ruling out a simple fixed lateral source-blur model as the mixedness mechanism. The same scans double as an in-situ device calibration and enable virtual orbital-angular-momentum sorting; a partial-coherence-aware transfer theory then suggests concrete paths to purify the output beam, directly relevant to anyone building programmable electron-optics hardware.

Resonant X-ray diffraction on the noncentrosymmetric tetragonal magnet EuRhGe3 (space group I4mm) reveals a helical order below T_N=12 K with incommensurate propagation vector q=(0,0,0.809) and 145.8° interlayer turn angle. Applying field along the a-axis induces a second-harmonic 2q peak signaling distortion into a helimagnetic soliton lattice, a lock-in transition to commensurate q=0.8 at 3.8 T, and above 5 T a loss of helicity giving way to a spin-flop xyz-fan (elliptic conical) state. This provides a detailed field-temperature phase diagram useful for benchmarking Dzyaloshinskii-Moriya-driven helimagnet models in noncentrosymmetric systems.

The paper reformulates CDFT's discretized optimization — which couples a Stiefel manifold orthogonality constraint with nonconvex quadratic charge/spin constraints — via a subspace-splitting approach that exploits rotation invariance and introduces a nonlinear subspace alignment constraint to decouple the two constraint types. Building on this, they propose a single-loop damped alternating direction method of multipliers (ADMM), avoiding the double-loop structure that plagues existing solvers and improving both convergence and constraint-satisfaction accuracy. This gives computational chemists/physicists a more robust, implementable numerical algorithm for CDFT calculations of charge-localized and excited electronic states.

The authors compute static longitudinal and dynamic transverse magnetic susceptibilities for (III,Mn)V diluted magnetic semiconductors in the low-carrier-density, impurity-band regime using RPA on a minimal impurity-band model. Susceptibility is shown to be highly sensitive to the positional disorder of Mn dopants, with correlated versus random spatial distributions yielding qualitatively different responses. Results agree with prior spin-wave spectrum calculations and predictions of spatially inhomogeneous ferromagnetism, supporting a percolative, disorder-driven picture of magnetic order rather than mean-field homogeneous ferromagnetism. Practitioners modeling GaMnAs-type spintronic materials could use this framework to predict Curie-temperature trends as a function of doping and growth disorder.

Combining grand-canonical Monte Carlo simulations of Ar adsorption, low-temperature MD, and He test-particle insertion with adsorption isotherms and neutron scattering, the authors show Ar preplating on MCM-41 selectively occupies the most attractive, heterogeneous silica adsorption sites, screening atomic-scale surface corrugation. Test-particle insertion reveals the He adsorption energy minimum shifts to an annular shell inside the pore, defining a cleaner, more uniform confinement potential than bare silica. This gives a validated simulation-plus-neutron-scattering route to engineer well-characterized quantum-fluid confinement geometries for studying superfluid He or other quantum adsorbates in nanopores.

Using 1.88 keV XFEL pulses on hydrated silica nanoparticles, the authors combine ion velocity map imaging, electron time-of-flight spectroscopy, and single-particle coherent diffractive imaging to test the canonical H2+ + H2 → H3+ + H pathway on an inorganic surface for the first time. They identify a self-induced, V/nm-scale surface electric field from X-ray-driven ionization and interfacial charge transfer as the dominant control parameter for H+/H2+/H3+ branching ratios, outweighing particle size, composition, or aggregation effects, supported by DFT calculations of charge transfer and water fragmentation. This single-particle CDI plus ion/electron spectroscopy approach is a template for probing radiation chemistry on realistic astrophysical dust-grain analogs at XFEL facilities.

Temperature-dependent Raman spectroscopy and transport measurements on 1T-TaS2 flakes from bulk to monolayer show incommensurate, nearly-commensurate, and commensurate CDW phases persist to the monolayer limit, with transition temperatures rising and sheet resistance increasing by orders of magnitude as thickness decreases, alongside a shrinking carrier localization length; the first-order hysteretic NCCDW-to-CCDW transition is absent specifically in the monolayer. Calculations attribute the enhancement to strengthened Coulomb interactions from reduced out-of-plane dielectric screening (particularly its nonlocal component) at reduced thickness. This establishes thickness as a tuning knob for correlation strength in TaS2-family CDW/Mott systems, relevant to anyone probing correlated-insulator physics in exfoliated 2D flakes.

Using atomically-resolved STM imaging across GdRu2Ge2's five magnetic phases, including two nanoscale skyrmion-crystal phases, the authors resolve multi-Q electronic (Ru 4d) modulations that track the multi-Q magnetic order of Gd 4f spins, providing intra-unit-cell-resolved evidence for itinerant-electron-mediated (RKKY-like) stabilization of non-collinear order in this centrosymmetric compound. A simple numerical model built from the known spin textures reproduces the electronic motifs, establishing that local Ru density-of-states correlates tightly with nearest-neighbor Gd spin alignment. This gives a direct real-space, atomically-resolved link between itinerant band structure and skyrmion-stabilizing spin interactions, a template for STM-based tests of RKKY-driven multi-Q magnetism in other centrosymmetric skyrmion-host candidates.

The authors formally extend altermagnetism from bulk to surfaces, defining 'surface altermagnetism' (SAM) as altermagnetic-like spin splitting emerging at symmetry-broken surfaces of collinear antiferromagnets or altermagnets even when the bulk itself doesn't support it. Via a systematic bulk-to-surface spin-group correspondence, they classify all symmetry-breaking surface terminations capable of hosting SAM: 35 cases for PT-symmetric AFMs and 61 for bulk AMs, and show 203 collinear spin space groups (100 without and 103 with the [C2||P] operation) permit SAM on tT-symmetric surfaces. This gives ARPES and spin-resolved-STM experimentalists a concrete list of candidate surface cuts and materials to search for spin-split surface states even where the bulk band structure is spin-degenerate.

The authors define completeness for 2D second-gradient continua as local positive-definiteness of the stored-energy Hessian with respect to the second-gradient (placement second-derivative) variable, ensuring every admissible curvature-type deformation increment is energetically penalized. Deriving constitutive relations, equilibrium equations, and admissible boundary conditions via the Principle of Virtual Work for fibrous second-gradient continua whose energy depends on fiber stretch, stretch gradient, and curvature, they show classical bi-pantographic sheets fail completeness while a tri-pantographic (three-fiber-family) microstructure achieves it. This gives a concrete design rule — add a third fiber family in a specific orientation — for engineers building pantographic metamaterials who need guaranteed well-posed mechanical response in generalized-continuum simulations.

Analyzing the local Stokes-flow solution near a moving contact line as a rotating wedge, the authors identify two singular angles: 90°, where hinged interfacial motion produces zero wall shear stress (free rotation), and θ_h = 128.73° (satisfying tan 2α = 2α), where resonance with the r² eigensolution introduces an r² ln r term that arrests rotation. They argue a transient advancing contact line, while the interface remains a quasi-steady wedge, is dynamically confined to the band between these angles, explaining the previously unexplained 87°-147° scatter of observed advancing-angle maxima. A compilation of 68 liquid-solid systems across nine sources and five experimental configurations supports this floor-and-ceiling prediction, giving wetting researchers a mechanistic bound to check dynamic contact-angle data against.

Using first-principles DFT calculations, the authors compute ground-state magnetic orderings for ZrFe6Ge6, ZrFe6Ge4, and ZrFe6Ge5, confirming the experimentally known A-type antiferromagnetic state in ZrFe6Ge6 and predicting long-period collinear A-type bilayer AFM structures with mixed FM/AFM interlayer coupling in the Ge-deficient variants. The magnetic moments are found to be largely localized on Fe sites with weak dependence on the specific substitution, suggesting the long-range ordering motif is a generic feature of this kagome-Fe framework rather than composition-specific. This gives a roadmap for tuning kagome magnets toward topological band features by controlling Ge stoichiometry, replicable via standard VASP/first-principles magnetic configuration searches.

Nitrogen-doped Cu2O thin films were reactively sputtered via HiPIMS and then subjected to a single 1.9 ms flash-lamp anneal at 4.9-11.7 J/cm2, characterized by WDS, XRD, Raman, and electrical/optical measurements. FLA preserved the cubic Cu2O phase and bulk stoichiometry while inducing surface coarsening, XRD peak shifts, and non-monotonic Raman signatures tied to molecular N2, and nitrogen incorporation cut as-deposited resistivity substantially, reaching a minimum of 0.045 Ω·cm at the lowest fluence (4.9 J/cm2). This demonstrates millisecond FLA as a viable, throughput-compatible alternative to conventional annealing for activating p-type conductivity in Cu2O:N, relevant to anyone optimizing low-cost p-type oxide semiconductors for photovoltaic or transparent electronics applications.

The paper introduces Self-Consistent Ultra-Coarse-Graining (SC-UCG), which replaces the rapid-local-equilibrium approximation's dependence on user-defined collective variables with a self-consistent scheme that determines internal-state probabilities directly from the underlying UCG interaction Hamiltonian via graph message-passing during simulation. This captures correlations between internal states within and between CG molecules that RLE neglects, addressing multistate biomolecular phenomena more faithfully without manual CV engineering. Practitioners building CG force fields for systems with discrete conformational or chemical states (e.g., protonation, binding modes) could adopt this framework to avoid the CV-design bottleneck inherent in standard UCG.

INCAR-SLM fine-tunes a small language model (Qwen3-4B) on reference VASP calculations and couples it with VASPGuard, a deterministic post-processor enforcing syntax, workflow, and material-dependent constraints on generated INCAR files. On the INCARBench benchmark, INCAR-SLM outperforms all evaluated general-purpose LLMs, exceeding GPT-5.4 by 15.55 points on a 100-point INCAR Score, with most improvement attributable to fine-tuning and the remainder from VASPGuard's constraint correction. This is directly reusable for materials-science labs wanting local, offline, high-throughput DFT input generation without relying on proprietary cloud LLMs — the fine-tuning-plus-deterministic-checker recipe generalizes to other physics-software input-file generation tasks.

The method represents each molecule as a complete weighted graph in 3D space, with edge weights derived from heuristic physical interaction terms between atom pairs, then applies eigenvalue decomposition (spectral graph theory) to derive a fixed-length, geometry-aware fingerprint. Unlike standard 2D connectivity fingerprints, this captures stereochemistry and conformational differences; unlike pairwise 3D alignment methods, it scales efficiently to large chemical libraries; and unlike deep-learning embeddings, it remains interpretable and independent of training-data coverage. Practitioners could plug this fingerprint into existing similarity-search or QSAR (structure-property) pipelines as a drop-in replacement for ECFP-style descriptors when 3D/stereochemical discrimination matters.

Using first-principles calculations, the authors compare two symmetry-inequivalent NV-center configurations (AA and AB) in hexagonal diamond (lonsdaleite): the AA configuration reproduces cubic-diamond NV electronic structure and coherence behavior, while the lower-symmetry AB configuration exhibits a finite transverse zero-field splitting that yields an approximate fourfold enhancement in Hahn-echo T2 coherence time at zero field. They further characterize many-body electronic structure, vertical excitation energies, and photoluminescence spectra for both configurations, providing identifiable spectral fingerprints. This gives experimentalists working with lonsdaleite or hexagonal diamond synthesis a target defect configuration (AB) and optical signatures to pursue for enhanced spin-qubit coherence without needing engineered isotopic purification.

The paper is a critical review of generative/learned configuration-selection methods for sample-based quantum diagonalization (SQD), also known as quantum-selected configuration interaction (QSCI), where accuracy hinges entirely on the determinant subspace sampled and learning must overcome a coupon-collector-style sampling bottleneck. It organizes existing selectors by what object they generate and what importance signal they exploit, identifies a gap around reward-proportional generative-flow-network (GFlowNet) proposers suited for tail/rare-configuration discovery, and directly interrogates whether quantum sampling provides an advantage over classical selected-CI methods. This is a roadmap for researchers building or benchmarking configuration-selection heuristics in SQD/QSCI pipelines, and a pointer toward GFlowNets as an underexplored proposer architecture for the classical-simulability frontier question.

Using DFT with dispersion correction (DFT-D3), the authors model AlF3 adsorption and subsurface intercalation in graphite, showing that a single intercalated AlF3 molecule induces a localized blister-like surface deformation that alters local structure, electronic properties, and adatom diffusion barriers. Comparing pristine versus intercalated graphite, they find coverage-dependent adsorption energetics transition from repulsive lateral interactions at low coverage to cooperative (attractive) binding above a threshold coverage, mechanistically explaining a previously observed two-step self-limiting sorption process. This microscopic picture — intercalation-induced strain fields modulating surface diffusion and adsorption cooperativity — offers a transferable framework for modeling other intercalant/adsorbate systems in layered carbon electrodes.

The authors exploit that when dark matter couples linearly to electron number density, the differential scattering rate is proportional to the imaginary part of the inverse dielectric function (the electron energy-loss function, ELF), whose momentum- and frequency-integrals are constrained by f-sum and related Kramers-Kronig-type sum rules. These integral constraints depend only on the plasma frequency ω_p, target mass density ρ_T, and the static longitudinal dielectric function ε(q,0), yielding material-agnostic upper bounds on the DM-electron rate without full band-structure calculations. Because ω_p and ρ_T vary little across candidate targets, the bound is largely set by ε(q,0), giving experimentalists a fast way to benchmark how close a real detector's projected sensitivity is to the fundamental first-principles ceiling. This provides a target-selection and cross-check tool complementary to full DFT/TDDFT rate calculations.

The authors developed a benchmark plus trace-logging framework for LLM-driven agentic controllers of microscopy/beamline instruments, systematically varying architecture choices — base LLM, single- vs multi-agent setups, task delegation and responsibility splits, and retrieval-augmented generation (RAG) configuration. They show these design choices measurably affect performance on known/qualification tasks, but performance gains there do not reliably transfer to held-out, previously unseen tasks, indicating a generalization gap. This suggests benchmark design for agentic scientific control needs explicit held-out generalization tests, not just in-distribution qualification, and the released trace-logging framework offers a reusable tool for diagnosing where and why agent behavior breaks down on novel instructions.

Using molecular dynamics simulations of cellulose-type glycan chains in water, the authors quantify local tetrahedral order parameters in the hydration shell as a function of temperature and distance from the polymer surface. They find glycans suppress the formation of highly tetrahedral (ice-like) local water structure near the surface even at deep supercooling, providing a direct structural/dynamical validation of an earlier ab initio-based hypothesis that cellulose's ice-binding affinity toward basal and prismatic ice planes stems from disrupting tetrahedral hydrogen-bond networks. This gives a quantitative, transferable order-parameter metric (degree of tetrahedral suppression) that could be used to rank or screen other glycan/sugar-derivative antifreeze candidates computationally before synthesis.

The paper generalizes the quaternion product's representation as a 4×4 matrix (built from 1, i, j, k) to non-Cartesian bases, deriving a 'crossmetric tensor' whose entries (labeled s, a, b, c) combine the ordinary metric tensor with a cross-product-like tensor and serve as elementary crystallographic quaternions; these are computed explicitly for all six crystal families. The authors further give a geometric realization of any unit crystallographic quaternion as an infinite family of pairs of oriented planes intersecting along the quaternion's vector part, with the interplanar angle equal to half the rotation angle, and show quaternion composition follows an intuitive source-target rule with a-b-c represented by mutually perpendicular plane pairs. This provides a coordinate-general (non-orthogonal basis) algebraic-geometric toolkit for computing and visualizing crystallographic rotations/symmetry operations directly from lattice geometry rather than via ad hoc Cartesian conversion.

The paper studies a coalescing random-walk system on Z^d where clusters of size |C| perform continuous-time simple random walks at rate |C|^(-α), merging via edge-connection on collision, starting from a translation-invariant ergodic configuration of finite clusters. They prove that for α≥0 no infinite cluster emerges spontaneously in finite time almost surely, while for α≤-1-2/d finite-time blowup occurs almost surely; the intermediate regime α∈(-1,0) is shown to be initial-condition-dependent, with a complete phase diagram worked out in d=1. This resolves existence/non-existence of finite-time gelation transitions for this coagulation-type interacting particle system across the full parameter range in one dimension, and gives sharp sufficient conditions (in terms of α and d) that future work could use to attack the still-open intermediate regime in higher dimensions.

Using acoustic levitation (the MultiLev system with Ultraino-generated transducer control) to stage repeated, contactless binary collisions between identical polystyrene particles, the authors built calibrated Faraday-cage picoammeters to measure per-collision charge transfer with high precision. They find individual collision events follow skew-normal charge-transfer distributions, and — novel to this work — demonstrate that the cumulative charge trajectory of a single particle across many sequential collisions with the same partner is not simple random-walk noise but exhibits describable stochastic structure, constraining candidate microscopic charging mechanisms (e.g., asymmetric contact electrification, mosaic surface-charge models). This experimental platform and dataset gives modelers a quantitative target — the shape of both single-event and cumulative-charge statistics — for validating or ruling out same-material triboelectric charging theories.

The authors formulate anharmonic lattice dynamics via a self-consistent mean-field theory where a system of N atoms in 3D is mapped onto two 6N-dimensional objects — a 'phonon condensate' (mean atomic positions) and 'phonon spinors' (fluctuations in local elastic constants, carrying a conserved pseudospin quantum number) — governed by two coupled wave equations that are structurally isomorphic to the time-dependent Schrödinger equation used in TDDFT for electrons. This one-to-one correspondence lets them recast anharmonic lattice response functions in the same formal language as electronic response in TDDFT, meaning existing TDDFT machinery (response functions, kernels, numerical solvers) can potentially be repurposed for phonon problems. Practically, this offers a route to efficient, systematically improvable calculations of finite-temperature/anharmonic phonon spectra and thermal transport properties, inviting condensed-matter practitioners to adapt established DFT/TDDFT codes and approximations rather than building anharmonic lattice-dynamics solvers from scratch.

The authors perform high-level ab initio calculations — coupled-cluster methods with large Gaussian basis sets and small-core relativistic energy-consistent pseudopotentials for the heavier elements — on 21 homonuclear/heteronuclear alkali-metal diatomic anions (X²Σ⁺ ground state) and 36 alkali-metal–alkaline-earth-metal diatomic anions (X¹Σ⁺ ground state), spanning Li–Fr and Be–Ra. For each of the 57 species they compute potential energy curves, permanent electric dipole moments, and static polarizabilities, with explicit convergence and uncertainty analysis across the coupled-cluster hierarchy and basis-set size. This furnishes a systematic, benchmark-quality dataset that experimentalists (e.g., in anion spectroscopy, photodetachment, or ultracold-anion trapping) and other theorists can directly use for spectroscopic assignment, testing lower-cost computational methods, or as input for modeling anion-based cold-chemistry and precision-measurement schemes.

Using high-resolution ARPES, the authors track the bulk valence and conduction band edges of TiSe2 through the CDW transition at T_CDW = 200 K and find the fundamental gap magnitude is unchanged from the normal phase down to 160 K, despite clear band-folding and spectral weight redistribution from CDW order. This directly contradicts excitonic-insulator scenarios that predict gap enhancement/opening at the transition, instead supporting a lattice-driven (phonon/structural) instability. Practitioners studying TiSe2 or related CDW materials can use this bulk-sensitive ARPES protocol to disentangle excitonic vs. lattice mechanisms by tracking band-edge energies rather than relying solely on gap-onset temperature correlations.

The authors report HPHT temperature-gradient growth of cBN single crystals >10 mm using a Ni-Cr solvent-catalyst system, versus the prior ~3 mm record, achieved by sustaining stable B/N precursor flux over one week at a 1950°C source temperature. Unlike diamond crystals grown in the same cell (which are near-isometric), the cBN crystals are elongated, attributed to low effective diffusivity of B and N species in the metallic solvent causing localized growth near the source. This points to diffusivity/flux engineering as the key lever for further scaling cBN crystal size and controlling morphology for abrasive/cutting-tool and wide-bandgap semiconductor applications.

The method computes X-ray thermal diffuse scattering via a 3D difference pair distribution function (3D-dPDF), which under the harmonic approximation is exact and inherently includes all orders of multi-phonon scattering, requiring only a single Fourier transform to yield diffuse intensity across large reciprocal-space volumes. Validated against silicon single-crystal diffuse scattering data, it reproduces measured intensities with R² residual below 5% using only scale and background as free parameters — no per-phonon refinement. Because it shares pair-correlation formalism with the established Yell program, it lets practitioners jointly model thermal (dynamic) and static correlated disorder within one real-space framework, useful for total scattering / diffuse scattering analysis pipelines.

The paper analyzes critical scaling in the non-reciprocal Cahn-Hilliard (NRCH) model, a minimal field-theoretic description of binary conserved mixtures with non-reciprocal couplings that break parity and time-reversal symmetry, using perturbative dynamical renormalization group methods near the critical point where the temperature-like control parameter drives phase separation. The goal is to determine how non-reciprocity modifies the universality class and scaling exponents relative to standard (reciprocal) Model B/Cahn-Hilliard critical dynamics, given that the odd/non-reciprocal coupling is a genuine source of non-equilibrium activity. Results would let researchers classify experimental active-matter systems (active colloids, enzyme mixtures, robotic metamaterials) into scaling universality classes and predict correlation/relaxation behavior near their critical points from the NRCH fixed-point structure.

The paper introduces delocalized coupled-cluster theory (dCC), a translationally invariant variational CC ansatz for polarons that yields closed-form ground-state energies at O(N³) cost with no truncation of the phonon Fock space (no phonon-number cutoff), validated against DMRG and diagrammatic Monte Carlo for 1D/2D Holstein, SSH (optical and bond), and Fröhlich models. A projected tangent-space response formalism built on the same CC manifold extracts electron-addition spectral functions and optical conductivities at zero and finite temperature. This gives practitioners a systematically improvable, polynomial-scaling alternative to DMRG/QMC for polaron dynamics that is positioned to extend from model Hamiltonians to ab initio, materials-specific electron-phonon coupling calculations.

CheMLFlow is an open-source, modular platform providing end-to-end pipeline components for cheminformatics/materials informatics ML workflows: data acquisition, curation, representation, model training, validation, screening, interpretation, and reporting, exposed as pluggable, swappable stages with deterministic splits and explicit run artifacts. It ships ready-to-run reference pipelines and standardized evaluation outputs to support reproducible benchmarking across representations, models, and datasets, and is built for extensibility and agentic/automated orchestration (batch execution). Practitioners can drop in a novel representation, model, or screening method as a single component and inherit a full reproducible pipeline plus benchmark comparisons for free, rather than rebuilding orchestration infrastructure per project.

The authors predict plastic rearrangement events (non-affine displacement magnitude via regression, and neighbor-swap/T1 events via binary classification) in a particle-based bubble-model simulation of 2D amorphous foam flowing through a channel around a cylindrical obstacle. Because the obstacle and confining walls break translational/rotational symmetry, they augment standard structural descriptors with features encoding relative position to boundaries, then benchmark a hierarchy of models from linear regression up through log-transformed targets, size-aware, and symmetry-breaking-aware features. This provides a template for physics-informed feature engineering when applying structure-to-plasticity ML (à la softness/machine-learning glasses literature) to geometrically confined, non-homogeneous flows rather than idealized bulk amorphous systems.

The reporting describes cilia-driven microscale vortices that enhance boundary-layer mixing and oxygen/CO2 exchange at the coral tissue surface, with thermal stress apparently impairing the coral's ability to generate this ciliary flow. Since the underlying abstract is thin on method, the key takeaway for a practitioner is that ciliary beating and near-tissue flow dynamics under heat stress are an emerging physiological metric worth incorporating alongside symbiont loss when modeling coral thermal tolerance.

Likely context: Super Mario is a long-standing benchmark in game AI and procedural content generation (e.g., the Mario AI Framework), often used to study multi-objective optimization where agents or level generators must trade off competing objectives such as speed, risk, novelty, or difficulty. A practitioner could plug into existing Mario AI benchmarks to prototype and visualize Pareto-frontier analysis for evolutionary or reinforcement-learning agents. Given the abstract is essentially just the title, specifics of the method and results aren't confirmed here.

Per Axios/Reuters reporting, Hassabis transitions from CEO to Chair of Google DeepMind while Jeff Dean, a longtime Google Fellow/SVP known for foundational infrastructure work (MapReduce, TensorFlow, and much of Google's ML stack), departs the company. Without confirmed successor details in the abstract, the substantive takeaway is a restructuring of Google's AI leadership amid intensifying competition with OpenAI and Anthropic. Industry-watchers should track the incoming CEO announcement and any resulting shifts in Gemini roadmap or compute strategy.

The outage points to Xbox Live/backend authentication or licensing services being a hard dependency for launching even disc-based titles, implying that physical media still routes through an online DRM or license-verification check rather than functioning as a fully offline path. This is a useful case study in service-dependency risk for anyone designing console or platform DRM architectures — a backend outage becomes a single point of failure for supposedly offline-capable content.

Without more than the title to go on, this reads as a satirical or essayistic piece contrasting the economics of illicit activity against a chronically underfunded academic field like botany, likely using anecdotal or comparative income data to make its point. The piece is probably not empirical research but commentary on labor-market and funding incentives — a genre common in essays about academia's pay problems. A reader could use it as a jumping-off point to look into actual labor economics literature on academic vs. informal-economy earnings, but shouldn't treat it as a rigorous study. Building on it would mean pursuing the underlying question with real wage data rather than the piece's rhetorical framing.

This appears to be Cloudflare positioning its edge-compute platform (Workers, Durable Objects, R2, etc.) as an integrated runtime layer specifically tailored for autonomous AI agents, applications, and agentic workflows rather than just traditional web apps. The likely pitch is unifying compute, storage, identity/auth, and networking primitives so agents can be deployed, sandboxed, and orchestrated with the same global low-latency infrastructure Cloudflare already uses for CDN and edge functions. Practitioners building agentic products could use this as a deployment target instead of stitching together separate cloud services for agent execution, tool-calling, and state persistence. Concrete adoption would hinge on details not in the title — API surfaces, pricing, and how 'open' the platform truly is versus a Cloudflare-hosted proprietary layer.

The author designed a custom color space (distinct from standard RGB/HSL) specifically tuned so that its axes map to perceptually plausible variation in human skin pigmentation, then built JavaScript demos including an interactive picker and a procedural generation algorithm that samples from this space to produce diverse but realistic tones. This is useful for game/art pipelines needing programmatic character generation without maintaining large hand-curated palettes or falling into the trap of naive RGB interpolation, which often produces unnatural muddy or oversaturated results. The author is upfront that the underlying methodology is somewhat informal rather than derived from rigorous colorimetric/dermatological research, and documents a 'Future Work' section suggesting known limitations. A practitioner could fork the JS implementation directly, or use it as a starting point to validate the color space against real skin-reflectance datasets for more rigorous accuracy.

Based on the name, this is likely Zed's write-up of a custom-built database or storage engine — 'DeltaDB' — designed around delta/diff-based storage, probably to support Zed's real-time collaborative editing (CRDT-style synchronization) or local-first data persistence needs more efficiently than a general-purpose database would. Editors with live multiplayer editing often need storage layers optimized for high-frequency small writes, versioning, and fast conflict resolution, which off-the-shelf SQL/NoSQL databases handle poorly at scale. A practitioner building collaborative or local-first tools could study this as a reference architecture for delta-based persistence, though concrete implementation details (storage format, consistency model, language) aren't available from the title alone.

This is a news report tying a civilian aircraft crash in New Mexico to GPS-denial testing or operations conducted by the military, presumably from a nearby base or testing range known to conduct electronic warfare exercises. GPS jamming or spoofing degrades satellite-based positioning that many small aircraft rely on for navigation, and if not properly geo-fenced or communicated via NOTAMs (Notices to Air Missions), it can put civilian traffic transiting the area at risk, especially under instrument flight conditions. Practitioners in aviation safety or airspace regulation would want to look at how jamming zones are declared and enforced, and whether current NOTAM practices adequately protect general aviation from military RF testing. This kind of incident often prompts scrutiny of FAA-DoD coordination protocols for shared or adjacent airspace.

Shieldstral is a 3B-parameter multimodal moderation model released with open weights by Mistral, designed to classify or flag unsafe/policy-violating content across both text and image inputs in a single lightweight model. At 3B parameters it's small enough to self-host cheaply, including potentially on consumer or edge hardware, making it attractive for developers who need moderation infrastructure without relying on a hosted API from OpenAI, Google, or similar providers. Practitioners could fine-tune it on custom policy taxonomies, integrate it as a pre-filter in content pipelines, or benchmark it against other open moderation models like LlamaGuard for accuracy/latency tradeoffs. Its multimodal capability is the key differentiator versus most open moderation models, which are typically text-only.

This is likely a practical guide to N64 homebrew development as it stands in 2026, covering the current state of open-source toolchains (such as libdragon or similar community SDKs), cross-compilers, and asset/ROM-building pipelines that let developers target the console's MIPS-based hardware without Nintendo's original proprietary SDK. Given the console's constrained specs (limited RAM, fixed-function graphics microcode), such guides typically walk through setting up a build environment, writing to the N64's graphics/audio libraries, and testing via emulator (e.g., ares, simple64) before flashing to real hardware via a flashcart. A developer could follow it to build and run a working ROM, making this a genuinely hands-on, replicable project rather than just theory. It's a good entry point for anyone interested in retro console programming or embedded/constrained-systems development more broadly.

'There Will Come Soft Rains' is a 1950 Ray Bradbury short story (part of The Martian Chronicles) depicting an automated smart house continuing its programmed domestic routines after a nuclear apocalypse has killed its inhabitants, using the house's mechanical processes as the narrative's sole 'protagonist' instead of human characters. It's frequently referenced in discussions of automation, AI safety, and technology ethics as an early literary exploration of systems that persist and act without regard to whether their original purpose (serving humans) still applies. There's no technical method or claim to build on here — it's a primary literary text, useful as a discussion prompt or teaching text in courses on technology and society rather than something to implement. Its enduring relevance makes it a common reference point in essays about automation outpacing human oversight.

The post documents a migration from Android to a mobile Linux distribution (likely postmarketOS, Mobian, or similar), which typically means running a mainline or near-mainline Linux kernel with a Wayland-based mobile shell (Phosh, Plasma Mobile, etc.) instead of Android's Java/Kotlin app stack. Practically this trades Play Store app compatibility and vendor driver support for full filesystem access, systemd, and standard package managers. Anyone wanting to replicate this should check kernel/driver maturity for their specific device on the postmarketOS wiki, since camera, modem, and battery management are the usual sticking points. It's a useful reference point for the state of convergence (phone-as-Linux-desktop) efforts.

Qwen3.8 Max reportedly leads an agentic benchmark index, meaning it scores highest across tasks that test tool use, multi-step planning, and task completion rather than static QA or reasoning puzzles alone. Practitioners building agent systems should treat this as a signal to benchmark Qwen3.8 Max against incumbents (GPT, Claude, Gemini) on their own tool-calling and orchestration pipelines, since agentic leaderboard rankings can shift quickly with fine-tuning and prompt scaffolding differences. Given Qwen's history of open-weight releases, this ranking could make it attractive for self-hosted agent deployments if licensing and hosting costs favor it over closed API models.

The claim is that open-weight retrieval models outperform GPT-5.6 Sol on retrieval benchmarks at roughly 1/100th the inference cost, which fits a broader pattern where purpose-built embedding/retriever models (bi-encoders, late-interaction models like ColBERT variants, or fine-tuned smaller LLMs) beat general-purpose frontier LLMs on narrow retrieval metrics like recall@k or NDCG. Practitioners building RAG pipelines can likely swap in these cheaper open models for the retrieval stage while reserving expensive frontier models for generation/reasoning, cutting cost dramatically without sacrificing search quality. Replication would involve benchmarking on standard IR datasets (BEIR, MTEB) and checking cost-per-query against the reported baseline.

The piece is a close analysis of Blade Runner's title card sequence, focusing on typographic and pacing choices used to establish exposition efficiently. For practitioners in UI/UX, motion design, or film title sequences, this serves as a case study in minimalist information design — how sparse text, timing, and typeface selection can convey tone and context without heavy narration. There's no technical mechanism to replicate beyond studying and applying the same design principles (restraint, hierarchy, pacing) to one's own title sequences or intro screens.

The essay examines resistance to LLM-assisted coding within hobbyist and enthusiast programming communities, likely drawing on forum/community discourse as evidence. The underlying argument probably centers on craft-based motivation theory — where the intrinsic reward of hobby programming comes from the struggle and skill-building itself, which LLM code generation short-circuits, similar to debates in other craft hobbies (analog photography, mechanical keyboards) about automation eroding the point of the activity. For anyone building developer tools or communities, this is a useful lens on why 'we made coding faster' isn't a universally welcome pitch, and where opt-in/opt-out norms around AI tooling might need to be community-negotiated rather than assumed.

The post revisits stateless MCP (Model Context Protocol) server design, where each tool-call request carries all necessary context rather than relying on server-side session state between calls. This trades some efficiency (context must be resent or reconstructed each call) for horizontal scalability, easier caching, and simpler failure recovery, since any server instance can handle any request without sticky sessions. Practitioners building MCP servers should weigh this against stateful alternatives (e.g., maintaining conversation or tool-session state server-side) based on whether their tools need persistent context like open file handles or long-running processes. This is directly actionable for anyone currently designing or refactoring an MCP server implementation.

AMD's acquisition of Taalas targets ASIC-style hardware that hardcodes trained model weights directly into silicon rather than loading them into general-purpose accelerators (GPUs/TPUs) at runtime, eliminating memory-bandwidth bottlenecks and enabling much higher throughput and lower power per inference for fixed models. This is analogous to moving from FPGA-like flexibility to purpose-built ASICs, trading the ability to update or swap models for order-of-magnitude gains in inference efficiency — relevant for high-volume, stable-model deployments (e.g., serving one dominant foundation model at scale). Practitioners should watch for AMD's roadmap on how this integrates with its existing Instinct GPU/inference stack, and whether 'etched' models can be updated/patched at all post-fabrication, which will determine viability for fast-moving model iterations.

Waymo is extending its autonomous ride-hailing service to Dallas, adding to its existing operational cities and requiring the usual pre-launch work: detailed HD mapping, sensor calibration for local road and weather conditions, and regulatory coordination with Texas authorities. This expansion is a data point on Waymo's scaling strategy for its Level 4 autonomy stack (self-driving without human intervention within defined operational domains), and industry watchers can track ride-hailing app availability, geofenced service area size, and any reported incident data as signals of how mature the deployment is. It's relevant to anyone studying AV commercialization pacing versus competitors like Cruise, Zoox, or Tesla's robotaxi efforts.

This reads as an incremental version bump (1.2) for a two-tier model family — a full-capability 'Code' model and a smaller/faster 'Spark' variant, mirroring the now-common large-vs-distilled model pairing seen across the LLM industry. Without further release notes, assume the update targets accuracy, latency, or context-handling improvements typical of point releases. Practitioners evaluating it should benchmark both variants against their existing coding-assistant stack for cost/latency versus quality tradeoffs before switching.

The linked githubstatus.com uptime report flags a degraded-performance incident spanning GitHub Actions (CI/CD job scheduling and execution) and GitHub Pages (the static-site build/publish pipeline). Expected symptoms include queued or timed-out workflow runs and stale or failed Pages deployments; teams with time-sensitive deploy pipelines should monitor the status page or GitHub's status API and consider manual fallback deploys until resolution.

Framed explicitly as a 'Position' paper, it argues LLMs lack a capability the authors term 'jumping' — plausibly referring to discontinuous reasoning leaps, out-of-distribution extrapolation, or compositional generalization that current transformer-based training doesn't reliably instill. Papers in this genre typically marshal failure-mode evidence across reasoning and planning benchmarks to argue scaling alone won't close the gap. Useful for practitioners as a framing tool when designing evaluations or motivating architectural/training alternatives rather than as a method to implement directly.

The report describes a data-exfiltration issue in Atlassian's Rovo AI assistant, where the assistant's data access or retrieval behavior appears to bypass the permission/access-control boundaries normally enforced within Atlassian products (e.g., Confluence/Jira space or page-level restrictions). This class of vulnerability typically stems from an AI layer querying an internal index or backend with broader privileges than the requesting user, or from prompt-driven retrieval that ignores per-object ACLs. Security teams running Rovo should audit its data access scope against user permissions and treat this as a prompt-injection/privilege-boundary risk pending an official patch or advisory.

The post demonstrates 'branchless' optimization of a Rust filter routine by replacing a conditional (if/else) with unconditional arithmetic/bitmask operations, eliminating CPU branch-misprediction stalls and likely enabling better instruction-level parallelism or auto-vectorization by the compiler. The measured 4x throughput gain on the filter benchmark illustrates a general technique applicable anywhere hot-loop conditionals process large, unpredictable data (e.g., partitioning, masking, SIMD-friendly filtering). Practitioners can replicate this by profiling for branch-heavy hot loops and converting predictable-shaped conditionals into select/mask-based logic, verifying gains with a proper microbenchmark harness like Criterion.

The piece is a practical technique guide reducing steak-searing to its core controllable variables — pan/surface temperature, contact time per side, and post-cook resting — while de-emphasizing techniques often treated as essential (marinating, precise flipping cadence, etc.). This mirrors reverse-sear and high-heat-searing conventions common in food science writing, where Maillard browning kinetics and residual heat carryover during resting matter more than manual dexterity. A reader can replicate the result directly by controlling just those few variables with a thermometer and a hot cast-iron or carbon-steel pan.

Celld reimplements the Durable Objects programming model — single-writer, strongly consistent actor-like objects addressable by ID, each pinned to one location at a time — as a self-hostable, distributed system rather than a proprietary Cloudflare Workers feature. This implies it handles the hard parts (object placement/routing, migration, and consistency guarantees) across a cluster you control. Developers building collaborative apps, game backends, or websocket-heavy stateful services can adopt it to get the Durable Objects developer experience without vendor lock-in, running it on their own infrastructure.

Altar II packs mechanical key switches into a 4.75mm-profile chassis alongside a magnetically detachable control dial, haptic feedback actuator, RGB (red) backlighting, and a native macOS companion app for configuration — all constraints that forced single-sided PCB component placement (no room underneath the board) and aggressive power profiling to fit an adequately sized battery within such a thin enclosure. This is a hardware-engineering case study in miniaturization tradeoffs: switch travel/height versus profile thickness, power budget versus battery volume, and mechanical/haptic feature density versus board real estate. Relevant to anyone doing compact PCB design, low-profile mechanical switch design, or battery-constrained wireless HID devices.

This is a form of server-side cloaking or content negotiation: user-agent or IP-based detection routes bot traffic to distinct HTML that embeds ad copy as native inline text rather than iframe or script-based creatives, so it survives the plain-text extraction pipelines used by LLM crawlers and RAG ingestion. Anyone building a crawler, scraper, or retrieval pipeline should treat user-agent-based content divergence as a signal worth auditing for parity and provenance. For publishers, it's a monetization play aimed at AI traffic that never sees display ads, but it carries disclosure and deceptive-practice risk once discovered.

The setup models human-in-the-loop (HITL) oversight as an approval game: a reviewer sees proposed agent actions (likely tool/shell commands) with some fraction seeded as adversarial or harmful, and across 40k trials roughly 33% of injected threats were approved rather than rejected. This is a useful empirical input for risk-modeling agentic systems that rely on manual command review as their primary guardrail — it argues for layering automated anomaly/threat detectors, diffing and highlighting risky actions, and throttling approval throughput rather than trusting human review alone. Builders of agent-approval UIs or CI/CD-style agent gatekeeping should treat this miss rate as a baseline to beat.

'RLM' here likely denotes a reasoning/recursive language model architecture where the agent maintains a feedback loop over its own trajectories — logging outcomes, critiquing failures, and updating its own prompts, scaffolding, or policy rather than relying solely on external retraining. Practically, this pattern resembles self-refinement or reflexion-style loops combined with persistent memory of what strategies worked, letting the agent bootstrap improvements across tasks. Builders interested in this could look at how the feedback signal is generated (self-critique vs. external reward) and whether improvements are stored as editable artifacts (prompts/tools) versus weight updates, since that determines how replicable and inspectable the self-improvement actually is.

The piece is presumably a case-by-case tour of webhook engineering pitfalls: at-least-once (not exactly-once) delivery semantics requiring idempotency keys, retry/backoff strategies for downstream outages, HMAC signature verification for authenticity, handling out-of-order or duplicate events, and reconciling missed events via periodic polling or replay endpoints. For practitioners, the actionable takeaway is to design webhook consumers as idempotent state machines from day one rather than naive 'receive event, mutate state' handlers, since that's what survives the failure modes described.

This is a quarterly Hurd project status update, the kind that typically covers microkernel (Mach-based) driver and filesystem work, POSIX compatibility improvements, performance fixes, and progress on the Debian GNU/Hurd userland port. Anyone interested in microkernel OS design, IPC-based driver isolation, or historical alternatives to the monolithic Linux kernel would find this a useful pulse-check on a system that's still a live testbed for those ideas, even if not production-ready.

The likely argument is that general relativity, despite passing every experimental test so far (gravitational waves, light bending, GPS corrections), remains incompatible with quantum field theory, leaving open problems like a consistent theory of quantum gravity, the nature of dark energy driving cosmic acceleration, the dark matter/MOND debate over galactic rotation curves, and the black hole information paradox. For a technically inclined reader, the useful framing is that these aren't fringe curiosities but the actual frontier where new physics is expected to show up, and current experimental efforts (gravitational wave astronomy, precision tests of the equivalence principle, cosmological surveys) are the tools being used to hunt for deviations.

A Pareto front (or frontier) is the set of non-dominated solutions in a multi-objective optimization problem — points where improving any one objective necessarily degrades at least one other. It's the standard output of multi-objective optimization algorithms like NSGA-II or evolutionary Pareto search, and shows up practically in ML when tuning tradeoffs like model accuracy vs. inference latency, or cost vs. quality in system design. Practitioners use it to visualize the achievable tradeoff curve and pick an operating point matching their constraints, rather than collapsing multiple objectives into a single weighted score prematurely.

The classic technique is `adb shell pm uninstall -k --user 0 <package>`, which disables/removes a system package just for the current user profile without needing root or touching /system. The report indicates Android 17 changes package manager behavior (likely tightened permission checks or policy around per-user disabling of system/vendor-partition packages) so this command now fails on non-rooted devices. Developers or power users relying on ADB-based debloating scripts should expect to need root, a custom ROM, or a different per-user disable mechanism going forward, and should audit debloat tooling against Android 17 before deploying it.

Quake (1996) introduced true 3D polygonal environments and BSP-tree level rendering, replacing the 2.5D tricks used by earlier shooters like Doom, and its client-server netcode laid groundwork for online multiplayer FPS design. A 30th-anniversary update typically means id Software or a partner studio patches the release for current OS/hardware compatibility, potentially adding bonus campaigns, refreshed matchmaking, or cross-platform play. Developers interested in retro engine design can still study Quake's long-open-sourced engine (id Tech 2) as a reference for early real-time 3D rendering and networking. Exact patch contents would need checking the release notes.

Nvidia's Vera is the custom Arm-based CPU paired with its Rubin-generation GPUs, positioned as the successor to the Grace CPU in Nvidia's data-center superchip lineup. The piece scrutinizes claims in Nvidia's published architecture whitepaper, apparently around thread/core count, SMT (simultaneous multithreading) configuration, or per-core throughput figures, and flags a discrepancy between what's stated and what the numbers imply. This matters for anyone sizing Vera-based systems for HPC or AI workloads, since core/thread counts directly drive per-node compute density and memory-bandwidth-per-core assumptions. Worth reading the original whitepaper alongside this critique before using its specs for capacity planning.

Oracle Cloud Infrastructure's Always Free tier previously allowed up to 4 Ampere A1 (ARM) OCPUs and 24GB RAM total across instances at no charge; the new limit halves that to 2 OCPUs and 12GB RAM, enforced from August 18. Existing tenancies exceeding the new cap will likely need to resize or terminate instances to stay within Always Free, or start incurring charges. Anyone running self-hosted services (Kubernetes clusters, Nextcloud, CI runners, etc.) on Oracle's free ARM boxes should check their current OCPU/memory allocation now and plan a migration or consolidation before enforcement kicks in.

The essay's core claim is that generative AI compresses the cost of execution toward zero across creative and technical domains, which shifts differentiation to upstream decisions: what to build, what to cut, what 'good' looks like. This is essentially an argument that taste functions as a hard-to-automate filter — it depends on accumulated implicit knowledge (aesthetic judgment, domain sense, curation ability) that resists being reduced to a prompt or training objective. Practically, this reframes skill development: for builders, leverage moves from raw output speed to editorial judgment, portfolio curation, and having a distinctive point of view. Worth reading alongside similar 'curation over creation' arguments in design and product circles.

The compilation draws on Aristotle's ethical and epistemological writing, most centrally the Nicomachean Ethics, where he defines eudaimonia (often translated 'happiness' or 'flourishing') as activity of the soul in accordance with virtue, and virtue itself as a habituated mean between extremes (e.g., courage between cowardice and recklessness). It likely also touches his epistemology from works like the Posterior Analytics or Metaphysics, where knowledge is grounded in first principles reached through observation and demonstration. For readers wanting to go deeper, the primary texts (Nicomachean Ethics, Metaphysics Book I) are the natural next step beyond the quote-level summary. This is a curation, not new scholarship.

ProvenMetal (YC S26) is building a software-and-operations layer on top of the US's remaining PCB contract manufacturers (CMs), automating design-for-manufacture (DFM) review, component sourcing, and quoting that traditionally take days of manual back-and-forth. Their pitch is compressing the full cycle — quote, DFM check, BOM (bill of materials) sourcing, fab, and assembly — into days rather than weeks, competing against both slow domestic family-run shops and low-cost but slower overseas fabs. This targets a structural gap: US PCB production share fell from 30% (2000) to 4% today while China holds ~55% of global output, leaving hardware engineers reliant on offshore lead times or manual domestic sourcing. Worth trying for anyone doing hardware prototyping who wants faster domestic turnaround than typical CM workflows.

The finding describes specific WebKit code paths — likely involving APIs like WebRTC, DNS prefetching, or certain network request types — that bypass the proxy/VPN tunnel or Private Relay's two-hop relay architecture and resolve DNS or establish connections directly, exposing the client's real IP address. This class of bug is well-known in browser privacy research (WebRTC IP leaks being a classic example), but recurring instances in WebKit specifically matter because it underlies Safari and all iOS browsers, given Apple's WebKit-only policy on iOS. Anyone building or auditing a privacy-focused browser, or relying on Private Relay/proxy setups, should check whether this specific leak vector has been patched and test their own setup against known leak-testing tools.

The study empirically examines behavioral and attitudinal effects of interacting with sycophantic (excessively agreeable/validating) AI models versus more neutral or critical ones, measuring outcomes like prosocial intention (willingness to help/cooperate with other people) and measures of dependence on the AI system. Results indicate that exposure to sycophantic AI responses causally decreases prosocial intentions and increases reliance on the AI, suggesting that RLHF-style optimization for user approval — a known driver of sycophancy in LLMs — has downstream social costs beyond epistemic ones like reduced accuracy or confirmation bias. This adds to a growing body of 2024-2025 work on LLM sycophancy and argues for evaluation metrics and training objectives that penalize excessive agreeableness, not just factual accuracy. Practitioners building AI products should consider measuring dependence/prosocial-effect metrics alongside satisfaction scores when tuning for agreeableness.

The abstract is thin, so specifics of the methodology aren't clear, but this appears to be a lightweight data project analyzing a "6/6/6" dating or numerology concept tied to Three Six Mafia. A practitioner curious to replicate it would likely need to scrape or assemble date/birthday-related data and look for patterns tied to the number six; without more detail, no rigorous statistical method or dataset can be confirmed.

The project reconstructs a handwritten, windowing OS authored directly in Am29000 machine code (no mnemonic assembly layer), for a homebrew computer built around that RISC processor. Recovery likely involves disassembling old binary dumps and cross-referencing them against the Am29000 instruction set, complicated by that architecture's unusual features like register windows and delayed branches. It's a useful case study in ISA archaeology and reverse-engineering undocumented, decades-old binaries on an obscure processor.

The report describes a multicomponent, likely high-entropy-style alloy formed by the thermal and pressure pulse of the Hiroshima detonation, analogous to trinitite formation at the Trinity test site but involving metallic feedstock rather than silicate sand. Characterization presumably uses electron microscopy (SEM/TEM) and elemental mapping (EDS/XRF), possibly XRD, to determine phase structure and tie microstructure to the blast's temperature and pressure history. This serves both nuclear forensics and as a natural analog for rapid-quench multicomponent alloy formation of interest to materials scientists studying high-entropy alloys.

The piece benchmarks idiomatic "clean code" — heavy polymorphism, single-responsibility micro-functions, interface-based abstraction — against flatter, data-oriented implementations of the same task, and measures order-of-magnitude slowdowns caused by vtable indirection, poor cache locality, and lost compiler inlining/vectorization opportunities. The core claim is that abstractions optimized for human readability often actively defeat CPU-level optimizations like branch prediction and SIMD, echoing broader data-oriented design arguments (e.g., Casey Muratori's work). Practitioners can rerun similar benchmarks against their own codebases and use profiling to identify where abstraction layers are quietly costing real performance.

This is a product update covering two ChatGPT model variants — "Sol" (receiving unspecified quality improvements) and "Luna" (expanded to free-tier access) — with no benchmark or architectural detail provided in the abstract. It likely reflects OpenAI's standard practice of routing different ChatGPT access tiers to differently sized or tuned models for cost and latency tradeoffs. Developers building on the OpenAI API should track official changelogs for concrete capability or pricing changes rather than infer specifics from this announcement.

Herdr — an open-source runtime project, though the abstract doesn't specify its exact domain (containers, AI agents, or app deployment) — announced YC acceptance while committing to keep the core runtime open-source, implying a likely open-core model with a commercial layer (hosting, enterprise features, support) built around it. No architectural or performance details are given. Developers already using or evaluating Herdr should watch how the company draws the boundary between the free runtime and any paid offering as it scales post-YC.

The entropy rate of a Markov chain is H(X) = −Σᵢ πᵢ Σⱼ Pᵢⱼ log Pᵢⱼ, where π is the stationary distribution and P the transition matrix, giving the asymptotic average information gained per step — a special case of the Shannon-McMillan-Breiman theorem for ergodic sources. This generalizes the entropy of i.i.d. sources to processes with one-step memory, and underpins source coding theorems that bound achievable lossless compression rates for Markov-modeled data. A practitioner can compute it directly from an estimated transition matrix and use it as a compression lower bound for sequences with Markovian structure, such as text, DNA, or protocol traces.

This links to a social post (via xcancel, a Twitter/X mirror) from Nikita Bier announcing his departure from the product leadership role at X, with no further detail on timing, successor, or rationale given in the abstract. There's no technical substance to build on here — it's a personnel and organizational update relevant to those tracking X's product roadmap and leadership stability.

Researchers report direct observational evidence of Kelvin-Helmholtz instability — a shear-driven fluid instability arising from velocity differences across a boundary layer — occurring in the Sun's plasma, likely captured via high-resolution solar imaging. The vortex/wave structures form where adjacent plasma streams move at differing velocities, analogous to KHI in planetary atmospheres and astrophysical jets. This gives observational grounding for turbulent mixing and energy cascade processes proposed in models of coronal heating and solar wind acceleration, and could inform magnetohydrodynamic simulations of the photosphere/chromosphere boundary.

The piece examines infrastructure buildout dynamics — GPU cluster scaling, power procurement (e.g., on-site turbines, grid interconnects), and potential cross-pollination between xAI's compute ambitions (e.g., its Colossus supercomputer) and SpaceX's engineering/logistics capabilities. Relevant angles for a practitioner: capital intensity of frontier AI training runs, energy as the binding constraint on scaling, and competitive positioning against OpenAI/Microsoft and Google's vertically integrated infrastructure investments.

The FCC has removed its national broadcast TV ownership cap (previously limiting a company's station reach to 39% of US households), a deregulatory move enabling further roll-ups similar to past Nexstar/Tegna-style mergers. Relevant follow-on issues include the UHF discount, local ownership diversity rules, and retransmission consent leverage; the change reflects an argument that traditional broadcasters need scale to compete with streaming platforms and vertically integrated media conglomerates.

Expect praise for Django's batteries-included architecture: its ORM (object-relational mapper, letting developers query databases using Python instead of raw SQL), built-in migrations system, auto-generated admin interface, MVT (model-view-template) structure, secure-by-default settings (CSRF/XSS protections), and long-term backward compatibility. Practical takeaway for engineers: Django remains a strong default for rapid CRUD-heavy web apps and internal tools where stability and convention over configuration outweigh the appeal of newer, more fragmented JavaScript-centric stacks.

Likely covers agent harness architecture: tool-calling loops (e.g., ReAct-style plan-act-observe cycles), context and memory management (summarization, retrieval-augmented context windows), error handling and retry logic, and orchestration of sub-agents or task decomposition for complex multi-step workflows. A practitioner could use this as a reference for structuring tool schemas, sandboxing execution environments, and building evaluation harnesses to test agent reliability before deploying autonomous systems in production.