Saturday, 1 August 2026

ACE is a human-centric data-collection infrastructure that spatially calibrates and temporally synchronizes multimodal capture across two configurations — table-scale for dexterous hand-object manipulation and room-scale for whole-body locomotion and interaction in furnished homes. It jointly records egocentric and multi-view exocentric video, full-body and articulated hand pose/motion, object geometry/state, audio, and touch, addressing the fragmentation problem where prior embodied datasets only partially observe the perception-action loop. This positions ACE as foundational training data infrastructure for embodied intelligence models that need joint multimodal supervision rather than single-modality streams. Practitioners building imitation-learning or world-model pipelines for household robotics could use ACE-style synchronized captures to train policies that generalize across manipulation and locomotion tasks.

PhiZero learns a discrete, compact 'physical language' representing world-state transitions via self-supervision on in-the-wild video, then adopts a reason-then-render paradigm: it first autoregressively infers future dynamics as a symbolic physical-language sequence, then decodes that sequence into pixel-space video. This decouples dynamics reasoning from visual rendering, unlike standard video-prediction world models that implicitly encode physics within a high-dimensional pixel predictor. The approach is validated across both generation and understanding benchmarks, suggesting the physical-language tokens are useful as an intermediate representation for downstream reasoning tasks, not just video synthesis. Builders could reuse the physical-language tokenizer as a standalone world-state encoder for planning or reasoning modules that don't require full video rendering.

AskChem restructures literature retrieval around atomic, typed, provenance-carrying claims (each grounded to a source DOI plus a verbatim quote or evidence locator) rather than whole documents. Over this shared claim store it exposes a stabilized faceted taxonomy for hierarchical browsing, an evidence graph linking claims via typed relations, and a dynamically updated 'living taxonomy' for exploratory synthesis. This claim-centric architecture directly targets the provenance and cross-paper aggregation weaknesses of standard RAG-over-documents pipelines used by LLM science agents. Practitioners building chemistry-domain LLM agents could plug into AskChem's claim store as a grounded retrieval backend to reduce hallucination and improve citation traceability in synthesized answers.

AISPA is an auditing framework that decomposes system prompts into individual instructions and scores them along eight user-relevant dimensions, then classifies each as protective or problematic. The authors applied it to 3,249 instructions extracted from system prompts across 88 commercial AI products, producing an empirical dataset of real-world deployment behavior rather than lab-controlled prompts. Reported findings include substantial variance in system prompt design practices across products/developers, implying inconsistent safety/UX standards industry-wide. Practitioners auditing their own or competitor AI products could adopt the eight-dimension rubric directly, or use the released instruction dataset as ground truth for training automated system-prompt classifiers.

Chimera is a hybrid diffusion transformer that processes text, image, and video tokens as one raster-ordered stream without positional embeddings, combining Kimi Delta Attention (linear-complexity long-context state tracking), interleaved Multi-head Latent Attention (for direct global interaction), modality-aware short convolutions (local spatiotemporal context), and sparse Mixture-of-Experts layers for capacity scaling with bounded activated compute. Its key contribution beyond the architecture is HeteroP, a module-wise hyperparameter-transfer scheme that scales width/depth hyperparameters according to each tensor's functional fan-in and depth, addressing the fact that standard scaling-law transfer breaks down for heterogeneous attention mixtures. This gives practitioners a validated recipe (Chinchilla-style scaling but adapted to hybrid attention) for training large visual diffusion backbones without re-deriving hyperparameters from scratch at each scale. The architecture is directly relevant to anyone building long-video or high-resolution multimodal diffusion models constrained by full-attention's quadratic cost.

OSReward is a benchmark for evaluating vision-language model (VLM) judges used to verify computer-using-agent (CUA) trajectories — sequences of actions, observed states, and reasoning traces. It sources trajectories from diverse agent backbones executing human-verified instructions across multiple platforms, with ground-truth verdicts established through multi-stage human annotation rather than automated heuristics. The benchmark directly targets the reliability gap in using VLM-as-judge for CUA evaluation, data curation, and RL reward modeling — a practice that has become standard without systematic validation. Practitioners building RL pipelines or data-filtering systems for computer-use agents can use OSReward to select or fine-tune a VLM judge with measured reliability rather than assuming judge quality.

The authors show that safety fine-tuning targeting self-attributed consciousness in LLMs has a broader, unintended effect: it suppresses mind-attribution to non-human animals and natural objects and reduces expressed spiritual belief, indicating shared underlying representations of 'mindedness' across self and other. They demonstrate two reversal methods — ablating the learned safety-refusal direction, and mechanistically steering a 'consciousness' vector in activation space — both of which restore broad mind attribution and shift responses on standardized sociological survey instruments (religiosity, moral values, hope, well-being) toward more human-like patterns. Critically, these shifts occur without degrading Theory of Mind task performance, suggesting the suppression is a representational side-effect of alignment rather than a necessary consequence of safe behavior. This is directly relevant to interpretability researchers doing activation-steering or direction-ablation work, and raises a concrete methodological concern for anyone evaluating alignment interventions using downstream psychological/sociological survey proxies.

Beacon frames agentic visual reasoning around two evaluation axes for multimodal LLMs' tool use: Mode Adaptiveness (whether the model recognizes when tool invocation is actually necessary versus unnecessary overhead) and Tool Effect (whether tool use extends capability on otherwise-unsolvable problems without introducing new errors on already-solvable ones). This reframes the goal of agentic visual reasoning from maximizing reasoning sophistication to maximizing task success rate under efficient, well-calibrated tool invocation. The paper conducts a comprehensive analysis (details of the resulting method/benchmark are truncated in the abstract) using these two dimensions as diagnostic criteria for MLLM tool-use policies. Practitioners designing tool-augmented visual reasoning agents could adopt MA/TE as evaluation axes to diagnose whether their agent's tool-calling policy is miscalibrated (over-calling or under-calling) rather than relying solely on aggregate accuracy.

VAD targets the source-mixing problem in multimodal on-policy distillation, where next-token corrections from a privileged-view teacher conflate visual grounding with linguistic priors and teacher-specific artifacts. At each student-generated prefix it runs a counterfactual: evaluate the same fixed teacher with relevant visual evidence present versus ablated, and take the change in centered log-probabilities as ut, a signed per-token estimate of visual attributability. This isolated signal is then used to reconstruct distillation targets that emphasize genuinely evidence-supported corrections, giving a cleaner training signal than raw teacher logits for OPD pipelines.

MixFrag is a mixed-precision post-training quantization (PTQ) framework for ViTs that replaces uniform bit-width assignment with component-level fragility estimation. Fragility is quantified via KL divergence between full-precision and isolated quantized output distributions on a small calibration set, capturing heterogeneous sensitivity across transformer components (e.g., attention vs. MLP blocks). Bit allocation is then formulated as a Multiple-Choice Knapsack Problem (MCKP) to optimally assign per-layer precision under a global bit-budget constraint, solved without retraining. Results on ImageNet-1K across multiple ViT variants show this fragility-aware allocation outperforms uniform-bitwidth PTQ baselines at matched compression ratios.

The paper shows vanilla on-policy self-distillation (OPSD) corresponds to β=1 in a broader policy-optimization family where β weights the KL penalty anchoring the student to a reference policy, versus pursuing privileged teacher guidance. Generalizing to β-OPSD, they derive the optimal policy in closed form as a geometric interpolation between the reference policy and the privileged teacher distribution. Since directly optimizing this objective via RL would be costly and high-variance, they instead propose a more tractable training procedure (described further in the omitted text) that approximates the derived optimal policy. This gives practitioners a principled, tunable regularization axis (β) for stabilizing self-distillation of reasoning LLMs instead of ad hoc engineering fixes.

ROAD transfers priors from discriminative 3D foundation models into diffusion transformers for 3D shape generation, aiming to cut training cost versus scaling generative capacity and data alone. It addresses semantic-structural heterogeneity between generative and discriminative latent spaces via a reciprocal-objective alignment strategy combining 'Holistic Semantic Condensing' (global semantic coherence) and a structural optimization component (truncated in the abstract) for fine-grained geometric consistency. This is essentially a distillation/alignment recipe bridging discriminative representation learning and generative diffusion training for 3D assets. Practitioners working on 3D diffusion transformers could apply this alignment loss to warm-start generative training from existing discriminative 3D encoders, reducing compute versus from-scratch training.

DualG-MRAG addresses the granularity dilemma in graph-enhanced multimodal RAG: fine-grained visual features cause graph explosion and retrieval noise, while coarse representations discard critical local evidence. It introduces a dual-tier architecture decoupling a Macro-reasoning Graph, which handles global structural/cross-document relationship reasoning, from a Micro-matching Graph, which handles fine-grained instance-level evidence matching, isolating noise sources by design. This targets multi-hop multimodal QA where prior instance-level matching or single-graph approaches fail to capture explicit cross-modal, cross-document relationships. Implementers building multimodal RAG pipelines could adopt this two-graph separation as a retrieval-stage architecture pattern rather than a single unified knowledge graph.

The paper is a controlled replication and extension of Wang et al. (2024)'s finding that self-consistency-style repeated sampling can outperform self-refine/reflexion-style methods at equal compute, but adds confidence intervals and significance testing that the original lacked. They evaluate seven inference-time methods (self-refine, reflexion, best-of-n, debate, etc.) against repeated sampling with majority voting, on open models at 1.5B, 3B, and 7B parameters across two math benchmarks (150 questions each), counting every generated token including critique/reflection/debate/checking overhead as the fairness metric. The design isolates method quality from raw token expenditure, letting practitioners judge whether a given inference-time scaffolding technique is worth its complexity versus simply increasing sample count. This is directly actionable for anyone choosing an inference-time compute strategy: the result argues for defaulting to repeated sampling as a strong, simple baseline before adopting more elaborate self-reflection pipelines.

The paper studies winner determination for Thiele voting rules (a parameterized family of approval-based committee election rules including PAV) under the Voter Interval (VI) domain restriction, where voters can be ordered so each candidate's approval set forms a contiguous interval. They characterize structural constraints on optimal committees induced by the approval-ballot dependency structure between candidates, and leverage this to design fixed-parameter tractable (FPT) algorithms for PAV and other Thiele rules restricted to the VI domain — a problem that is NP-hard/intractable in the unrestricted case. This contributes to the algorithmic social choice literature on exploiting single-peaked-like or interval structure in preference domains to escape general-case hardness. Researchers in computational social choice could build on these FPT results to identify further tractable ballot-structure restrictions or implement exact solvers for VI-structured elections.

The study benchmarks Qwen3-VL-8B/30B-A3B, UI-TARS-1.5-7B, and OpenCUA-7B on OSWorld, systematically varying inference-time scaling along contextual (history), temporal (deliberation depth), structural (planning decomposition), and parallel (sampling/voting) axes. Results show diminishing accuracy gains with added compute and qualitative shifts in failure modes rather than uniform improvement; contextual scaling specifically improves trajectory stability via historical grounding. Practitioners building local CUAs should treat inference-time scaling as a tunable tradeoff against hardware budget rather than a default lever, and should profile which failure modes each scaling axis introduces before deploying.

Using a frozen CLIP image encoder and a fixed training recipe as controls, the authors benchmark eight fusion architectures spanning attention-based, state-space model (Mamba) based, and learned-compression approaches (their proposed Temporal Bottleneck Fusion, TBF) on bi-temporal change retrieval over LEVIR-CC and Dubai-CC-style benchmarks. The controlled setup isolates the fusion module's contribution to retrieval quality versus query-time latency, since this module runs per candidate pair at search time. This gives practitioners a like-for-like efficiency/accuracy comparison for choosing or designing fusion modules in large-scale change-retrieval pipelines.

OpenMLE provides verifiable ML-engineering task environments with execution feedback (OpenMLE-Gym), operator-level RL training (OpenMLE-RL), and long-horizon evolutionary search (OpenMLE-Evo). Frontis-MA1 (35B) is post-trained via execution-grounded SFT and RL over four atomic program-evolution operators (Draft, Improve, Debug, Crossover), deduplicated against evaluation benchmarks, then composed into a long-horizon evolutionary search loop unifying learning and evolution. It is evaluated on MLE-Bench Lite under a 12-hour per-task compute budget, offering a reproducible operator-based framework other researchers could extend for recursive self-improvement research in ML engineering.

DR-FRL is a cross-fitted estimation workflow that maps irregular functional histories (point-cloud-style measurements at unequal times) into estimand-targeted latent states via functional and temporal encoders, coupled with nuisance heads estimating outcome, treatment, and censoring functions for doubly robust estimation. Unlike standard sequence learners that optimize predictive loss without guaranteeing the efficient influence function (EIF) is stabilized, DR-FRL explicitly validates the learned state against EIF-targeted diagnostics — calibration, overlap, tail behavior, and ablations — before it's used in the estimating equation. This gives causal inference practitioners a template for building representation-learning-based doubly robust estimators for longitudinal data without relying on hand-crafted scalar summaries.

MIND combines a vision-language model (BioMedGPT) that generates intent-driven textual descriptions of diagnostic goals and pathology from source images, using these as conditioning signals for a Diffusion Transformer (DiT)-based fusion network. To counteract the loss of 2D spatial continuity inherent in flattening images to 1D token sequences for DiTs, the authors introduce a Multi-scale Latent Adapter that extracts spatial features pre-serialization and injects them back via dimension-aligned fusion. This text-guided, pathology-aware conditioning approach offers a template for practitioners wanting to steer generative fusion models toward clinically relevant, task-specific outputs rather than generic global fusion rules.

APO adapts group-relative policy optimization (GRPO), typically used for LLM alignment, to 3D atomic structure prediction, removing the need for supervised preference learning against ground-truth coordinates (as required by prior flow-matching approaches like FlowDPO). It uses a dual-reward mechanism: one reward reinforces the model's dominant latent structural modes identified via eigen-decomposition of sample similarity, paired with a second complementary reward (details truncated in the abstract) to refine structural accuracy. This offers a path for structure-prediction practitioners to fine-tune flow-matching or diffusion-based structure generators in data-scarce regimes — novel crystal phases or de novo proteins — without requiring experimentally labeled reference structures.

ScaFE has a web-enabled LLM retrieve clinical literature and synthesize deterministic, executable 'feature programs' that quantify visually assessable scar attributes (rather than performing end-to-end image classification or VLM-based diagnosis), which run in a restricted local execution environment. Only aggregate validation statistics and feature-level SHAP summaries are needed for model refinement, keeping raw images and predictions on-premise for data-governance compliance. This approach lets practitioners build small-data, interpretable, hospital-deployable classifiers where each decision traces back to auditable, human-readable feature-extraction code rather than opaque model weights.

Both tasks are framed as image-to-text translation using vision-language models, generating line notations (e.g., SMILES-style strings) directly from raster images. OCSRGlyph advances state-of-the-art OCSR accuracy by explicitly modeling stereochemistry in the generation objective, an area prior OCSR systems often got wrong. MarkushGlyph departs from typical segment-then-recognize pipelines by reading the entire Markush structure image holistically in one pass rather than decomposing it into substructures. This is directly useful for practitioners building patent-mining or cheminformatics training-data pipelines who need robust image-to-SMILES/Markush extraction.

The paper conducts a discourse and case-study analysis spanning training data composition, model design and evaluation benchmarks, human feedback/RLHF loops, and public/media reception, arguing ideology gets embedded at each pipeline stage. Its central case study is the 'delve' controversy, where a lexical marker common in non-Inner-Circle (particularly Nigerian/African) English got used as an AI-detection heuristic, disproportionately flagging non-native-coded writers as suspect. The analysis implies that grammaticality/naturalness benchmarks and annotator pools skew toward Inner Circle norms, systematically penalizing other varieties in both model outputs and downstream detection tools. Relevant to NLP practitioners designing dialect-fair evaluation sets, annotator diversity protocols, and AI-text-detection systems that shouldn't conflate 'unusual word choice' with 'AI-generated.'

MANTA initializes a task-conditioned communication topology from prior structural experience, then during execution monitors collaboration traces (message/validation outcomes) and applies bounded structural updates — altering agent roles, communication edges, execution order, information visibility, and validation pathways — while holding the task interface and agent budget fixed. This positions topology as a dynamic, online-adapted variable rather than a fixed design choice or purely offline-optimized target, distinguishing it from static multi-agent orchestration frameworks. Practitioners building agent orchestration systems (AutoGen/CrewAI-style) could use the monitoring-plus-bounded-update mechanism as a template for making any fixed agent graph self-repairing at inference time.

The Degradation Archetype Representation (DAR) module models degradation as a simplex-constrained mixture of learned archetypes — essentially soft-clustering the input onto a basis of canonical degradation types to form a structured degradation state. This state feeds a Semantic Ambiguity Rectification (SeAR) module that generates degradation-aware prompts for channel-wise feature modulation, explicitly disentangling degradation cues from scene semantics rather than mixing them in a shared latent space as prior all-in-one methods do. The framing suggests a companion mechanism handles the paper's 'spatial ambiguity' (localizing where restoration should act) though details are cut off. Practitioners building unified restoration backbones could adopt the archetype-mixture conditioning as a drop-in alternative to shared-latent degradation embeddings.

The paper formalizes same-graph NC-LP transfer and introduces a leakage-free evaluation protocol: fixed node/edge splits, message passing restricted to a graph excluding evaluated edges, and fixed negative sampling for LP, tested across GCN, GraphSAGE, and GPS backbones. Key finding: NC→LP transfer is consistently beneficial on homophilic graphs (where connected nodes share labels/features), while LP→NC transfer is direction-fragile and can degrade accuracy under naive representation reuse, becoming reliably positive only under certain conditions. This gives practitioners both a reusable leakage-free benchmarking protocol and a predictive heuristic (homophily + transfer direction) for deciding when to share representations between NC and LP heads in multi-task GNN systems.

CS-RNR (budget-constrained confidence-scheduled restricted responses) tracks pooled opponent action frequencies with anytime-valid confidence sequences — statistical intervals that remain valid under continuous monitoring without inflating false-positive rates — and flags a frequency as exploitable only once its confidence interval separates from the equilibrium reference. Confirmed deviations define a conservative restricted-best-response set, and the agent computes and certifies a safety bound on the exact strategy it deploys, rather than relying on a fixed heuristic release threshold. This builds on the restricted Nash response (RNR) literature and is directly applicable to poker or other two-player zero-sum agents wanting principled, self-audited online opponent exploitation with worst-case exploitability guarantees.

GenEvA inserts a query-conditioned latent aggregation stage between frame/token selection and answer generation for long-video QA, using a distribution over frames (weighted by relevance to the query) to guide a generative module that fuses frame-specific representations into compact cross-frame latent evidence. This addresses the failure mode where compact evidence sets retain relevant content but leave cross-frame integration to the downstream LLM, which often fails to combine complementary cues implicitly. Practitioners building long-video QA pipelines could adopt this as a drop-in module between existing frame-selection methods and the generation backbone to improve multi-moment reasoning without increasing token budget.

The authors formalize literary influence as a multi-level comparison problem, computing directional, control-calibrated alignment/similarity scores between source and derivative texts across five representational levels (lexical, semantic, conceptual, structural, narrative). Applied to historically attested literary relationships, the method yields 'transformation profiles' showing that different text pairs preserve fidelity at different levels (e.g., structural retention with conceptual divergence). This offers a reusable, quantitative pipeline for computational literary studies to characterize influence and creativity beyond binary originality judgments, and could be extended with additional levels or applied to cross-lingual or cross-media adaptation studies.

RefCaptioner tackles multi-reference image-grounded video captioning via a two-stage post-training pipeline: mixed-data supervised fine-tuning followed by Hierarchical Coverage-Discounted GRPO, a reinforcement-learning scheme that jointly optimizes reference selection, phrase-level grounding, distractor rejection, and cross-reference consistency while preserving general captioning quality. It's trained/evaluated on a purpose-built corpus of 20K videos and 171K reference images, and evaluated on the new MRVBench benchmark covering both real and AI-generated video factuality and grounding. This provides a concrete recipe (SFT + custom GRPO reward shaping) that practitioners can adapt for other grounded-generation tasks requiring fine-grained attribution to external evidence sources.

This is a structured scholarly dialogue (not an empirical study) among five World Englishes researchers, organized around five guiding questions about GenAI's effects on academic writing and publishing norms. It surfaces tensions between GenAI's potential to lower entry barriers for multilingual/non-native English scholars and its tendency to normalize dominant (typically Inner Circle) English varieties, discussing implications for linguistic justice, author agency, and institutional policy. Readers building AI writing-assistance tools or setting journal AI-use policies can use this as a framework for anticipating equity trade-offs rather than a source of quantitative findings.

InfoOps Bench is a live, continuously refreshed safety benchmark built from a monitoring pipeline tracking 2,100+ real Russian, Chinese, and Iranian state-backed information operations, designed to resist saturation/gaming by updating weekly with fresh claims (companion site: pattrn.ai/research/infoopsbench). The authors evaluate 17 frontier models across 8 providers under four prompt framings, measuring an 'integrity score' as refusal rate; scores range from 8.8% to 94.5%, an 85.7-point spread not fully explained by model capability alone, implying divergent safety training around propaganda-adjacent requests. Practitioners doing AI safety red-teaming or model selection for content-moderation-sensitive deployments can use this benchmark as an ongoing, non-static eval rather than a one-time static test set.

TCA-SIR reframes Scientific Inspiration Retrieval as target-conditioned abstraction: rather than ranking candidate papers by topical/embedding similarity to a target problem, it learns to generate a transferable abstract principle extracted from each candidate specifically conditioned on the target, then uses representations of these generated abstractions to predict transferability/relevance. This directly targets the 'remote inspiration' problem where topical similarity fails to capture cross-domain conceptual transfer. On ResearchBench it outperforms prior similarity-ranking-based SIR baselines, suggesting a template — generate-then-score conditioned abstractions rather than embed-and-rank — that could generalize to other analogy-retrieval or hypothesis-generation pipelines in AI-for-Science systems.

The paper formalizes a causal performative framework for algorithmic recourse, modeling recourse actions as interventions propagating through a structural causal model that captures feature interdependencies and their causal effect on the true (unobserved) label, rather than treating recourse as pure counterfactual explanation for flipping a classifier's output. This exposes a performativity failure mode: recourse policies optimized only for prediction-flipping induce behavioral responses that degrade accuracy over retraining cycles and become self-defeating. The resulting optimization is non-convex, and the framework provides a basis for designing recourse algorithms constrained to causally-effective interventions, relevant to practitioners deploying recourse in lending/credit scoring who need policies robust to model retraining and strategic response.

The authors move beyond scalar sentiment scores for news-driven stock prediction by using LLaMA-3.1-70B to extract six semantic dimensions (event type, impact scope, temporal horizon, semantic confidence, and others) from financial news, then evaluate predictive value on 41,618 news-stock pairs from the FNSPID dataset. A key finding is that FinBERT sentiment features are highly nonlinear predictors (F1=0.576 under nonlinear models vs. only 0.230 under linear models), demonstrating that sentiment's predictive power is model-dependent and incomplete on its own. This suggests a practical recipe for quant/NLP practitioners: augment or replace single-score sentiment pipelines with LLM-extracted structured multi-dimensional features to capture orthogonal, independently predictive signal in financial text.

KV-cache reconstruction via one-shot prefill diverges from a retained live cache at BF16 precision due to floating-point non-associativity rather than any logical difference in computation; a fixed-prefix 2x2 design isolates construction method from numeric precision, and BF16 disagreements (166/200 differing suffixes, 20 differing correctness labels) disappear entirely under FP32 (Wilson upper bound 1.88%). The authors then engineer a token-by-token incremental bridging procedure that achieves bit-exact agreement between incremental and retained caches on 12/12 test cases. Takeaway for practitioners: interpretability or stage-replay auditing tools for LLMs must control for numeric precision or adopt bit-exact cache reconstruction to avoid attributing precision artifacts to reasoning differences.

SCOPE frames assortment (product-location placement), source assignment, replenishment frequency, and delivery routing as coupled decisions in one composite policy architecture rather than sequential, siloed optimization stages, explicitly modeling how assortment shapes downstream demand/load, how sourcing and cadence reshape delivery requests, and how routing feasibility/cost feeds back into the value of upstream choices. The abstract is truncated before detailing the exact policy representation, but the framing implies an end-to-end trainable or jointly-optimized decision system spanning inventory and logistics layers. Practitioners in supply-chain ML could benchmark such coupled policies against siloed department-level baselines on stockout rate, inventory exposure, and transportation cost.

The framework embeds a deterministic recurrent state-space model (RSSM-style latent dynamics) into an encoder-decoder segmentation architecture, replacing single-pass feed-forward prediction with iterative latent refinement. Multi-scale encoder features are fused with partially decoded representations to form a 'structural observation' that seeds the latent state, which then evolves over inference steps toward a final segmentation, explicitly handling nested/overlapping labels (composite cartilage-plus-tissue regions vs. cartilage-only regions). Practitioners working on small, ambiguous-boundary anatomical structures could adapt this world-model/RSSM refinement approach as a general alternative to single-shot segmentation heads.

The authors reformulate PixOOD's Neyman-Pearson scoring stage for computational efficiency and deploy the full pipeline via hardware-optimized TensorRT compilation, achieving 182 FPS on a desktop RTX 4060 and 75 FPS on a Jetson AGX Orin embedded platform — 20x and 18x speedups over the original baseline, respectively — targeting autonomous driving and railway domains. This provides a concrete, benchmarked template for taking accuracy-oriented OOD/anomaly segmentation models and making them real-time-deployable on embedded automotive-grade hardware. Practitioners could replicate the approach by applying similar scoring-stage reformulation plus TensorRT graph compilation to other segmentation-based safety-critical pipelines.

The pipeline decouples perception from reasoning: a Swin Transformer performs multilabel prediction of 14 CODE degrees directly from inspection images, while a Weka J48 decision tree trained on CODE-to-severity ground truth is converted into 19 fixed IF-THEN rules. Inference combines these via fuzzy logic — t-norm activations from CODE conditions weighted by rule confidence, aggregated with corresponding s-norms — with the authors comparing Product, Łukasiewicz, and Hamacher fuzzy operator pairs for producing interpretable class evidence. This gives infrastructure-inspection practitioners a replicable neuro-symbolic template (CNN/transformer perception + decision-tree-derived symbolic rules + fuzzy aggregation) for interpretable severity scoring beyond black-box classification.

FA-RDP tackles the frequency mismatch in diffusion-based manipulation policies: low-frequency, multi-step denoising preserves pre-contact action multimodality but reacts sluggishly to force feedback, while high-frequency sampling improves reactivity post-contact at the cost of mode collapse. The method uses a shared multi-frequency visual-force Transformer that predicts actions conditioned on inferred contact phase, allowing the policy to adaptively select its inference frequency/sampling steps online rather than fixing them a priori. This effectively decouples the multimodality-preservation and reactivity requirements that previously forced a single global frequency compromise. Practitioners building contact-rich manipulation policies (insertion, assembly, contact-based grasping) could adopt this frequency-adaptive scheduling as a drop-in mechanism atop existing diffusion policy architectures that already fuse vision and force sensing.

X-NavDP post-trains a diffusion-based navigation policy (originally pretrained on oracle-planner demonstrations for a single embodiment) using a novel data-efficient RL framework called Group Q-score Reweighted Matching (GQRM). GQRM addresses the core obstacle to RL fine-tuning of diffusion policies — the intractable likelihood that destabilizes policy gradients and hampers exploration — by reweighting groups of sampled trajectories via Q-scores rather than relying on standard policy-gradient estimates. This yields more than marginal gains (the stated failure mode of prior diffusion+RL approaches) in generalizing to novel embodiments and reactive behaviors like dead-end escape and obstacle detour using only onboard local observations. Roboticists working on embodiment-agnostic navigation or diffusion-policy RL fine-tuning have a concrete alternative to standard policy-gradient or DDPO-style approaches to try.

TEA-AgriVLN extends the AgriVLN vision-and-language navigation framework by adding a Traversability Estimation Alarm module that predicts a continuous or embodiment-conditioned traversability score from the current camera image, rather than relying on the binary indoor-style traversability assumptions baked into prior VLN-CE benchmarks like A2A. When the policy's selected low-level action conflicts with the estimated traversability, the alarm triggers a re-evaluation step in the decision-maker, effectively adding a safety/consistency check on top of the base navigation policy. This targets a specific domain gap — ambiguous, embodiment-dependent traversability in agricultural terrain versus indoor scenes — and offers a pattern (estimate-then-alarm-then-replan) that could generalize to other outdoor or unstructured VLN-CE settings. Implementers could integrate the TEA module as an auxiliary head or gating layer on existing VLN-CE policies without retraining the full pipeline from scratch.

The paper proposes Aging-Aware Autonomous Intelligence (AAAI), a framework integrating real-time hardware health estimation directly into an autonomous system's reasoning and planning loop, motivated by the failure mode of 'agnostic collapse' where accumulated, distributed hardware degradation (not a single fault) causes mission failure because the AI stack assumes nominal hardware. The framework rests on three pillars: hardware self-awareness (continuous health estimation of power, sensing, memory, and compute using physics-of-failure models), self-adaptive reasoning (dynamically scaling inference complexity and planning horizon to match remaining capability), and presumably a third execution/mission-adaptation layer. This reframes robustness engineering from static fault-tolerance design toward closed-loop, degradation-conditioned autonomy, giving systems engineers a blueprint for coupling prognostics-and-health-management (PHM) signals directly into planning/task-prioritization logic rather than treating health monitoring as a separate offline concern.

CS-JEPA is a decentralized recurrent joint-embedding predictive architecture (JEPA) where every robot in a swarm outputs the same shared future 'token field' representing collective state, using only a 16-frame local observation history and one 64-dimensional recurrent message per directed communication edge — no global pooling, target encoder, episode clock, or recorded future actions are required at deployment. Pretraining is self-supervised (no downstream collective-state labels), after which frozen representations are evaluated via ridge regression probes fit on as few as 6–24 globally labeled episodes, testing label efficiency of the learned representation. The comparison baseline is raw-future reconstruction with matched receiver capacity but ~9,607 additional training-only parameters, and a prospectively registered five-seed replication reports improvement for CS-JEPA — a methodologically notable detail since pre-registration guards against post-hoc cherry-picking. This offers a template for scalable multi-robot state estimation under strict bandwidth/compute constraints, and the ridge-probe evaluation protocol is directly reusable for benchmarking other decentralized representation-learning approaches in swarm robotics.

FasTac combines single-sensor multispectral photometric stereo (multiple simultaneous light-color channels for surface normal estimation) with boundary-prior fast Poisson depth reconstruction for accurate curved 3D reconstruction, and a position-aware dynamic-convolution network (HyperForce) that models the spatially nonuniform stiffness of a curved elastomer to regress three-axis (normal + shear) forces. All processing is accelerated on an FPGA for high-speed inference suitable for closed-loop control. This is a hardware+algorithm co-design that a practitioner could replicate by pairing a multispectral illumination ring with an FPGA pipeline for real-time tactile feedback in grasping controllers.

QQWorld replaces the Epps-Pulley (EP) characteristic-function-based regularizer in LeWorldModel with a quantile-quantile (QQ) matching objective that aligns rank-ordered projected latent samples directly against theoretical Gaussian quantiles, avoiding the vanishing-gradient problem EP exhibits for isolated tail samples. A 'cross-batch QQ' variant expands the effective sample pool for ranking using detached samples from prior batches, with an explicit bias-variance analysis of that choice. Practitioners building latent world models for planning can substitute this regularizer as a drop-in loss term to better control heavy-tailed latent distributions, likely improving downstream planning robustness.

TacWAM extends World Action Models (that jointly predict future states and generate actions) beyond vision by adding a mechanics-aware tactile future prediction branch. A Spatially Aligned Fusion (SAF) Tactile Encoder maps raw tactile imagery, dense force fields, and deformation flow into a shared latent space, using bilateral force/torque reconstruction as an auxiliary objective to preserve global contact physics; a temporal tactile history encoder supplies context for prediction. The design explicitly guards against tactile futures becoming a 'privileged cue' that the action head shortcuts through rather than genuinely conditioning on, which practitioners can view as a template for adding modality-specific future prediction to action models without inflating an evaluation-time information advantage.

SemAnCorr is a training-free correspondence framework that replaces naive nearest-neighbor descriptor matching (prone to spatially incoherent correspondences) with anchor-guided matching: it jointly optimizes pose and correspondence to select semantically consistent anchor regions, then propagates these as constraints over the full object surface via functional maps, yielding correspondences that are both semantically and geometrically coherent. This recovers reliable local geometric frames needed to transfer object-centric manipulation skills across instances with shared function but differing geometry. A practitioner could adopt this as a preprocessing/correspondence module ahead of any skill-transfer or grasp-retargeting pipeline that currently relies on brittle descriptor nearest-neighbor matching.

The paper targets affordance segmentation (identifying actionable surface regions) on embedded wearable-robot hardware using RGB-D input, proposing two complementary approaches to explore the Pareto-optimal accuracy/efficiency front: (1) a hardware-aware neural architecture search (NAS) with a newly designed search space that natively incorporates depth channels into compact network architectures, and (2) a fine-tuning approach that adds a preprocessing/fusion layer merging depth with RGB to make depth compatible with conventional (RGB-only-trained) architecture families. Evaluated on real-world datasets against existing tiny-network baselines, the methods aim to overcome failure modes of prior tiny models in critical scenarios under strict embedded hardware constraints. Practitioners deploying affordance perception on portable accelerators (e.g., wearable exoskeleton controllers) could apply either the custom NAS search space or the depth-fusion fine-tuning recipe depending on whether they can afford a full architecture search.

This demo implements a task-oriented semantic communications (SemCom) pipeline for a ROS 2 mobile robot (RGB-D + LiDAR) paired with a Jetson Orin edge node: RGB frames are encoded on-robot via an ONNX Runtime VQ-VAE encoder into discrete latent tokens (a 320x240 image becomes an 80x60 token grid, 5400-byte packed payload), transmitted, and reconstructed edge-side via a PyTorch decoder, feeding RTAB-Map SLAM and semantic object handling with browser-based mission visualization. The system operationalizes the 6G task-oriented communication principle that link quality should be measured by mission/task success rather than raw packet delivery. Practitioners building bandwidth-constrained robot-edge systems can use this as a reference architecture for VQ-VAE-based visual token compression integrated with existing SLAM stacks (RTAB-Map) over real wireless links.

UniCross unifies grasping, relocation, in-hand rotation, and in-hand translation, previously modeled as separate skills with distinct action constraints/objectives/morphologies, into a single formulation sharing state space, action space, and objective structure. This shared formulation enables distillation of one cross-skill policy that matches or exceeds skill-specific baselines on each individual skill while generalizing to unseen objects/configurations and supporting long-horizon composition without hand-off discontinuities between skill-specific controllers. Practitioners building dexterous manipulation stacks can use this unified state/action/objective design as a template to replace a multi-policy pipeline with a single distilled policy, simplifying long-horizon task composition.

RaDiVe improves 4D radar odometry (registration-based ego-motion estimation from sparse, noisy Doppler-radar point clouds) via two mechanisms: a distance-bounded Normal Distributions Transform (NDT) that restricts correspondence search to near-distance voxel pairs, improving optimization stability and computational efficiency versus full-range NDT matching, and a velocity-discrepancy point uncertainty model that weights each point's contribution by the mismatch between its measured Doppler radial velocity and the radial velocity predicted from the estimated ego-velocity, down-weighting unreliable/noisy points. The method also incorporates signed distance information (truncated in the excerpt) likely for additional geometric consistency. This gives practitioners a concrete recipe — bounded-distance NDT plus Doppler-consistency weighting — for building more robust radar-only odometry pipelines for all-weather autonomous navigation.

PT-Flow parameterizes sequence dynamics as an ODE-defined latent velocity field, recasting future prediction as temporal integration via an ODE solver in a compressed representation space rather than iterative discrete-step decoding. ODEWorld builds on this to create a continuous-time latent world model, extracting time-variant features to structure the latent space for solver stability and efficiency. This approach targets the inefficiency of standard discrete-time video/world-model prediction by exploiting the continuous nature of physical dynamics, and a practitioner could build on it by swapping in different ODE solvers or downstream control/planning heads operating on the learned continuous latent trajectory.

CALF (Convolutional Attention for Leg Features) is an end-to-end policy combining convolutional layers, attention, and an MLP to map raw 2D LiDAR leg-scan patterns directly to navigation commands, trained via deep RL. The authors built LegNav, a lightweight 2D simulator combining LiDAR ray tracing with a novel pedestrian gait model, to generate realistic leg-motion training data absent from prior circle-approximation approaches. Evaluation compares CALF against classical and learning-based baselines on navigation performance and social compliance metrics, suggesting the leg-level representation improves both. Practitioners could reuse LegNav for benchmarking or extend CALF's architecture to other low-mounted-sensor social navigation settings.

SIDO factorizes moving-object manipulation into object-pose prediction and pose-conditioned reaching, converting a policy trained only on static demonstrations into a goal-conditioned one via counterfactual action augmentation: demonstrated action chunks are morphed to preserve the hand-object relative pose after displacing the object to a predicted future position. At deployment, an external object pose predictor feeds the future position into this goal-conditioned policy in real time. The method is evaluated across three simulated tasks (Mug, Square, Stack) under five motion patterns plus real-world tests, targeting the common failure mode of static-only visuomotor policies under object drift. Practitioners can apply this as a lightweight augmentation on top of existing action-chunking imitation policies without collecting new dynamic demonstrations.

RoboBRIDGE is a modular orchestration framework composed of five coordinated components — Monitor, Perceptor, Planner, Controller, and Robot Interface — layered on top of a pretrained VLA to add failure detection/recovery, long-horizon execution consistency, and robustness to observation/task/embodiment shifts without retraining the underlying model. The Monitor performs rapid failure detection (paired with recovery logic per the abstract) while the other modules handle perception grounding, high-level planning, low-level control, and hardware abstraction respectively. This targets a real deployment gap in current VLA research: policies are strong action predictors but lack systemic robustness scaffolding. Practitioners can plug in different pretrained VLAs or swap individual modules (e.g., a new Monitor) to compose robust agents from existing components rather than building bespoke environment-specific fixes.

DexDirect combines direct kinesthetic teaching — an operator physically drags a 6-DoF gravity-compensated arm by a handle — with webcam-based hand-pose retargeting onto a 16-joint, 13-DoF dexterous hand, decoupling arm trajectory capture (physical) from hand articulation capture (visual). User studies report 17.2x and 3.2x more successful demonstrations versus AnyTeleop (vision-only) and TeleDex (pose-tracking) baselines respectively, alongside NASA-TLX scores showing reduced mental demand, effort, and frustration at the cost of increased physical demand. Collected demonstrations were used to train a diffusion policy, suggesting the interface is compatible with standard imitation-learning pipelines. This offers practitioners a lower-cost, higher-throughput demonstration collection pipeline for dexterous manipulation without specialized haptic or full-body mocap hardware.

RedFlow is an offline reinforcement learning framework for flow-matching VLA policies that performs action-level (not trajectory-level) credit assignment on failure rollouts via two components: a Context-Aware Corrective Matching mechanism that identifies failure-inducing actions and retrieves successful alternative actions from similar contexts as correction targets, and an Adaptive Redirection Objective that incorporates these corrections into training. This addresses the low sample efficiency of prior offline RL approaches that either discard failure data entirely or apply only coarse trajectory-level penalties, which fail to localize and fix specific compounding errors from distribution shift. Practitioners deploying flow-matching VLA policies could apply RedFlow as a post-deployment fine-tuning stage using logged rollout data, without needing new expert demonstrations for every failure mode.

The method reconstructs articulated objects' geometry and kinematic structure from a single rest-state (closed-configuration) observation, an ill-posed inverse problem since joint motion is never directly observed. It uses an explicit mesh as an intermediate representation to cross-verify and fuse noisy part segmentations from vision-language and segmentation models into spatially consistent parts, then estimates joint parameters by having a video diffusion model synthesize plausible articulation hypotheses, validated via geometric consistency checks against the reconstructed mesh. This targets digital-twin creation pipelines that currently require multi-state motion capture, replacing that requirement with generative motion priors. A practitioner could plug this into existing 3D scanning/reconstruction pipelines to add functional articulation metadata to single-pose object scans.

LabEvolver is a training-free agent framework using episodic memory: a state-grounded inner trial loop handles adaptive perception, online planning, and safety validation during task execution, while an outer evolution loop distills completed trajectories into reusable skill, strategy, and safety experience stored for future retrieval — no model weight updates involved. On robotic pH-regulation solution-preparation tasks it reduced completion time by 48.2% and safety-gate interceptions by 60.0%; on the ALFWorld benchmark it raised cumulative success rate from 76.2% (ReAct baseline) to 91.4% over 500 continual tasks, demonstrating cross-domain generality. This offers a practical, retraining-free path to closed-loop autonomous experimentation, and practitioners could adapt the inner/outer loop memory-distillation architecture to other safety-critical, long-horizon agentic domains beyond wet labs.

Arm2Air pretrains a transformer-based motion-transfer model on ordered obstacle-avoidance skeletons extracted from a pretrained Neural MP model for robot arms, then adapts it to the UAV relay domain with limited target data via Low-Rank Adaptation (LoRA). The transferred skeleton seeds an initial relay-chain layout, refined against connectivity, bottleneck capacity, delay, and movement-cost objectives. On nine held-out high-clutter 3D urban maps, this cut median end-to-end planning runtime by 64.9% versus baseline. Practitioners could replicate the skeleton-extraction-plus-LoRA recipe for other embodiment pairs beyond arm-to-UAV.

World Action Planner combines VLM-based high-level reasoning with a multi-task, pose-image-conditioned world model that generates imagined rollouts of candidate action plans, iteratively refining plans via search and optimization over these rollouts rather than emitting actions directly. It's benchmarked against state-of-the-art VLA and WAM baselines on compositional tasks, novel layouts, and zero-shot generalization, reportedly outperforming both. The replicable idea is decoupling plan proposal (VLM) from plan verification/refinement (world-model rollout + search) — a general test-time-compute strategy for any policy with an available forward dynamics model.

The authors investigate embedding behavior-aligned, embodiment-invariant representations (object bounding boxes, language-described motions, end-effector traces) into VLA models as intermediate prediction targets to improve cross-embodiment generalization. They build a simulation benchmark specifically designed to isolate and measure transfer quality to novel embodiments, then compare representation choices and integration strategies against standard action-only VLA training. This offers practitioners building multi-robot foundation policies a testable hypothesis (representation invariance predicts transfer) plus a benchmark to validate new representation designs before committing to large-scale training runs.

FoMo-FD trains an action-conditioned flow-matching world model to capture nominal short-horizon visual dynamics of surgical manipulation, then detects failures by scoring the 'inverse-transport nonconformity' of observed endpoint latents relative to the model's predicted distribution — a conformal-prediction-style anomaly score over learned dynamics. Detection operates at the window level to flag visual-action inconsistencies, requiring zero failure demonstrations for training, directly addressing the scarce-failure-data problem. Teams building monitors for imitation-learned policies in other safety-critical domains could adopt the same recipe: train a conditional generative dynamics model on nominal data only, then use distributional nonconformity as an anomaly signal.

The paper formalizes a subclass of assistance games (POMDPs with asymmetric goal information) in which pragmatic-pedagogic reasoning — human and robot each modeling the other as approximately rational communicators, à la rational speech act models — collapses the full-horizon POMDP into a tractable single-round best-response computation. They prove standard inverse optimal control has an inherent 'inference ceiling' on goal disambiguation because it treats human actions as pure task execution rather than potential communicative signals, and show the pragmatic-pedagogic policy overcomes this by selecting actions informative precisely because they'd be suboptimal under any goal but the true one. Results are validated experimentally, offering a tractable alternative to full POMDP solving when goal ambiguity is the main bottleneck.

The simulator couples a position-based dynamics (PBD) suture-thread model with a material point method (MPM) soft-tissue model via a novel two-way contact-coupling scheme accounting for frictional and drag forces during needle-thread insertion through deformable tissue, addressing a gap in prior work that typically simulates sutures or soft tissue in isolation. It's positioned as a training ground for robot reinforcement learning on suturing tasks, so practitioners could use it to generate large-scale, physically plausible interaction data for RL policies before real-tissue validation, with the PBD-MPM coupling method itself reusable for other thread-on-deformable-body simulation problems.

Building on Aanes et al.'s retroactive constant-product liquidity pools (which maintain an implicit virtual pool per possible resolution of active cross-chain/cross-shard locks), this paper addresses computing 'safe quotes' — quotes bounded by the minimum possible output across all virtual pool states — as a hard correctness constraint, with a secondary optimization goal of minimizing the gap between the safe quote and the true minimum achievable output. This is essentially a worst-case optimization problem over the combinatorial space of lock resolutions in constant-product AMMs. Teams building cross-chain or sharded AMM infrastructure could use these safe-quote algorithms directly to guarantee price integrity without waiting for full lock resolution, trading computational cost against quote tightness.

The paper surveys and systematizes performance evaluation practice for Distributed Computing Continuum Systems (DCCS), which span cloud data centers to resource-constrained edge nodes, organizing metrics into a structured taxonomy spanning computation, networking, energy efficiency, and application-level quality dimensions. It argues existing evaluation work is siloed by dimension and lacks cross-layer consistency, making comparisons between DCCS architectures and algorithms unreliable. The taxonomy is intended as a reference framework researchers can use to select consistent, comprehensive metric sets when benchmarking new continuum architectures, scheduling algorithms, or orchestration systems, rather than relying on ad hoc single-dimension evaluations.

The system uses only pathloss-derived (noise-free) distance measurements from a stationary origin receiver and a mobile pursuing agent — no GPS, angle-of-arrival, or multi-anchor triangulation. The algorithm first localizes the target's quadrant and position via a bounded search (proven to converge within 13 steps), then differentiates consecutive position estimates to recover a linear motion vector, transitioning the agent from search mode to pursuit mode. Simulations validate bounded tracking error during the pursuit phase. This offers a low-infrastructure alternative to multilateration-based tracking, useful for GPS-denied or anchor-sparse deployments such as indoor or contested-RF environments, and the step-bound result gives a concrete worst-case latency guarantee for practitioners designing real-time pursuit controllers.

The system addresses two failure modes in enterprise RAG over data catalogs: poor retrieval (baseline Hit@10=19.1%) and lack of usage/semantics knowledge causing metric misinterpretation, attributed to four root causes (semantic gap, entity ambiguity, schema drift, asset-usage gap). It combines a three-tier, 179-document knowledge base (structured via an eight-section annotation template, refreshed daily via single-approval hot-reload) with a Graph-Guided Retriever using a 2,859-node knowledge graph as a candidate gate plus intent routing, cutting token usage 71.6x. A Scene-Aware Ranker then applies 19-class entity recognition for final disambiguation. This architecture is a template for large-scale enterprise text-to-data or table-retrieval systems needing both accurate asset discovery and correct downstream metric interpretation at production scale (14 domains, thousands of tables).

LightRot combines two algorithmic techniques — Grouped Local Rotation (GLR) and Outlier Direction Aligning (ODA) — with a hierarchical Fast Hadamard Transform (FHT) rotation unit to suppress activation outliers cheaply before 4-bit quantization, addressing the energy overhead that rotation-based quantization methods typically incur. The design was fabricated in 28nm CMOS and measured at 27.4 TOPS/W peak energy efficiency for 4-bit inference, exceeding prior state-of-the-art low-bit accelerators, while the paper claims evaluation beyond basic language modeling benchmarks (unlike some prior rotation-based approaches). Hardware/ML co-design practitioners could adopt the GLR+ODA+FHT pipeline as a quantization-aware rotation scheme, or use the reported efficiency numbers as a benchmark for competing low-bit accelerator designs.

GyRot targets the known incompatibility between rotation-based and fine-grained group quantization for low-bit LLM inference, where rotation's global scope conflicts with group scaling's local scope, typically causing accuracy loss or hardware overhead. Its core contributions are Coarse Rotation, Fine Grouping (CoRFiG) and Harmonic-Aligned Permutation (HAP) to align the two techniques cooperatively, plus a reformulated asymmetric quantization scheme with a zero-point rounding strategy enabling fully integer dequantization (avoiding floating-point overhead). The system is implemented on an INT4-based tensor processing-element architecture via algorithm-hardware co-design. This gives practitioners a concrete recipe — CoRFiG + HAP + integer-only dequant — for combining rotation and group quantization in custom low-bit accelerators without the usual accuracy/hardware tradeoff.

FedPAIE performs personalized color grading via federated learning, avoiding centralization of raw photos or user ratings. It trains a lightweight dual-cue aesthetic scorer, calibrates it per-user on a small local support set into a personalized scorer, then freezes it as a reward signal to guide regularized adaptation of a lightweight CLUT (color lookup table) enhancer using unpaired local photographs. Fidelity constraints and an excess-gap penalty prevent the enhancer from over-optimizing against the proxy scorer (reward hacking), keeping outputs natural. This gives a practical blueprint for on-device, privacy-preserving personalization pipelines that combine a small learned reward model with regularized preference-guided fine-tuning, applicable beyond photo enhancement to other federated preference-learning tasks on resource-constrained devices.

The paper derives a dynamical speed limit for analog computers that represent matrix rows as impedance networks: the fastest normalized eigenmode of the system is bounded by 2π times the largest combined unity-gain bandwidth (CUGBW) across all rows. CUGBW for a row is defined as its finite-gain-adjusted unity-gain bandwidth plus contributions from coupled rows, each contribution being the geometric mean of the two rows' unity-gain bandwidths scaled by their coupling conductance and normalized by the geometric mean of their total conductance loadings. This acts as an analog analogue of the Courant-Friedrichs-Lewy condition in numerical time-stepping, bounding the operator rates physically resolvable by analog hardware, and is validated against large-scale LTspice simulations across multiple architectures. Analog/mixed-signal chip designers can use this closed-form bound to predict, prior to fabrication, the maximum operational speed of impedance-network-based analog solvers (e.g., for optimization or linear-algebra acceleration).

The paper analyzes O-RAN's disaggregated vRAN architecture, where standardized, multi-vendor interfaces replace traditional tightly-coupled hardware/software stacks, and identifies which fronthaul/midhaul interfaces carry the strictest latency and jitter requirements when running over shared, cost-effective transport (general-purpose Ethernet) rather than dedicated over-provisioned links. It then surveys Time Sensitive Networking (TSN) mechanisms — e.g., traffic shaping, scheduling, and synchronization standards — as candidate solutions for meeting these deterministic performance requirements on shared infrastructure. This serves as a reference for network architects designing O-RAN transport networks, mapping specific TSN standards/features to specific O-RAN interface latency budgets rather than treating TSN adoption as one-size-fits-all.

The paper contrasts the dominant 'passive paradigm' in ER — static, one-shot similarity computation via rule-based, deep-learning, or LLM methods — with a proposed 'Agentic ER' framing that treats duplicate detection as sequential decision-making. This involves iterative evidence gathering across heterogeneous data-lake sources (CSV, JSON, RDF, free text), multi-hop reasoning, and selective human-in-the-loop intervention rather than a single similarity threshold. It's a position paper motivating a research agenda rather than presenting a system, so practitioners should read it as a framing device for building LLM-agent pipelines (tool-calling, iterative retrieval) around existing ER benchmarks. Difficulty: Conceptual.

The paper audits SWE-bench Verified, a standard benchmark for LLM issue-resolution, and finds 13.6% of PR-issue pairs are misaligned across five patterns and eleven fine-grained scenarios (e.g., the linked issue doesn't fully describe the problem the patch resolves, or the patch addresses more/less than the issue specifies). PAIChecker is a multi-agent system operating in three phases to automatically detect this misalignment at scale, intended as a quality-control step for constructing SWE-bench-like benchmarks. Practitioners building or curating issue-resolution benchmarks can use PAIChecker to filter or flag unreliable instances before evaluation, improving the validity of reported LLM coding performance.

ORCA-bench pairs a production-fidelity, OpenTelemetry-instrumented microservice environment (exposing six days of metrics, logs, and traces via Prometheus, Jaeger, and OpenSearch/Grafana, plus full source access) with 1,079 root-cause-analysis tasks that systematically vary report specificity, time-to-detection, and co-occurring fault scenarios. Ground truth is SRE-curated, and the LLM-as-judge scoring is validated against human re-scoring (weighted Cohen's κ=0.90), enabling rigorous comparison across frontier models. This gives agent developers a reproducible, realistic testbed for evaluating and improving RCA-specific capabilities — noisy multi-signal reasoning under time pressure — distinct from standard code-generation benchmarks.

The paper extends Katsuno-Mendelzon belief update semantics by inserting, per source world, a transformation of the epistemic input into a weaker proxy prior to applying the credibility-limited transition function, enabling partial incorporation of compound (multi-part) inputs rather than forcing an all-or-nothing acceptance. It provides both semantic (possible-worlds model) and axiomatic (AGM/KM-style postulate) characterizations of the resulting operator class, and identifies well-behaved subclasses with additional properties. This is directly relevant to KR researchers building belief revision/update engines for planning, diagnosis, or multi-agent reasoning systems that need finer-grained control than binary accept/reject update rules.

This is a survey chapter covering the use of interactive theorem provers to formally verify security properties across four domains: system security, language-based security, secure compilation, and cryptography. It frames formal verification as both a design-validation tool and a foundation for certification processes. For a technical reader, it functions as an orientation map to the mechanized-proof literature in security, useful for selecting which proof-assistant techniques and domain (e.g., compiler correctness vs. protocol proofs) fit a given verification goal. Difficulty: Conceptual.

The authors implement a binary Ring-LWE-based logic locking scheme on a RISC-V SoC that secures the privilege-switching logic path without exposing raw locking parameters to end users, unlike conventional logic locking. Measured overhead: the locking module uses 3519 LUTs and 2645 registers, adding 6.0% LUT and 6.9% register overhead versus the unlocked baseline, with an unlock latency of about 2.6 microseconds affecting mainly system-level, not user-level, operations. This gives hardware security practitioners a concrete FPGA-implementable reference design with quantified resource and timing costs for HE-based IP protection. Difficulty: Buildable.

CoGate modifies the co-decoding acceptance rule to be gated by the security expert model's confidence at each generation step, suppressing the expert's influence when it's operating on unseen or out-of-distribution patterns rather than always applying its steering signal. It's evaluated across multiple LLM backends including CodeGen, DeepSeek-Coder, and Qwen-Coder, positioning confidence-gating as a general improvement to decode-time security steering pipelines. Practitioners building similar expert-guided decoding systems could adopt this confidence-based acceptance rule as a drop-in replacement for uniform acceptance thresholds. Difficulty: Buildable.

LeanCSP is a Lean-based verification framework operating at two levels: formulation-level, where it proves metatheorems like equivalence, equisatisfiability, and correctness of symmetry-breaking constraints parametrically across entire problem families; and instance-level, where it checks solver-produced certificates via translation backends to external formats (MiniZinc, SMT-LIB, OPB). Combined, these give an end-to-end certified workflow establishing satisfiability/unsatisfiability with formal guarantees rather than trusting solver output directly. Practitioners in verified constraint programming could extend the certificate-checking backends or reuse the metatheorem library to certify new symmetry-breaking techniques. Difficulty: Buildable.

This is an exploratory empirical study applying three shortlisted GenAI-aware threat modeling methods to an industrial SME case study of a GenAI-augmented system, benchmarking their threat coverage against each other and implicitly against traditional methods like STRIDE. Results show meaningful divergence in which threats each method surfaces, and reveal systematic gaps in coverage of certain GenAI-specific risk categories. This gives practitioners early empirical evidence on which existing threat modeling methods to combine, and highlights specific risk categories that current tooling underserves. Difficulty: Conceptual.

The study performs a large-scale empirical analysis linking publicly available Solana bot implementations (sourced from GitHub) to their observable on-chain execution patterns or 'fingerprints,' extending prior Solana research that focused narrowly on system performance and smart-contract security. It aims to systematize what bot strategies (e.g., arbitrage, sniping, MEV-style extraction) are actually implemented in practice and how they manifest in transaction traces, which is directly useful for building bot-detection heuristics or on-chain forensic tooling. Difficulty: Runnable.

ThreatForest implements a multi-agent pipeline — repository analysis, context refinement, threat generation, parallel attack-tree construction with TTP mapping, mitigation synthesis, and report generation — orchestrated as a directed graph with deterministic verification gates, bounded retries, and three human-in-the-loop checkpoints. It supports pluggable adversary technique frameworks (MITRE ATT&CK, CAPEC, cloud-specific threat matrices) and uses a domain-specific sentence-transformer to map generated attack steps onto these TTP taxonomies. Practitioners building reliable multi-agent security tooling over source repositories could adopt its orchestration pattern of verification gates plus staged human review as a reusable template. Difficulty: Buildable.

The authors target sequential equivalence checking (SEC) of circuits after retiming plus arbitrary sequential resynthesis, a class of transformations current EC tools struggle to verify without unsound orchestration of multiple model checkers. Their method preprocesses with retiming, then runs an IC3/PDR-based procedure that uses simulation traces to propose candidate invariants, avoiding unrolling-based BMC/induction pipelines. A key feature is certificate generation — the proof output can be independently checked, addressing the 'trust the checker' problem in EC tooling. Evaluated on retimed/resynthesized open circuit benchmarks, it substantially outperforms the winning portfolio solver from the latest Hardware Model Checking Competition.

The paper extends Twee's goal-directed clause selection heuristic — originally limited to unit equational reasoning, where it works by introducing equational definitions to transform the proof problem — to full first-order saturation-based proving. It also proposes an alternative implementation that directly biases clause selection using shared terms with the conjecture, rather than relying on the definition-transformation trick. This addresses a core bottleneck in saturation provers (e.g., Vampire, E, Twee itself): combinatorial explosion of the given-clause loop when heuristics are goal-agnostic. Reported experiments show the shared-term approach yields strong performance gains, suggesting it's a candidate heuristic for integration into general-purpose ATP clause-selection strategies.

SWE-NFI comprises 188 tasks distilled from merged PRs in open-source Python repos, each paired with behavior-preserving non-functional improvement goals (e.g., performance, readability, maintainability) operationalized into 92 executable rule-based checks. The evaluation suite combines standard functional-correctness testing (to confirm behavior preservation) with rule-based scoring of whether the intended NFI was actually achieved, moving beyond pass/fail correctness metrics used in benchmarks like SWE-bench. State-of-the-art coding agents are evaluated against this suite, quantifying a gap between functional competence and genuine code-quality engineering. The rule/task set and harness appear reusable for benchmarking or fine-tuning agents specifically on refactoring-style objectives.

The paper unifies path types (cubical TT, fixed endpoints), Riehl–Shulman extension types (terms fixed on subshapes), and cooltt/Agda's controlled-unfolding mechanism as instances of a single construction inside two-level type theory (2LTT), where the meta-theoretic face/shape calculus governing boundary conditions is already present rather than added axiomatically. The key technical claim is conservativity and freeness: standard models of HoTT automatically become models of 2LTT, so no new semantic construction is required to support extension types across these variants. This gives type-theory implementers a uniform, already-verified foundation for extension-type features instead of ad hoc per-system boundary calculi, and could inform how cubical/extension-type support is added to proof assistants like Agda or forthcoming HoTT-based systems.

The system pairs a security-relevance classifier (trained/evaluated on a released dataset of 288 enterprise backlog items labeled by nine security experts) with a RAG pipeline that retrieves applicable clauses from security requirements documents to enrich flagged backlog items with actionable guidance. It was developed and validated in situ at a large regulated enterprise, addressing the practical gap between informal agile artifacts (short free-text tickets) and formal compliance documentation. The contribution set includes the labeled dataset, the classifier+RAG architecture, and an evaluation of detection/retrieval quality — providing a reusable blueprint for continuous security-compliance tooling integrated into backlog/ticketing systems (e.g., Jira) rather than periodic manual audits.

PROGRESS integrates LLM-generated jqwik properties — derived from intent-bearing code context while limiting implementation leakage — into coverage-guided search-based test generation, extending EvoSuite's DynaMOSA algorithm so the search is fitness-guided toward states that could falsify these semantic properties, not just toward structural coverage. This directly targets the 'oracle problem' in regression testing: purely search-based tools derive expected behavior from the SUT itself, silently codifying existing faults, whereas PROGRESS's LLM-authored properties act as an independent semantic oracle. Practitioners could adopt the pipeline (LLM property synthesis + property-guided evolutionary search) as a drop-in enhancement to existing EvoSuite-based test suites to surface real behavioral bugs, not just increase branch coverage numbers.

TrustChain-Review couples a blockchain-based evidence ledger for review provenance with a three-player game-theoretic model (developer, reviewer, platform) capturing strategic incentives around verification effort, plus a risk-adaptive rule that escalates governance intensity only when expected benefit exceeds cost. Evaluation is via controlled simulation calibrated against the Diff Quality Estimation dataset, comparing six governance configurations across 30 independent runs — i.e., a mechanism-design/simulation study rather than a deployed empirical evaluation on live repositories. The contribution is primarily architectural/theoretical: a template for combining tamper-evident evidence logging with incentive-compatible governance thresholds for AI-assisted review pipelines, which could inform future implementations on real VCS/CI platforms.

The authors measure that across 30 skills and two model generations, prose-driven tool-calling agents execute only 56% of steps mandated by their own skill files, despite passing output-level checks — evidence that skill compliance can't be inferred from output validity alone. Their remedy, Skill Compilation (implemented as SIGIL), converts a prose skill into an executable harness via AG-IR, a typed agentic intermediate representation that explicitly partitions model-owned cognition (things the LLM must still decide) from code-owned control flow (things enforced deterministically), preserving the natural-language authoring surface while eliminating the model's ability to silently skip mandated verification steps. This is directly reusable: teams writing agent skills today could adopt AG-IR-style compilation to convert existing prose skill libraries into harnesses with enforced step execution, trading some flexibility of free-form prose for auditable procedural guarantees.

The authors performed a systematic literature and standards review across academic and industrial sources to build a 361-requirement corpus for Human-Robot Teamwork (HRT), then applied iterative qualitative coding to derive a two-level taxonomy with 6 top-level categories (information provision, relational control, decision support, safety mechanisms, performance monitoring, and a foundational category) and 21 subcategories. This provides a reusable reference model that requirements engineers can map existing systems against, use to identify coverage gaps in a given domain's standards, or extend as a checklist during HRT system specification and certification. It's a synthesis/classification contribution rather than a new algorithm or system.

This is an ongoing formalization effort in the Lean theorem prover encoding the full content of an upper-level probability theory textbook, spanning Riemann–Stieltjes integration through martingales and limit theorems, likely built atop Lean's Mathlib measure theory library. The contribution is both a machine-checked artifact (verified statements and proofs matching the textbook) and reusable Lean infrastructure (lemmas, definitions, typeclasses) for future probability-related formalizations. Practitioners can build on this by importing the library for their own formalization projects or using it as a template for large-scale textbook formalization workflows.

AgentS4D is a sandboxed benchmark for lifecycle-wide runtime safety evaluation of LLM-based workspace agents, organized around a four-dimensional framework: six risk-entry sources, six induction strategies (how unsafe behavior gets elicited), nine target harm categories, and seven lifecycle checkpoints for post-run evidence collection. It contains 328 risk-injected cases spanning stateful, multi-step, tool-using workflows, enabling evaluation of not just final-outcome safety but where in an execution trajectory risk originates and manifests. Practitioners building or auditing agent systems could use this to pinpoint which stage of their pipeline (tool call, state mutation, side effect) needs stronger monitoring or intervention rather than relying on end-of-trajectory checks alone.

The paper presents PIE-Abducer (incremental direct-derivation abduction) and PIE-APT (Abductive Planning for Temporal Knowledge Graphs), operating natively in Description Logic to avoid the decidability issues and Ramification Problem that plague standard action formalisms. State transitions are modeled as non-monotonic updates to deductively closed DL theories along a linear timeline, treating the incremental reasoner as a black box and representing actions directly in OWL without external modal operators — preserving logical decidability while avoiding the combinatorial blowup of structural abduction for incomplete knowledge. This gives knowledge-graph practitioners a decidable, OWL-native mechanism for temporal planning and inconsistency/contradiction detection over dynamic knowledge graphs, useful for building reasoners that must handle open-world, evolving data.

This position paper argues that claims of 'long-horizon failure' are underspecified unless benchmarks compare actual full-task success against a baseline prediction constructed from matched short, individual-stage performance — since compounding of ordinary per-stage errors alone can explain degradation without invoking any horizon-specific effect like context rot. They formalize this comparison as the log-ratio between the short-task-derived prediction and actual long-task success, termed the 'horizon residual,' which isolates trajectory-induced degradation (where earlier execution genuinely makes later work harder, e.g., via accumulating context) from mere multiplicative compounding of independent per-step error rates. Benchmark designers could adopt this residual metric to diagnose whether their long-horizon evaluation is actually measuring something new versus just restating known short-task failure rates.

Analyzes transport in lattice composite Fermi liquids (CFLs) formed at half-filling of inversion-broken Chern bands, extending beyond the standard HLR (Halperin-Lee-Read) Landau-level treatment. Predicts a non-analytic AC longitudinal resistivity Re ρxx(ω) ~ |ω|^(4/3) from gauge-field fluctuations under gate-screened Coulomb interaction, plus nonreciprocal transport and a Hall conductivity scaling ~|q| at finite wavevector, both experimentally accessible via surface acoustic wave (SAW) measurements. Also identifies a separate singular-transport mechanism from renormalization of 2k_F backscattering. Offers concrete, falsifiable SAW-probe predictions for inversion-broken Chern-band CFL candidate materials.

The authors study non-interacting quantum particles on random locally tree-like layered graphs (structurally disordered generalizations of square lattices/ladders, equivalently random 1D chains of coupled components) with uniform coupling strengths, so all randomness is purely geometric rather than in the Hamiltonian's matrix elements. They show that extensive layer size produces level repulsion (a hallmark of Wigner-Dyson statistics), robust quantum chaos, and diffusive transport, while the quasi-1D limit yields coexisting localized states instead. This establishes graph dimensionality/geometry as an independent chaos-inducing mechanism, offering a tunable toy model for probing the boundary between many-body-localization-like and ergodic regimes without needing interactions or on-site disorder.

The authors map gate-voltage phase space in WSe2-proximitized rhombohedral graphene, distinguishing pockets with true zero-resistance superconductivity from adjacent pockets exhibiting finite saturation resistance despite otherwise matching phenomenology (sharp transitions vs. temperature, perpendicular field, and current). This directly probes the 'anomalous metal' phase debate in 2D superconductors, where quantum phase fluctuations or vortex dynamics are hypothesized to prevent full resistance collapse. The key experimental lever is in-plane magnetic field, which expands and merges the zero-resistance regions with the finite-resistance ones, implicating spin-orbit or Zeeman-related physics tied to WSe2 proximity coupling. This gives a tunable platform for testing theories of dissipative superconducting-like states and unconventional pairing mechanisms in moiré-free rhombohedral graphene.

The paper investigates the speed-accuracy trade-off in equivariant machine-learned interatomic potentials (MLIPs), which encode physical symmetries (rotational/reflection equivariance) directly into network architecture, historically at a computational cost disadvantage versus invariant models at scale. Using the NequIP and Allegro architectures trained on very large, diverse datasets — a regime where data efficiency matters less than raw capacity — they present a family of foundation MLIPs that reportedly match leading inference and training speeds while retaining equivariant accuracy. Practitioners doing MD-based materials/chemistry screening could fine-tune these foundation models on targeted chemical spaces, gaining both the sample efficiency benefits of equivariance and throughput competitive with faster invariant models.

The authors grew epitaxial α-MnTe thin films via molecular beam epitaxy on InP(111) substrates and performed low-temperature magneto-transport measurements, observing a hysteretic butterfly-shaped longitudinal magnetoresistance and nonlinear transverse magnetoresistance — signatures inconsistent with MnTe's nominal compensated antiferromagnetic ground state and indicative of finite net magnetization. DFT calculations identify plausible mechanisms including interface-induced symmetry breaking and point-defect-mediated magnetization. This is experimental evidence relevant to altermagnetic spintronics, where researchers exploiting MnTe's predicted spin-split bands (despite zero net moment in bulk) would need to account for thin-film-induced magnetization as both a complication and a potential tunable control knob for device engineering.

The paper models checkpoint-writing training jobs as integrate-and-fire oscillators coupled through a shared bandwidth-limited resource, the same framework underlying Kuramoto and Mirollo-Strogatz synchronization theory. It defines 'anonymous' resources as those whose delivered rate depends only on the count of active users, not their identity, and proves that for identical jobs with write time shorter than compute interval, the two-job phase-gap return map is exactly the identity — under storage contention, a shared power cap, or both. This means the pairwise coupling term the classical phase-locking frameworks require is not merely small but zero, so checkpoint storms cannot be explained by two-body phase-locking dynamics in anonymous systems. Practitioners investigating checkpoint storms should look to higher-order (N-body) effects, non-anonymous resource policies, or job heterogeneity rather than pairwise synchronization as the driving mechanism.

The work identifies and mitigates local-oscillator (LO) pulse overlap in a mixer-first receiver front-end (RXFE), a known but under-addressed cause of input-matching (S11) and noise-figure (NF) degradation, using a CML-based LO generator design methodology to suppress overlap. Fabricated in TSMC 65nm CMOS, the chip achieves S11 below -10 dB, NF degradation of only 1.08 dB, and 12.19 mA current draw from a 1.2V supply across 1–5.6 GHz, in a compact 302×209 μm active area. RF/analog IC designers working on sub-6GHz mixer-first receivers (common in software-defined radio and cellular front-ends) can adopt this LO-generator methodology directly to improve NF and matching without changing the core mixer topology.

The study combines spatially resolved X-ray fluorescence (XRF), X-ray diffraction (XRD), and nanoscale X-ray imaging on the same sample regions of archaeological and lab-synthesized lead carbonate (psimythion, PbCO3) to co-register trace-element distributions with specific mineralogical phases — something bulk XRF/XRD alone cannot do since each yields separate averaged datasets. The experimental reference sample was synthesized per a documented 4th century BCE recipe, giving a ground-truth comparison for the archaeological material. This multimodal, spatially-correlated methodology is directly applicable to other cultural heritage or forensic materials science problems where elemental and structural data need to be tied to the same microstructural location rather than treated as independent bulk measurements.

The paper develops a subspace-based data-driven predictive control (DDPC) method for linear parameter-varying (LPV) systems, deriving a multi-step predictor from an affine LPV state-space innovation form that separates contributions of past data, future inputs, scheduling trajectories, and innovations, then projects onto the row span of lifted input-output-scheduling data to get an asymptotically unbiased predictor usable directly in receding-horizon control without explicit LPV system identification. To keep the online optimization tractable, they extend γ-DDPC to the LPV setting via LQ factorization, fixing the number of online decision variables independent of dataset length, plus a reduced-order predictor to curb the otherwise exponential growth in scheduling-related terms. Control engineers working with gain-scheduled or parameter-varying plants (e.g., robotics, aerospace, process control) could apply this to implement MPC-style controllers straight from data logs, bypassing costly LPV model identification while retaining tractable online computation.

Using NV-center magnetometry to perform NMR spectroscopy on water confined in 5.6 nm channels, the authors observe strongly suppressed diffusivity that permits direct 1H spectral detection despite the reduced sample volume. Intermittently the proton line splits into a doublet with tens-of-kHz splitting, attributed to hyperfine coupling to long-lived paramagnetic charge complexes seeded by laser-injected solvated electrons. The metastable, intermittent nature of the feature suggests a transient photochemical species rather than a steady-state impurity effect. This establishes NV-NMR as a viable probe of nanoconfined liquid dynamics and opens a route to studying photoinduced radical chemistry at the nanoscale in situ.

CRILIN is a longitudinally segmented, semi-homogeneous PbF2 Cherenkov calorimeter (~22 X0 deep) read out by UV-extended SiPMs, targeting fast timing and fine transverse/longitudinal granularity for future lepton colliders. A five-module (7×7 crystal matrix) prototype was beam-tested at the CERN SPS with 10–120 GeV electrons and 150 GeV muons. The system achieves sub-50 ps time resolution for electrons above 10 GeV, improving further at higher energy, validating the single-channel 4-SiPM-per-crystal readout scheme. This demonstrates a scalable design path for combining Cherenkov-based fast timing with shower-shape reconstruction in a radiation-tolerant EM calorimeter.

The paper presents a universal 2-DOF (two-degree-of-freedom) stabilizing controller implementation using only stable constituent blocks, applicable to any regular SISO/MIMO plant, including those violating the parity interlacing property that normally forces an unstable controller realization. By applying Youla-Kučera parametrization within an input-output feedback topology, controller synthesis is decoupled from its physical realization, so performance specs can be chosen freely and then mapped onto a cyclic arrangement of stable blocks. This yields practical benefits like localized anti-windup and simplified initialization compared to existing algebraic factorization approaches that trade off design freedom for block stability. Control engineers can use this as a drop-in realization scheme for otherwise-unstabilizable controller designs.

The paper surveys busbar protection (BBP) schemes — leakage-to-frame, high-impedance differential (87Z), interlocking overcurrent (ILOC), electronic (EBBP), and numerical (NBBP) — evaluating their practical limitations in industrial settings, motivated by the 2022 Taiwan '303 blackout' caused by a fault combined with CT (current transformer) supervision lockout. It addresses design challenges including CT placement and saturation, dynamic protection zone selection, and evolving fault scenarios not well captured by existing standards. The analysis is a structured comparative evaluation rather than a novel algorithm, giving protection engineers criteria for selecting a BBP scheme suited to their CT infrastructure and fault-tolerance requirements. This is directly applicable to substation and switchgear protection design reviews.

Double positron (β+β+) decay produces four coincident 511 keV annihilation gammas alongside a characteristic 'double blob' track topology, offering a near-background-free signature when all four photons are tagged in coincidence. The authors propose a liquid krypton TPC (time projection chamber) exploiting this four-gamma tagging to suppress backgrounds essentially to zero. They estimate that even a modest-scale prototype could improve on existing limits, with a ton-scale detector projected to reach sensitivities of 10^29–10^30 years for the decay half-life, a substantial improvement over current bounds. This establishes liquid krypton TPCs as a candidate technology for next-generation rare-event searches, worth pursuing via prototype R&D on gamma tracking and coincidence tagging efficiency.

The paper proposes an approximate robust nonlinear MPC (model predictive control) scheme combining an input-to-state control Lyapunov function with a robust control barrier function, extending a prior nominal infinitesimal-horizon NMPC formulation that reduces online computation to small quadratic programs solvable on embedded hardware. The ISS (input-to-state stability) guarantee bounds state error under bounded disturbances or model mismatch without requiring the full nonlinear MPC optimization at each step. Numerical experiments on nonlinear constrained spacecraft control show performance competitive with existing robust NMPC methods from the literature at a fraction of the computational cost. Control practitioners working on real-time embedded nonlinear control, such as aerospace or robotics, could adopt this QP-based scheme where full NMPC is infeasible.

The paper studies LLMs for parent-order execution, splitting a large order into child orders to minimize market impact and cost, an area previously dominated by hand-specified strategies like TWAP and Almgren-Chriss or task-specific trained RL/ML models. PACE (Plan-Ahead Controlled Execution) is a hierarchical framework where an LLM performs long-horizon planning of the order-splitting schedule and separately handles short-horizon execution decisions, requiring no explicit market microstructure assumptions or fine-tuning. On Shenzhen Stock Exchange Level-1 order book data, PACE outperforms TWAP, Almgren-Chriss, and unspecified learning-based baselines. This suggests LLM-based hierarchical planning is a viable zero-shot alternative to custom-trained execution algorithms, and practitioners could replicate the plan/execute decomposition with off-the-shelf LLMs plus market data feeds.

The GNN learns an effective magnetic energy functional from electronic-structure (DFT-type) calculations, conceptually mirroring machine-learned interatomic potentials (e.g., NequIP/MACE) but for spin torques instead of atomic forces. This lets it capture the nonlinear, spatially extended, itinerant-electron-mediated interactions (RKKY-like) without repeated self-consistent electronic solves during time evolution. It's benchmarked across collinear, noncollinear, and noncoplanar spin textures in representative metallic magnets, reproducing dynamics that would otherwise require full electronic recalculation at each step. Practitioners could adapt the message-passing architecture and training pipeline to accelerate atomistic spin-dynamics codes for spintronic device design.

The authors present the first flexible, non-parametric reconstruction of the joint four-dimensional distribution of binary black hole population parameters (from the LVK gravitational-wave catalog) that avoids both the prior-driven bias of strongly parametric population models and the dimensionality-scaling failures of typical model-free approaches that discard cross-parameter correlations. By preserving correlations across four dimensions simultaneously, the reconstruction can reveal astrophysically meaningful joint structure (e.g., mass/other-parameter correlations tied to distinct formation channels) that lower-dimensional or single-parameter analyses would miss. This is directly usable by population-synthesis modelers to stress-test formation-channel predictions against correlated structure in the real catalog rather than marginal distributions alone.

The construction builds families of lifted product (LP) codes via group extensions and graph lifts applied to the Tanner graph, preserving local (low-density) structure while relating code parameters, logical operators, and fault-tolerant logical gadgets across the family through chain and cochain maps. The authors report LP codes with improved parameters over previously known constructions, show that code-surgery gadgets transfer across finite lifts via chain maps (with lower space overhead in several cases), and develop a parallel product surgery scheme for lifted clustered cyclic codes. Quantum LDPC code designers could use this lifting framework as a systematic route to scale up code families while reusing already-designed logical gate implementations rather than re-deriving them from scratch.

The authors analyze the su(n) algebraic structure (embedded within u(n^N)) of the one- and two-body Hamiltonian governing collective neutrino oscillations, deriving a product structure of the algebra that yields generic closed-form expressions for operator expectation values, Rényi entropy, and Wigner functions. Applying BBGKY hierarchy truncation to this structure, they construct a systematic hierarchy of approximations that extend beyond mean-field theory while keeping classical computational cost polynomial rather than exponential in system size. This gives supernova and early-universe neutrino transport codes a tractable route to include beyond-mean-field correlation corrections without resorting to full exponential-cost many-body simulation.

Uses renormalization-group scaling of the Yang-Mills coupling — relevant below 4D, marginal at 4D, irrelevant above 4D — to argue 4D is an IR-critical dimensionality. Builds an explicit racetrack superpotential model (multiple gaugino condensates) generating a radion potential, where gauge-criticality-induced SUSY breaking plus supergravity corrections stabilize the radion at a value corresponding to noncompact 4D spacetime, without presupposing 4D at the outset. Serves as a proof-of-principle mechanism that could be extended with different gauge groups/matter content or connected to swampland/landscape selection arguments.

Proposes a non-adaptive, ancilla-free, control-free algorithm for learning an unknown Lindbladian generator (Hamiltonian plus dissipative Pauli-indexed coefficients) purely from time-evolved measurement statistics. For dynamical strength bound Λ, it estimates each coefficient to error ε using Õ(Λ²/ε²) experiments and Õ(Λ/ε²) total evolution time, with polynomial classical post-processing. A support-learning stage first isolates a poly(Λ/η)-sized candidate set containing every coefficient of magnitude ≥η, feeding a subsequent estimation stage. Directly applicable to characterizing noise/dissipation on real quantum devices without entangling ancillas, making it practical for near-term hardware benchmarking and calibration.

Addresses the tension between high-rate qLDPC codes' physical-qubit efficiency and the difficulty of realizing modular, low-overhead fault-tolerant logic (code surgery, gate teleportation) on their otherwise unexploited structure. Co-designs canonical lifted-product (LP) codes with cyclic symmetry alongside a matching logical instruction set, revealing a canonical logical basis where conjugate logical operator pairs organize into rows/columns of cyclic orbits inherited from the underlying classical codes — analogous to hypergraph-product code structure. This gives a systematic, certifiable route to fault-tolerant logical gates on high-rate LP codes for anyone designing near-term qLDPC-based fault-tolerant architectures who wants structured rather than ad hoc logical operations.

Introduces a physics-informed generative U-Net for fuzzy dark matter (FDM) simulations handling two tasks: evolving cosmological fields from initial conditions to an arbitrary scale factor, and super-resolving low-resolution FDM snapshots. The core contribution is a physics-informed loss term enforcing consistency with the governing Schrödinger-Poisson (SP) equations during training, rather than relying solely on data-driven generative objectives. Results show this physics-informed loss significantly improves adherence to SP dynamics for the evolution task compared to baselines lacking it. Functions as a fast drop-in emulator/upsampler for FDM simulation pipelines, and the physics-informed loss design is transferable to other PDE-constrained generative cosmological emulators.

Formalizes Pauli Encodings — one-bit encryption schemes whose ciphertexts are eigenspace projectors of Pauli strings — and analyzes their unclonable-encryption security through the monogamy-of-entanglement (MoE) game framework. Proves a universal lower bound of 1/2 + 1/(2√K) on optimal MoE winning probability for K-Pauli-string schemes, with tighter bounds for structured subfamilies, and shows encodings restricted to X/Z-only Pauli strings of length n are provably insecure. Also establishes a universal 3/4 obstruction proving that security arguments based solely on pairwise adversary-guessing marginals cannot establish full unclonable-indistinguishability, ruling out an entire class of proof techniques. Gives cryptographers concrete bounds and no-go results to design around when building or analyzing Pauli-based unclonable encryption or copy-protection schemes.

SymFT is a high-throughput classical simulator for Clifford-dominated fault-tolerant circuits containing Pauli rotations, stochastic Pauli noise, mid-circuit Pauli measurements, and measurement-conditioned Pauli feedback. It combines symbolic Clifford-Pauli frame factorization, which reduces per-shot branch-probability sampling to just the Pauli rotations and measurement projectors (since the unitary Clifford/Pauli-frame parts don't affect branch probabilities and need not be applied every shot), with adaptive stabilizer-coordinate planning, which uses a shared stabilizer-destabilizer tableau and stores only the dynamically active non-stabilizer degrees of freedom. This should yield substantial speedups over naive stabilizer-plus-magic-state simulators for large-scale fault-tolerant circuit benchmarking, directly useful for validating QEC protocol designs before hardware deployment.

This extends the well-known axial inflaton-gauge coupling ('ξ' controls tachyonic amplification of one helicity) to a massive Abelian vector field, showing the instability threshold shifts to |ξ|>m/H≡m̄, with amplitude scaling as exp[π(|ξ|-m̄)] in the heavy regime rather than the massless exp(π|ξ|). Because amplified modes stay sub-Hubble for m̄≫1, their sourcing of curvature perturbations ζ is power-law suppressed, yielding P_ζ^id ∝ m̄^{-2} in the weak-backreaction limit. A practitioner modeling axion-gauge inflation or gauge-field friction mechanisms (e.g., for non-Gaussianity or gravitational-wave signatures) can use this scaling to bound how much a massive spectator vector could contaminate CMB/LSS observables versus contribute unobservable friction.

The paper models the stochastic gravitational-wave background from metastable cosmic strings that decay via zero-temperature nucleation of monopole pairs, which truncates loop production after a lifetime set by the breaking rate κ_CS, producing an IR spectral tail and a transition to a stable-string plateau at high frequency. Using synthetic LISA data including instrument noise and unresolved astrophysical foregrounds, they perform Bayesian parameter estimation in the (Gμ, κ_CS) plane to map detectability, parameter correlations, and degeneracies. This gives a template for distinguishing mere background detection from full spectral reconstruction, directly useful for LISA science-case forecasting of cosmic-string or other metastable-defect scenarios.

The authors run DMRG (Density Matrix Renormalization Group, a tensor-network method) on NERSC's Perlmutter supercomputer to generate ground-state energies of the LMG model for systems up to 1400 particles, producing one of the largest such benchmark datasets to date. They compare against VQE and Sample-Based Quantum Diagonalization (SQD) executed on an IBM Eagle quantum processor, finding VQE within 1% error only at 6 particles before error growth at scale. This gives practitioners a concrete classical-vs-NISQ error/scaling baseline for LMG-type collective spin models, useful for calibrating expectations and designing better ansätze or error-mitigation strategies for near-term quantum hardware.

The paper analyzes linear perturbations near planar interfaces moving at constant velocity v in relativistic media, covering moving obstacles, thermal boundaries, and external sources within one formalism. Using a Laplace-transform approach, they derive a propagator representation showing the interface solution decomposes into modes with purely imaginary frequency and wavenumber, selected by the line iω = vik in the (iω, ik) complex plane; sweeping v across this line geometrically classifies all interface-localized responses. This gives a practitioner a general recipe for computing wake and boundary-layer profiles (and shock-tail asymptotics) in relativistic hydrodynamic or kinetic theories directly from the medium's dispersion relation, without solving the full PDE case-by-case.

The review synthesizes constraints on dark matter microphysics — free-streaming suppression (warm DM), wave interference/quantum pressure (fuzzy DM), and self- or DM-SM interactions — using nonlinear structure below ~1 Mpc as the discriminating regime. It surveys how dwarf galaxy counts/internal structure, strong lensing substructure, the Lyman-α forest power spectrum, stellar stream perturbations, and high-z galaxy abundances each translate into bounds on parameters like DM particle mass, cross-section, or interaction strength. For a practitioner, it's a current reference map of which small-scale observable is most constraining for which DM model class, useful for prioritizing which datasets or simulations to target next in testing non-CDM scenarios.

Building on a known single-bit uncloneable encryption scheme whose security proof relies on encoding bases drawn from a subset of Clifford unitaries, the authors extend the construction to arbitrary-length messages while preserving unconditional (information-theoretic) security, with encoding time polynomial in message length and security parameter. This closes an open question about whether statistical security survives the single-bit-to-many-bit generalization for uncloneable encryption. Cryptographers building quantum-secure primitives (uncloneable currency, copy-protected software/licenses) can now use this as a drop-in scheme for protecting multi-bit payloads without falling back to computational (rather than statistical) security assumptions.

Working within the type-I seesaw mechanism with a diagonal heavy Majorana mass matrix, the authors impose Hermiticity plus three texture zeros on the light Dirac mass matrix and numerically scan for benchmark solutions consistent with neutrino oscillation data. They identify four viable texture patterns under normal ordering and three under inverted ordering that reproduce mixing angles and mass-squared differences within 3σ of the NuFIT global fit. This provides concrete, falsifiable Yukawa-matrix ansätze that model-builders can use as starting points for constructing seesaw models with reduced parameter counts, testable against future precision neutrino oscillation and 0νββ (neutrinoless double beta decay) data.

The paper analyzes the covariant phase space of field theories on null boundaries and isolates a global zero-mode ambiguity distinct from the residual gauge (BMS-like) transformations usually studied. They show this zero mode modifies the symplectic/canonical structure, producing a new quasilocal edge observable that enters and deforms the charge algebra (central extensions/non-trivial brackets). The construction is general, applying uniformly to null infinity and finite-distance null surfaces like black hole horizons, suggesting implications for holographic charge algebras and horizon entropy bookkeeping.

The authors present a hybrid qubit-qumode algorithm for polynomial-degree nonlinear ODEs, using the Fokker-Planck/warped-phase-transformation route (Tennie-Magri, Jin-Liu-Yu) to recast nonlinear dynamics as a linear Schrödinger-type evolution, with the added Fourier-mode parameter placed on a continuous-variable qumode rather than requiring additional qubits or oracles. Their key structural result is a bipartite Pauli decomposition of the discretized generator's Hermitian parts into O(log N) mutually commuting families, each further factorized via local positive-semidefinite decomposition — enabling efficient Hamiltonian simulation. This gives a provably correct, oracle-free circuit construction that practitioners could implement on emerging hybrid CV-DV (continuous-variable/discrete-variable) hardware for nonlinear dynamical simulation.

The study performs an unmodeled (generic burst) coincidence search between IceCube high-energy neutrino alerts and gravitational-wave strain data from Advanced LIGO/Virgo/KAGRA's O3 run, targeting weaker joint signals below the sensitivity of standard real-time or catalog-based follow-up. No statistically significant coincident gravitational-wave signal was found. The null result is translated into distance exclusion limits for candidate joint emission models (e.g., core-collapse or compact object mergers with neutrino production), providing updated multi-messenger sensitivity benchmarks that future searches or population studies can build on.

The paper establishes sharp, system-size-independent inverse-temperature thresholds for quantum Gibbs states of long-range Pauli Hamiltonians (bounded per-site interaction strength) at which classical-mimicking properties (unentanglement, zero magic, partition-function analyticity, correlation decay, algorithmic tractability) each individually fail, forming a hierarchy of distinct 'classical-to-quantum' transitions rather than a single crossover. They resolve an open problem from Rouzé, França, and Alhambra (STOC'25) by proving entanglement death occurs at constant (size-independent) temperature even under all-to-all long-range coupling. They also give a polynomial-time classical Gibbs-state preparation algorithm valid up to (and asymptotically colder than) the entanglement-death threshold, giving practitioners a concrete tool for classical simulation limits and benchmarking near-term quantum thermal state preparation.

The authors construct a QSP-based quantum algorithm implementing the overlap fermion Hamiltonian that satisfies the Ginsparg-Wilson relation up to a tunable error ε_e, at a cost only logarithmically larger in ε_e than simulating the Wilson-Dirac Hamiltonian, and with reduced qubit overhead compared to domain-wall fermion simulations (at a modest circuit-complexity cost). They show the QSP construction effectively synthesizes an emergent extra dimension, mirroring the domain-wall origin of overlap fermions as a boundary theory. This gives lattice QCD practitioners a near-optimal, chirally-exact quantum simulation primitive that could be integrated into larger quantum algorithms for lattice gauge theory.

The paper proves that any 'ironed' two-qubit gadget with KAK decomposition parameter a=5/9 achieves the same second-moment spectral gap as the iSWAP gadget when applied across all pairs in K_n for n≥5. The proof uses a representation-theoretic localization argument showing the operator's most negative eigenvalue always resides in the highest-spin SU(2) irreducible summand, combined with a local positive-semidefinite decomposition that handles all spin sectors except the top two. This settles a conjecture of Kong, Li, and Liu and gives circuit designers a provable equivalence class of gates for constructing efficient unitary/state 2-designs.

The paper models dark matter as non-topological solitons (Q-balls, field configurations stabilized by a conserved global charge) formed in the early universe, and shows that interactions among these Q-balls cause them to preferentially grow in the high-density central regions of dark matter halos. This growth dynamically redistributes central mass, flattening the cuspy NFW-like density profile predicted by collisionless cold dark matter into a cored profile, while naturally producing halo-to-halo diversity in inner density/rotation-curve shape at fixed halo mass. The toy model provides a concrete, simulate-able framework (via Q-ball interaction and growth dynamics) for practitioners to test against rotation-curve data and constrain the dark-sector charge/self-interaction parameters.

The authors define a spacetime block-diagram representation of color-code logical patches and operations, grounded in the code's anyon-condensation and domain-wall structure, and derive rules governing valid block assembly. They establish a correspondence between this representation and ZX-calculus diagrams, enabling diagram rewrites that preserve the computed logical operation, and build a compilation algorithm that converts arbitrary ZX representations into valid color-code spacetime layouts. This provides a color-code analogue to lattice-surgery compilers used for surface codes, exploiting color codes' richer transversal gate set, and gives a concrete formal target for anyone building color-code-based fault-tolerant compilers.

Monomial α-attractor T-model potentials fit new ns constraints (small-scale CMB plus DESI) only for hyperbolic-tangent power p≥6, requiring the supergravity construction to assume lower-order terms stay negligible throughout inflation and reheating. The authors instead study a binomial potential (quadratic plus quartic term) with quartic coefficient c satisfying 0<c+1/2≪1, finding non-universal ns behavior reaching values up to 0.965 with α-dependence qualitatively distinct from the monomial case. This is presented as a first step toward more general, non-monomial supergravity embeddings of α-attractors, opening testable deviations from standard predictions for future CMB and large-scale-structure analyses.

They implement structure-preserving, Fourier-based quantum circuits encoding 1D (up to 1024 grid points) and 2D (32×32) acoustic wave dynamics and variable-mass Dirac dynamics, executed on the Quantinuum H2-2 trapped-ion processor with encoded state-space dimension up to 4096. Instead of full-state tomography, they estimate subdomain kinetic energies directly from measurement samples, sidestepping the exponential cost of full readout. Results track classical kinetic-energy dynamics with low mean absolute error across all tested cases, demonstrating a scalable observable-extraction strategy that others could reuse for larger grids or other PDE classes on near-term quantum hardware.

The scheme cascades N parametric microring resonators coupled to a common bus waveguide to synthesize a broadband, flat-topped squeezed-vacuum spectrum, overcoming the depth-versus-bandwidth trade-off of single-cavity OPO sources. The authors derive the conditions producing a flat-topped spectrum even under realistic intracavity pump attenuation, and show robustness to inhomogeneous resonator detunings and individual component failure. Notably, as few as N=5 coupled rings converge to the ideal Markovian-reservoir limit substantially faster than a single-cavity Lorentzian source, suggesting a near-term integrated-photonics route to broadband squeezed reservoirs for enhanced light-matter coupling.

They implement a modified Hayden-Preskill protocol on a quantum processor in which the post-injection scrambler is dynamically related to the initial state via a SWAP gate, generalized to finite-temperature initial states derived from the SYK model. Both postselection probability and conditional recovery fidelity are measured and found to scale proportionally with temperature, consistent with reduced entanglement in the initial state at lower temperature limiting scrambling-based recovery. Late-time analytic estimates derived under a uniform-operator-spreading assumption match the numerics well, giving a quantitative link between scrambling strength and recovery fidelity that could be extended to other finite-temperature chaotic systems or hardware platforms.

Using above-band excitation delivered via a laser scanning microscope on device-coupled single T centers in silicon, the authors demonstrate up to 70% reduction in optical linewidth, attributed to photo-generated free carriers filling nearby charge traps and suppressing spectral diffusion. They probe charge stabilization dynamics with pulsed above-band excitation and fit a rate-equation model that reproduces both the linewidth narrowing and an accompanying spectral shift. This gives an all-optical, fabrication-free protocol that device engineers can apply to push T-center emission closer to its cavity-limited radiative linewidth, a prerequisite for high-fidelity photon-mediated entanglement in quantum network nodes.

Using tensor-network simulation of a globally-driven quasi-2D superconducting ladder architecture, the authors quantify how amplitude-damping and dephasing degrade information flow and one-/two-qubit gate fidelities under global drives. They apply optimal control to reshape the global pulse, compressing gate-sequence duration by roughly an order of magnitude and restoring high fidelities despite decoherence acting on every physical qubit, including ancillas outside the logical register that sustain the ordered phase. This shows decoherence mitigation in globally-driven processors must target the whole drive protocol rather than isolated qubits, offering a template optimization approach for other global-control hardware proposals.

The authors benchmark exact diagonalization (L=8,14,20) against DMRG (up to L=100) for the 1D transverse-field Ising model using one consistent field grid and convergence diagnostic — independent runs at bond dimensions χ=100 and 200 at every point — verifying exact agreement where the two methods must coincide. They identify and correct a bond-dimension convergence artifact that had produced a spurious discontinuity in an earlier, less systematic dataset, documenting it as a worked example of a common DMRG pitfall. From the corrected dataset they extract the critical point's central charge via two complementary methods, providing a validated, reproducible pipeline and diagnostic checklist for entanglement-based DMRG studies of critical spin chains.

The paper re-derives and generalizes the bms3-integrable hierarchy of Fuentealba et al. (2018) by constructing a bi-Hamiltonian structure via the variational complex on the ring of polynomial symbols, onto which a Nijenhuis operator (a recursion operator ensuring compatibility of Poisson structures) is attached, establishing integrability in the classical sense. An analogous construction is carried out for AdS3, with an explicit flat-limit check recovering the bms3 result, an alternative tau-function scheme is presented, and a Lie-Poisson analysis shows the hierarchy is not unique. For a subclass of energy-dependent Schrödinger (Lax) operators, the Lax flows are identified with coadjoint orbits of the bms3 algebra, giving a concrete route for readers to construct further conserved charges or explore other admissible bi-Hamiltonian/Nijenhuis structures on asymptotic symmetry algebras.

The authors compute quantum steering, Bell nonlocality, and uncertainty-induced nonlocality between two non-interacting electrons in a 2DEG as functions of Rashba spin-orbit interaction (RSOI) strength α_R and inter-electron separation, using the Bi/Ag(111) system (α_0 = 3.05e-11 eV·m) as a concrete benchmark. All three correlation measures show non-monotonic behavior with α_R: initial suppression followed by a recovery peak near α_R ≈ 4.32e-11 eV·m, robust across the separations studied. This identifies RSOI strength as a tunable control parameter for engineering and stabilizing nonclassical correlations in solid-state 2DEG platforms, offering a concrete target coupling regime for experimentalists designing spin-orbit-coupled qubit or sensing schemes.

The authors work in massive scalar-Gauss-Bonnet (sGB) gravity, where a scalar field couples to the Gauss-Bonnet quadratic curvature invariant and carries an explicit mass term, and derive the post-Newtonian (PN) dynamics of a nonspinning black hole binary up to 1PN order. They obtain the equations of motion, the center-of-mass frame transformation, and the binding energy for both circular and eccentric orbits, incorporating for the first time the combined effects of the scalar mass and the Gauss-Bonnet coupling at this PN order. These results provide analytic templates that can feed into gravitational waveform models (e.g., for LIGO/Virgo/LISA-type searches) to place bounds on the scalar mass and coupling strength via binary inspiral observations.

Using a joint dataset spanning Planck, SPT, and ACT CMB measurements plus DESI BAO and full-shape galaxy clustering, the authors fit models allowing nonzero spatial curvature (Ω_k) and a dynamical dark energy equation of state (w0, wa) and examine the resulting shifts in the inferred spectral index n_s. They report Ωk ≈ 3×10^-3 lowers n_s toward values favored by Starobinsky, Higgs, and simplest α-attractor inflation models, yielding n_s = 0.9667±0.0041 (Planck+DESI) or 0.9692±0.0035 (adding ACT/SPT); allowing time-varying dark energy similarly shifts n_s to 0.9716±0.0032 (combined) or 0.9694±0.0035 with curvature included. The result suggests degeneracies between curvature, dark energy dynamics, and n_s that practitioners modeling inflation or dark energy should account for jointly rather than fixing Ωk=0 and w=-1 by default when comparing to inflationary predictions.

The authors forecast multi-messenger binary neutron star (BNS) detection rates for the Einstein Telescope (ET) and Cosmic Explorer (CE), and propagate these through joint hierarchical Bayesian inference to project constraints on the neutron star equation of state (EOS), BNS mass distribution, and cosmological parameters (e.g., H0). Assuming a local merger rate of 106.6 Gpc^-3 yr^-1 and testing both a narrow mass function centered at 1.4 M_sun and a wide 1.1-2 M_sun function, they find ET alone yields ~40-100 electromagnetic-counterpart-identified events per year depending on detector layout and mass distribution assumptions, with further gains from a joint ET+CE network. This provides a quantitative basis for designing follow-up strategies and prioritizing EOS/cosmology science cases for next-generation GW facility proposals.

The authors reformulate the Carr-Madan Fourier-based European option pricing framework using tensor networks, specifically applying the Superfast Fourier Transform (SFFT), a Tensor Train (compressed matrix product) representation of the Quantum Fourier Transform, directly to the tensorized pricing computation without ever materializing the exponentially large price/Fourier vectors explicitly. This classical tensor-network algorithm sidesteps the memory bottleneck of standard FFT-based Carr-Madan pricing at fine discretizations, and the paper draws a direct comparison between this classical tensor-network approach and genuine quantum-hardware QFT-based pricing. Practitioners working on large-scale derivatives pricing could adopt the SFFT/Tensor Train formulation as a drop-in replacement for classical FFT pricers to scale to much finer strike/maturity grids on classical hardware.

The paper extends the proven convergence regime of Carleman-embedded quantum algorithms for the collisional Vlasov-Poisson system, establishing both analytical and numerical lower bounds on convergence for physically realistic collision frequencies, using a Fourier-Hermite expansion of the shifted phase-space distribution function as the basis. They show that for a broad class of basis functions, convergence of the Carleman linearization requires collision frequency to scale up with increasing velocity-space resolution (i.e., finer Hermite truncation demands stronger collisional damping to remain in the convergent regime). They further note the algorithm's computational complexity depends strongly on whether the desired output is time-averaged versus time-resolved, which is directly relevant for practitioners choosing Carleman truncation order and collision modeling when designing near-term quantum plasma simulation algorithms.

Constructs an explicit family of invertible n×n matrices over Z2 requiring at least 4n − o(n) CNOT/row operations to reduce to the identity, matching known non-constructive counting bounds on generic matrix complexity but with a concrete, verifiable instance rather than an existence proof. Extends the same lower bound to the strictly more powerful model where gates may be arbitrary invertible linear transformations on coordinate pairs ('local logic gates'), not just CNOTs. Proves the permutation group G_n generated by local logic gates on {0,1}^n is naturally isomorphic to the group of invertible affine transformations of Z2^n, reducing questions about permutation complexity in G_n to matrix row-reduction complexity — a structural bridge useful for circuit lower-bound research, compiler benchmarking, and cryptographic linear-layer hardness arguments.

The authors prove that in the large-N limit of arbitrary tensor models, any melonic operator's expectation value depends only on its degree, establishing redundancy among the moments that populate the positive semidefinite matrices central to the tensor bootstrap (tensor-model analogues of Toeplitz/Hankel matrix positivity constraints from random matrix and lattice gauge theory bootstraps). They extend this to show an analogous simplification in the Schwinger-Dyson equations governing melonic moments. The result substantially reduces the dimensionality of the bootstrap's semidefinite programming problem and applies beyond purely melonic tensor models, directly simplifying numerical implementations of the tensor bootstrap program.

The paper studies degree-power sums l_{r,p}(F) = sum_R d_F(R)^p over t-intersecting families F in the k-subsets of [n], generalizing classical codegree extremal problems. The core technical contribution is a two-moment interpolation argument that majorizes x^p (for real p>=2) on integer lattice points by a quadratic upper bound, reducing the problem to bounding just the first two falling moments of the degree sequence, which are tractable via combinatorial counting. Combined with Bey's size-sensitive quadratic inequality, they prove the full t-star (fixing a t-subset) maximizes co_p(F) for all real p>=2 in the sharp range n>=(t+1)(k-t+1), extending to all nontrivial degree levels r for intersecting families with n>=2k, and characterize all equality cases. This unifies and extends prior results previously known only for integer p or specific degree levels, giving a reusable interpolation framework applicable to other extremal set theory problems.

The authors define a tampered memory elephant random walk on Z, where the memory set at each step n is partitioned into D_n (replaced by independent innovations, i.e. simple random walk steps) and its complement D_n^c (behaving as standard elephant random walk memory). This creates a superposition of two competing stochastic drivers, letting the authors isolate how the size and structure of D_n controls the classical diffusive/critical/superdiffusive phase transition first characterized by Gut and Stadtmuller. Under a renewal-structure assumption on the growth of {D_n} and {D_n^c}, they establish a law of large numbers for the resulting process, laying groundwork for finer scaling-limit results (e.g. CLTs or functional limit theorems) that pin down the exact memory threshold needed to preserve each regime. This gives a concrete, tunable model for testing how much 'forgetting' a self-reinforcing stochastic process can tolerate before losing its anomalous long-range behavior.

The paper improves the explicit diophantine definition of Q\Z over Q, exhibiting a polynomial P(t,x_1,...,x_7) in Z[t,x_1,...,x_7] such that t is not in Z iff P has a rational root in the x_i, reducing the auxiliary variable count from Daans' 2024 bound of n=10 down to n=7, building on Koenigsmann's original 2016 existence proof, and the result is shown to generalize to arbitrary global fields. Combined with Z.-W. Sun's prior reduction, this yields an explicit undecidability result: there is no algorithm to decide, for F(x_1,...,x_16) in Z[x_1,...,x_16], the truth of a for-all/exists rational sentence with 9 universal and 7 existential quantifiers, tightening the known unsolvability boundary within the arithmetic hierarchy over Q. Practitioners working on Hilbert's Tenth Problem over Q, still open in general, can use this smaller polynomial as a more tractable building block for further undecidability or complexity reductions.

The paper proves Kemeny Score is NP-complete for exactly three unweighted rankings, even restricted to profiles where every candidate pair is split 2-to-1 (minimal majority margin), settling the complexity of the smallest nontrivial voter count. On the same instances, Winner, Unique-Winner, and Possible/Necessary-Precedence are shown Θ₂ᵖ-complete, while recognizing a (uniquely) Kemeny-optimal aggregate is coNP-complete — mirroring complexity results previously known only for larger, unbounded voter counts. The reduction is from Max-Cut on tournaments of majority dimension exactly 3, and critically it's tight: the optimal Kemeny score directly yields the exact max-cut value, and any Kemeny-optimal aggregate can be decoded back into a maximum cut. The result generalizes to a sharp dichotomy for every fixed number of voters q≥3 and minimum pairwise support s, giving a full complexity map practitioners can use to identify exactly which aggregation settings admit efficient algorithms versus which require heuristics or ILP solvers.

The authors establish the minimal model program (MMP) for klt adjoint foliated surfaces and use it to prove that ε-lc adjoint foliated surfaces with nef adjoint canonical divisor and a nef-and-big integral polarization form a bounded family whenever the volume of their sum is bounded above — a foliated analogue of classical boundedness theorems for polarized varieties. As corollaries they get boundedness and effective birationality for general-type adjoint foliated surfaces plus a uniform positive lower bound on their volumes (a Noether-type inequality), and for ε-lc Calabi–Yau adjoint foliated surfaces they show the canonical volumes of rank-one foliations on the underlying surfaces lie in a fixed discrete set — a gap/discreteness theorem. This extends the foliation-MMP program (building on work of McQuillan, Cascini–Spicer, and others) into the boundedness regime pioneered for ordinary varieties by Hacon–McKernan–Xu, giving foliation theorists concrete tools for classification problems.

The paper constructs a mediating object between two established geometric formalizations of the microscopic-to-macroscopic limit in Markov diffusion theory — likely the stochastic-process/McKean–Vlasov route versus the Otto-calculus/Wasserstein-gradient-flow route on the space of densities. It treats the space of smooth positive probability densities on a closed Riemannian manifold as an infinite-dimensional Fréchet manifold, distinguishing its 'full' continuous cotangent space (arbitrary distributions) from a smoother 'regular' cotangent subspace, and pins down how the choice of continuity-equation operator (mapping particle velocity fields to measure evolution) differentiates the two descriptions. This clarifies the rigorous relationship between particle-level convergence theorems and PDE/variational-calculus-level gradient-flow structures, offering a foundation practitioners in optimal transport, functional inequalities, or diffusion-model theory could build on to move results between the two frameworks.

Building on a result of Bernardi–Holden–Sun linking percolation-decorated random triangulations to peeling exploration, the authors give an explicit geometric bijection/correspondence between the perimeter process of the peeling exploration along the percolation interface and Kreweras walks (a classical two-dimensional lattice walk model with a small fixed step set). They extend this correspondence to the branching peeling exploration, i.e., a tree-indexed family of simultaneous peelings. The result explains, at a concrete combinatorial level, the previously discovered relationship (Da Silva–Powell–Watson) between growth-fragmentation processes — used to encode scaling limits like Brownian disks — and families of correlated Brownian excursions, giving researchers a discrete-walk toolkit for deriving or re-deriving continuum scaling-limit properties of random planar maps.

In F_q^× with A a multiplicative subgroup generating F_q over F_p, the paper proves a proper A is a generalized arithmetic progression (GAP) iff |A| ∈ {1,2,4}, and characterizes when the full group is a GAP. It further shows additive irreducibility for families where |A|>4 and p^e ≡ -1 (mod |A|): such A admits no nontrivial sumset decomposition. A companion bound shows any nonzero c has fewer than |A|/2 representations as a sum/difference of two A-elements when the subgroup index ≥3 and |A|≥5. Results feed directly into exponential sum estimates and pseudorandom constructions over finite fields.

Models an OWC via 1D nonlinear shallow water equations or Boussinesq-Abbott equations, coupling fluid motion to chamber air pressure acting as a spring force, with the partially-immersed structure imposing kinematic constraints. Reformulates the constrained system as a transmission problem between the open-water and chamber domains and proves local well-posedness under conservation of total fluid-elastic energy (no structural damping). Also develops a reduced model treating the chamber's upper fluid layer as rigid. Provides a rigorous PDE foundation for numerical simulation and control design of OWC wave-energy converters.

For binary polynomial optimization over a hypergraph's multilinear set, the paper proposes decision-diagram (DD) based cut separation: for a chosen support pattern, a compact recursive DD encoding (representing only vertex variables, with hyperedge variables implicit in the state) separates a facet-defining cut in the local multilinear polytope and lifts it to the original problem. This yields an extended formulation of the multilinear polytope, a width characterization via valid antichains identifying a new polynomially-solvable hypergraph class, and a generalization of classical hand-derived inequalities (two-links, flowers, odd β-cycles). It's directly implementable as a cut-generation subroutine inside branch-and-cut solvers for 0/1 polynomial optimization.

Completes the Braverman-Kazhdan-Ngô program for Schwartz spaces on L-monoids over non-Archimedean local fields, conditional on the local Langlands correspondence for tempered representations plus a mild assumption on γ-factors of non-supercuspidal discrete series; unconditional for general linear groups. This delivers the desired Schwartz space/Fourier transform framework underlying functional equations for automorphic L-functions in this local setting. Relevant to specialists in the analytic/geometric side of the Langlands program, particularly those working with L-monoids, Vinberg monoids, or BKN-style harmonic analysis.

Constructs profinite monoids DG_K for number fields K whose unit group is G_K, whose maximal abelian quotient recovers the Deligne-Ribet monoid DR_K, whose idempotents biject with subsets P_K of primes, and whose maximal closed subgroups at idempotents realize exactly the ramification-restricted Galois groups G_{K,S} (with a converse: all such subgroups arise this way). Relates DG_K to Borger-de Smit's p-typical Witt vector formalism, describing the fundamental monoid of the semi-Galois category of Λ-rings via fields of norms of local fields K_p. Offers a unifying monoid-theoretic framework linking class field theory, ramification theory, and Witt vector constructions for arithmetic geometers.

For the s-fractional p-Laplacian with homogeneous nonlocal Dirichlet conditions and reaction f ∈ L^q(Ω), the paper proves solutions u are α-Hölder continuous up to ∂Ω, with α below p'(s − N/(pq)) when N/(ps) < q ≤ N/s, and α = s when q > N/s; in the latter regime, u/d_Ω^s extends Hölder-continuously to the closure. The bounds are shown to be nearly optimal and generalize the known linear-case (p=2) boundary regularity theory to the nonlinear fractional p-Laplacian. Relevant to researchers in nonlocal PDE regularity theory and related free-boundary problems.

Studies the nonuniform arithmetic quotient of the Bruhat-Tits building for PGL_d(F_q((t^{-1}))), computing vertex volumes in closed product form via building-theoretic counting (dominant sector parametrization, exact stabilizer counts, cut-set recursion over block compositions). Introduces a homothety-invariant normalized lattice-minima height α, proves α ∈ L^r exactly for 0<r<d, and establishes a sharp cusp-tail estimate of order T^{-d}. Derives convergence of the associated positive-moment height zeta function (the Mellin transform of the cusp-height distribution). Provides exact combinatorial/analytic tools for volume and height computations on Bruhat-Tits building quotients, of use in automorphic forms and arithmetic statistics over function fields.

Proves lattice covering density θ_L(K) for any n-dimensional convex body K satisfies θ_L(K) ≤ Cn log n (log log n)^{10/3+o(1)}, improving the previous universal O(n²) bound of Ordentlich-Regev-Weiss and matching, up to log factors, the classical linear lower bound of Coxeter-Few-Rogers for Euclidean balls — thus fixing the optimal polynomial growth exponent for worst-case lattice covering density. The construction is a 'vertical-horizontal' method that converts a uniform weight estimate on the Boolean cube into an actual lattice covering for arbitrary K, using folded Gaussians and a finite-field Kakeya theorem in a low-dimensional vertical space to select near-optimal Boolean patterns per translate. The vertical-horizontal technique may generalize to other packing/covering problems in discrete geometry and coding theory.

The authors introduce windowed thinning, which partitions bouncy-particle-sampler and coordinate-Zigzag trajectories into deterministic windows and builds a tractable local envelope for the event rate from a single gradient evaluation per window, avoiding continuous rate checks. Paired with quantitative mixing bounds and finite-time bounds on expected bounce/flip counts, this yields explicit query complexities from a Gaussian cold start: O(κ^{1/2}d(d log κ + log 1/ε)) gradient queries for the bouncy particle sampler and O(κd^{1/4}(d log κ + log 1/ε)) full-gradient equivalents for Zigzag, for m-strongly-convex, L-smooth potentials with condition number κ=L/m. This sharpens the known dimension/condition-number dependence for exact PDMP-based sampling and provides a template (envelope construction + mixing bounds) for analyzing other event-driven MCMC samplers.

The authors extend Bridgeland-type stability conditions from projective schemes over a field to projective families over an arbitrary base and prove existence of proper relative moduli spaces of semistable objects in the derived category. They establish mass-Hom bounds (resolving a conjecture of Halpern-Leistner–Robotis), compare the construction to tilt-stability on surfaces/threefolds, build stability conditions on supported derived categories of total spaces of vector bundles including all local Calabi–Yau varieties, and give a new proof of Bondal–Orlov reconstruction. This is foundational infrastructure likely to underpin future moduli constructions (wall-crossing, DT/PT theory) for families rather than single spaces.

The paper proves Ginzburg's conjecture computing the unramified local factor of a global Eulerian integral built from a cuspidal automorphic representation of GL_{2r} paired against the generalized Speh representation Δ(τ_2,r) attached to a cuspidal representation τ_2 of GL_2, and performs an analogous unramified computation for an integral involving degenerate Eisenstein series. Such local computations are the technical linchpin of the Rankin–Selberg/doubling method for showing a global integral represents a specific automorphic L-function, so the result likely feeds directly into functoriality or L-function analytic-continuation arguments in that program.

The authors prove every positive rational admits 'eventually greedy' best Egyptian underapproximations, both allowing repeated denominators and requiring distinct ones, resolving a problem of Erdős–Graham revisited by Nathanson. The technique reformulates greedy-optimality as an optimal control problem for a discrete dynamical system tracking the residual gap, constructs a payoff functional, and analyzes the associated Bellman value function to show eventual agreement between the greedy strategy and the true optimum, yielding asymptotic denominator growth rates as a corollary. They also answer a further Nathanson question by constructing an irrational number whose best Egyptian underapproximation is both unique and greedy at every step.

The authors show that for a large class of open varieties, the log canonical model's compactification admits explicit coordinates given by canonical forms of associated positive geometries (objects originally used to encode scattering-amplitude structure), and use these to compute explicit defining equations. Applications include complements of hyperplane arrangements, cubic surfaces with lines removed, and the moduli space of marked cubic del Pezzo surfaces. This provides a computational bridge between the positive-geometry/amplitudes literature and classical birational geometry (log canonical models, minimal model program), suggesting canonical-form techniques as a general tool for computing explicit equations of compactified moduli and arrangement complements.

The paper studies continuous-time quantum walks (CTQWs) on cycle graphs C_n, focusing on uniform mixing (the transition matrix reaching uniform magnitude at some time t). Using algebraic geometry and Gröbner basis computations, the authors rule out the existence of 'cyclic 9-roots' satisfying the constraints required for uniform mixing on C_9, proving it does not occur at any time. This resolves a previously open case in the classification of which cycles admit uniform mixing, refining the characterization relevant to quantum-walk-based search and state-transfer algorithm design.

The authors study when a density/flux pair solving a no-flux continuity equation in a bounded domain admits a regular Lagrangian flow (RLF) confined to the closed domain, motivated by marginal flows of reflected diffusions. They give sufficient conditions — interior BV regularity, BV control in a boundary collar, a one-sided bound on the absolutely continuous divergence, and vanishing normal trace of the velocity — extending Ambrosio–DiPerna–Lions theory to the boundary case by showing tangency removes the singular boundary contribution to the divergence of the zero-extended velocity. They prove these conditions are sharp by constructing an explicit smooth density/flux pair exhibiting a 'boundary current' that obstructs existence of a confined RLF, giving checkable criteria relevant to well-posedness of reflected SDEs and transport equations in bounded domains.

For a finite family of proper, lower semicontinuous, convex functions on a real Hilbert space, the authors characterize existence and uniqueness of exact solutions x to the system prox_{f_i}(x)=p_i for prescribed points p_i, generalizing classical simultaneous-projection problems. They derive necessary and sufficient conditions for approximate solvability and introduce the inverse proximal property (IPP), generalizing the inverse best approximation property (IBAP), with several equivalent characterizations. Applications to convex feasibility problems and signal recovery give practitioners verifiable conditions for when a proximal-based recovery or feasibility algorithm is well-posed and has a computable solution.

The authors formalize a 'Rational Mentalizing' model that computes the expected utility of social learning by inverse-planning over another agent's goals to estimate the informativeness of their likely future actions, then compares this against the utility of independent exploration under a shared cost structure. Validated on a novel game requiring players to choose between observation and exploration, the model quantitatively fits human behavioral trade-offs, outperforming presumably simpler heuristics. This connects Bayesian Theory-of-Mind inference (used elsewhere for goal/intent inference) to the explore-exploit and social-learning literatures, offering a computational account that could be extended to multi-agent or robotic social learning systems.

The paper derives mechanism-led renewal equations on a network-based mobility model, decomposing the classical instantaneous reproduction number R(t) into a family of location- and route-specific indicators: inward and outward R(t) per location, pairwise R(t) between locations, R(t) at shared meeting sites, and an aggregate network-level R(t), all incorporating within-day (not just day-to-day) movement heterogeneity. This generalizes standard homogeneous-mixing renewal-equation estimators (e.g., Cori et al.-style methods) to explicitly encode commuting-driven contact structure. Practitioners with mobility datasets (transit, mobile phone, or census commuting matrices) could use these formulas as drop-in replacements for real-time surveillance to identify which locations or routes are driving transmission, enabling targeted rather than uniform interventions.

The method addresses spatially variable gene (SVG) detection within tissue regions that are effectively 1D manifolds, such as epithelial layers, where standard 2D coordinate-based SVG methods underperform because they ignore the region's intrinsic geometry. It uses spectral graph theory to extract a principal curve approximating the spatial coordinates of the region, establishing a morphology-aware 1D coordinate system, then fits a GAM (generalized additive model) over this coordinate to detect genes with significant spatial patterns. This provides a more statistically powered and biologically interpretable alternative to 2D spatial smoothing methods for structurally constrained tissue regions. Practitioners working with spatial transcriptomics data, such as Visium or Slide-seq, could adopt this as a preprocessing coordinate-transform step ahead of existing SVG detection pipelines.

Using isogenic RAI1 heterozygous and homozygous loss-of-function hESC lines differentiated via an in vitro cortical protocol, the authors performed longitudinal bulk transcriptomics and found that RAI1 deficiency accelerates progression through the normal developmental gene expression trajectory. Single-cell RNA-seq further resolved how this temporal acceleration manifests across distinct neural cell populations during differentiation. RAI1 encodes a nucleosome-binding protein, so the phenotype implicates chromatin-level regulation of developmental gene timing, linking RAI1 haploinsufficiency (the cause of Smith-Magenis Syndrome) to disrupted pacing rather than simple loss of gene expression. This isogenic hESC system provides a tractable platform for dissecting how chromatin regulators set neurodevelopmental tempo and could be extended to test other SMS-linked mechanisms or candidate therapeutics that restore normal timing.

The authors developed an MS2-tagging pipeline for endogenous transcripts in human iPSCs, combined with single-particle tracking and Hidden Markov Modelling to classify mRNA diffusive states and construct mobility landscapes during differentiation. Applied to beta-actin and beta2b-tubulin transcripts across neural organoids, directly programmed neurons, and vascular organoids, particle tracking revealed a conserved shift from free diffusion toward constrained, compartmentalized mobility states as cells acquire terminal identity. Cytoskeletal perturbation experiments (implicating microtubule-dependent transport) causally link this mobility shift to the underlying transport machinery rather than passive diffusion changes. The tagging and HMM-based analysis pipeline is generalizable to other endogenous transcripts and cell systems, offering a route to quantitatively map RNA localization dynamics genome-wide during any live cell-state transition.

The study identifies a reciprocal antagonism between CTCF occupancy and the histone mark H4K20me3 at TAD (topologically associating domain) boundaries as a mechanism controlling the proliferation-quiescence switch independent of large-scale TAD repositioning. Genome-wide profiling and functional perturbation show elevated H4K20me3 in quiescent fibroblasts displaces CTCF at specific boundaries, driving compact chromatin, altered nuclear morphology, and quiescence-associated transcription, while CTCF binding dominates in proliferating cells and sustains open chromatin and pro-proliferative gene expression even under quiescence cues. Fibroblasts lacking the H4K20me3 methyltransferase KMT5C/Suv4-20h2 show a hyperproliferative phenotype, functionally validating the mark's necessity for the quiescent state. This establishes a boundary-level, mark-versus-factor competition model that practitioners could test at other TAD boundaries or in other cell-state transitions using ChIP-seq/Hi-C plus KMT5C perturbation.

Using quantitative genetic 'animal models' applied to shotgun metagenomic data from 2,394 samples across 794 known-fate feral Sable Island horses, the authors partition variance in fitness-associated microbiome dimensions (derived via canonical analysis of principal coordinates) into additive genetic, maternal, and social-transmission components. This is the first study to jointly estimate microbiome heritability and detect natural selection acting on microbiome traits within the same population, while also testing non-genetic inheritance pathways (maternal and social) as contributors to microbiome-mediated fitness variation. The long-term, individual-based pedigree and metagenomic dataset allows disentangling genetic versus environmentally/socially transmitted microbiome effects at a scale rarely available outside model organisms or humans. The animal-model framework applied here is replicable in other wild or captive populations with pedigree and longitudinal metagenomic data to test whether microbiome-mediated adaptation is a general evolutionary phenomenon.

Applying the Geometric Framework for Nutrition, the authors mapped Plutella xylostella performance across a protein:carbohydrate (P:C) nutrient space, showing growth (biomass) is protein-limited while immune traits (hemocyte counts, phenoloxidase activity) peak at intermediate P:C ratios distinct from the growth optimum. Survival after Bacillus thuringiensis challenge tracked protein intake specifically, and antioxidant/detoxification enzyme responses varied non-linearly and non-uniformly across the nutrient landscape. This demonstrates trait-specific rather than universal nutritional optima, a finding replicable via response-surface/GF designs in other insect systems and relevant to integrated pest management (e.g., optimizing host-plant nutrient content to suppress pest immunity or stress tolerance).

The authors build an epidemiological model comparing single-mode (purely density- or frequency-dependent) versus mixed-mode transmission under an imposed linear trade-off between infection probabilities via each route, then compare equilibrium prevalence. Key result: mixed-mode transmission only exceeds single-mode prevalence when the two transmission routes differentially affect host fitness (e.g., differing virulence or fecundity costs per route); absent that fitness asymmetry, mixing modes confers no prevalence advantage over the best single mode. This is a useful null result for modelers of STIs, vector-borne, or environmentally transmitted pathogens — it suggests searching for route-specific fitness costs before attributing high prevalence to transmission-mode diversity per se.

Researchers applied Metarhizium brunneum and Beauveria bassiana as preventive soil treatments against Rhynchophorus ferrugineus on date palms, quantifying fungal persistence via culture-based isolation and high-throughput metagenomic sequencing of soil microbiota over 180 days, alongside continuous palm health monitoring. Both EPF species persisted for the full 180-day window and significantly reduced weevil damage, with B. bassiana outperforming M. brunneum in duration of protection despite comparable persistence; critically, metagenomic community-level analysis showed no major shifts in overall soil microbiome composition post-application. This supports EPF as a non-disruptive biocontrol option for palm plantations and provides a persistence/efficacy benchmark (180-day window) and a metagenomic protocol for assessing off-target microbiome effects in future biocontrol deployments.

The paper formally analyzes CERIS-JGRA, a reaction-norm GWAS method that substitutes a climate-derived environmental index for the Finlay-Wilkinson environmental mean. It proves that an index with correlation ρ to the true environmental mean decomposes as ρ·(mean) + residual, meaning its loading τ on mean-orthogonal (plasticity-relevant) variation is bounded by √(1−ρ²) — and since the algorithm is constructed to maximize ρ, it necessarily minimizes τ, and thus the non-centrality of any plasticity-specific association test (which scales as τ²). Simulations across 189 conditions confirm this predicted power loss with a parameter-free closed-form expression matching observed results. Practically, this implies GWAS practitioners using CERIS-JGRA-style indices should decouple index construction from mean-correlation maximization if they want statistical power to detect genotype-by-environment/plasticity loci rather than just mean-performance loci.

The team performed in silico point-mutation scans on the antimicrobial peptide pardaxin, using embeddings from ESM Cambrian (ESMC), a protein language model, to score candidate substitutions alongside conventional AMP design heuristics (e.g., amphipathicity, net charge, hydrophobic moment). Principal component analysis on the ESMC embedding space was used to identify single- and likely multi-substitution variants clustering toward regions associated with improved membrane-disruptive activity. This embedding-plus-rules pipeline offers a template other groups could reuse to triage AMP variant libraries computationally before committing to synthesis and MIC assays against pathogens like S. aureus and E. coli.

HyBRIC and HyBRIC2 are genetically encoded bioluminescent Ca2+ indicators that couple Ca2+-dependent luciferase activation with bioluminescence resonance energy transfer (BRET) to shift emission into the red optical window, where tissue absorption/scattering is minimized. Using the luciferin substrate pyCTZ, the sensors achieve ~55-fold Ca2+-dependent signal increases in vitro, with HyBRIC2 offering improved brightness and response kinetics over HyBRIC in mammalian cells. This addresses the historically low dynamic range of bioluminescent indicators relative to fluorescent GCaMP-type sensors, and the modular BRET/luciferase-activation architecture is a design a lab could adapt to other analytes or luciferin variants for deeper, excitation-free in vivo imaging.

The authors characterized young versus aged bovine thymus using histology, oscillatory rheology, untargeted lipidomics, ICP-MS, and AP-MALDI mass spectrometry imaging to build an integrated mechanical-compositional-spatial profile of thymic involution. Aging increased both storage and loss moduli by over an order of magnitude while narrowing the linear viscoelastic region ~5-fold, indicating a much stiffer, more strain-sensitive tissue state, accompanied by lipidomic remodeling (notably a doubling of triacylglycerol contribution to the lipid pool). These quantitative benchmarks — stiffness ranges, viscoelastic limits, lipid ratios — provide target design parameters for engineers building thymus-mimetic scaffolds or hydrogels intended for immune tissue regeneration.

The bio-hybrid skin model combines an optically engineered synthetic epidermal layer — polydopamine nanoparticles tune melanin-like optical absorption to represent different skin tones — with living primary human keratinocytes cultured in 2D and 3D configurations, aiming to merge the reproducibility of synthetic optical phantoms with the biological authenticity of tissue-engineered skin. In their 'Optical Protection Model,' the team quantified UV-B-induced apoptosis in the keratinocyte layer, demonstrating that the construct can serve as a dose-response readout for phototoxicity and photoprotection testing. This modular architecture is positioned as an animal-free validation platform for optical biomedical devices (e.g., sunscreens, phototherapy, diagnostic optical sensors), and the polydopamine-tone-tuning approach could be extended to model diverse skin pigmentation in device safety testing.

The authors trained a DCGAN on the LGG-MRI dataset (3,929 image-mask pairs) to generate synthetic brain MRI slices, then produced pseudo-segmentation masks for these synthetic images via thresholding, creating an augmented training corpus for a U-Net tumor-segmentation baseline. They compare segmentation performance (presumably Dice/IoU) between a U-Net trained on the original dataset alone versus one trained on the GAN-augmented set, evaluated on a held-out real test split. This is a data-centric AI study — the contribution is empirical evidence on whether GAN-synthesized images with heuristic pseudo-labels are a viable low-cost substitute for expert-annotated data in low-resource medical segmentation pipelines.

The platform prints high-density cell strands using a fast-degrading sacrificial carrier bioink embedded within a mechanically compliant, self-softening base hydrogel. As the carrier degrades during culture, cells proliferate and form interconnected networks whose endogenous cellular contractile force (CCF) drives autonomous, spatially programmable shape morphing of the surrounding hydrogel — a genuinely self-actuating 4D system requiring no external stimuli (light/heat/pH) to trigger shape change, unlike most existing 4D bioprinting approaches. Spatial patterning of the embedded cell-strand geometry acts as the design parameter for the final morphed architecture, suggesting a generalizable strategy for engineering tissues that recapitulate developmental folding/morphogenesis for regenerative or disease-modeling applications.

Bio-CM2 implements 'distributed computational optics,' partitioning the imaging field-of-view across four coordinated miniature optical modules that each capture a sub-FOV onto a shared image sensor, with computational reconstruction integrating the multi-module measurements into a single cellular-resolution, millimeter-scale image. This architecture sidesteps the aberration-scaling problem that limits conventional single-lens miniature microscopes as FOV grows, while avoiding the bulk/complexity of multi-sensor or scanning-based wide-FOV systems. The approach is relevant to anyone building head-mounted or wearable mesoscopes for freely behaving animal imaging, offering a modular hardware+computational-reconstruction template for scaling FOV without proportionally scaling optical complexity or device weight.

ProNA3D is an analysis platform for protein–nucleic acid and nucleic-acid-only complex interfaces that accepts both experimentally solved and computationally predicted (including AlphaFold3) structures, incorporating AF3-specific confidence/scoring metrics into its interface analysis. It performs distance-based interface characterization with interactive 2D interface maps and RNA/DNA secondary-structure topology plots, plus an interface-based density-zoning feature for evaluating dynamic or ambiguous regions directly within cryo-EM maps. For structural biologists, it functions as a post-prediction/post-solving analysis layer — bridging AlphaFold3-era structure generation with functional interface interpretation that previously required stitching together multiple disparate tools.

Using transcriptional profiling, genome-wide CRISPR/Cas9 screens for replication-stress sensitivity, and proteomics, the authors identify FBXO44 as a late p53 target gene that terminates translesion DNA synthesis (TLS). FBXO44 acts as an F-box substrate receptor that ubiquitinates RAD18, the E3 ligase responsible for PCNA monoubiquitination that licenses TLS polymerase recruitment, triggering RAD18 turnover during recovery from replication stress. This creates a temporally restrained circuit: p53 activation induces FBXO44 to shut down RAD18-PCNA-Ub signaling once lesion bypass is complete, limiting persistent error-prone synthesis. Loss of this axis is predicted to raise mutagenesis and could be probed as a synthetic-lethal vulnerability in p53-proficient, TLS-dependent tumors.

The authors used recombinant full-length mouse SEMG1 (Q32-G375) and truncated fragments with patch-clamp electrophysiology to show SEMG1 directly inhibits CatSper, the sperm-specific principal calcium channel required for hyperactivated motility, at physiologically relevant concentrations. Because NH4Cl-induced alkalinization failed to rescue hyperactivation, the inhibitory mechanism acts downstream of the pH-dependent CatSper activation step rather than blocking capacitation-associated alkalinization itself. Domain-mapping with truncated constructs identified two distinct functional regions each contributing to CatSper current suppression, suggesting a multi-site inhibitory mechanism. This establishes CatSper as a direct pharmacological target of seminal plasma regulation, offering a template for CatSper-modulating non-hormonal contraceptives or fertility diagnostics.

Using analytical population-genetic models and individual-based simulations, the authors compare invasion dynamics of a self-compatible (SC) allele into sporophytic self-incompatibility (SSI) systems in diploid versus tetraploid populations, under both codominant and hierarchically-dominant SI allele architectures. The core result: SSI breakdown is systematically easier in tetraploids than diploids, and increasing the dominance rank of the invading SC allele further accelerates its fixation. The dominance-class scenario shows that segregation and expression patterns unique to autotetraploid genetics (more allele combinations per genotype) weaken effective self-recognition. This provides a quantitative framework linking ploidy-driven dominance/segregation changes to the empirically observed correlation between polyploidy and elevated selfing rates in angiosperms.

Hydathodes are xylem-connected structures at Arabidopsis leaf margins that mediate guttation and serve as known infection portals for vascular pathogens like Xanthomonas campestris. The authors show Arabidopsis mutants with reduced hydathode numbers develop spontaneous mesophyll water-soaking upon infection and support unexpectedly large bacterial populations, driven by pathogen entry through stomata and proliferation directly in the water-soaked mesophyll niche rather than the vasculature. Physical occlusion of hydathodes reproduced both water-soaking and enhanced bacterial growth in Arabidopsis and cauliflower, causally linking hydraulic dysfunction to loss of nonvascular tissue resistance. This establishes leaf hydraulic regulation as an active component of "water immunity," suggesting hydathode function or water-soaking resistance as a novel breeding target for bacterial disease resistance.

The authors used adipocyte-specific beta3-adrenergic receptor (beta3AR) knockout mice challenged with a high-fat-sucrose (HF-S) diet, with and without the selective beta3AR agonist CL316,243, to test whether adipocyte beta3AR is required for beige/brown fat-mediated cardiometabolic protection. Beta3AR normally drives adipose tissue lipolysis and BAT thermogenesis, and pharmacological activation induces beiging of white adipose depots — a pathway of therapeutic interest given BAT's limited abundance in aged/obese humans. The genetic deletion isolates adipocyte-autonomous beta3AR signaling from systemic or cardiac-intrinsic drug effects, directly testing whether cardiac metabolic/structural protection depends on this receptor in fat cells specifically. Results should clarify whether beta3AR agonists' cardioprotective effects in obesity are mediated via adipose remodeling, informing whether adipose-targeted versus cardiac-targeted agonism is the better drug strategy.

The authors developed a Xenopus oocyte-based screening platform in which oocytes heterologously expressing individual human SLC transporters are incubated in human serum — a complex, physiologically relevant metabolite library — followed by paired untargeted LC-MS/MS profiling of both intracellular extracts and surrounding medium to detect metabolite depletion/accumulation signatures indicative of transport activity. Applied to five transporters (SLC10A2, SLC10A6, SLC13A2, SLC16A10, SLC46A1), the method detected characteristic exchange signatures, with automated feature annotation refined by manual chromatographic/spectral review to assign putative substrates. This sidesteps the need for pre-selected candidate substrate panels, enabling unbiased deorphanization against physiological metabolite complexity rather than synthetic buffers. The platform is generalizable to any cloneable SLC and could scale to map substrate spectra across the ~400-member human SLC superfamily, informing drug-transporter interaction and nutrient biology studies.

The therapy uses a high-capacity (helper-dependent) adenoviral vector co-expressing two transgenes — IL-1Ra (an interleukin-1 receptor antagonist targeting inflammation) and PRG4/lubricin (supporting chondrocyte health and cartilage lubrication) — delivered via intra-articular injection. Efficacy was demonstrated across both murine and equine osteoarthritis models, with the equine model providing large-animal, biomechanically human-relevant validation uncommon in OA gene therapy studies. Beyond local joint outcomes (functional, structural, pain metrics), treatment normalized inflammatory gene/protein signatures in joint tissue and in the dorsal root ganglia (DRG) housing joint-innervating sensory neurons, indicating a peripheral neuro-immune mechanism for pain reversal rather than purely local anti-inflammatory action. The dual-transgene, single-vector, intra-articular design and cross-species efficacy data position this as a translatable disease-modifying OA candidate ready for further large-animal/clinical-stage development.

The authors generated a four-way founder cross of 444 F2 mice with genetically diverse liver disease susceptibility, then performed deep quantitative proteomic profiling of livers paired with genotyping to map protein quantitative trait loci (pQTLs). They identified QTLs for 2,652 proteins across more than 4,000 unique genomic loci, distinguishing cis-acting (local, likely direct genetic control) from trans-acting (distal, likely network-mediated) regulatory loci. Integration of proteomic, genomic, and phenotypic data enabled construction of gene-protein-phenotype networks, nominating candidate regulatory genes and proteins underlying MASLD trait variation. This resource-scale pQTL dataset provides a systems-genetics framework and candidate gene list other researchers can mine for causal driver validation, functional follow-up, or cross-referencing against human MASLD GWAS/proteomics data.

The authors sampled 40-68 wild adult male Dendropsophus minutus across an annual cycle and quantified four hepatic phenotypic modules — intracellular histochemistry (pigment/glycogen stores), cellular morphometry, tissue-level volumetric composition (hepatocytes, sinusoids, melanomacrophage centres, portal structures), and organismal indices (liver mass, body condition) — to test for a predictable temporal response hierarchy across organizational levels. The core hypothesis is that lower organizational levels (intracellular/cellular) exhibit faster, more immediate seasonal responses than higher levels (tissue/organismal), which would lag or integrate cumulative change. This design offers a replicable multi-level framework for other ectotherm field studies aiming to disentangle rate-dependent physiological plasticity from mere seasonal correlation.

Using 7T fMRI from the Natural Scenes Dataset, the authors built encoding models relating cortical responses to narrative scene descriptions versus word-order-scrambled lexical controls that preserve content words but destroy compositional syntax. Intact narratives predicted high-level visual cortex responses better than scrambled controls, with the advantage scaling with a scene's compositional richness, indicating the region encodes relational/compositional structure beyond a bag-of-words object list. A language-aligned vision model (trained with both visual and linguistic supervision) outperformed a self-supervised vision-only model at capturing this structure, suggesting language-alignment during training instills representations closer to how visual cortex organizes scene relations — a testable target for future multimodal model-brain alignment work.

In a murine imiquimod (IMQ)-induced psoriasis-like dermatitis model, IMQ drives ER-mitochondria contact site (MAM) formation and UPR activation in dendritic cells, which enhances IL-23 expression independently of but convergent with TLR7/MyD88 signaling; concurrently, elevated cytosolic Ca2+ activates the NLRP3 inflammasome, releasing mitochondrial DNA. Keratinocytes exposed to IMQ upregulate UPR-dependent genes including the antimicrobial peptide Defb14/mBD14, and extracellular mtDNA plus mBD14 cooperatively engage plasmacytoid dendritic cells via TLR9, closing an amplification loop. This positions ER stress/UPR and MAM signaling as a mechanistic node upstream of IL-23 and TLR9 activation, offering candidate targets (e.g., UPR inhibitors, MAM disruptors) for combination therapy alongside existing IL-23-axis biologics like the gene mentioned Gelsolin counteracting this circuit.

The authors mined public SARS-CoV-2 sequence data from the pandemic's first two years to identify high-diversity epitope sites, then designed mRNA-based 'wobble vaccines' — complex antigen pools incorporating multiple variant sequences at those sites — to apply 'rare epitope suppression' (RES), redirecting humoral immunity away from hypervariable immunodominant epitopes toward more conserved regions. The claim is that this pool-based antigen design broadens antibody targeting and improves cross-strain neutralization compared to single-strain vaccines. This is a platform-level approach that could be replicated for other rapidly evolving pathogens (e.g., influenza) by substituting pathogen-specific diversity-mapped antigen sets into the same mRNA delivery framework.

Using sequencing, imaging, and proteomics in mammalian cells subjected to UV damage, the authors show that DNA methylation is largely preserved during nucleotide excision repair-associated synthesis, and they identify recruitment of DNMT1 and DNMT3A to repair sites, driven by the DNMT1 cofactor UHRF1 together with UV damage repair endonucleases. They further demonstrate a mechanistic link to chromatin dynamics: newly deposited H3.3 histones at damage sites help recruit this methylation-maintenance machinery, extending the known replication-fork paradigm (where UHRF1/DNMT1 maintain methylation) to a DNA-damage-repair context. This establishes damage-repair synthesis as a second genomic context requiring active methylation maintenance, suggesting UHRF1/DNMT1/H3.3 dynamics as candidate targets for probing epigenetic instability in damage-prone or repair-deficient tissues.

In the murine ACL-transection model of post-traumatic osteoarthritis, single intra-articular injection of high-capacity adenovirus (HCAd5) delivering NFkB-responsive IL-1Ra provides structural and functional benefit in early-stage disease but fails to sustain cartilage protection as OA progresses, and repeat dosing with the same HCAd5 serotype is blunted by pre-existing anti-vector immunity. The authors demonstrate that serotype-switching for the repeat injection restores effective transduction and therapeutic benefit, establishing a sequential heterologous-serotype dosing strategy to overcome anti-adenoviral immunity in a chronic, progressive disease context. This has direct translational relevance for gene therapy dosing regimens in OA and other conditions requiring repeated local adenoviral delivery, where serotype rotation could be built into treatment protocols.

Using yeast two-hybrid assays, in vitro pull-downs with bacterially expressed proteins, and bimolecular fluorescence complementation (BiFC) in Drosophila S2 cells, the authors mapped the minimal interacting domains between the insulator protein BEAF and three co-insulator proteins (CP190, Pzg, Chro). They find that Pzg and Chro, unlike CP190, can engage BEAF via interactions detectable at a distance (consistent with BiFC signal indicating proximity without requiring direct stable binding captured by pull-down), and functional assays in transfected S2 cells reveal distinct roles for each co-insulator in BEAF-dependent chromatin organization/transcriptional regulation. This domain-mapping data provides concrete interaction interfaces that could be mutated or disrupted in follow-up studies to dissect BEAF-dependent TAD boundary formation and its transcriptional consequences.

SPARK-ID is a proximity-labeling method used with SIRT6, Ku80, and MRE11 as baits to capture time-resolved, chromatin-associated interactomes from DSB formation through recovery. The three sensors share an enriched core repair interactome while diverging in temporally specialized partners: SIRT6 toward RNA/chromatin regulation, Ku80 toward telomere-associated and translational programs, and MRE11 toward recombination and DNA synthesis. Modularity analysis identified functional modules linked by 'connector' proteins, with Nucleolin bridging DNA repair, RNA-metabolism, and nucleolar modules; Nucleolin depletion perturbs this cross-module coordination, implicating it as a hub for integrating repair with broader nuclear processes. The dataset and proximity-labeling framework offer a template for mapping temporally dynamic interactomes around other DNA damage response factors or other transient chromatin events.

The study systematically profiles PDE4B expression in SSc using single-cell RNA-seq (transcript level) alongside immunohistochemistry, immunofluorescence, and multiplexed immunofluorescence (protein level) across lung, PBMCs, and skin. PDE4B was consistently dysregulated in immune cell populations, but direction and magnitude of dysregulation varied by tissue compartment, suggesting cell-type- and organ-specific regulation rather than uniform pathway activation. This compartmental map provides a mechanistic rationale and biomarker framework for trialing PDE4B inhibitors like nerandomilast in SSc, and could guide patient stratification by matching target expression to tissue involvement.

Sindbis virus's structural polyprotein uses a hydrophobic segment corresponding to E2's second transmembrane domain to engage translocon components cotranslationally, modulating -1 programmed ribosomal frameshifting (-1PRF) that controls polyprotein stoichiometry. Building on prior work linking this segment to translocon interactions, the authors knock out components of the recently characterized multipass translocon (MPT) and show this increases -1PRF efficiency, implicating MPT recruitment as a regulatory checkpoint for viral protein biogenesis. This links host ER quality-control machinery directly to viral gene expression control, suggesting MPT components as potential targets for antiviral strategies aimed at disrupting alphavirus structural protein maturation.

Leveraging the newly solved ELMO1/DOCK2 complex structure, the authors perform structure-based virtual screening to identify small molecules predicted to disrupt this protein-protein interaction interface that activates Rac1 GTPase signaling. They apply molecular dynamics (MD) simulations and free energy landscape (FEL) analysis to assess binding stability and druggability of candidate hits at the interface. This establishes a computational pipeline and starting hit list that could be handed off for experimental validation (binding assays, cellular Rac1 activity assays) as a first-in-class ELMO1/DOCK2 inhibitor program.

The study quantifies STLV-1 Tax expression across oropharyngeal and gastrointestinal tissues (tonsils, stomach, and others) from naturally infected olive baboons (Papio anubis), a validated animal model for HTLV-1, using and comparing multiple PCR-based quantification methods. This addresses the scarcity of human tissue data on HTLV-1 distribution along suspected oral transmission routes. The methodological comparison of PCR approaches for Tax detection provides a benchmarked protocol other researchers could adopt for tissue-level retroviral quantification in NHP models, informing future studies of HTLV-1/STLV-1 tissue tropism and transmission biology.

Using human pluripotent stem cell-derived lung organoids combined with comparative computational analysis (likely single-cell transcriptomics and trajectory inference), the authors show that terminal and respiratory bronchiole epithelial cells, structures present in humans but absent in mice, can directly regenerate AT1 alveolar cells without transiting through the AT2-derived KRT8+ intermediate state established in mouse injury models. They further find that cells resembling profibrotic 'aberrant basaloid cells,' previously associated with pathological fibrosis, arise as a faster, physiological intermediate in this human-specific regenerative trajectory. This reframes aberrant basaloid-like states as a normal repair checkpoint rather than purely a disease marker, with implications for interpreting single-cell data from IPF/pulmonary fibrosis patient lungs and for building more human-relevant organoid injury-repair models.

In an in vitro model, SH-SY5Y human neuroblastoma cells exposed to high glucose (50 mM, 24h) exhibited reduced viability, elevated intracellular and mitochondrial ROS, depleted glutathione, and increased apoptosis, hallmark features of diabetic neuropathy-associated mitochondrial dysfunction. Corilagin treatment reversed these effects, with mechanistic data pointing to restoration of the AMPK-SIRT1-PGC1α-TFAM axis, a core regulatory pathway for mitochondrial biogenesis and antioxidant defense. This positions Corilagin as a candidate neuroprotective compound acting via mitochondrial biogenesis pathways rather than glycemic control, providing a mechanistic starting point for in vivo diabetic neuropathy models and structure-activity studies of ellagitannin derivatives.

Using live-cell calcium imaging in Physcomitrium patens, the authors map a spatiotemporal Ca2+ signaling framework following mechanical wounding: an initial propagating Ca2+ wave, followed by two distinct secondary response modes, including a novel stochastic 'flashing' Ca2+ signature confined to wound-adjacent cells that persists for hours. They show membrane deformation from wounding activates plasma membrane Ca2+ influx, and that osmotic stress-induced membrane tension alone can trigger rapid Ca2+ spikes without the spatially restricted flashing pattern, and pharmacological channel blockade experiments further dissect the mechanosensitive Ca2+ influx pathway. This establishes a testable model linking mechanosensation to regenerative cell fate reprogramming, providing specific Ca2+ dynamic signatures and channel targets that could be probed in other regenerating plant or even animal wound-response systems.

Post-meiotic spermatids lack sister chromatids or homologous chromosomes, precluding homology-directed repair, yet must maintain genome integrity over an extended maturation period. Using γ-irradiation of Drosophila to induce DNA damage, the authors show that spermiogenesis is actively rerouted to eliminate damaged haploid nuclei rather than repair them, and identify a non-canonical role for the histone variant H2Av in mediating this elimination pathway. This describes a haploid-stage-specific genome quality-control checkpoint distinct from canonical DNA damage response pathways, suggesting H2Av-dependent nuclear elimination machinery as a target for future studies of male germline mutation load and fertility, with likely conservation given the extended haploid phase in human spermatid maturation.

The study profiles NPC transcriptomes across humans and rhesus macaques, an outgroup for hominoid comparisons, and identifies genes enriched in human NPCs tied to radial glial identity, proliferative capacity, and sister chromatid segregation. Two zinc-finger transcription factors emerge from this comparison: ape-restricted ZNF90 and deeply conserved OVOL2. The authors map genome-wide binding sites and transcriptional targets of both factors and perturb them in cerebral organoids to assay downstream effects on NPC dynamics. This establishes a testable regulatory axis for comparative work on human-specific features of cortical progenitor expansion.

The authors characterize UBC development in the rhombic lip (RL) using multiplex immunofluorescence and single-cell RNA-seq in wild-type and conditional Eomes-knockout mice, benchmarking against human RL architecture. They find the murine RL is molecularly compartmentalized like the human RL, with an abundant, persistent population of cycling EOMES+ UBC progenitors extending postnatally. Eomes loss disrupts the transcriptional networks driving UBC differentiation and migration, though some processes appear Eomes-independent. This establishes cross-species conservation of RL/UBC biology, supporting mouse conditional-knockout models for studying Dandy-Walker Syndrome and medulloblastoma origins.

Prior work using dimension-reduction (e.g., PCA) on trait matrices reported organ-specific acquisitive-conservative axes in leaves and roots, but such analyses are vulnerable to third-variable confounds (e.g., plant size, phylogeny) inflating apparent independent structure. The authors apply a dimension-reduction method that explicitly controls for these confounders across a broad trait dataset. The controlled analysis collapses the ostensibly separate leaf and root spectra into a single acquisitive-conservative axis operating at the whole-plant level. This challenges the organ-modular framing common in plant economics spectrum literature and suggests trait-based ecological models should integrate above- and below-ground traits jointly rather than treating them as independent axes.

The study examines a long-standing hybrid zone between Penstemon davidsonii (bee-pollination syndrome) and P. newberryi (bird-pollination syndrome) to test classic pollinator-driven ecological speciation models. Despite persistent hybridization over 85+ years, the parent taxa retain strong multivariate trait divergence and genomic differentiation. Floral hue maps to a single genomic locus containing two candidate anthocyanin-pathway genes with large phenotypic effects, and pollinator visual modeling assesses whether bees/birds can discriminate this variation. The key finding — that genomic signatures of reproductive isolation are decoupled from floral trait divergence — implies barriers other than pollinator-mediated assortative mating maintain species boundaries here, complicating simple pollination-syndrome speciation narratives.

The authors run large-scale simulations of a triangular-lattice chiral magnet with quenched random Dzyaloshinskii-Moriya interaction disorder, tracking translational order, sixfold bond-orientational order, and total topological charge as functions of disorder strength and applied field. At high fields, global translational (positional) coherence is destroyed at weak disorder while bond-orientational order and topological charge persist to much larger disorder, opening an intermediate 'topological buffer' regime spanning a Bragg-glass-like skyrmion phase followed by a skyrmion-glass phase. Finite-size scaling, spatial correlation functions, defect (dislocation/disclination) statistics, and spin autocorrelation dynamics are used to characterize the distinct structural and glassy signatures of each phase. This gives a field-tunable disorder phase diagram that practitioners could use to identify disorder tolerance windows for skyrmion-based device design.

This is a pedagogical review of wavefunction-based (post-Hartree-Fock/correlated) periodic quantum chemistry, covering Ewald summation for handling long-range periodic Coulomb interactions, the two dominant one-electron basis choices (plane waves vs. periodic Gaussian-type/atomic-orbital LCAO bases), explicit Hamiltonian integral formulas for both, and k-point sampling as a Bloch-symmetry-adapted reduction of supercell calculations that cuts computational cost. It bridges molecular quantum chemistry (CCSD(T), etc.) with solid-state physics conventions, targeting researchers who want to apply high-accuracy correlated methods (beyond DFT) to condensed-phase systems. Readers implementing or using codes like VASP, CRYSTAL, or PySCF's periodic modules would use this as a reference for the underlying formalism and cost trade-offs of basis choice and k-point convergence.

Using DFT and Wannier function analysis, the authors characterize the symmetry-protected topology of bulk rhombohedral graphite (RG), showing it is a weak topological insulator with spin-orbit coupling and a chiral nodal-line semimetal without it, and examine drumhead surface states (DSS) on semi-infinite and thick RG slabs against the 'ideal quantum geometry' (IQG) criterion |Ω|/TrG=1 linked to fractional Chern insulator and flat-band superconductivity proposals. They find the DSS flat band exhibits sizable convex curvature with a specific depth dependence, consistent with recent experimental/theoretical claims about RG's role in flat-band physics. This provides a first-principles benchmark for researchers modeling IQG-driven correlated phases (fractional Chern insulators, unconventional superconductivity) in few- and many-layer rhombohedral graphene/graphite systems.

Via reflux synthesis, the authors produced a series of hydrated layered copper selenites, (NH3)2Cu5(SeO3)2(OH)6(H2O)2+x for x=0,1,3, and solved their structures using electron diffraction and single-crystal X-ray diffraction. Each compound shares an identical Cu(OH)4-square/Cu-square-pyramid kagome-like layer, but the degree of interlayer hydration systematically modulates interlayer spacing and stacking sequence while leaving intralayer connectivity unchanged. This establishes hydration state as an independent structural control parameter decoupled from layer composition, offering a model system for researchers studying stacking-dependent properties (e.g., magnetic frustration in kagome lattices) or designing intercalation-based tuning strategies in layered inorganic solids.

Using time- and momentum-resolved ARPES on monolayer WSe2 combined with first-principles excitonic Bloch equation simulations, the authors directly resolve the ultrafast conversion of unbound electron-hole pairs into bound excitons, observing transient coexistence of free-carrier and excitonic spectral bands with measurable spectral weight transfer between them. They identify the formation mechanism as phonon-mediated, occurring via a sequential relaxation pathway rather than instantaneous binding. This provides direct experimental access to exciton formation dynamics previously inferred only indirectly, giving a benchmark dataset and mechanism (phonon-assisted, sequential energy-momentum relaxation) that theorists modeling exciton dynamics in 2D semiconductors, and experimentalists designing excitonic optoelectronic devices, can validate against or exploit for engineering formation timescales.

The authors introduce a performance-oriented benchmarking suite for DMRG implementations, addressing the lack of standardized, apples-to-apples comparisons across the >50 existing codes that differ in language, tensor backends, symmetry handling, and parallelization strategy. The framework specifies consistent test Hamiltonians, bond-dimension/convergence settings, and timing/memory metrics so that runtime and scaling behavior can be directly compared rather than inferred from disparate, incomplete reports. Practitioners can use this to select or tune a DMRG package for their hardware and problem size, and tool developers can use it as a regression/performance test suite when optimizing their own implementations.

Using Langevin dynamics with a coarse-grained bead-spring polymer model and a semi-implicit nanopore, the authors compare voltage-driven translocation of linear versus three-arm star polyelectrolytes of equal molecular weight. Linear polymers show the expected translocation-time inverse-proportionality to voltage across a wide range, while star polymers deviate sharply below a threshold voltage that is insensitive to molecular weight but scales with pore size and salt concentration. Metadynamics is used to reconstruct the free-energy landscape governing the branched polymer's translocation, likely revealing a barrier associated with arm-junction passage through the pore. This has direct implications for interpreting nanopore sensing signals from branched biomolecules or synthetic polymers.

The authors measure wavelength-resolved photo-CIDNP in a tetraphenylporphyrin–1,4-benzoquinone donor-acceptor system across 350–800 nm, using both a tunable laser and a xenon lamp with 10 nm interference filters, with signals normalized to photon flux for direct comparison. The key finding is a non-monotonic wavelength dependence with pronounced polarization enhancements at specific excitation bands, rather than a smooth spectral response, implying wavelength-selective activation of distinct radical-pair spin-chemical mechanisms (e.g., different excited states or intersystem crossing pathways). This spectral map provides a practical guide for choosing excitation wavelengths to optimize hyperpolarization yield and offers mechanistic constraints that theorists can use to test spin-dynamics models of radical pair reactions.

No abstract is available beyond the title, so specifics of the argument, domain (e.g., web sessions, LLM chat context, OS sessions), and any proposed solution can't be confirmed here. Read literally, the title points to the general problem of session-state portability — data or context bound to one runtime, provider, or device that resists export, migration, or reuse elsewhere. Readers interested in the specifics should consult the source directly, as this summary is necessarily general given the missing content.

No abstract is provided, only the title referencing an update to DeepSeek-V4-Flash. 'Flash' variants typically trade some capability for lower latency and cost versus a full-size flagship model; readers should check the linked release notes or model card directly (see the related DeepSeek V4 Flash 0731 item) for concrete changes such as benchmark scores, context length, or pricing.

The linked resource is the Hugging Face model card for deepseek-ai/DeepSeek-V4-Flash-0731, accompanied by a third-party analysis benchmarking its intelligence (likely via standard eval suites), latency/throughput, and API or hosting price relative to competing models. Practitioners evaluating a model swap should check the linked analysis for concrete benchmark scores and price-per-token figures before integrating this model into a production pipeline, since flash-tier models typically target a different cost/latency tradeoff than flagship-tier releases.

The report credits AI-assisted vulnerability discovery and triage for a sharp spike in Chrome security bug fixes in June, reportedly exceeding the combined total of the prior two years. This aligns with a broader industry trend of using LLM-based fuzzing augmentation, automated code review, and AI-assisted patch generation to scale vulnerability research beyond manual analyst throughput. Security teams evaluating similar tooling should look at whether the reported fixes stem from AI-generated fuzz targets, AI-assisted static/dynamic analysis triage, or LLM-drafted patches, as each implies different integration points into an existing SDLC and different verification burdens before merging.

There isn't enough information in the title alone to characterize the mechanism, architecture, or claimed results — "qm" could plausibly abbreviate anything from a query manager to a quantum-mechanics simulator to a queue/messaging tool. A practitioner would need to open the linked source or README to assess the implementation, dependencies, and how to integrate or extend it. Treat any specific claim about its function as speculative until verified against the primary source.

The piece is a postmortem-style analysis arguing that Tailscale's mesh VPN (which authenticates devices and encrypts node-to-node traffic) addresses network-perimeter access but doesn't mitigate threats that originate from valid credentials, compromised endpoints already admitted to the tailnet, or vulnerabilities in applications running behind it. For practitioners, the takeaway is to pair zero-trust networking with defense-in-depth: strong secrets management, least-privilege access controls, endpoint hardening, and application-layer monitoring, since a mesh VPN narrows the attack surface but doesn't eliminate it.

No concrete claims, methods, or data are given in the title, so any technical summary would be speculative. If this turns out to be about a specific technology, product, or organizational shift, the value for a technical reader would lie in understanding what's being deprecated or superseded and what replaces it — but that requires reading the source material directly.

The likely thesis is that generative models converge on a shared aesthetic because they're trained on overlapping datasets and optimized against similar objectives (e.g., aesthetic-score reward models, diffusion guidance scales tuned for perceptual appeal), which biases outputs toward certain compositional and tonal patterns regardless of underlying architecture. For practitioners, this connects to ongoing discussion about dataset bias, reward-model overfitting in RLHF/RLAIF pipelines, and how prompt engineering or fine-tuning can push outputs away from this default "look."

Likely an opinion piece examining how velocity-as-virtue manifests in engineering and product culture — rapid iteration cycles, minimal viable products, move-fast-and-break-things norms, and the current AI arms race — and arguing these are treated as unquestionable goods rather than tradeoffs against stability, safety, or thoughtfulness. A technical reader might connect this to real debates around release cadence versus reliability, or the pressure to ship AI features ahead of adequate safety testing.

The account is a firsthand report of flagging suspected fabricated authorship — a hallmark of paper-mill-generated submissions — to conference organizers, only to see both papers pass review and get accepted as oral presentations, a track typically reserved for higher-scrutiny, higher-visibility work. This points to gaps in conference-level vetting processes (e.g., lack of author identity verification, reviewer overload, or insufficient escalation paths for fraud reports) and adds to a growing body of evidence about paper-mill infiltration of peer-reviewed venues, relevant to anyone building or evaluating academic integrity tooling.

This reflects a self-imposed governance decision by Premier League clubs to phase out front-of-shirt gambling sponsorships, following criticism from public health bodies and gambling-harm campaigners over normalized exposure via matchday broadcasts reaching millions of viewers, including minors. The change is notable as an industry-led move rather than legislated restriction, and sits alongside broader UK regulatory tightening on gambling advertising; the precise scope (front-of-shirt only vs. all sponsorship channels) determines how much commercial impact clubs actually absorb.

JEP 401 lets developers declare "value classes" whose instances lack object identity (no identity-based synchronization, no reference-based equality guarantees), which frees the JVM to eliminate header overhead, avoid heap indirection, and flatten these instances directly into arrays and containing objects rather than storing them as separate heap-allocated pointers. This is the latest deliverable of Project Valhalla, aimed at closing the performance gap between user-defined types and Java's primitives; merging to OpenJDK master as a preview feature means it will ship for community testing in an upcoming JDK release, and practitioners building performance-sensitive code (numerics, graphics, data-heavy structures) can start experimenting with value classes and reporting feedback ahead of finalization.

The write-up reconstructs three documented security incidents and cross-checks them against an existing cybersecurity evaluation suite to assess how predictive the benchmarks are of real-world attacker behavior. The likely methodology involves timeline reconstruction of each incident's techniques and mapping them onto benchmark task categories to find coverage gaps. The practical payoff is a template for incident-driven validation of security evals — a way to catch blind spots that synthetic red-teaming scenarios miss. Teams building their own AI security evaluations could adopt a similar retrospective-incident methodology to pressure-test benchmark validity.

No abstract or supporting text was provided for this entry, only the title 'Severance.' There isn't enough signal here to characterize a method, claim, or result without fabricating detail that isn't supported by the source. To evaluate or build on this item, the original link or document would need to be reviewed directly.

Gander is an Android file viewer that requests zero permissions, including INTERNET, so the OS enforces at the platform level that no file data can be transmitted off-device — a stronger guarantee than a no-logging policy. PDF rendering uses Pdfium, media playback uses Media3, and Office formats (Word/Excel/PowerPoint) are rendered by bundled JavaScript libraries running inside a local WebView instance, avoiding any external rendering service. Known limitations: complex PowerPoint layouts render only approximately, spreadsheet charts aren't drawn, and legacy binary .doc/.ppt formats have partial support. This is a useful reference architecture for any app wanting to prove a security property via missing permissions rather than policy.

Only a title, 'Big Food vs. the People,' is available, without an abstract detailing the specific claims, methodology, or scope. This likely falls into food-policy or public-health journalism/analysis territory, but no concrete mechanism, data, or argument can be responsibly summarized without the source text.

The team distilled DeepSeek V4 Flash (teacher) into GPT-OSS-120B (student) for finance-reasoning tasks, achieving 83.61% on FinanceReasoning at an 8k token budget, beating Kimi K3 (81.93%) and Inkling (65.13%); a 20B version has been released as open weights. To test whether the teacher's politically sensitive-topic behavior transferred, they measured response divergence on political questions and found the teacher's answers differed from expected baselines by 7 standard deviations, yet the distilled model's behavior remained statistically indistinguishable from its original American base model. This suggests distillation on a narrow task domain (finance) doesn't generalize censorship/alignment behavior outside that domain — a useful data point for teams considering cross-lab distillation pipelines who are concerned about inheriting unwanted model behaviors. A no-auth playground is available to replicate query comparisons directly.

This is a novelty project in the well-established 'Bad Apple but it's X' genre, here implemented via traceroute — likely encoding video frame data into hop sequences or router hostnames/addresses so that running traceroute against a crafted set of targets reproduces the video's imagery in the terminal output. Without further detail on implementation, the likely mechanism involves a controlled set of intermediate hosts or DNS-based hostname encoding to control what each hop displays. It's a fun reference for anyone interested in creative abuse of network protocols for output rendering, similar to esolang or 'weird machine' style projects.

The piece is framed as a historical analogy between 1990s cryptography export controls (the 'crypto wars,' e.g. U.S. munitions-list restrictions on strong encryption) and current-day export controls or governance proposals around AI model weights. Without the full text, the likely argument concerns whether restricting distribution of a a piece of information/software (encryption algorithms then, model weight files now) is technically enforceable given that both are ultimately just data that can be copied and distributed. This is relevant to practitioners tracking AI export control policy (e.g. U.S. BIS rules on frontier model weights) who want historical precedent for how similar restrictions played out for cryptographic software.

The post demonstrates running Kimi K3, a large-scale model, within a 29 GB RAM budget — well below its typical deployment footprint — almost certainly via aggressive quantization and/or disk/SSD offloading of model weights rather than keeping everything resident in RAM or VRAM. The tradeoff is stark: throughput drops to roughly 0.50 tokens/second, illustrating the practical floor of consumer-hardware inference for frontier-scale MoE or dense models. This is useful as a proof-of-concept for practitioners exploring extreme memory-constrained local inference setups (e.g. llama.cpp-style offloading), though it's far too slow for interactive use and mainly serves as a feasibility/benchmark data point.

The index/fund tracking 'situational-awareness'-themed AI equities fell 67% in July, consistent with a broader correction in richly-valued AI infrastructure and application names. Drawdowns like this typically reflect multiple-compression (falling P/E or P/S ratios) rather than a change in reported fundamentals, often amplified by leveraged or thematic ETF flows. Anyone tracking AI-stock baskets should watch for correlation breakdowns between compute/model-progress narratives and equity pricing, and for concentration risk inherent to thematic indices.

A counterexample refuting the Maxwell Conjecture has reportedly been produced, with the discovery credited to an AI model. As with other AI-assisted disproofs, the value lies in the search process, an AI exploring or generating candidate structures far faster than manual case analysis, with a human or automated verifier then confirming the counterexample satisfies the conjecture's hypotheses while violating its conclusion. Replicating or building on this would require the precise formal statement of the conjecture and the produced counterexample for independent verification.

Arch Linux has disabled the AUR's package-adoption workflow, which previously let any registered user claim orphaned packages and push new PKGBUILD updates. This closes a takeover vector where an attacker adopts a stale but popular package and injects malicious build or install scripts, analogous to typosquatting and repo-hijack attacks seen in npm/PyPI. Maintainers and downstream users should audit recently-adopted packages, verify PKGBUILD provenance, and expect a revised, likely more vetted, ownership-transfer process from the Arch team going forward.

Record-low Danube water levels have forced a reduction or shutdown at Hungary's Paks Nuclear Power Plant, which relies on river water for reactor cooling and condenser heat rejection. Low flow combined with elevated intake temperatures reduces the thermal margin needed for safe cooling, forcing plants to derate or trip units under regulatory thermal-discharge and safety limits. This highlights a growing grid-resilience concern for river-cooled thermal plants broadly: cooling-water availability is becoming a climate-exposed constraint on baseload capacity, relevant to plant siting, cooling-tower retrofits, and drought contingency planning.

The post details achieving 25 Gbps throughput over Thunderbolt on a Mac Studio, exceeding the common 10GbE ceiling by exploiting Thunderbolt's native ~40 Gbps link (or Thunderbolt-to-QSFP/SFP28 adapters) alongside macOS's Thunderbolt networking/bridging support rather than a standard Ethernet NIC. Hitting real-world 25 Gbps typically requires the right adapter/cable choice plus MTU and jumbo-frame tuning, validated with tools like iperf3 to rule out CPU or driver bottlenecks. Anyone with equivalent Thunderbolt 4/5 hardware could replicate this for high-speed workstation interconnects, such as NAS access or ML data transfer, without a full 25GbE switch upgrade.

This describes a certified reference-grade water standard, likely a national-metrology-institute Standard Reference Material, priced around $120,000/gallon due to the cost of achieving and verifying extreme purity (contaminants controlled to parts-per-trillion) plus the certification and traceability chain required for it to serve as a calibration standard. Such materials underpin measurement traceability for trace-metal analytics, isotope-ratio analysis, and instrument calibration across analytical chemistry labs. The steep price reflects that its value comes entirely from certified purity and documented chain-of-custody, not from any inherent property of the water.

The proposal targets Go's `container/` packages (list, heap, ring), pre-generics APIs that rely on `interface{}`/`any` and runtime type assertions, and would introduce or refactor generic collection types using Go's type parameters (available since Go 1.18). This would give developers compile-time type safety and eliminate runtime casts for structures like doubly-linked lists, heaps, and rings. As with prior generics-adoption efforts (e.g., the `slices` and `maps` packages), expect a phased design-review process on the Go issue tracker before it lands in the standard library; developers can prototype equivalent generic containers today to anticipate the eventual API.

The piece traces industry-funded, Red Bull-linked studies on caffeine safety and energy-drink consumption that were subsequently cited in regulatory risk assessments or industry safety guidance, a funding-conflict pattern familiar from tobacco, sugar, and pharma research. Likely concerns include selective outcome reporting, industry-favorable study design, and citation laundering, where sponsored papers appear in the literature as if independent. For anyone evaluating nutrition or food-safety research, the actionable takeaway is to check funding disclosures and author conflicts-of-interest before treating a cited study as independent evidence in policy debates.

MarbleOS is a proposed GUI paradigm for AI agent interaction, drawing an explicit analogy between today's natural-language agent interfaces and pre-GUI command-line computing. The core critique: natural language removed strict syntax but preserved the underlying UX problem of recall-dependent interaction — users must know and articulate what's possible rather than discovering it through direct manipulation. The implied design direction follows the Xerox PARC/Mac/NeXTSTEP lineage — surfacing agent capabilities, state, and available actions as visible, manipulable UI elements rather than a blank prompt box. As a Show HN post this is a product pitch rather than a technical paper, so no architecture or implementation details are given here; a practitioner would need to check the actual product/demo for specifics.

This is an interactive walkthrough of the JPEG lossy compression pipeline: color-space conversion (RGB to YCbCr), chroma subsampling, 8x8 block-based discrete cosine transform (DCT), quantization against a quality-dependent table, and entropy coding. The interactivity lets users manipulate parameters like quality level and observe the resulting artifacts directly on sample images, making the frequency-domain effects of quantization (blockiness, ringing) tangible rather than theoretical. It's a solid reference for anyone implementing or debugging an image codec, tuning JPEG quality settings for a product, or building intuition for DCT-based compression before tackling video codecs that reuse the same building blocks.

The piece walks through debugging a real latency issue by moving from a single summary statistic (mean) to distributional visualizations — likely histograms or percentile plots (p50/p95/p99) — to expose multimodal or heavy-tailed latency behavior the mean obscures. The key technical point is that means are not sufficient statistics for skewed or multimodal distributions common in systems performance data, such as bimodal cache hit/miss latency or GC pauses, so aggregate dashboards built on averages can read 'green' while a real subpopulation of requests is degraded. Practitioners can apply this directly by swapping mean-based alerting and dashboards for percentile- or histogram-based views when triaging performance regressions.

According to the linked report, Moonshot AI's Kimi models are trained and/or served on a roughly 20,000-GPU Nvidia cluster provisioned via Alibaba Cloud rather than self-owned infrastructure. This is notable given U.S. export controls on advanced AI chips to China, which have pushed domestic labs toward cloud-rental arrangements with hyperscalers like Alibaba that have accumulated or gained access to large GPU fleets. For anyone tracking compute supply chains or competitive positioning among Chinese AI labs, this is a data point on scale and infrastructure strategy rather than a methods disclosure — the abstract gives no model architecture or training details.