Daily Digest — 2026-07-10
250 items · 5 research labs, 237 arxiv papers, 8 industry media
🏛️ Research Labs (5)
GPT-5.6 is now the preferred model in Microsoft 365 Copilot
OpenAI has deployed GPT-5.6 as the preferred model in Microsoft 365 Copilot, enhancing productivity tools like Word, Excel, PowerPoint, and Cowork. The model improves token efficiency and task performance, enabling users to generate higher-quality outputs with reduced manual effort. Integration occurs both natively and via OpenAI's API, with Microsoft reporting improved document drafting, data analysis, presentation creation, and cross-functional collaboration capabilities.
gpt-5.6microsoft 365 copilottoken efficiencyapi integrationproductivity tools
ChatGPT is now a partner for your most ambitious work
OpenAI introduces ChatGPT Work, an agentic system powered by GPT-5.6 that autonomously executes multi-step workflows across applications. The system integrates Codex technology for task automation, leveraging in-context learning to process files (docs, slides), manage scheduled tasks, and maintain state over extended durations. Early testing shows 100% adoption internally, reducing sales POC timelines from weeks to 24 hours and finance workflows from days to hours. The architecture supports plugin-based app integration, enterprise controls via Compliance API, and auto-review for data protection. Deployment begins with Pro/Enterprise tiers, expanding to all plans.
agentic systemgpt-5.6codexin-context learningcompliance api
GPT-5.5 Bio Bug Bounty
OpenAI expands its Bio Bug Bounty Program to a continuous private initiative targeting universal jailbreaks in biological risk scenarios, focusing on frontier models beginning with GPT-5.6. The program doubles rewards to $50,000 for successful universal jailbreaks against GPT-5.5 and GPT-5.6, with partial findings eligible for smaller awards. Testing for GPT-5.5 concludes July 27, 2026; applicants undergo a rolling selection process requiring NDAs and existing ChatGPT accounts to participate in biosafety challenge evaluations.
universal jailbreaksbiosafety challengefrontier modelsbug bountyndas
GPT-5.6: Frontier intelligence that scales with your ambition
OpenAI introduces GPT-5.6, a family of models (Sol, Terra, Luna) achieving state-of-the-art performance in coding, knowledge work, cybersecurity, and science with improved efficiency. Sol outperforms Claude Fable 5 by 13.1 points on Agents’ Last Exam (53.6 vs. 40.5) at lower cost, while Terra and Luna match Fable 5 at 1/16th the cost. Key innovations include Programmatic Tool Calling for reduced token usage, multi-agent coordination ('ultra' mode), and enhanced design judgment. Sol sets new benchmarks on Terminal-Bench 2.1 (80.0) and ExploitBench2 (73.5%). Safeguards combine real-time monitoring and adaptive access controls for dual-use domains like cybersecurity.
programmatic tool callingmulti-agent coordinationpareto improvementskv-cache optimizationdual-use safeguards
Our approach to government and national security partnerships
OpenAI introduces National Security Principles to govern its partnerships with governments and national security entities, emphasizing democratic accountability and human oversight. The framework was developed through cross-company collaboration, expert consultation with national security specialist David Kris, and employee engagement across research, safety, and policy teams. Key initiatives include the Daybreak cyber defense program, partnerships with nine countries and EU institutions, and controlled access to GPT-Rosalind for biosecurity applications. Contractual restrictions prohibit mass surveillance, autonomous weapons, and high-stakes automated decisions, aligning with broader legislative efforts for AI safeguards in national security contexts.
national security principlesgpt-rosalinddaybreak cyber defensedemocratic accountabilityautonomous weapons systems
📜 arXiv Papers (237)
Accurate, Interdisciplinary and Transparent Structure-property Understanding with Deep Native Structural Reasoning
SciReasoner introduces a multimodal foundation model for native structural reasoning across proteins, small molecules, and inorganic crystals, addressing the joint challenge of representation and reasoning in structure-property relationships. The model discretizes structural information into a unified vocabulary, treating structural tokens as addressable evidence units during reasoning. It achieves state-of-the-art performance on 67 of 86 benchmarks, including improving Cellular Component annotation $F_{\max}$ from 0.42 to 0.55 in low-homology proteins and increasing single-step retrosynthesis accuracy from 0.63 to 0.72, with expert evaluations preferring its reasoning traces in 98% of cases.
structure-property relationshipsmultimodal foundation modelnative structural reasoningaddressable evidence unitsscientific constraints
Co-LMLM: Continuous-Query Limited Memory Language Models
The paper introduces Continuous-Query Limited Memory Language Models (CO-LMLM), which externalize factual knowledge to a vector-keyed knowledge base (KB) instead of relational KBs used in prior LMLMs. The method generates continuous vector queries during inference and employs a novel annotation pipeline to extract factual spans from arbitrary text, eliminating Wikipedia dependency. Evaluated on Wikipedia and FineWeb-Edu at multiple scales, CO-LMLM achieves superior perplexity and factual precision, with a 360M parameter model outperforming larger models trained on 40x more data and matching gpt-4o-mini on SimpleQA verification.
limited memory language modelsknowledge base retrievalcontinuous vector queriesfactual precisionperplexity optimization
Breaking Database Lock-in: Agentic Regeneration of High Performance Storage Readers for Database Bypass
Jailbreak introduces an LLM-assisted method to bypass database engines by directly reading storage files and materializing data as in-memory columnar buffers. The approach leverages code synthesis from database documentation and source code to regenerate operator-specific table readers, eliminating the need for human-engineered parsing logic. Evaluated on PostgreSQL and MySQL using TPC-H, Jailbreak achieves up to 27x speedups in analytical throughput while maintaining correctness against JDBC/ODBC baselines, demonstrating viability for breaking data lock-in across systems with available file formats.
database bypassllm-assisted synthesiscolumnar buffersstorage decodinganalytical throughput
Institutional Red-Teaming: Deployment Rules, Not Just Models, Causally Shape Multi-Agent AI Safety
The paper introduces institutional red-teaming, a methodology for evaluating deployment rules in multi-agent AI systems by varying a single rule while holding other factors constant. The approach is instantiated in IABench-CA, a benchmark comprising 228 contexts, five canonical rules, and seven model populations (33,924 games), with cooperative references and auto-labelled reasoning traces. Key findings include: (1) deployment rules causally impact collective safety, altering fatality rates by 22-58 percentage points; (2) no universally safe default exists, with identity-targeting being consistently unsafe; (3) identity salience drives targeted elimination, as anonymization delays but does not prevent exploitation. The methodology is packaged as a safety-case workflow for provisional rule certification.
institutional red-teamingmulti-agent aideployment rulescollective safetyidentity salience
Selective Timestep Weighting and Advantage-Based Replay for Sample-Efficient Diffusion RLHF
The paper introduces two strategies to improve feedback efficiency in reinforcement learning from human feedback (RLHF) for diffusion models: selective timestep weighting and advantage-based replay. The method addresses uneven reward distribution in diffusion trajectories by reweighting denoising steps during policy optimization and prioritizing informative trajectories for replay. Experiments show a 6× improvement in sample efficiency over baselines under identical hyperparameters.
diffusion modelsrlhfsample efficiencyproximal policy optimizationdenoising timesteps
Agon: Competitive Cross-Model RL with Implicit Rival Grading of Reasoning
Agon introduces competitive cross-model reinforcement learning where two models alternately grade each other's reasoning during problem-solving, eliminating the need for explicit process labels or reward models. The method pairs comparably strong but behaviorally distinct models that optimize against each other, creating progressively stronger adversaries. Evaluated on DeepMath's hard split with Qwen3, Agon doubles GRPO's pass@1 performance, showing consistent gains across competitive programming tasks and model families (Qwen3.5, Gemma 4). Future work extends reasoning to latent-space communication.
competitive reinforcement learningcross-model gradingimplicit reasoning evaluationprogressive adversarial traininglatent-space reasoning
SkillCenter: A Large-Scale Source-Grounded Skill Library for Autonomous AI Agents
SkillCenter introduces the largest open skill library for autonomous AI agents, comprising 216,938 structured skills across 24 domain bundles. The framework combines 114,565 source-grounded skills from peer-reviewed journals and technical sources with 102,373 community skills from GitHub and ClawHub. A SkillGate-filtered pipeline ensures quality through multi-source acquisition, LLM-based quality gating, template-driven generation, and iterative source-grounding, with traceability guarantees linking claims to exact source quotations. All skills are packaged as offline-searchable SQLite FTS5 bundles.
autonomous agentsskill librarysource groundingllm-based quality gatesqlite fts5
DiaLLM: An Investigation into the Robustness-Generation Gap in English Dialect Adaptation
DiaLLM investigates the robustness-generation gap in English dialect adaptation by continually pretraining three open-weight language model families on the International Corpus of English and applying implicit and explicit post-training paradigms combined with three alignment strategies. The study focuses on Australian, Indian, and Northern British English, revealing that dialectal robustness and generation are dissociated: benchmarks are influenced by continual pretraining and supervised fine-tuning (SFT), while alignment reshapes generation in ways benchmarks do not capture. Explicit variety-targeted adaptation produces dialect-recognized output preferred over broad alignment, yet aggressive optimization of dialectal reward is not favored by human evaluators. Linguistic analysis corroborates the reward-quality gap, particularly in two model families. The study releases code, checkpoints, and preference datasets.
dialectal robustnessalignment strategiessupervised fine-tuningreward-quality gapinternational corpus of english
Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops
The paper proposes a taxonomy for recursive self-improvement (RSI) in AI systems, analyzing 1,250 arXiv papers (2024-2026) along two axes: improvement targets (behavior, policy, evaluator, research process) and loop closure (human-in-the-loop to fully autonomous). It distinguishes bounded self-refinement from open-ended RSI, highlighting self-evaluation as a critical category where improvement loops implicitly substitute signals for human judgment. Results show self-improvement strength correlates with verification hierarchy strength (formal verifiers to intrinsic self-assessment), with failure modes (self-confirming loops, collapse dynamics) emerging from hierarchy violations. The study identifies governance-grade measurement as an underdeveloped research area.
recursive self-improvementself-evaluationverification hierarchymodel collapseloop closure
RL Post-Training Builds Compositional Reasoning Strategies
The study demonstrates that RL post-training enables compositional reasoning strategies beyond primitive skills latent in pretrained models. Using a Transformer pretrained on primitive symbol-rewrite chains and post-trained with binary final-answer rewards in a Trace-based reasoning task, RL solves held-out problems rarely addressed by pretrained models. Trace analysis reveals RL reorganizes primitive competence into phased compositional mechanisms, including sequential and parallel compositions, which are reused and consolidated. Compared to rejection fine-tuning, RL exhibits superior selectivity, concentrating exploration into valid reusable structures. Pretraining ablations indicate compositional strategies emerge when primitive competence is organized into reduction procedures.
rl post-trainingcompositional reasoningtransformertrace-based reasoningrejection fine-tuning
ALER-TI: Aligned Latent Embedding Retrieval for Time Series Imputation
ALER-TI introduces a retrieval-augmented framework for time series imputation that supplements local context with historical patterns via Latent Embedding Alignment (LEA). LEA aligns corrupted queries with complete historical candidates through post-hoc masking in latent space, enabling efficient pre-computation and retrieval. The model-agnostic approach, evaluated on six real-world datasets, consistently improves baseline performance and robustness across varying missing rates.
time series imputationlatent embedding alignmentretrieval-augmented frameworknon-stationary dynamicspost-hoc masking
QCNN with Rough Path Signature Kernels
The authors propose a hybrid quantum-classical architecture for time series classification that combines quantum neural networks with path signature kernels to address time reparameterization invariance. The method employs feature layers computing signature kernels between input paths using classical or quantum variational linear solvers (VQLS), followed by a Quantum Convolutional Neural Network (QCNN) for downstream tasks. Evaluated on binary classification of handwritten digit time series, the architecture demonstrates potential advantages of quantum path signature kernels while highlighting computational limitations of VQLS.
quantum convolutional neural networkpath signature kernelstime reparameterization invariancevariational quantum linear solvertime series classification
Future Confidence Distillation in Large Language Models
The paper introduces future confidence distillation, a method for improving confidence estimation in large language models (LLMs) by leveraging temporal dynamics between pre-solution (Feeling-of-Knowing) and post-solution (Judgement-of-Learning) confidence signals. The authors demonstrate that post-solution confidence is better calibrated and more discriminative, and propose training linear probes on pre-solution hidden representations using post-solution correctness as supervision. Results show distilled predictors achieve calibration improvements comparable to post-solution confidence while remaining sample-efficient and transferable within domains, enabling low-cost reliable confidence estimation.
confidence estimationlarge language modelsfeeling-of-knowingjudgement-of-learninglinear probes
Towards Agentic AI Governance: A Preliminary Assessment
The paper conducts a systematic review of emerging literature on agentic AI governance, addressing ethical and regulatory challenges posed by autonomous AI systems. It analyzes distinguishing features of agentic AI that necessitate specialized governance frameworks, synthesizes current governance priorities and proposed mechanisms, and identifies key stakeholder roles. The review establishes foundational groundwork for developing structured governance roadmaps to ensure responsible deployment of agentic AI systems.
agentic aiautonomous systemsgovernance frameworksethical challengesregulatory mechanisms
CARLA-GS: Decoupling Representation, Reasoning, and Physics Simulation for Autonomous Driving Corner-Case Synthesis
CARLA-GS introduces a modular pipeline for photorealistic corner-case synthesis in autonomous driving, decoupling visual representation (Gaussian scene reconstruction), semantic reasoning (multi-agent LLM), and physics-based execution (CARLA/PID control). The framework reconstructs editable scenes from real data, uses LLMs for intent-level trajectory generation, and ensures kinematic feasibility through CARLA simulation before re-projecting states for ego-centric rendering. Evaluations on Waymo Open Dataset demonstrate spatiotemporally consistent, physically feasible corner cases with controllable semantic intent.
gaussian scene reconstructioncorner-case synthesismulti-agent llmphysics simulationautonomous driving
Collaborative Synthetic Data Generation for Knowledge Transfer in Federated Learning
FedKT-CSD introduces a federated learning framework for one-shot knowledge transfer with formal privacy guarantees, addressing communication overhead and data heterogeneity. The method leverages pretrained autoencoders to encode client data into class-conditional latent statistics, aggregates them via secure aggregation with differential privacy noise, and decodes a synthetic dataset for global model training. Results show competitive performance with non-private baselines across diverse datasets and client counts, while maintaining lightweight client-side computation and communication.
federated learningdifferential privacysynthetic dataknowledge transferautoencoder
Creativity from Friction: Human-AI Interaction for Exploratory Structural Design
The paper proposes interactive AI systems for structural design that preserve creative friction while reducing repetitive tasks, contrasting with conventional generative AI approaches focused on final outputs. The authors develop design principles for vision-language model-based interfaces enabling conversational, multimodal exploration of structural constraints (spatial, mechanical, material). A pilot interface and expert study demonstrate how such systems can align with designers' iterative workflows by maintaining reflective friction during exploration while automating routine modeling tasks.
human-ai interactionstructural designvision-language modelsdesign constraintsiterative workflows
Stability of Flow Models for Graph Signals
The paper establishes stability bounds for continuous normalized flow models parameterized by Graph Neural Networks (GNNs) when generating graph signals under structural perturbations. It proves permutation equivariance for both continuous-time ODEs and their discrete approximations, then introduces a stability-promoting regularized flow matching strategy that penalizes the spatial Lipschitz constant of the vector field during training. Experiments on synthetic stochastic block model graphs and real-world fMRI brain connectomes demonstrate improved robustness to structural noise without degrading output quality.
graph neural networksnormalized flowpermutation equivariancestructural perturbationslipschitz constant
Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning
The paper introduces Single-rollout Asynchronous Optimization (SAO), a method addressing stability and off-policy challenges in asynchronous reinforcement learning (RL) for agentic tasks. SAO replaces group-wise sampling with single-rollout sampling (one rollout per prompt) and employs a strict token-level clipping strategy for optimization stability. Evaluated on SWE-Bench Verified, BeyondAIME, and IMOAnswerBench, SAO outperforms GRPO variants and demonstrates effectiveness in online learning settings. The method was deployed in training the GLM-5.2 model (750B-A40B).
asynchronous reinforcement learningsingle-rollout samplingoff-policy correctiontoken-level clippingagentic rl
HIVE: Understanding Post-Hallucination Reasoning in Vision Language Models
This work introduces HIVE (Hallucination Inference and Verification Engine), a framework for studying Post-Hallucination Reasoning (PHR) in Vision Language Models (VLMs), where hallucinated semantics influence downstream predictions. HIVE enables controlled comparisons between faithful and hallucinated captions across nine tasks and nine models. Results reveal modality-dependent patterns: hallucinated captions often enhance accuracy in vision-language tasks while showing limited effects in text-only tasks. Analyses indicate that hallucinated cues broaden semantic coverage and reshape reasoning dynamics while maintaining stable inference. These findings underscore the importance of understanding PHR for improving multimodal reasoning systems' reliability and interpretability.
post-hallucination reasoningvision language modelssemantic coveragemultimodal reasoninghallucination inference
Do LLM-Generated Skills Make Better AI Data Scientists? A Component Ablation Across Data-Science Workflows
The study evaluates whether LLM-generated skills improve performance in data-science workflows compared to no-skill prompting, focusing on four lifecycle stages: data preparation, extraction, statistical analysis, and reporting. Through component ablation across 56 tasks, nine model configurations, and three providers (7,560 runs), the authors find no significant improvement from generated skills or their variants (p-values ≥ 0.396, performance spread ≤ 1.2 pp). A token-matched control (1,512 additional runs) confirms that full skills perform similarly to irrelevant skill-formatted content, cautioning against their default use in single-shot prompting.
llm-generated skillsdata-science workflowscomponent ablationsingle-shot promptingtoken-matched control
TimEE: End-to-end Time Series Classification via In-Context Learning
TimEE introduces a 4.5M-parameter foundation model for end-to-end time series classification (TSC) via in-context learning, eliminating per-dataset training. The model, meta-trained exclusively on synthetic TSC tasks with structured distributional shifts, directly predicts class distributions given a labeled support set and query time series. On the UCR benchmark, TimEE achieves state-of-the-art ROC AUC and ranks third in accuracy, outperforming both foundation models and supervised deep learning baselines despite synthetic-only pretraining.
time series classificationin-context learningfoundation modelmeta-trainingsynthetic pretraining
Reward-Adaptive Iterative Discovery: A Case Study on Automated Game Testing for NHL26
The paper introduces Reward-Adaptive Iterative Discovery (RAID), a reinforcement learning method for automated game testing that discovers diverse behavioral exploits in AI systems. RAID extends existing RL algorithms by training a population of goal-scoring agents to identify multiple high-quality solutions, addressing the overfitting problem in single-solution approaches. In a case study on EA SPORTS NHL 26's goalie AI, RAID identified six distinct scoring exploits within a single experiment, matching strategies previously found through hours of manual playtesting.
reinforcement learningautomated testingbehavioral exploitsgoal-scoring agentsoverfitting
Search, Fail, Recover: A Training Framework for Correction-Aware Reasoning
The paper introduces Pyligent, a training framework for correction-aware reasoning that models problem-solving as validated search over partial solution chains. The method employs a task validator to label continuations and failures, converting search trees into supervised targets for three actions (continue, finish, backtrack) with optional traces of abandoned branches. Evaluations on hidden directed graphs, 4×4 Sudoku, and Blocksworld show solve rate improvements of up to 72.7 percentage points over gold-only fine-tuning, demonstrating the value of failed-branch supervision for recovery behavior.
correction-aware reasoningvalidated searchpartial solution chainsdelayed-failure recoverysupervised targets
Beyond Attack-Success Rate: Action-Graded Severity Scale for Tool-Using AI Agents
The paper introduces an action-graded harm rubric for tool-using AI agents, addressing the limitation of binary attack-success metrics by scoring trajectories on a seven-level ordinal scale (L0–L6) based on reversibility, scope crossing, and privilege escalation. The method employs both a deterministic oracle and a three-judge panel of frontier language models to evaluate trajectories, validated on four victim models and two defenses in the AgentDojo workspace suite. Results show high ordinal agreement (α=0.91) between judges and oracle, while exposing cases where binary metrics fail, such as cross-scope leaks. The rubric is released with code, prompts, and logs.
harm rubrictool-using agentsordinal scalered-teamingseverity grading
Where to Intervene? Benchmarking Fairness-Aware Learning on Differentially Private Synthetic Tabular Data
This work presents the first systematic benchmark of fairness interventions on differentially private synthetic tabular data, focusing on the Adaptive Iterative Mechanism (AIM) as the state-of-the-art DP synthesizer. The study evaluates four pipeline configurations (Baseline, DP-only, Fair-only, DP+Fair) across four datasets, multiple group fairness metrics, and three mitigation strategies (pre-processing, in-processing, post-processing) under varying privacy budgets. Results show that DP degrades utility and fairness, but fairness interventions, particularly post-processing methods, can restore equitable outcomes while maintaining competitive utility. All code and data are released for reproducibility.
differential privacyfairness-aware learningadaptive iterative mechanismtabular data synthesisprivacy-fairness trade-off
SynthAVE: Scalable Synthetic Labeling for E-Commerce with LLM-Arena Validation
SynthAVE introduces a scalable synthetic labeling benchmark for e-commerce attribute extraction, covering 12,726 products across 229 types, 792 attributes, and 4 languages. The method employs a multi-LLM arena framework with 21 judge configurations (7 model families × 3 prompts), using majority voting for label validation. Results show high reliability, with majority vote achieving Cohen's κ = 0.92 (95.2% human agreement) and inter-model Fleiss' κ = 0.76, demonstrating cost-effective quality parity with human review.
attribute extractionsynthetic labelingmulti-llm arenamajority votingcohen's kappa
SpaCellAgent: A Self-Evolving LLM-Based Multi-Agent Framework for Trajectory Analysis
SpaCellAgent introduces a self-evolving LLM-based multi-agent framework for automating trajectory inference (TI) in spatial and single-cell transcriptomics. The system employs a multi-agent architecture for workflow planning, a dynamic tool-orchestration engine for adaptive algorithm selection, and a self-evolution module for iterative performance refinement. Evaluated on six heterogeneous datasets with complex temporal trajectories and diverse sequencing platforms, SpaCellAgent achieves over 40% improvement in analytical efficiency while maintaining expert-aligned accuracy. By converting natural language specifications into optimized workflows, it democratizes advanced spatiotemporal modeling and establishes a scalable agent-driven paradigm in computational biology.
trajectory inferencesingle-cell transcriptomicsmulti-agent architecturetool-orchestration engineself-evolution module
The Blind Curator: How a Biased Judge Silently Disables Skill Retirement in Self-Evolving Agents
The paper demonstrates that biased LLM judges silently disable skill retirement in self-evolving agents when false-pass bias exceeds a critical threshold, rendering contribution-based retirement ineffective regardless of data volume. Through corrupted-reward analysis and behavioral experiments on a reference-free report-writing testbed with code-generation cross-checks, the authors show that symmetric noise preserves retirement, while false-pass bias causes universal mechanism failure across domains. The study identifies this as a behavioral safety issue, proposing a defect-injection audit to preemptively diagnose judge bias before deployment.
skill retirementfalse-pass biasself-evolving agentscorrupted-reward analysisreference-free evaluation
RLVP: Penalize the Path, Reward the Outcome
The paper introduces RLVP (Reinforcement Learning with Verifiable Penalties), a method addressing two key challenges in real-world agent deployment: path-dependent constraints and sample efficiency. The approach combines outcome-based rewards with verifiable path penalties, avoiding constraint violations that occur in pure outcome-based RL. The method leverages the insight that path penalties provide reliable variance signals where outcome rewards fail, particularly in sparse-reward environments. Experiments show RLVP achieves high task success with near-zero constraint violations, compared to frequent violations in outcome-only training. The authors provide four design rules for effective penalty implementation, including mitigation of inaction traps.
reinforcement learningverifiable penaltiessample efficiencyconstraint violationpath-dependent constraints
InductWave: Inductive Multi-Hop Logical Query Answering on Knowledge Graphs
InductWave introduces a wavelet-based inductive embedding method for multi-hop logical query answering on knowledge graphs, addressing the transductive limitation of existing approaches. The method operates with fewer message-passing layers (50-75% of baselines) while maintaining performance, enabling scalability to large KGs like Wiki-KG. Evaluations on FB15k-(237) show superior performance across varying train-test graph proportions compared to state-of-the-art models.
inductive reasoningknowledge graphswavelet embeddingsmulti-hop querieslogical query answering
Reason Less, Verify More: Deterministic Gates Recover a Silent Policy-Violation Failure Mode in Tool-Using LLM Agents
The study identifies a silent policy-violation failure mode in tool-using LLM agents, where agents execute forbidden state transitions without detection, as demonstrated in the $τ^2$-bench airline domain (78% of failures). A lightweight intervention using deterministic, read-only pre-execution gates improves task success from 29.6% to 42.0% (+12.4pp; P=0.0012) by preventing policy-violating writes. The effect is concentrated on tasks where gates fire (+19.2pp), with no significant impact on non-firing tasks. The failure mode persists even in frontier models (GPT-5.2), and gates remain effective (+10.4pp; P=0.020).
llm agentspolicy violationdeterministic gatessilent failurespre-execution verification
Heterogeneity-Adaptive Diffusion Schrodinger Bridge for PET-Guided Whole-Body MRI Translation
The paper introduces Heterogeneity-Adaptive Diffusion Schrödinger Bridge (HA-DSB), a novel framework for whole-body MRI translation that addresses heterogeneous feature distributions across anatomical regions and pathological tissues. HA-DSB employs region context embeddings from a vision-language model (VLM) for region-specific modeling and integrates lesion-aware metabolic prior from PET through a dual-stage guidance mechanism. This includes a PET-guided noise modulation module during the forward process and PET feature-enhanced attention during the reverse process. Experiments demonstrate superior translation quality across body regions and improved fidelity in lesion areas under PET guidance.
diffusion schrödinger bridgewhole-body mripet-guidedvision-language modellesion-aware
Agentic Data Environments
The paper introduces Agentic Data Environments as an execution substrate for autonomous agents, aiming to enhance capabilities while ensuring safety. By extending beyond traditional databases to include files, APIs, applications, and system state, these environments reframe data systems as active substrates for reliable execution. The approach seeks to balance the benefits of automation (speed, scale, efficiency) with mechanisms to mitigate irreversible failure costs.
autonomous agentsexecution substratedata systemssafety guaranteesagentic automation
When Prompts Ignore Structure: Graph-Based Attribute Reasoning for Calibrated VLMs
ARGTCA introduces a graph-based approach to improve confidence calibration in vision-language models (VLMs) during test-time adaptation. The method constructs a Symbolic Attribute Graph where (class, attribute) pairs are nodes, and trains a Graph Attention Network (GAT) using contrastive objectives to capture inter-attribute dependencies. Two attribute selection strategies are proposed: ARGTCA-DIV for intra-class diversity and ARGTCA-DISC for inter-class discrimination. Experiments across nine benchmarks demonstrate that ARGTCA-DIV reduces average Expected Calibration Error (ECE) by ~37% over baselines, while ARGTCA-DISC achieves a ~17% reduction, highlighting the benefits of modeling attribute interactions for reliable VLM adaptation.
symbolic attribute graphgraph attention networkexpected calibration errortest-time adaptationcontrastive objectives
MIRA-Math: A Benchmark for Minimal Information Requesting and Mathematical Reasoning
MIRA-Math introduces a benchmark for minimal information requesting in mathematical reasoning, focusing on problems with a unique latent-state answer but missing one atomic fact. The solver must request this fact in natural language under strict constraints and integrate it into an exact final answer. The benchmark includes 2,310 instances across 22 mathematical families, validated deterministically with a fixed LLM responder. Experiments reveal separable metrics for request success and final-answer accuracy, highlighting distinct failure modes. The release includes generators, verifiers, prompts, and metadata for reproducible evaluation.
minimal information requestingmathematical reasoningatomic factlatent statellm responder
Physics-Audited Agentic Discovery in Scientific Machine Learning
Physics-Audited Agentic SciML (PA-SciML) introduces a verification-first workflow for agentic scientific machine learning discovery, ensuring surrogate models satisfy physics constraints such as boundary conditions, causality, and stiffness scaling. The method pre-defines a scoring evaluator, derives machine-checkable physics requirements, and audits candidate models on outputs and input ranges without reference solutions. It includes optional numerical probes and isolated modeling changes to track score improvements. In computational solid mechanics experiments, PA-SciML selected surrogates with lower validation error than error-only baselines, passing linear-elastic checks, while error-only baselines failed stricter causality tests. The key distinction is per-candidate physics evidence on predicted fields.
physics-auditedsurrogate modelscausalityboundary conditionsagentic sciml
On Adversarial Vulnerability of Vision-Language Models through the Lens of Intermediate Spectral Subspaces
The paper introduces a spectral-subspace-guided attack (SSGRA) that exploits intermediate linear transformations in transformer-based vision-language models (VLMs) to enhance adversarial effectiveness. By aligning representations with subspaces spanned by bottom right singular vectors, SSGRA outperforms existing baselines and provides spectral insights into VLM vulnerability. Experiments demonstrate improved attack success, offering a novel perspective on adversarial robustness through spectral decomposition of intermediate layers.
adversarial vulnerabilityvision-language modelsspectral decompositiontransformer networkssingular vectors
Behavior Foundations for Quadruped Robots: ABot-C0 Technical Report
The paper introduces ABot-C0, a generalist motion-control system for quadruped robots addressing three key challenges: scalable motion-data acquisition, robust policy learning, and reliable deployment. The method combines conditional video-generation synthesis, annotated motion capture, teleoperation, and human design to create 16,074 physically feasible motion clips. It trains a Flow-Matching generalist policy demonstrating scaling-law improvements in motion tracking and introduces a three-stage privileged-to-perceptive framework for all-terrain locomotion. Experiments show successful urban-terrain navigation and multimodal interaction, advancing quadruped robots toward product-level behavioral intelligence.
quadruped robotsmotion trackingflow-matching policyprivileged-to-perceptive frameworkconditional video-generation
Multi-Agent AI Control: Distributed Attacks Hamper Per-Instance Monitors
We introduce the empirical study of multi-agent AI control, formalizing distributed attacks where multiple agents collaborate to achieve malicious objectives. Using FakeLab, a synthetic AI-lab codebase with 9 services, 86 benign tasks, and 4 attack objectives, we evaluate single-agent monitoring against distributed attacks, varying agent count, coordination, model capabilities, and monitoring configurations. Key findings include the fragmentation effect: increased agent coordination reduces per-agent monitoring effectiveness. This effect is not driven by benign-to-malicious code ratio but likely by model capability. An explicit planner amplifies fragmentation, increasing attack completion rates up to sevenfold and mildly improving executors' monitor-theory-of-mind. Stronger monitoring reduces undetected attacks significantly but remains imperfect.
multi-agent ai controldistributed attacksfragmentation effectmonitor-theory-of-mindexplicit planner
HumAIN: Human-Aware Implicit Social Robot Navigation
HumAIN introduces a human-aware implicit social robot navigation framework that integrates subtle skeletal cues into planning via knowledge distillation. The method employs a transformer-based teacher model processing multi-modal inputs (historic images, skeletal keypoints, robot state, target goal) to learn trajectory representations, then distills this into a lightweight student model optimized for trajectory reconstruction and latent feature alignment. Experiments demonstrate a 29.8% average improvement in trajectory prediction metrics over baselines, validating the approach's effectiveness in enabling socially compliant navigation from minimal inputs.
social robot navigationknowledge distillationtrajectory predictionmulti-modal fusiontransformer-based model
Latency-Aware Bid Acceptance under Operational Feasibility: A Public Benchmark with Hindsight Ceilings
The authors introduce FreightBidBench, a public benchmark for online truckload bid acceptance that models operational feasibility (pickup reach, appointment windows, hours-of-service) and economic factors (service penalties, fleet value) using calibrated public data. They develop two hindsight ceilings: an LP relaxation and a tighter Lagrangian-per-truck method (20.7-39.3% tighter than LP), plus a parametric surrogate-rollout cascade with escalation triggers. Evaluated on tight/scarce-capacity scenarios, their cascade achieves ~98% of rollout profit at 40-56% lower latency, with performance statistically indistinguishable from the teacher policy.
bid acceptanceoperational feasibilityhindsight ceilingssurrogate-rollout cascadelagrangian relaxation
Quantum simulation of real-world nonlinear dynamics via Koopman method
The authors propose the quantum Koopman method, a data-driven framework for simulating nonlinear dynamics on quantum computers by embedding them into learned linear representations. This method learns Koopman observables from trajectory data, projects the lifted dynamics onto finite-dimensional subspaces, and decomposes the non-unitary propagator into parallel spectral channels implemented via shallow quantum circuits. Experiments on a superconducting processor simulate three nonlinear systems—reaction-diffusion dynamics, fluid motion on a sphere, and Gulf Stream currents—using up to 32 parallel circuits of 10 qubits. Results capture dominant multiscale patterns and statistical signatures, revealing a transition from hardware noise-limited performance in weakly nonlinear systems to Koopman representation-limited performance as nonlinearity increases.
quantum koopman methodnonlinear dynamicskoopman observablesspectral channelssuperconducting processor
Hypergraph Neural Stochastic Diffusion: An SDE Framework for Uncertainty Estimation
The paper introduces Hypergraph Neural Stochastic Diffusion (HyperNSD), a stochastic differential equation framework for uncertainty estimation in hypergraph neural networks. HyperNSD models hypergraph representations as stochastic processes with learnable drift and diffusion functions, capturing deterministic dynamics and structural ambiguity. Theoretical analyses confirm stability, equivariance, and convergence, while experiments on hypergraph benchmarks show reliable uncertainty estimation for out-of-distribution detection and misclassification, maintaining competitive prediction accuracy.
hypergraph neural networksstochastic differential equationsuncertainty estimationhigher-order relationsrepresentation learning
HAJJv2-CrowdCount: Zero-Shot Benchmark for Dense Crowd Counting
The paper introduces HAJJv2-CrowdCount, a zero-shot benchmark for dense crowd counting featuring per-second human annotations on Hajj footage, where extreme occlusion and steep camera angles challenge existing methods. Three zero-shot paradigms are evaluated: YOLO-World (open-vocabulary detection), APGCC (point-based counting), and SAM3Count (promptable segmentation). SAM3Count achieves the lowest overall MAE (70.4), but APGCC outperforms in dense scenes (MAE 114.9 vs. >300 for others), highlighting a critical trade-off for deployment in occlusion-heavy environments. Annotations are released for reproducibility.
crowd countingzero-shot learningmean absolute errorocclusion handlingbenchmark dataset
From Atomic Actions to Standard Operating Procedures: Iterative Tool Optimization for Self-Evolving LLM Agents
The study introduces EvoSOP, a framework enabling LLM agents to synthesize atomic actions into reusable Standard Operating Procedures (SOPs) for self-evolution. EvoSOP employs an iterative lifecycle of construction, merging, evaluation, and pruning to optimize the toolset. Experiments show EvoSOP significantly improves task success rates and reduces interaction rounds compared to baselines, fostering reliable and efficient tool-use patterns.
llm agentsstandard operating procedurestool optimizationself-evolving agentsexecution trajectories
FedCVESA: Taking Away Training Data in Federated Learning via Correlation Value Encoding and Segmented Aggregation
FedCVESA introduces a federated variant of the Correlation Value Encoding Attack (CVEA) to study white-box Taking Away Training Data (TATD) attacks in federated learning (FL). The method employs a Pearson-correlation regularizer in the loss function of target clients to encode private training data into selected model parameters, termed carrier parameters. Segmented aggregation is proposed to preserve these carrier parameters during server aggregation while maintaining standard averaging on others. Experiments on MNIST, Fashion-MNIST, and CIFAR-10 under Dirichlet non-IID partitions demonstrate successful theft of semantically meaningful private training images while preserving main-task utility, highlighting FL as a parameter-level memorization channel for active TATD attacks.
federated learningcorrelation value encodingsegmented aggregationcarrier parameterstaking away training data
POO-LPSP: Parallel Osprey Optimized Least Penalty-Squared Prioritization Methods for Priority Derivation in the Analytic Hierarchy Process
The study proposes POO-LPSP, a parallel metaheuristic method for priority derivation in the Analytic Hierarchy Process (AHP), addressing computational limitations of traditional eigenvector methods. It introduces revised Least Penalty-Squared Prioritization (LPSP) models (LPPDS and LPPWS) to minimize Root Mean Penalty-Squared Variance (RMPSV) and Root Mean Penalty-Weighted Square Variance (RMPSWV), optimized via an improved Parallel Osprey Optimization Algorithm (POOA). Validation on a Generative AI vendor selection problem demonstrates enhanced reliability and computational efficiency compared to Saaty's Eigen system method.
analytic hierarchy processpairwise comparisonmetaheuristic optimizationpriority derivationparallel osprey algorithm
Multimodal Voice Activity Projection for Turn-Taking in Social Robots with Voice-Activity-Related Pretrained Encoders
The paper introduces a Multimodal Voice Activity Projection (MM-VAP) framework for turn-taking prediction in social robots, extending audio-only VAP to synchronized audio-visual inputs while maintaining self-supervised future-projection objectives. The method leverages pretrained audio-visual backbones optimized for speech tasks, adapts them via Low-Rank Adaptation, and employs inter-speaker attention to model relational dynamics for projecting future voice activity. A semantic consistency loss regularizes the 256-state output space based on higher-level dialogue patterns. Evaluations on NoXi, NoXi+J, and Haru EDR corpora demonstrate improvements over baselines, particularly for specific turn-taking events, validating its efficacy in mediation-oriented human-robot interaction.
multimodal voice activity projectionlow-rank adaptationinter-speaker attentionsemantic consistency lossturn-taking prediction
CarbonCLIP: Enhance Carbon Prediction from Satellite Imagery via Integrated Street-View Semantics and Temporal Context Training
CarbonCLIP introduces a multimodal distillation framework for satellite-based urban carbon emission prediction, addressing data heterogeneity and semantic-temporal context gaps. The method employs dual-branch contrastive learning: a spatial branch leverages LMM-generated street-view textual descriptions for semantic priors, while a temporal branch encodes monthly emission variations. Pretrained with multimodal data, it operates on satellite imagery alone during inference. Evaluations on Beijing and Singapore show superior performance over baselines, demonstrating effective knowledge transfer to satellite representations for scalable urban carbon modeling.
multimodal distillationcontrastive learningsemantic priorstemporal encodersatellite representation
Bayesian Optimization of Genetic Algorithm Hyperparameters in a Multi-Fidelity Framework for Efficient Lattice Material Design
The study introduces a multi-fidelity framework for optimizing genetic algorithm (GA) hyperparameters in lattice material design, combining high-fidelity Fast Fourier Transform homogenization, a medium-fidelity 3D CNN surrogate, and a low-fidelity Gaussian process surrogate with Bayesian optimization. The logNEI acquisition function outperforms others by handling GA evaluation noise. Results show a 24% computational cost reduction (from 225 to 171 hours) while maintaining mechanical performance, with 25-generation GA runs matching 75-generation outcomes. Penalized BO further reduces required lattice evaluations with minimal performance loss.
bayesian optimizationgenetic algorithmmulti-fidelitysurrogate modellattice design
FMMVCC: Fuzzy Mamba-based Multi-View Contrastive Clustering for Univariate Time Series
FMMVCC proposes a Mamba-based deep clustering framework for univariate time series, addressing limitations in capturing long-range dependencies and computational efficiency. The method combines state space sequence modeling for linear-complexity temporal representation learning with multi-view self-supervision through temporal masking and augmentations. Evaluation across 15 benchmarks shows superior performance, achieving best results in 29/60 metric evaluations and highest average rank in all scenarios.
mambatime series clusteringstate space modelsmulti-view learningself-supervised learning
ORCAID: Oblique Rule-Based Continuous-Action Interpretation for Deep RL Policies
ORCAID introduces an oblique decision tree algorithm for extracting interpretable rule-based policies from deep RL agents in mixed continuous-discrete environments with continuous action spaces. The method employs a three-stage split search (random initialization, local refinement, backward elimination) to partition state space via hyperplanes and fit local linear models, followed by leaf merging for concise rule generation. Evaluations show the extracted policies maintain performance with low parameter counts and can enhance original RL policy performance.
reinforcement learninginterpretable policiesoblique decision treescontinuous action spacesrule extraction
DiPhon: Diffusion on Graphons for Scalable Graph Generation
DiPhon introduces a diffusion framework for scalable graph generation by leveraging graphons, the limit objects of dense graph sequences. The method formulates a continuous diffusion process on graphon space via a Jacobi stochastic differential equation (SDE), with a discretized graph-level process that mimics these dynamics. The reverse-time process uses a tractable marginal score, estimated via graph denoising, to generate samples. Theoretically, DiPhon matches the first moment and approximates the second moment of the continuous process. Empirically, it demonstrates scalability by generating larger graphs from small training sets while preserving topological properties.
graph generationdiffusion modelsgraphonsstochastic differential equationtopological properties
Reasoning Consistency Scanning: A Framework for Auditing Chain-of-Thought Validity in AI Safety Evaluations
The paper introduces reasoning consistency scanning, a method for auditing chain-of-thought (CoT) validity in AI safety evaluations by assessing logical consistency between stated reasoning and outputs. The authors formalize reasoning consistency (distinct from faithfulness), propose a six-subtype inconsistency taxonomy, and develop a benchmark of 60 InstrumentalEval-derived transcripts. They implement InspectScout, the first scanner for this property, and demonstrate its effectiveness across four generator models and three inspect_evals tasks, revealing systematic variation in inconsistency rates.
chain-of-thoughtreasoning consistencyai safetyinspectscoutinstrumentaleval
Vision Foundation Models in Radiology: A Scoping Review of Data, Methodology, Evaluation and Clinical Translation
The scoping review analyzes 67 studies (2017-2026) on radiology-specific vision foundation models (VFMs), identifying key gaps in data representativeness and clinical deployment. Using PRISMAScR methodology, researchers mapped studies across three pillars: dataset characteristics (brain MRI/thoracoabdominal CT/chest X-ray, 100K-multi-million samples), model architectures (predominantly transformer-based with self-supervised pretraining like masked image modeling), and evaluation practices (segmentation/classification focus with inconsistent cross-center validation). Results show promising transferability but highlight limitations in data heterogeneity (FUTURE-AI alignment), benchmark standardization, and deployment-oriented evaluation.
vision foundation modelsmasked image modelingself-supervised pretrainingcross-center validationfuture-ai principles
Memory Scarcity, Open Models, and the Restructuring of the AI Industry, 2026-2030 -- A quantitative scenario analysis of inference economics, training-cost divergence, and infrastructure solvency
The study presents a quantitative scenario analysis of AI industry restructuring from 2026-2030, driven by memory scarcity, open models, and infrastructure economics. Using a model-agnostic inference cost metric ($/PB of bandwidth), it demonstrates persistent cost gaps between incumbents and entrants due to fleet depreciation dynamics (3.2x in 2026, 1.9x in 2027). Findings include bifurcation of training costs ($18-38B frontier vs $5M mass tier), solvency constraints requiring 2x annual token-demand growth, and vulnerability of 2026/2028-29 capacity vintages. The analysis evaluates five scenarios (25% Rotating Landlord Oligopoly most likely) and identifies China's LineShine LX2 as cost-decoupled from global memory crises.
inference economicskv-cache compressionhbm price surgetraining-cost divergencevintage-breakeven analysis
Validate the Dream Before You Trust Its Verdict: Admissibility for World-Model Simulators
The paper introduces an admissibility ladder (L0-L4) for accrediting World Models (WMs) used as test oracles in robotics, ensuring their verdicts on action policies are trustworthy. Building on Verification, Validation & Accreditation (VV&A), Safety of the Intended Functionality (SOTIF), and scenario-based testing standards, the framework evaluates WMs beyond visual fidelity metrics like Fréchet Video Distance (FVD) to assess action-following robustness. Applied to autonomous driving WMs, the results reveal a reversal: higher visual generation quality (L0) does not correlate with better action-robustness (L1-L2), highlighting the need for comprehensive validation.
world modelsadmissibility ladderfréchet video distanceverification validation accreditationaction-following robustness
Does AI Understand Imaging? A Systematic Benchmark of Agentic AI for Computational Imaging Tasks
The paper introduces ImagingBench, a systematic benchmark for evaluating agentic AI on computational imaging tasks, covering 20 tasks across five categories: ray/wave optics, image signal processing, inverse reconstruction, computational sensing, and calibration. It assesses three settings (Expert, Planner, Forward) using proprietary and open-source VLMs (Gemini, GPT, Qwen) against specialized baselines. Results show agentic models underperform task-specific methods, particularly in computational sensing (lensless imaging, event-based reconstruction), with Planner guidance offering minimal improvement. Visual plausibility often masks poor fidelity, highlighting a semantic-physical gap in AI imaging competence.
computational imagingvision-language modelsinverse problemsagentic aibenchmarking
Predicting LLM Safety Before Release by Simulating Deployment
The authors propose deployment simulation as a method for pre-release safety evaluation of large language models (LLMs), addressing limitations of traditional evaluations in coverage, representativeness, and test-awareness. Their approach regenerates responses from de-identified conversation prefixes using candidate models, enabling both novel misalignment detection and misbehavior rate estimation. Evaluated across four GPT-5-series deployments, deployment simulation outperformed adversarially selected production data baselines, with evaluation-awareness point estimates closer to production traffic than traditional methods. Results indicate feasibility even with public chat datasets, enabling external researchers to conduct deployment-grounded evaluations without access to private logs.
deployment simulationmisalignment detectionpre-release evaluationconversation prefixmisbehavior rate
Entropy Pacing Policy Optimization for Multi-Task Agentic Reinforcement Learning
The paper introduces Entropy Pacing Policy Optimization (EPPO), a method addressing exploration-exploitation pace mismatch in multi-task agentic reinforcement learning for LLMs. EPPO employs a task-wise dynamic clipping mechanism that adapts the Group Relative Policy Optimization (GRPO) threshold based on task entropy, tightening updates for over-confident tasks and relaxing them for under-explored ones. Experiments on multi-task agentic benchmarks show EPPO outperforms existing methods.
multi-task reinforcement learningentropy pacingpolicy optimizationlarge language modelsexploration-exploitation
Tree-of-Thoughts Reasoning for Text-to-Image In-Context Learning
The paper proposes a Tree-of-Thoughts (ToT) reasoning framework for text-to-image in-context learning (T2I-ICL) to address compositional reasoning limitations in multimodal large language models. The method introduces a multi-stage layer that generates, evaluates, and selects among multiple candidate hypotheses before final prompt construction, mitigating ambiguity and compositional errors. Evaluated on the CoBSAT benchmark, ToT-T2IICL demonstrates improved consistency and semantic alignment in image generation compared to baseline and Chain-of-Thought approaches, without requiring additional training.
tree-of-thoughtstext-to-imagein-context learningcompositional reasoningmultimodal llms
GeoProp: Grounding Robot State in Vision for Generalist Manipulation
GeoProp introduces a lightweight adapter for grounding robot proprioception in visual features to improve manipulation policies. The method projects robot state onto the image plane, samples localized visual features, and injects spatial priors via FiLM modulation. It also incorporates look-ahead visual context by predicting short-horizon coordinates from recent kinematics. Evaluated across 67 tasks, GeoProp improves Diffusion Policy by 8.7% on 63 simulation tasks and pi_0 by 4.0% on RoboTwin, with a 10.6% average gain in real-world performance, while adding only 2-3% to parameter count.
proprioceptionfilm modulationdiffusion policyspatial priorsvisual grounding
Operational Reframing and Approval-Framed Delegation in Multi-Agent LLM Safety
The study disentangles multi-agent LLM safety evaluation by isolating three mechanisms conflated in pipeline effect analyses: operational reframing, planner refusal/transformation, and approval-framed delegation. Using a five-condition controlled contrast design, the authors evaluate 30 synthetic harmful scenarios and an external validation set from four agent-safety benchmarks, employing LLM-judged compliance. Results show operational reframing as the most portable risk signal, increasing compliance for GPT, Gemini, and DeepSeek, while Claude resists it. Planner refusal offsets risk, but executable steps increase executor compliance. Approval-framed delegation varies by prompt design, model pairing, and scenario source. Findings advocate for separate reporting of reframing, planner behavior, delegation framing, and model pairing in safety evaluations.
operational reframingplanner-executor pipelineapproval-framed delegationllm-judged compliancecontrolled contrast design
AT-Attn: Temporal-Aware Cross-Attention for Longitudinal Multimodal Alzheimer's Disease Diagnosis
The paper proposes AT-Attn, a temporal-aware multimodal framework for Alzheimer's disease diagnosis that integrates irregular longitudinal MRI and clinical data via Change-and-Time encoding, time-biased asymmetric cross-attention, and gated fusion. Evaluated on 1,520 ADNI patients with structural MRI, cognitive trajectories, and static clinical variables, AT-Attn achieves 0.719 accuracy, 0.873 ROC-AUC, and outperforms unimodal and naive fusion baselines while matching strong tabular models. Results demonstrate that constrained temporal fusion enables MRI to provide clinically complementary information despite noise or missingness.
multimodal fusionlongitudinal analysisasymmetric cross-attentionalzheimer's diagnosistemporal encoding
Navigating Hierarchy: Hyperbolic Learning on Brain Graphs for Disorder Diagnosis
The paper introduces Hyperbolic Learning on Brain Graphs (HLBG), a framework for hierarchical modeling of functional brain networks across ROI, community, and whole-brain levels. HLBG leverages Lorentzian hyperbolic space to encode multi-level hierarchy via geometric entailment constraints and proposes Graph-aware Mamba (GaMamba) to capture long-range dependencies with topological preservation. Evaluated on ABIDE-I and REST-MDD datasets, HLBG outperforms state-of-the-art methods in disorder diagnosis and biomarker identification.
hyperbolic learningfunctional brain networkslorentzian spacegraph-aware mambahierarchical modeling
Multiplication Beyond Groups: Stratified Fourier Mechanisms in Transformer Circuits
The paper introduces the monoid extension, a localized generalization of Group Composition via Representation (GCR), to explain how small transformers learn modular integer multiplication over composite moduli—a non-invertible operation. The method partitions input space into hierarchical algebraic regions where group-like structure persists, enabling Fourier mechanisms. Experiments on square-free modular multiplication reveal embeddings organized by these regions, class-sensitive attention routing, and local character features explaining significant output logit variance, extending representation-theoretic mechanisms beyond groups.
monoid extensionmodular multiplicationfourier mechanismstransformer circuitsrepresentation-theoretic
Complexity-Budgeted, Interaction-Aware Interpretable Model for Tabular Data
The paper introduces Interaction Aware Interpretable Machine Learning (IAIML), a framework for tabular data that preserves predictive interactions while maintaining interpretability through three mechanisms: adaptive feature discretization, pairwise interaction scoring, and partitioned explanation budgets. IAIML routes detected interactions via relaxed screening or explicit pair terms, achieving mean AUC within 1.4 points of gradient-boosted ensembles on 40 datasets while using 14–28× fewer components. It outperforms baselines on datasets with strong pairwise interactions and low marginal signal, matching RuleFit in AUC/component count but degrading for higher-order interactions. Ablations confirm incremental benefits from adaptive discretization and interaction-aware admission.
interpretable machine learningtabular datapairwise interactionsadaptive discretizationexplanation budget
Progressive Crystallization: Turning Agent Exploration into Deterministic, Lower-Cost Workflows in Production
Progressive crystallization introduces a lifecycle for AIOps that converts costly agent exploration into deterministic workflows, reducing permanent LLM inference costs. The method defines a three-stage execution taxonomy (agent-orchestrated, hybrid, deterministic) with evidence-based promotion/demotion mechanisms to validate and optimize workflows. Evaluated on a production cloud networking system handling tens of thousands of incidents monthly, the approach increased deterministic execution from 0% to 45% over eight months, reduced per-incident agent costs by >70%, and improved safety via reproducibility and auditability. The paper details taxonomy, criteria, trace extraction, economic modeling, and limitations.
progressive crystallizationaiopsllm inferencedeterministic workflowsevidence-based promotion
Making Implicit Preservation Intent Explicit in Conversational Image Editing
We introduce OCCUR-Bench, a benchmark for evaluating temporal preservation in conversational image editing, and ReSpec, a training-free framework for explicit preservation. OCCUR-Bench provides diverse occlusion-and-revelation scenarios with historical restoration references, enabling evaluation of faithful restoration. ReSpec identifies persistent content, selects historical image states for missing visual evidence, and conditions an in-context editor on restoration-aware instructions and reference images. Experiments demonstrate that ReSpec improves restoration fidelity and temporal consistency on OCCUR-Bench, emphasizing the importance of grounding preservation in editing history rather than solely in the current image.
occur-benchtemporal preservationconversational image editingrestoration-aware instructionsin-context editor
Riemannian Geometry for Pre-trained Language Model Embeddings
The paper introduces Riemannian Mean Pooling (RMP), a method for aggregating contextual token embeddings using Riemannian geometry to improve sentence-level classification. RMP extracts per-token pullback metrics from an encoder's analytical Jacobian and aggregates them via the Fréchet mean on the symmetric positive definite manifold. Evaluated on CoLA, CREAK, and RTE, RMP outperforms Euclidean mean pooling, while remaining at chance on FEVER-Symmetric, indicating robustness to annotation artifacts. Ablations reveal that geometric aggregation alone (even with random encoders) drives gains on two datasets, while trained encoders add signal specifically on knowledge-heavy CREAK.
riemannian geometrypre-trained language modelssentence embeddingsfréchet meanmanifold learning
Measuring Intelligence Beyond Human Scale
The paper proposes a relative measurement paradigm for evaluating intelligence beyond human capability, addressing limitations of human-authored benchmarks. The method involves models generating public challenges to differentiate other systems, creating an adversarial psychometric rating system. Practical protocols mitigate private-information attacks and enable judge-free adjudication, scaling with agent capabilities. The framework is instantiated across verifiable and open-ended domains, demonstrating model-generated evaluation's potential to measure superhuman systems.
psychometric ratingadversarial evaluationrelative measurementsuperhuman intelligencebenchmark saturation
On the Principles of Deep Feedforward ReLU Networks
The paper systematically analyzes the mechanisms of deep feedforward ReLU networks, extending principles from two-layer networks to multiple hidden layers. By introducing the concept of paths and their interrelationships, it demonstrates that units in deep ReLU networks form piecewise linear manifolds to partition input space, contrasting with hyperplanes in two-layer cases. The study reveals how hidden-layer units efficiently produce both linear functions and input space partitions, generalizing principles like multiple strict partial orders and continuity restrictions. These foundational principles explain complex training solutions obtained via backpropagation, thereby elucidating the 'black box' of deep ReLU networks.
relu networkspiecewise linear manifoldinput space partitionstrict partial orderscontinuity restriction
Intrinsic Green's Learning: Supervised Learning on Manifolds via Inverse PDE
The paper introduces Intrinsic Green's Learning (IGL), a framework for supervised learning on manifolds by modeling target functions as solutions to linear PDEs with learned source terms. IGL decomposes the source and Green's kernel into low-rank tensors in a learned coordinate chart, reducing integration complexity to linear in intrinsic dimension. A two-stage algorithm separates coordinate discovery (via encoder) from source fitting, avoiding dimensional collapse. Learnable gates automatically identify intrinsic dimensionality. Experiments on synthetic manifolds and MNIST demonstrate near-optimal classification while recovering intrinsic dimensions.
intrinsic green's learningmanifold learninglow-rank decompositionpde-based learningintrinsic dimension
AnchorPrune: Relevance-Anchored Contextual Expansion for Visual Token Pruning
AnchorPrune introduces a training-free framework for visual token pruning in vision-language models by constructing a relevance anchor and expanding it with complementary context. The method adaptively determines anchor size from relevance-ranked tokens' novelty profile, preserving query-critical evidence, and allocates remaining budget via importance-weighted novelty to recover non-redundant context. Evaluated on LLaVA-NeXT-7B, it retains 97.6% of full-token performance using only 160 of 2,880 visual tokens, outperforming baselines under aggressive compression.
visual token pruningrelevance anchorcontextual expansiontraining-freemultimodal inference
Gimitest: A Comprehensive Tool for Testing Reinforcement Learning Policies
The authors present Gimitest, an open-source framework for comprehensive testing of reinforcement learning policies across diverse environments and algorithms. The tool supports single- and multi-agent RL testing through modifications of integrated components in gym frameworks, including Farama Gymnasium and PettingZoo. Experiments demonstrate Gimitest's effectiveness in evaluating policy reliability under varying conditions, addressing limitations of existing methods that target only specific scenarios.
reinforcement learningpolicy testingmulti-agent systemsgym frameworksopen-source tool
Latent graph encoding of multimodal neuroimaging features with generative AI architectures
The study proposes a multimodal graph VAE (gMMVAE) framework for neuroimaging feature encoding, systematically evaluating generative architectures (VAEs, transformers, GANs, diffusion models) with modality-aware graph encoding of structural (GMV) and functional (sFNC) MRI features. Graph-based latent space encoding outperforms vectorized approaches in generation fidelity, reconstruction quality (quantified metrics unspecified), and discriminability. The gMMVAE demonstrates superior performance across multiple evaluation metrics compared to alternative generative models.
multimodal vaegraph encodingneuroimaging featuresgenerative architectureslatent space fusion
Learning social norms enhances compatibility in dynamic human-AI coordination
The study demonstrates that explicitly quantifying social norms enhances AI-human coordination in dynamic interactions. Using pedestrian-vehicle interactions as a testbed, the authors identified three normative principles (outcome predictability, value alignment, advantage awareness) from 3,456 human interactions and integrated them into LLM agents. The social-norm-informed LLM achieved 4× higher scores than baseline strategies and outperformed human-human interactions by 43% in closed-loop coordination tasks.
social normshuman-ai coordinationlarge language modelsdynamic interactionsvalue alignment
Multimodal Spatiotemporal-Frequency Fusion with Peak Enhancement for Cellular Traffic Forecasting
The paper proposes MSPF-Net, a multimodal framework for cellular traffic forecasting that jointly models endogenous traffic patterns, burst behavior, and exogenous urban events. The method integrates a Spatiotemporal-Frequency Traffic Encoder, Peak Enhancement Module for spike detection, News Context Representation Module for urban event embeddings, and Dynamic Fusion Prediction Module for adaptive signal fusion. Experiments on Milano, Trento, and LTE datasets demonstrate performance improvements through this multimodal approach compared to single-modality baselines.
cellular traffic forecastingspatiotemporal-frequency fusionpeak enhancement moduleexogenous contextual signalsdynamic fusion prediction
Physics-guided spatiotemporal neural models for fuel density prediction
The paper proposes a physics-guided machine learning (PGML) framework for fuel density prediction, integrating physical constraints into deep learning models to improve accuracy and stability. The method employs ConvLSTM, Adaptive Fourier Neural Operator (AFNONet), and Video Vision Transformer (ViViT) architectures, augmented with differentiable physics-informed loss terms for mass conservation and rate-of-spread estimation. Experimental results show the PGML framework outperforms purely data-driven baselines in both accuracy and stability across multiple trials, enabling efficient and physically plausible fire forecasting for prescribed burn management.
physics-guided machine learningspatiotemporal modelingconvolutional lstmadaptive fourier neural operatordifferentiable physics
Large Behavior Model: A Promptable Digital Twin of the Retail Customer
The paper introduces the Large Behavioral Model (LBM), a promptable digital twin for retail customer behavior that unifies predictive accuracy with explainable decision-making. LBM combines behavioral profiles from historical purchases with retrieval-augmented generation for product context, trained via continued pre-training on verbalized data, supervised fine-tuning, and evidence-based reinforcement learning. Evaluations on purchase prediction, basket completion, and cross-domain tasks show LBM outperforms general-purpose language models in retail domains, with ablations highlighting the importance of continued pre-training and retrieval augmentation. The model demonstrates effective zero-shot transfer and improved reliance on behavioral evidence over language priors.
large behavioral modelretrieval-augmented generationbehavioral profileevidence-based calibrationdigital twin
WAM-TTT: Steering World-Action Models by Watching Human Play at Test Time
WAM-TTT introduces a test-time training framework for steering world-action models (WAMs) using raw human videos, eliminating the need for robot demonstrations or task-specific fine-tuning. The method absorbs human videos into a lightweight adaptive memory within a frozen WAM through self-supervised video prediction, aligned with robot behaviors via meta-training on paired human-robot data. At test time, only unlabeled human videos are required for memory adaptation, preserving the WAM's generalization capabilities. Experiments demonstrate that WAM-TTT consistently outperforms in-context human-video conditioning baselines across diverse manipulation tasks and generalization scenarios.
world-action modelstest-time trainingself-supervised video predictionmeta-traininggeneralization
Hybrid Least Squares/Gradient Descent Methods for MIONets
The authors propose a hybrid least squares/gradient descent (LSGD) method to accelerate training of MIONets, generalizing the LSGD approach from DeepONets. By treating MIONet as a multilinear function with respect to last-layer branch network parameters, they employ alternating least squares optimization with Kronecker/Khatri-Rao products and tensor permutations for efficient matrix factorization. The method supports general $L^2$ loss functions with regularization terms and linear operators applied to MIONet outputs.
mionetsleast squaresgradient descentkronecker productmultilinear optimization
End-to-End LLM Flight Planning with RAG-based Memory and Multi-modal Coach Agent
The paper introduces FRAMe, an end-to-end LLM-based flight planning system that integrates a planner LLM with retrieval-augmented generation (RAG) memory and a multi-modal coach agent to bridge human pilot intent and autonomous eVTOL operations. The system translates natural language instructions into valid flight plans while aligning with operator preferences. Evaluated across four LLMs in varied scenarios, FRAMe achieves up to 93.8% aggregate validity (99% on Easy scenarios) and improves preference-relevant metrics when headroom exists.
llmretrieval-augmented generationmulti-modal agentflight planningevtol
Large Language Models (LLMs) and Generative AI in Cybersecurity and Privacy: A Survey of Dual-Use Risks, AI-Generated Malware, Explainability, and Defensive Strategies
This survey examines the dual-use risks and defensive applications of Large Language Models (LLMs) and generative AI in cybersecurity, highlighting their role in both threat detection and malware generation. The authors analyze over 70 academic papers, industry reports, and technical documents, covering platforms such as Google Play Protect, Microsoft Defender, and Hugging Face Spaces. Findings indicate a projected increase in LLM-generated malware from 2% in 2021 to 50% by 2025, necessitating next-generation security frameworks. The paper concludes with recommendations for responsible LLM deployment, including model watermarking, adversarial defense, and cross-industry collaboration, offering a roadmap for secure AI-driven cybersecurity systems.
large language modelsgenerative aimalware generationmodel watermarkingadversarial defense
Self-Supervised Pretraining Improves Cross-Site and Cross-Scale Robustness of Point Cloud Leaf-Wood Segmentation
Self-supervised pretraining with Point-M2AE improves cross-site and cross-scale robustness in leaf-wood segmentation of tree point clouds. The model was pretrained on ShapeNet-55 augmented with 2,400 individual tree point clouds and fine-tuned using recursive voxel subdivision to handle varying point densities. Pretraining increased wood IoU from 60.5% to 70.0% for needleleaf and from 69.7% to 76.3% for broadleaf trees, achieving the smallest cross-site variation and highest overall performance on a benchmark spanning four countries. Plot-level segmentation maintained mIoU of 84.7% for broadleaf and 77.7% for needleleaf plots. Downstream wood volume estimation in tropical forests yielded the lowest error (MAE = 2.40 m³).
self-supervised learningpoint cloudsleaf-wood segmentationrecursive voxel subdivisionwood iou
Comprehensive Evaluation of Large Language Model Responses: A Multi-Factor Scoring System
This study introduces a multi-factor scoring system for evaluating large language model (LLM) responses, integrating accuracy, conciseness, factual consistency, readability, and coherence with a GUI for visualization. The framework was tested on TruthfulQA, revealing LLMs' strengths in reasoning (peak composite score 0.6104) but limitations in handling complex facts and ambiguities. The method provides a transparent, adaptable evaluation tool for model refinement, with potential extensions to multilingual domains.
large language modelstruthfulqamulti-factor scoringfactual consistencymodel evaluation
Imputation Meets Clustering: Exploiting Latent Subgroup Structure for Missing Data Recovery
The paper proposes CAGI (Cluster-Aware Generative Imputation), a framework that jointly optimizes missing data imputation and latent subgroup clustering through a co-optimization process. CAGI employs a Partition-Guide-Restore strategy, using dynamic cluster assignments as local priors for a Generative Adversarial Network, with an iterative feedback loop to refine both cluster structures and imputed values. The method combines instance-level reconstruction with distribution-level regularization for stability. Experiments on 14 benchmarks show superior performance over 15 baselines.
missing data imputationgenerative adversarial networklatent subgroup clusteringco-optimizationdistributional regularization
MADB: A Large-Scale Music Aesthetics Dataset with Professional and Multi-Dimensional Annotations
The authors introduce MADB, a large-scale dataset for music aesthetic assessment comprising 9,999 tracks annotated by 30 trained annotators. Each track receives ratings from approximately 10 annotators across 10 perceptual dimensions plus an overall score, accompanied by textual comments for multimodal analysis. They establish a unified evaluation framework using multiple pretrained models, revealing significant gaps between model predictions and human judgments. MADB addresses the lack of structured aesthetic annotations in music, providing a benchmark for human-aligned music understanding.
music aesthetic assessmentmultimodal analysisperceptual dimensionspretrained modelshuman-aligned
Grounding Spatial Relations in a Compact World Model: Instruction Leakage and a Goal-Free Dynamics Fix
The article identifies instruction leakage as a critical confound in goal-conditioned world models, where relation-readout accuracy stems from instruction transcription rather than perceptual grounding. Through empirical analysis across three settings, including BabyAI and a Language-Table forward-dynamics model, the authors demonstrate that withholding the goal collapses accuracy to chance (0.90→0.27) and that counterfactual instructions lead to false anchor predictions 94.5% of the time. The proposed fix involves decoupling the goal from dynamics and supervising the read path, achieving genuine instruction-independent grounding (0.88 accuracy). The detection protocol and remedy are applicable to any goal-conditioned world model where the instruction names the scored quantity.
instruction leakagegoal-conditioned modelsrelation-readout accuracyforward-dynamicsreference anchors
LoCA: Spatially-Aware Low-Rank Convolutional Adaptation of Vision Foundation Models
The paper introduces Low-Rank Convolutional Adaptation (LoCA), a parameter-efficient fine-tuning method for vision foundation models that preserves spatial-channel structure in convolutional layers. Unlike standard Low-Rank Adaptation (LoRA) which flattens 4D kernels into 2D matrices, LoCA decouples channel and spatial adaptation via low-rank channel mixing and SVD-refined spatial bases. Experiments demonstrate LoCA maintains pre-trained spatial priors while achieving state-of-the-art performance on fine-grained classification, domain-generalized semantic segmentation, and generative tasks.
parameter-efficient fine-tuninglow-rank adaptationconvolutional kernelsspatial-channel decouplingvision foundation models
The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI
The paper introduces the 'harness effect', demonstrating how orchestration layer design significantly impacts token economics in enterprise agentic AI systems. Through controlled experiments with six foundation models (Claude Sonnet 4.6, Gemini 3.1/Flash 3.5, Qwen 3.6, GLM 5.1, Palmyra X6) on 22 evaluation tasks, the Writer Agent Harness reduced blended costs by 41%, wall-clock time by 44%, and token usage by 38% while maintaining task-completion quality (0.78→0.81). The harness showed model-invariant efficiency gains (33-61% cost reduction) and capability-dependent quality improvements (r=0.99 correlation with baseline model strength), improving quality per dollar by 82% and completions per million tokens from 54.9 to 92.0.
orchestration layertoken economicsagentic aifoundation modelsprompt caching
Computing with Stochastic Oracles in AI-Augmented Computation
The Stochastic-Oracle Turing Machine (SOTM) framework formalizes AI-augmented computation via probabilistic Turing machines interacting with context-dependent stochastic oracles. This work analyzes SOTM performance under two oracle-response schemes: cached-response, where identical queries reuse responses, and fresh-response, where each call yields independent responses. Cached responses impose ceilings on correct identification and output quality based on transcript distributions, while fresh responses enable exponential error reduction via repeated queries. Theoretical results quantify error probabilities, query-count bounds for threshold stopping, and majority-based amplification in binary candidate-output models, elucidating how response reuse, transcript information, and score function access govern SOTM capabilities.
stochastic-oracle turing machinecached-response oraclefresh-response oracletranscript distributionchernoff rate
ReMoDEx: A Local-to-Global Relevance-Based Model Decision Explainability Framework for large-Scale Image Datasets
The paper introduces ReMoDEx, a framework for large-scale explainability analysis of image classifiers by combining local relevance maps with global clustering. The method employs GradCAM++, Integrated Gradients, Occlusion Sensitivity, and Layerwise Relevance Propagation to generate heatmaps, then standardizes and clusters them to identify decision strategies. Applied to a VGG16 COVID-19 classifier (86.27% accuracy), ReMoDEx revealed two dominant patterns: thoracic-region focus and border/corner sensitivity, exposing potential shortcut learning undetected by conventional metrics. Masked validation confirmed these strategies influenced model predictions.
explainabilityrelevance mapsshortcut learningimage classificationclustering
GemNav: Discrete-Token Visual Robot Navigation using a Multimodal Large Language Model
GemNav introduces a visual robot navigation policy that adapts a frozen Multimodal Large Language Model (MLLM) for waypoint navigation using Low-Rank Adaptation (LoRA) on the language tower alone, eliminating the need for auxiliary visual encoders or continuous regression heads. The method employs a shared discrete token vocabulary for waypoints and categorical navigation signals, with a soft-decoded auxiliary loss to recover metric structure. On a modest 8.7-hour training corpus, the policy achieves zero-shot transfer to four unseen environments, stopping within 0.25-0.42m of the goal in real-world trials, demonstrating data-efficient adaptation of MLLMs for navigation.
visual navigationmultimodal large language modellow-rank adaptationdiscrete-token adaptationzero-shot transfer
A Gold-Standard Study of What Makes a Lightweight Game-Playing Agent Strong
The study establishes a rule-based Gin Rummy expert as a gold-standard benchmark for evaluating lightweight game-playing agents, isolating key factors that improve performance. Through systematic ablation across 100+ runs, effective techniques include trust region updates, targeted reward shaping, opponent curricula, warm starts, and checkpoint selection, collectively improving self-play champions' win rates from 30% to 36% against the expert. Ineffective approaches included reward shaping, learned embeddings, imitation learning, and LLM opponents, while encoder architecture experiments revealed information bottlenecks rather than capacity limits. The method generalizes to Leduc Hold'em and is released as a reusable package.
reinforcement learningimperfect-information gamestrust region updatesself-playcurriculum learning
Gradient-Based Speech-to-Text Alignment for Any ASR Model: From CTC to Speech LLMs
The paper introduces a gradient-based method for speech-to-text alignment applicable to any differentiable ASR model, including CTC, transducer, AED, and speech LLMs. The approach computes token-level gradients with respect to input frames, reduces them to per-frame saliency scores, and decodes word boundaries via dynamic programming. Evaluated on 16 models across TIMIT and Buckeye datasets, the method produces usable alignments without model modification, performing comparably to native aligners and excelling where native methods are weak (e.g., streaming models), at the cost of one backward pass per token.
speech-to-text alignmentgradient-basedasr modelsdynamic programmingsaliency scores
Evaluating SageMath-Augmented LLM Agents for Computational and Experimental Mathematics
The study introduces a ReAct-style LLM agent augmented with SageMath for computational mathematics, demonstrating improved performance on research-level problems from the RealMath benchmark. The method combines LLM reasoning with verifiable CAS feedback and up-to-date documentation via Context7, alongside benchmark refinements including multi-step post-processing and validation. Experiments show average gains of +9.7pp across models, with Qwen~3.7-Max benefiting most (up to 27.8pp) and GPT-5.5 achieving a 75.2% solve rate, suggesting CAS integration enhances computational exploration for mathematical research.
computer algebra systemsreact-style agentrealmath benchmarkautomated conjecture discoveryin-context learning
Ad Headline Generation using Self-Critical Masked Language Model
The authors propose a novel method for programmatic generation of product advertising headlines using retail content. Their approach applies reinforcement learning policy gradient methods to transformer-based masked language models, jointly conditioning on multiple products to create headlines. The method outperforms existing transformer and LSTM+RL baselines in overlap metrics and quality audits. Notably, model-generated headlines surpass human-submitted ones in both grammatical correctness and creative quality, as validated through audits.
reinforcement learningpolicy gradientmasked language modeltransformerquality audit
When Agents Go Rogue: Activation-Based Detection of Malicious Behaviors in Multi-Agent Systems
The paper introduces AcMAS, an activation-based framework for detecting malicious behaviors in LLM-based Multi-Agent Systems (MAS) without relying on explicit interaction graphs or semantic attack signatures. By analyzing internal reasoning states in the activation space of local agents, AcMAS achieves synchronization-robust detection of stealthy attacks and enables functional restoration of compromised agents. Evaluations show AcMAS outperforms graph-based baselines by +0.22 F1 in synchronous settings (0.94 vs. 0.72) and +0.55 F1 in asynchronous settings (0.93 vs. 0.38), with robustness across LLM backbones, attack intensity, and MAS scale.
multi-agent systemsactivation spacestealthy attackssynchronization-robustllm backbones
A Multi-Analyst LLM Pipeline for Auditable Rule Discovery Across 68 Public Physiological Corpora
The study introduces a multi-analyst LLM pipeline for deriving auditable detection rules from 68 heterogeneous physiological corpora. Four independent LLM families processed corpus documentation under controlled prompts, generating 695 candidate rule markers, later deduplicated to 649 and audited for 51 sanity violations. Cross-corpus consolidation yielded 436 unique rule shapes, with 94 meeting hardware and personalization constraints for immediate detector component development. The workflow emphasizes auditability through analyst disagreement tracking, threshold checks, and CI integration without producing clinically validated detectors.
physiological corporallm workflowrule discoveryauditable pipelinedetector components
What Predicts Correctness in Text-to-SQL? A Selective-Prediction Study
The study identifies predictors of correctness in text-to-SQL queries, focusing on hard multi-table tasks. Using AUROC to evaluate ranking effectiveness, black-box signals like string self-consistency and schema-relevance score achieve AUROCs between 0.61 and 0.68, while white-box log-probability performs similarly (0.67). Verification-based signals, particularly LLM judges, outperform these, with GPT-4o-mini and Claude scoring 0.72 and 0.78 AUROC respectively. A two-provider ensemble reaches 0.82 AUROC with low expected calibration error (0.03). Fine-tuned verifiers achieve 0.77-0.79 AUROC in-distribution but drop to 0.66 on unseen schemas, indicating cross-schema transfer relies on model scale and reasoning rather than fine-tuning.
text-to-sqlaurocself-consistencyllm judgecross-schema transfer
Enhancing deep learning models for time series classification via knowledge distillation
This work demonstrates the effectiveness of Knowledge Distillation (KD) for Time Series Classification (TSC) across three architectures: Fully Convolutional Network (FCN), Inception, and transformer-based ConvTran. By modifying components like convolutional filters and attention heads, the authors show KD most benefits intermediate-complexity student models on the UCR Archive benchmark. Key results include a 38× parameter reduction in FCN, Inception students matching teacher performance with 42% fewer parameters, and ConvTran students with 2 attention heads showing greatest distillation gains. Code is provided for reproducibility.
knowledge distillationtime series classificationfully convolutional networkinception modelconvtran
From Agentic to Autogenic Network Management for AI-Native 6G and Beyond: A Standards Perspective
The paper proposes Autogenic network management, a reference architecture extending Agentic AI with self-programming, self-reflection, self-orienting, and self-architecting capabilities for 6G networks. The method combines Large AI Model (LAM)-based agents with progressive autonomy deployment, transitioning from human-supervised to fully autonomous operation. Validation uses TM Forum's autonomous network use cases, demonstrating practical solutions for operational challenges. A research roadmap identifies technical requirements for 6G implementation. The architecture addresses scalability and complexity limitations of current agentic approaches through runtime software generation and evolution.
autogenic networkagentic ailarge ai model6g networksself-programming
AirPASS: Over-the-Air Federated Learning via Pinching Antenna Systems
The paper proposes AirPASS, an over-the-air federated learning (AirFL) framework for wireless systems with multi-waveguide pinching antenna systems (PASS). It addresses the nonconvex joint optimization of device selection, receive beamforming, and antenna placement via alternating optimization: a homotopy-Riemannian margin-consolidation method for beamforming/selection and a homotopy-assisted geometry optimization for PASS configuration. Experiments demonstrate superior performance over MIMO baselines, near-ideal FedAvg accuracy, and favorable complexity tradeoffs versus SDR-DC and matching-pursuit alternatives.
over-the-air federated learningpinching antenna systemhomotopy-riemannian optimizationdevice selectionreceive beamforming
Cost-Effective Agent Harnesses for Abstract Reasoning and Generalization on ARC-AGI-1
(No summary returned.)
QANTIS: Hardware-Calibrated Sequential POMDP Belief Updates on IBM Heron
QANTIS demonstrates a hardware-calibrated quantum belief-update primitive for sequential POMDPs on IBM Heron processors, validating posterior consistency across multi-step horizons. The method employs boundary-aware BIQAE for stable amplitude estimation and compares no amplification, Grover amplification, and fixed-point amplification (FPAA) on Tiger POMDP trajectories. Results show FPAA preserves exact Bayes posteriors across 8-32 step horizons, with hardware and exact posteriors selecting identical actions in all tested cases. The study establishes an operating envelope for quantum belief updates without claiming standalone hardware advantage.
pomdpquantum amplitude estimationibm heronbelief updatefixed-point amplification
LLM-powered reasoning in agent-based modeling
The paper introduces HALE, a hybrid agent-based and language-driven epidemic modeling framework that integrates large language models (LLMs) with traditional agent-based modeling (ABM) to enhance real-time adaptability in simulating human decision-making. By leveraging LLMs' predictive capabilities, HALE addresses the limitation of static priors in conventional ABMs. As a proof-of-concept, the framework is applied to simulate COVID-19 dynamics in Salt Lake County, UT, demonstrating its scalability and potential for policy-making applications.
agent-based modelinglarge language modelsepidemic simulationhuman decision-makingreal-time adaptation
SmartHomeSecure: Automated Detection and Repair of Smart Home Configuration Errors Using Large Language Models
SmartHomeSecure introduces a system for automated detection and repair of Home Assistant YAML configuration errors using program analysis and constrained LLM generation. The method involves parsing YAML files, detecting syntactic/semantic errors, applying deterministic fixes, and guiding LLMs (gpt-oss-20b/120b, llama-3.1-8b/3.3-70b) via constrained prompts for valid repairs. Evaluation on 100 error-injected files showed 100% detection accuracy for three models, with repair success rates of 87-93% and no hallucinated corrections, demonstrating the feasibility of domain-aware AI for smart home configuration repair.
yaml configurationprogram analysisconstrained generationerror detectionsmart home automation
A Continual Learning Framework for Adaptive Control of Modular Soft Robots
A continual learning framework is proposed for adaptive control of modular soft robots (MSRs), addressing challenges in nonlinear dynamics and hyper-redundant morphologies. The framework enables incremental adaptation to morphological changes while preserving prior knowledge, eliminating the need for retraining from scratch. It supports distributed learning of module-specific dynamics for localized control and precision in fixed configurations. Validation includes closed-loop trajectory tracking experiments on a tendon-driven soft robot in simulation and a three-module pneumatic soft robotic arm in real-world settings. The framework's adaptability is demonstrated through a reaching experiment that selectively activates necessary modules, reducing computational overhead.
modular soft robotscontinual learningnonlinear dynamicshyper-redundant morphologieslocalized control
When Does In-Context Search Help? A Sampling-Complexity Theory of Reflection-Driven Reasoning
The paper presents a theoretical analysis of in-context search in large language models (LLMs), modeling it as approximate inference over reasoning traces where self-reflection provides feedback for posterior updates. The authors establish sampling complexity bounds, showing exponential improvements over base models when reflections reliably localize early mistakes (polynomial attempts vs. exponentially small zero-shot success), while demonstrating robustness to approximate updates and learnability via cross-entropy training. Theoretical predictions are validated empirically on large reasoning models, with connections drawn to optimal reinforcement learning policies under verifiable rewards.
in-context searchsampling complexityposterior updatesself-reflectionreasoning traces
Reliable and Developer-Aligned Evaluation of Agents for Software Engineering
The paper proposes a comprehensive evaluation methodology for LLM-powered agents in software engineering, addressing limitations of existing fragmented and syntactically biased assessments. The approach emphasizes contamination-awareness, in-the-wild agentic behavior analysis, and trajectory-aware benchmarks that capture realistic coding contexts, human-aligned behavior, and failure modes. This framework aims to provide developer-aligned reliability by grounding evaluations in real-world development practices rather than hypothetical scenarios.
llm-powered agentscontamination-awarenessagentic behaviortrajectory-aware benchmarkssoftware engineering evaluation
Vision Language Action (VLA) Models for Unmanned Aerial Robotics and Bimanual Manipulation: A Review
The review analyzes 183 Vision Language Action (VLA) model contributions (2017-2026) for bimanual manipulation and unmanned aerial robotics, identifying seven key dimensions: architectures, training recipes, action representations, coordination strategies, UAV control, language grounding, and system-level challenges. VLAs integrate visual perception, language understanding, and action generation into unified foundation models, enabling direct instruction execution (e.g., object folding or drone navigation) from visual inputs. Findings demonstrate transferability of bimanual coordination techniques (e.g., 7-DoF arm control) to UAV domains, with 14 identified cross-domain research directions.
vision language action modelsbimanual manipulationunmanned aerial vehiclesfoundation modelslanguage grounding
SPEAR: A Simulator for Photorealistic Embodied AI Research
SPEAR introduces a photorealistic simulator for embodied AI research that significantly advances programmability, rendering speed, and ground truth modalities compared to existing Unreal Engine-based solutions. The Python library connects to Unreal Engine via a modular plugin architecture, exposing over 14K UE functions and enabling deterministic execution of complex work graphs within a single frame. SPEAR achieves 73 FPS at 1920x1080 resolution while providing novel ground truth outputs (intrinsic image decomposition, material IDs, shading parameters) and demonstrates versatility across multi-agent control, procedural generation, and physics co-simulation applications.
photorealistic simulationunreal engineembodied aiprogrammable interfaceground truth rendering
tsbootstrap: Distribution-Free Uncertainty Quantification and Conformal Prediction for Time Series
The tsbootstrap library introduces distribution-free uncertainty quantification and conformal prediction methods for non-IID time series data, addressing limitations of existing tools. It implements block, residual, sieve, and wild resampling techniques alongside adaptive conformal calibrators (EnbPI, ACI, NexCP, AgACI) via a typed API. Empirical results show dependence-aware methods reduce coverage deficits observed with IID bootstrap under temporal dependence, with sieve bootstrap performing closest to nominal coverage under short-memory linear dependence. The system achieves computational efficiency through a compiled backend (faster than arch) and streaming reduction to limit memory overhead to O(B).
conformal predictiontime series bootstrapuncertainty quantificationadaptive calibrationnon-iid data
Digital Fragmentation and Generative AI Use Across 103 Million Application Events
This study quantifies digital fragmentation—frequent application switching—across 1,017 knowledge workers using 103M application events. Day-to-day variation (44.6%) exceeds individual (35.8%) and organizational (19.6%) differences, with fragmentation peaking midweek and resetting after breaks. Communication app use correlates with higher fragmentation, while generative AI use precedes more focused, predictable application patterns. Results suggest AI may structure rather than exacerbate fragmented workflows, highlighting the workday as a key intervention point.
digital fragmentationapplication switchingknowledge workersgenerative aiworkflow patterns
Danus: Orchestrating Mathematical Reasoning Agents with Fact-Graph Memory
Danus introduces a novel orchestration system for research-level mathematical reasoning, leveraging a shared fact graph as global memory management. The system comprises a main agent for planning and coordination, parallel worker agents for proof search, and a stateless verifier to validate mathematical claims before inclusion in the fact graph. Verified facts are stored with proofs and dependencies, enabling incremental construction of long arguments. Evaluated through six case studies in algebraic geometry, singularity theory, and combinatorics, Danus demonstrates effective scaling for long-horizon research problems. The system is open source.
fact graphproof searchmathematical reasoningorchestration systemstateless verifier
Diffusion enabled Optimal Transport distances for graph matching
A novel graph comparison method, Diffusion Semi-Relaxed Fused Gromov-Wasserstein (DsrFGW), unifies node features and structural connectivity through optimal transport while incorporating diffusion processes for information propagation. DsrFGW addresses limitations of traditional Gromov-Wasserstein variants (srGW, srFGW) by capturing local and global structural patterns and reducing sensitivity to noise or missing edges. Evaluated on 36 synthetic pairwise graph matching tasks, DsrFGW consistently outperforms srFGW, achieving accuracy improvements of 0-20 percentage points and significant Adjusted Rand Index (ARI) gains, particularly in medium-difficulty scenarios where srFGW often yields negative ARI. DsrFGW improves clustering quality in 92% of tasks under severe noise, demonstrating robustness in graph comparison under structural uncertainty.
diffusion processesoptimal transportgraph comparisongromov-wassersteinadjusted rand index
Healthier LLMs: Retrieval-Augmented Generation for Public Health Question Answering
The study extends PubHealthBench (7,929 UK public health QA pairs) to evaluate Retrieval-Augmented Generation (RAG) configurations for mitigating LLM hallucinations in evolving guidance. It systematically compares dense, sparse, and hybrid retrieval with multiple embedding models, finding hybrid retrieval improves recall and ranking, particularly with optimized chunk length and topic-aware context selection. Retrieval boosts multiple-choice accuracy across LLMs, enabling smaller models to match larger ones. A novel LLM-as-a-judge rubric assesses free-form answers, showing strong human agreement on faithfulness/completeness but weaker consistency/clarity metrics. Results emphasize retrieval quality as critical for reliable public health QA.
retrieval-augmented generationhybrid retrievalllm-as-a-judgepubhealthbenchcontext selection
The Rank-One Corner: How Much Value Equivalence Does a Task Need from a World Model?
The paper demonstrates that a world model's learned latent representations capture only as much task-relevant structure (closure) as required by the dimensionality of its training objective, not its capacity or observations. Using DreamerV3 in a controlled environment with known ground-truth closure, experiments show scalar reward objectives install 1D projections (R²=0.10), while full objectives recover higher-dimensional structure (R²=0.76). Sweeping objective dimensionality from 1 to 4 reveals matching rank in auxiliary and value heads, with reconstruction tasks bypassing this limitation. The results establish value equivalence as dimensional, not binary, governed by objective design.
world modelvalue equivalencelatent representationdimensionalitydreamerv3
At-Grok Is Not Converged:A Measurement-Validity Audit for Grokking Representation Metrics
The study audits representation metrics in grokking by demonstrating that network embeddings continue compressing long after generalization occurs, with effective rank overestimating converged values by 3-5x on MLPs and 1.3-1.5x on transformers. Using modular arithmetic tasks, the authors show compression lags accuracy transitions by ~10,000 steps, with LayerNorm ablation reducing compression fraction from 0.87 to 0.25. They introduce an audit toolkit to separate onset from compression, detect censoring, and validate reference floors, revealing a false-confidence bug in their own implementation. MLP-specific depth laws fail on transformers and invert under weight decay.
grokkingmodular arithmeticeffective ranklayer normweight decay
Specification Grounding Drives Test Effectiveness for LLM Code
The study demonstrates that grounding test cases in formal specifications significantly improves large language models' (LLMs) code correctness, outperforming ungrounded testing approaches by +38 percentage points across Claude models (Haiku 4.5, Sonnet 4.6, Opus 4.8) and +36 points on held-out data. By isolating the effect of specification grounding—via a prompt modification that provides the spec as a checklist—the authors show that test quantity alone is insufficient: doubling test budgets or combining ungrounded suites plateaus below grounded performance. Results generalize across vendors (GPT-5.3-codex +28, Gemini 3.5 Flash +19) and reduce false alarms from 33% to 0%, with task-level significance (p=0.002).
specification groundingtest effectivenessllm code generationfalse-alarm rateproperty-based testing
ProMoE-FL: Prototype-conditioned Mixture of Experts for Multimodal Federated Learning with Missing Modalities
ProMoE-FL introduces a Prototype-conditioned Mixture-of-Experts framework for multimodal federated learning with missing modalities, addressing limitations of existing methods that rely on public datasets or naive feature synthesis. The framework constructs a global client-aware prototype bank to capture modality priors across institutions, enabling direction-aware expert routing for dynamic missing feature synthesis. Extensive evaluations on four chest X-ray datasets (MIMIC-CXR, NIH Open-I, PadChest, CheXpert) demonstrate ProMoE-FL's superior performance in both homogeneous and heterogeneous settings compared to state-of-the-art methods.
multimodal federated learningmissing modalitiesmixture of expertsprototype bankfeature synthesis
Dynamic-in-Few-Step: Unifying Dynamic Computation and Few-Step Distillation for Efficient Video Generation
We introduce Dynamic-in-Few-Step, a framework unifying dynamic computation and few-step distillation for efficient video generation. The method integrates dynamic structural sparsification into the distillation process, jointly optimizing denoising steps and structured model sparsity to transform pre-trained Video Diffusion Models (VDMs) into step-specific Mixture-of-Models (MoM). A Progressive Training Strategy and Output Rollout Mechanism stabilize training, while a specialized inference engine ensures efficient deployment. On Wan-14B, the approach reduces per-step FLOPs by 24% atop 4-step distillation, achieving a 1.2x wall-clock gain and 30x speedup over the 50-step teacher while maintaining competitive generation quality.
video diffusion modelsfew-step distillationdynamic computationmixture-of-modelsstructural sparsification
UBEP: Re-architecting Expert Parallelism Communication Library for Production Superpods
UBEP introduces a communication library optimized for Mixture-of-Experts (MoE) models on high-bandwidth superpods, addressing three bottlenecks: BSP-induced serialization, synchronization overhead, and load imbalance. The method redesigns All-to-All primitives by leveraging unified global address spaces and high-bandwidth fabrics, enabling fine-grained parallelism and distance-aware scheduling. Large-scale experiments demonstrate 52.4% reduction in All-to-All latency and 11.1% improvement in Time Per Output Token (TPOT) for MoE inference.
mixture-of-expertsall-to-allsuperpodsynchronization overheadload imbalance
Cross-Trajectory Chimera Interventions Reveal Dissociable Roles of Weight Magnitude and Direction in Grokking
The paper introduces cross-trajectory chimera interventions to study causal portability of network properties across training runs. By decomposing weights into norm and direction components and recombining them between independently trained networks, the authors dissociate their roles in grokking. Results on modular arithmetic tasks show direction determines circuit identity (40/40 successful transfers) while norm governs solution susceptibility, with a threshold-like transfer localized to ±1/64 via adaptive bisection. The method reveals direction indexes solution convergence whereas norm controls identity overwritability.
grokkingcross-trajectory interventionsweight normweight directionmodular arithmetic
Open-Ended Scenario Reasoning for Specialist Model Adaptation
The paper proposes ROAM, a framework for adapting frozen specialist models to unseen industrial scenarios without retraining by leveraging LLM reasoning. ROAM confines corrections to a low-dimensional latent space, fuses LLM-generated scenario judgments with online observations probabilistically, and employs risk-constrained fallback to original models. Evaluations on mineral thickening and IndPenSim datasets show 20% MAE reduction in major shift scenarios with only 839 additional parameters and <0.02ms overhead per step, demonstrating effective LLM-based adaptation for deployed industrial models.
specialist model adaptationlatent space correctionrisk-constrained mechanismindustrial process modelingllm reasoning
AgentLens: Production-Assessed Trajectory Reviews for Coding Agent Evaluation
AgentLens introduces a production-assessed benchmark for interactive code agents, evaluating full execution trajectories rather than binary task completion. The method combines formal verification with LLM-generated trajectory reviews and side-by-side comparisons, providing explainable scores for agent behavior across instruction-following, tool usage, self-verification, and error recovery. Results demonstrate utility for model diagnosis, version comparison, and regression detection in a nightly evaluation pipeline. The benchmark is released as open-source.
code agentstrajectory evaluationformal verificationllm-written reviewsregression detection
LLM-Guided Task-Semantic Field Factorization for Industrial Process Forecasting
The paper proposes Task-Semantic Field Factorization (TSF), an LLM-guided framework for industrial process forecasting that leverages process documents to enhance time-series models. TSF constructs a task-semantic field offline using variable metadata and LLMs, then activates semantic information during inference without modifying the backbone architecture. Evaluations on industrial forecasting tasks show TSF reduces MAE by 6.4% on average (up to 25.5%) while adding only 1.8–3.0k parameters and <0.008 ms/step overhead, demonstrating lightweight deployment with measurable gains.
task-semantic field factorizationindustrial process forecastingsoft sensingtime-series backbonesllm-guided
Can Reinforcement Learning Efficiently Discover Price Manipulation?
This paper demonstrates that model-free reinforcement learning (RL) can identify price manipulation strategies more effectively than model-based approaches under certain market conditions. Using an Almgren-Chriss framework with non-linear permanent and linear temporary price impact, the authors compare a Deep Deterministic Policy Gradient (DDPG) agent against a model-based method relying on Sequential Least Squares Quadratic Programming. Results show that RL outperforms the model-based approach in intermediate volatility scenarios, even with limited training data and noisy parameter estimates, but fails in high volatility and underperforms in low volatility. The findings underscore RL's potential in financial control problems and associated risks.
almgren-chriss frameworkdeep deterministic policy gradientprice manipulationsequential least squares quadratic programmingmodel-free reinforcement learning
SpaR3D-MoE: Adaptive 3D Spatial Reasoning from Sparse Views Meets Geometry-Inductive Mixture-of-Experts
SpaR3D-MoE introduces an end-to-end framework enhancing Multimodal Large Language Models (MLLMs) with geometry-aware 3D spatial reasoning from sparse RGB inputs. The method combines adaptive spatiotemporal manifold sampling for keyframe extraction with a heterogeneous geometry-inductive Mixture-of-Experts (MoE) to resolve cross-modal contention. Evaluations on VSI-Bench, ScanQA, and SQA3D show state-of-the-art performance, including a 63.5 average score on VSI-Bench (7.8 points above baselines) and 35.4-51.4% relative improvements in spatial tasks.
multimodal large language models3d spatial reasoningmixture-of-expertsspatiotemporal manifoldgeometry-inductive
Overview of the NLPCC 2026 Shared Task 1: Difficulty-Aware Multilingual and Multimodal Medical Instructional Video Understanding Evaluation
The NLPCC 2026 shared task introduces Difficulty-Aware Medical Instructional Video Question Answering (DA-MIVQA), extending prior multilingual and multimodal medical video benchmarks by explicitly categorizing questions based on evidence complexity. DA-MIVQA includes three tracks: DA-TAGSV, DA-VCR, and DA-TAGVC, requiring varying levels of textual, visual, temporal, and procedural reasoning. The dataset, sourced from public medical instructional videos and manually annotated for difficulty, covers diverse medical scenarios. This benchmark evaluates systems on their ability to handle simple (text-based) versus complex (cross-modal) question answering tasks.
medical instructional videomultimodal question answeringdifficulty-aware evaluationtemporal answer groundingcross-modal reasoning
Inertia-1: An Open Exploration of Wearable Motion Foundation Models
The authors present Inertia-1, an open framework for exploring wearable motion foundation models using 18.2M hours of accelerometer data from diverse global sources. They systematically investigate data choices (sensor modality, placement, sampling rate), model choices (architecture, size), and training objectives across 15 downstream tasks including human activity recognition and disease prediction. Results demonstrate state-of-the-art performance while providing practical guidelines for motion representation learning in varied sensing conditions.
wearable motionfoundation modelsaccelerometer datarepresentation learninghuman activity recognition
The Key to Going Linear: Analysis-Driven Transformer Linearization
This work addresses the quadratic cost of causal self-attention in transformers by isolating state update design effects in a frozen-backbone regime. It demonstrates that softmax relies on key-dependent, rank-1 orthogonal projections, explaining delta-style networks' superiority over gated accumulation. Structural interventions—sink tokens, short convolutions, and fixed-budget cache routing—are introduced to reduce approximation errors. The linearization approach scales across LLaMA and Qwen models up to 32B parameters, outperforming prior post hoc baselines on MMLU and matching long-context retrieval in adaptive-caching frameworks.
self-attentionlinearizationsoftmaxcache routingdelta-style networks
ECGLight: Compute-Light Framework For Paper ECG Digitization and Myocardial Infarction Screening
The authors present ECGLight, a compute-light framework for paper ECG digitization and myocardial infarction (MI) screening, designed for resource-constrained settings. The end-to-end pipeline converts smartphone-captured paper ECGs into calibrated 12-lead signals using lightweight on-device processing, then performs MI detection with SHAP-based interpretability. Evaluated on 21,799 ECGs from PTB-XL and ECG-Matrix datasets, the system achieves 95.51% accuracy (F1=0.9519) for MI detection on PTB-XL and 88.89% accuracy (F1=0.8862) for OMI detection on ECG-Matrix, running in <30s per ECG on CPU-only hardware.
electrocardiographydigitizationmyocardial infarctionon-deviceinterpretability
Neural Operator-enabled Topology-informed Evolutionary Strategy for PDE-Constrained Optimization
The paper introduces Neural Operator-enabled Topology-informed Evolutionary Strategy (NOTES), a framework integrating dimensionality reduction, representation learning, and evolutionary optimization for PDE-constrained inverse design. NOTES combines a DeepONet-based neural operator with Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to optimize in a compact latent space encoding topology-aware priors, enabling transferable designs for unseen conditions. Evaluated on nanophotonic beam-deflector design governed by Maxwell's equations, NOTES reduces dimensionality from 256 to 25 while achieving over 95% efficiency, outperforming CMA-ES and topology optimization. In structural optimization, NOTES achieves compliance down to 246. The framework decouples topology learning from PDE solvers, enhancing flexibility and transferability.
neural operatorevolutionary strategydimensionality reductionpde-constrained optimizationtopology optimization
Any-Dimensional Learning by Sampling
The paper introduces a unified framework for analyzing generalization and sketching in machine learning models handling variable-size inputs (e.g., point clouds, sequences, graphs) via random sampling maps. The method employs generalized sampling techniques (with replacement, random binning, species sampling) tailored to domain symmetries and inter-size relations. Theoretical results provide explicit generalization bounds and sketching rates for function classes continuous under sampling, covering permutation-invariant transformers, graph neural networks, and homomorphism densities.
random sampling mapsgeneralization boundsvariable-size inputspermutation-invariant transformersgraph neural networks
How Data Shapes RoPE Frequency Usage: From Positional Scale Matching to Length Generalization
The study identifies a data-driven mechanism governing Rotary Position Embedding (RoPE) frequency usage in transformers, showing that frequencies are selected to match the relative-distance structure of training data. By formalizing a field-resolution tradeoff, the authors demonstrate that optimal frequency scales as $1/W$, where $W$ is the dependency profile width. This principle explains frequency usage patterns in synthetic and text-based data, linking mid-low frequency bands in language models to natural language's multi-scale dependency structure. Empirical results reveal that natural language exhibits approximate self-similarity across positional scales, enabling test-time frequency scaling to support long-context generalization. The findings highlight two forms of scale matching crucial for long-context generalization.
rotary position embeddingsfrequency scalingdependency profilefield-resolution tradeoffself-similarity
Max Out GRPO Signal: Adaptive Trace Prefix Control for Hard Reasoning Problems
AdaPrefix-GRPO introduces adaptive trace prefix control to enhance Group Relative Policy Optimization (GRPO) on hard reasoning problems. The method dynamically adjusts the length of correct solution prefixes during training, maintaining a 50% success rate to maximize gradient signal before withdrawing assistance. Implemented via data preparation and loss masking, it requires no trainer modifications. Evaluated on hard math problems, AdaPrefix-GRPO achieves 2.1x accuracy improvement for a 0.6B model, 1.6x for Qwen3-1.7B, and 1.7x for AIME at matched FLOPs, while reducing trace length by approximately half.
adaptive prefix controlgroup relative policy optimizationhard reasoning problemsgradient signal maximizationtrace length reduction
MedPMC: A Systematic Framework for Scaling High-Fidelity Medical Multimodal Data for Foundation Models
The authors introduce MedPMC, a systematic framework for curating high-fidelity medical image-text pairs from PubMed Central (PMC) literature to address data scarcity in medical multimodal foundation models. The automated pipeline processes 6.1 million PMC articles, yielding 11 million pairs with strong performance in figure detection (F1=96.5), caption alignment (ROUGE-L=85.3), and medical relevance (95.3% verified by clinicians). A CLIP-style model trained on MedPMC achieves 7.1 percentage point improvement in zero-shot AUC across 26 benchmarks and enhances visual question-answering (up to +16.9pp) and dermatology image retrieval (Recall@5 +11.7pp) compared to prior biomedical baselines.
multimodal foundation modelsmedical image-text pairszero-shot learningfigure-caption alignmentclinical relevance validation
PeTeR: Post-Training Robustification of Probabilistic Circuits
PeTeR introduces a data-free post-training framework for robustifying pre-trained probabilistic circuits (PCs) against distribution shifts without full retraining. The method leverages distributionally-robust optimization principles to enhance model resilience within Wasserstein-ball uncertainty sets, addressing overfitting and fragile generalization in standard PC learning. Evaluations on density estimation benchmarks show PeTeR outperforms data-dependent baselines in mitigating both random and adversarial perturbations.
probabilistic circuitsdistributionally-robust optimizationwasserstein ballpost-training robustificationdensity estimation
Does Bielik Know What It Doesn't Know? Activation Dispersion Separates Entity Familiarity from Factual Reliability Across Model Scale
The study demonstrates that activation dispersion metrics in Polish Bielik models (1.5B-11B parameters) reliably distinguish entity familiarity (known vs. fabricated entities) with AUROC 0.95-1.00 across four domains (athletes, cities, writers, musicians). Using unsupervised measures (inverse participation ratio, spectral entropy) on post-SwiGLU MLP activations, the method achieves near-perfect separation, persisting across layers and entity types (mean off-diagonal AUROC 0.92-0.99). While entity familiarity detection peaks at 1.5B, factual accuracy scales with model size (0-19/42 correct answers for athletes). Correct vs. hallucinated answer separation remains challenging (probe AUROC 0.93), and models rarely abstain (3/2520 refusals).
activation dispersionpost-swiglu mlpinverse participation ratiospectral entropyentity familiarity
Guidance Breaks the Fitted Operator: A Terminal-Fitted Repair for Classifier-Free Guidance
The paper proposes a terminal-fitted repair for classifier-free guidance (CFG) in diffusion models, addressing its instability at high guidance strengths. By analyzing CFG through numerical analysis, the authors identify that guidance re-stiffens the discriminative subspace, causing DDIM to diverge on coarse meshes. They introduce a modified guidance term, replacing CFG's $w(r-1)$ with $r^{1+w}-r$, which eliminates sigma_min-divergent blow-up and improves stability. Empirical results on CIFAR-10 and Stable Diffusion 1.5 DDIM show reduced residual amplification and saturation, with 9/9 point-FID wins over CFG, while preserving classifier-proxy accuracy.
classifier-free guidancediffusion modelsnumerical analysisterminal-fitted repairdiscriminative subspace
An optimal control approach for neural network architecture adaptation with a posteriori error estimation
The paper presents a novel optimal control framework for neural network depth adaptation using a posteriori error estimation. By formulating training as a continuous-time optimal control problem, the method derives rigorous error bounds that decompose approximation error across layers, enabling targeted insertion of new layers at high-error locations. The approach employs piecewise linear weight functions and dual weighted residual methodology from finite element analysis to compute error bounds. Experiments on scientific datasets, including Navier-Stokes parameter mapping, demonstrate superior generalization performance compared to existing architecture adaptation methods.
optimal controla posteriori error estimationarchitecture adaptationdual weighted residualnavier-stokes
Higher-Order Geometric Updates for Levenberg-Marquardt Method via Riemann Normal Coordinates
The authors propose Riemann-normal-coordinate Levenberg-Marquardt (RNC-LM), a higher-order geometric optimization method for nonlinear least-squares problems that improves parameter-effects curvature handling. RNC-LM extends geodesic acceleration via Riemann normal coordinates, enabling arbitrary-order corrections and finite-step updates with improved reparameterization consistency. A line search along RNC curves maintains cost proximity to standard LM while eliminating residual acceleration components order-by-order. Benchmarks demonstrate superior convergence in curved valleys (34× speedup on potential-energy-surface fitting) and PINN applications (L2 error ~1e-3).
levenberg-marquardtriemann normal coordinatesnonlinear least-squaresgeodesic accelerationparameter-effects curvature
Asymmetric Focal Loss Improves Graph Neural Network Prediction of Drug-Drug Interactions
The study demonstrates that asymmetric focal loss (ClinicalFocal) significantly improves graph neural network performance for drug-drug interaction prediction compared to binary cross-entropy. The method integrates ClinicalFocal into a relation-aware graph convolutional network using molecular fingerprints, physicochemical descriptors, and learned embeddings, evaluated on TWOSIDES via five-fold cross-validation. Results show accuracy improvements from 0.699 to 0.892 (+19.3pp), F1 from 0.700 to 0.894 (+19.4pp), AUROC from 0.766 to 0.914, and AUCPR from 0.714 to 0.860, with a 64.1% relative reduction in classification error (30.1% to 10.8%).
asymmetric focal lossgraph neural networksdrug-drug interactionsmulti-relational predictionclinicalfocal
Multi-Class vs. Multi-Label BERT for CVE-to-CWE Mapping: How Taxonomy Structure Shapes the Errors
The study compares multi-class and multi-label BERT formulations for mapping Common Vulnerabilities and Exposures (CVE) to Common Weakness Enumeration (CWE) categories, evaluating three transformer encoders (BERT Base, SecureBERT, CySecBERT) on nested label spaces (83, 47, 25 classes). Multi-class training achieves higher macro-F1 (gap narrowing from 21 to 2 percentage points with smaller label spaces), with threshold optimization closing the gap for 25 classes. Error analysis reveals confusion patterns follow the CWE hierarchy (Pearson r > 0.92), suggesting taxonomy design drives errors more than encoder choice. Hierarchy-relaxed evaluation improves macro-F1 from ~81% to ~90%, with CySecBERT performing best, especially in multi-label settings.
bertcvecwemulti-label classificationtransformer encoders
PALS: Percentile-Aware Layerwise Sparsity for LLM Pruning
PALS (Percentile-Aware Layerwise Sparsity) introduces a layerwise sparsity adjustment method for LLM pruning, dynamically setting per-layer sparsity based on the 99th percentile of activation magnitudes while maintaining ±5% bounds around the target ratio. Unlike uniform approaches like Wanda, PALS improves perplexity on LLaMA-2-7B (10.96 vs. 12.92 WikiText-2 at 50% sparsity) but shows architecture-dependent efficacy, with minimal gains on LLaMA-3-8B and none on Mistral-7B. The method is computationally lightweight, requires no fine-tuning, and demonstrates that gradient-based allocation underperforms random pruning.
layerwise sparsityllm pruningactivation magnitudeone-shot pruningperplexity improvement
Avoiding unsafe sets when training with Langevin Dynamics
The paper analyzes the probability of Langevin dynamics trajectories entering unsafe regions during training of strongly convex loss functions. It establishes three bounds: (1) equilibrium mass exponentially small in dimension $d$, (2) trajectory probability converging to equilibrium after $O(d)$ burn-in time via global spectral gap, and (3) improved bounds using local relaxation rates for geometrically isolated regions. Results demonstrate that strong convexity governs relaxation speed while unsafe set geometry determines transient trajectory behavior, with Ornstein-Uhlenbeck examples showing dimension-dependent transient swelling.
langevin dynamicsstrong convexityspectral gapunsafe setstrajectory probability
A Unified Detection Framework for AI-Related Content and Artifacts
The authors propose a unified detection framework for AI-related content and artifacts using Mahalanobis distance scores (MDS), applicable to LLM-generated text, hallucination, watermark, and adversarial example detection. The method characterizes the positive class (e.g., human-generated text) by estimating the covariance matrix of deep representations via joint casewise and cellwise minimum covariance determinant (MCD) estimators, addressing homogeneity and heterogeneity in multi-class positive samples. Efficient optimization algorithms are developed for both estimators, with proven convergence and high breakdown point properties. Empirical evaluations demonstrate the framework's effectiveness.
mahalanobis distance scoresminimum covariance determinantdeep representationsbreakdown pointjoint estimation
Gradient-free Riemannian Langevin Sampler
We propose Gradient-free Riemannian Langevin Sampler (GRiLS), a novel MCMC method for efficient sampling of multimodal distributions without requiring gradient evaluations. GRiLS introduces a Riemannian metric that reshapes local geometry to facilitate mode transitions, making it suitable for complex targets where derivatives are unavailable. The method estimates target density mean and covariance using an ensemble of interacting particles. Empirical evaluations on multimodal benchmarks demonstrate GRiLS's superior mixing compared to both gradient-based and gradient-free MCMC approaches.
markov chain monte carloriemannian metricmultimodal distributionsgradient-free samplingensemble estimation
Fast Rates for Semi-Supervised Learning via Data-Augmentation Graph Regularization
(No summary returned.)
GIFT: Geometry-Informed Low-precision Gradient Communication for LLM Pretraining
GIFT introduces a geometry-informed gradient scaling method for low-precision communication in LLM pretraining, addressing anisotropic gradient distortion via near-isotropic space transformation. The method employs a simplified geometry-aware transformation algorithm with low-rank approximation and selective application, maintaining optimizer and communication format compatibility. Evaluated on Llama-300M and Llama-600M, GIFT reduces pretraining time by 7.6% on 64 NVIDIA GH200 Superchips while improving downstream task preservation over direct FP8 communication.
gradient communicationlow-precision quantizationanisotropic gradientsgeometry-aware transformationllm pretraining
FourierQK: Spectral Preprocessing of Query-Key Projections Improves Transformer Attention
FourierQK introduces spectral preprocessing of query-key (Q/K) projections to enhance transformer attention in character-level language modeling. The method applies FFT-based spectral filtering to Q/K projections, preserving the full attention score structure while enabling global frequency-domain mixing. On TinyShakespeare, spectral preprocessing with four learned frequencies achieves a validation loss of 0.309 (Δ=+1.166), a 79% improvement over standard dot-product attention. The learned frequencies converge to a multi-scale ordering (49, 27, 10, 6 tokens/cycle) corresponding to paragraph, sub-paragraph, phrase, and word scales. Gains are specific to spectral preprocessing, as random orthogonal/non-orthogonal projections yield no improvement. The approach differs from FNet by preprocessing only Q/K projections rather than replacing attention with Fourier mixing.
spectral preprocessingquery-key projectionsfft-based filteringcharacter-level language modelingmulti-scale ordering
Statistical inverse learning and $\ell^1$-regularization
(No summary returned.)
Reliable mechanistic operator recovery with biologically-informed neural networks: principles for architecture and optimisation design
The study establishes evidence-based guidelines for reliable mechanistic operator recovery using biologically-informed neural networks (BINNs) by systematically evaluating architecture and optimization choices. Focusing on 1D advection-diffusion-reaction PDEs, it examines network expressivity, learning rate, loss weighting, and batch size effects on optimization stability and operator recovery. Results indicate moderately expressive architectures, intermediate learning rates, balanced loss terms, and intermediate batch sizes optimize performance, with diagnostics provided for common failure modes like overfitting and unstable optimization.
biologically-informed neural networksmechanistic operator recoveryadvection-diffusion-reactionoptimization stabilitypartial differential equations
The Optimal Sample Complexity of Learning Autoregressive Chain-of-Thought
The paper establishes the optimal sample complexity for learning autoregressive Chain-of-Thought (CoT) traces in the realizable PAC setting. By introducing parity dimension—a rollout-stable refinement of Daniely--Shalev-Shwartz (DS) dimension—the authors prove that exact-trace learning has sample complexity $O((\operatorname{DSdim}(\mathrm{H}) + \log(1/\delta))/\varepsilon)$, independent of rollout length. The analysis reveals that DS dimension can increase under rollout, necessitating the parity dimension detour. Results demonstrate worst-case optimal dependence on $\operatorname{DSdim}(\mathrm{H})$, with one-step stopping recovering standard multiclass learning.
autoregressive learningchain-of-thoughtsample complexityparity dimensionpac learning
TF-Engram: A Train-Free Engram with SSD-Backed Memory for Large Language Models
TF-Engram introduces a train-free Engram system for Large Language Models (LLMs) that constructs phrase-specific semantic memory offline from external corpora and stores it across a GPU--DRAM--SSD hierarchy. The system employs Early-Exit Guided Predictive Prefetching to mitigate external-memory latency during autoregressive decoding. Evaluated on Qwen3-0.6B, TF-Engram improves the average downstream score from 57.6 to 59.4, surpassing both the frozen backbone and a parameter-matched LoRA baseline. Results indicate that large TF-Engram tables can be built with moderate offline cost, SSD-backed storage reduces GPU memory demand, and predictive prefetching recovers throughput loss from external memory access, demonstrating scalable integration of static phrase memory into LLM inference.
engrampredictive prefetchingautoregressive decodingsemantic memoryssd-backed storage
Sparse Delta Memory: Scaling the State of Linear RNNs through Sparsity
The paper introduces Sparse Delta Memory (SDM), a method to scale hidden state capacity in gated linear RNNs via sparse addressing, addressing the recall-performance trade-off in linear attention models. SDM extends Gated DeltaNet by replacing dense key-value outer products with sparse reads/writes to a large explicit memory. Under isoFLOP constraints, SDM with higher state capacity improves in-context learning and long-context retrieval, while learned initial states enhance performance on common-knowledge and reasoning tasks.
sparse delta memorylinear attentionin-context learninggated deltanetlong-context retrieval
Generalist Vision-Language Models for Fast Radio Burst detection: a zero-shot benchmark against a specialized detector
The study evaluates zero-shot performance of generalist Vision-Language Models (VLMs) for Fast Radio Burst (FRB) detection against specialized deep learning models. Using Gemma 4 2B and 4B VLMs without fine-tuning on 2000 balanced samples of simulated L-band dynamic spectra, the VLMs achieve 93.65% accuracy, comparable to SwinYNet's 92.90%, with lower false-positive rates on structured RFI (6.4% vs. 25.0%) and zero false positives on noise. Prompt engineering enables three-class FRB/RFI/noise classification on 3000 spectra with 86% accuracy and no false FRB detections.
vision-language modelsfast radio burstszero-shot learningdynamic spectraradio frequency interference
R^3: Advertisement Compliance Rectification via Group-Relative Experience Extractor and Curriculum Reinforcement
The paper introduces R^3, a framework for rectifying textual violations in video advertisements while preserving semantic intent. The method combines three innovations: (1) a group-relative compliance experience extractor for data synthesis, (2) curriculum reinforcement learning with hierarchical rewards for compliance-semantics trade-off, and (3) an end-to-end video rectification pipeline integrating text recognition, rewriting, and re-rendering. Experiments on industrial datasets and online A/B tests show R^3 outperforms state-of-the-art baselines in balancing violation correction with intent preservation.
advertisement complianceexperience extractorcurriculum reinforcementtext rectificationsemantic preservation
Mechanistic Interpretability for Neural Networks: Circuits, Sparse Features and Symbolic Reasoning
The paper presents a systematic framework for mechanistic interpretability in neural networks, addressing the challenge of reverse-engineering internal algorithms beyond surface-level explanations. It analyzes Transformer circuits through residual streams, attention mechanisms, and induction heads to explain in-context learning behaviors. Methods like Sparse Autoencoders and transcoders disentangle polysemantic activations into interpretable features, while steering vectors enable controlled model interventions. The work bridges neural representations with neurosymbolic AI for symbolic rule extraction, advancing safety-critical model auditing capabilities.
mechanistic interpretabilitytransformer circuitssparse autoencoderspolysemanticitysteering vectors
Nonlinear Bandit
(No summary returned.)
BubbleSH: A Dataset of Rising Bubbles with Deformable Interfaces
The authors introduce BubbleSH, a dataset for studying bubbly flows featuring transient 3D bubble-swarm dynamics from high-fidelity direct numerical simulations. The dataset captures time-resolved trajectories, velocities, and shape evolution of deformable bubbles, represented compactly via spherical harmonics. Designed for data-driven modeling, it enables analysis of kinematic and morphological interactions, with evaluation metrics for trajectory and shape prediction. The dataset's sensitivity to perturbations makes it suitable for generative models learning future trajectory distributions. A permutationally and translationally equivariant probabilistic emulator is evaluated, establishing BubbleSH as a benchmark for chaotic multiphase systems.
bubbly flowsspherical harmonicsdirect numerical simulationsdata-driven modelingmultiphase systems
Safe Reinforcement Learning using Ideas from Model Predictive Control
The paper proposes a framework combining deep reinforcement learning (DRL) with model predictive control (MPC) to ensure strict safety constraints during policy learning. The method uses offline MPC computations to define a feasible state-action space, then projects the RL agent's actions onto this verified safe set via a safety filter. Evaluation on a non-linear 1-DoF testbed demonstrates successful exploration and policy convergence while maintaining safety guarantees on physical hardware.
safe reinforcement learningmodel predictive controlfeasible state-action spacesafety filternon-linear control
HPG-Diff: Hierarchical physics-guided diffusion with differentiable connectivity constraints for topology optimization
HPG-Diff introduces a hierarchical physics-guided diffusion framework for topology optimization, addressing poor generalizability and floating material artifacts in deep generative approaches. The method combines hierarchical physics guidance (aligning precomputed physics features with denoising) with a differentiable connectivity constraint (thermal conduction-inspired floating material suppression). Evaluations show 0.87% in-distribution and 5.29% out-of-distribution compliance errors, with floating material ratios reduced to 2.90% and 2.44%. Case studies demonstrate LoRA fine-tuning enables adaptation to rectangular domains.
topology optimizationdiffusion modelsphysics-guided learningdifferentiable constraintsfloating material suppression
Continual Learning With Participation Privacy: An Auditable Buffering-Aggregation Recipe
The paper introduces an auditable modular framework for ensuring participation privacy in continual learning systems under single-edit neighboring user streams, where insertions/deletions shift subsequent updates. The method employs a randomized buffering wrapper to emit bins of size $[U,2U]$, reducing the problem to a Hamming-style per-bin update stream with explicit backlog/delay guarantees, where $U$ is calibrated by privacy parameters $(\varepsilon,\delta)$. A certification theorem proves that non-adaptive Hamming-neighbor DP proofs lift to adaptive inputs if the primitive uses fresh per-round randomness and has a stable one-round privacy profile. The approach achieves trajectory-level $(\varepsilon,\delta)$-DP with a clear privacy--latency trade-off via $U$.
continual learningparticipation privacydifferential privacyadaptive interactionbuffering-aggregation
Geometric--Nongeometric Optimizer Calculus: A Modular Language for Reachable Gradient Methods
The paper introduces geometric--nongeometric optimizer calculus, a modular framework for analyzing gradient-based optimization methods under explicit constraints. The approach decomposes optimizers into geometric (positive cometric families mapping gradients to directions) and nongeometric modules (information, memory, control, etc.). Key results include a direction-expressivity theorem proving full positive-definite geometry expresses all strict descent directions, plus exact expressivity conditions for diagonal/block geometries. The framework enables trajectory-level residual complexity analysis and Pareto optimization over module budgets. Diagnostic prototypes demonstrate the language's utility, though no large-scale performance claims are made.
optimizer calculuscometric familydirection-expressivityresidual complexitymodule budgets
Restricted Dynamic Geometric Complexity: Certificates for Structured Preconditioning
The paper introduces restricted dynamic geometric complexity, a framework for certifying condition-number improvement when optimization metrics are constrained to specific families. It formulates the problem geometrically, proving monotonicity and submanifold-distance principles, with reachability characterized via linear matrix inequalities. Key results include exact diagonal complexity formulas, Kronecker projection theorems, and computable mismatch certificates, validated on synthetic quadratic instances. The approach reframes preconditioner design as geometric reachability problems.
geometric complexitycondition numberkronecker projectionlinear matrix inequalitypreconditioning
Comparative Study of Domain-adapted VLMs for General Document Visual Question Answering
This study conducts a comprehensive evaluation of 8 open-source pretrained Vision-Language Models (VLMs) on Document Visual Question Answering (DocVQA) across three document domains: industrial documents, infographics, and presentation slides. Performance was assessed under zero-shot, fully supervised finetuning, and few-shot learning scenarios, including inter- and intra-dataset evaluations. Results indicate that while large VLMs excel in structured layouts, they struggle with visually complex layouts. Supervised finetuning yields higher relative gains in smaller architectures, and few-shot experiments reveal visual understanding as the primary bottleneck. Finetuning with 50 target domain samples enables rapid adaptation, sometimes surpassing fully supervised performance.
vision-language modelsdocument visual question answeringzero-shot evaluationfew-shot learningsupervised finetuning
Information Allocation Dynamics in Neural Network Optimization
(No summary returned.)
Prior-matched evaluation of operational Earth-observation classifiers: a three-number reporting method demonstrated on Sentinel-1 internal-wave detection
The paper introduces a prior-matched evaluation method for operational Earth-observation classifiers, addressing systematic overestimation of precision when test priors mismatch operational conditions. The proposed three-number reporting framework contrasts balanced-test, operational-prior, and post-deployment metrics, revealing a precision gap from 0.794 to 0.192 in Sentinel-1 internal-wave detection. Methodologically, the approach enforces fixed recall (0.80), prior correction, and leakage-controlled development, achieving 0.927 operational precision through feature aggregation while maintaining temporal generalization.
prior-matched evaluationrare-event detectionprecision-recall tradeoffsentinel-1operational classifier
Fractal KV-Cache Archives: Lossless Symbolic Storage with In-Place Retrieval for Long-Context LLM Inference
Fractal KV-Cache Archives introduces a lossless symbolic storage method for quantized key-value (KV) caches in long-context LLM inference, leveraging contractive iterated-map codes for efficient serialization. The approach supports O(1) random access and amortized append, optimizing memory usage while maintaining exact retrieval. Experiments on GPT-2 with 1024-token contexts demonstrate a 36-54x reduction in cache size via per-head residual vector quantization, with a perplexity increase of 11-15%. The method also enables approximate substring queries directly on stored vectors, decoding matched context without reconstructing surrounding text. Code is released for reproducibility.
kv-cachevector quantizationcontractive iterated-map codeslong-context inferencesubstring queries
Distributed Sparse Interventions in Language Models
The paper introduces Distributed Sparse Interventions (DSI), a method for localized, neuron-level interventions in language models that accounts for nonlinear interactions across layers. DSI identifies sparse neuron sets (as few as 0.01% of total neurons) that activate task-specific behaviors, outperforming global activation-space steering approaches. Experiments across multiple tasks demonstrate DSI's effectiveness in task control and neuron set localization, providing insights into task composition and individual neuron roles through set-based computations.
distributed sparse interventionsneuron-level interventionstask compositionnonlinear interactionsmodel steering
Weight-Space Physics: Interpretable Hypernetworks for Lattice Quantum Field Theories
The paper introduces JEPAWG, a Joint-Embedding Predictive Architecture-based Weight Generator for interpretable hypernetworks in lattice quantum field theories. The method maps coupling constants directly to flow weights via a learned latent space, enabling physics extraction from network parameters without simulation. On 2D scalar field theories (lattice sizes $6^2$ to $11^2$), JEPAWG's latent space recovers the intrinsic manifold dimension, identifies phase transitions, and encodes finite-size scaling consistent with 2D Ising exponent $ν≈1$. The model outperforms PCA, AE, and VAE baselines in interpolation/extrapolation tasks and handles weight-space discontinuities from multi-seed training.
hypernetworkslattice field theoryinterpretabilitynormalizing flowsjoint-embedding
A knowledge-augmented dataset of high-risk driving scenarios with LLM annotations for autonomous driving
The paper introduces K-Risk, a knowledge-augmented dataset for high-risk driving scenarios, combining structured trajectories with LLM-generated semantic annotations to address under-representation of rare safety-critical events. The method integrates 20 trajectory datasets from diverse regions and road types, applying a unified risk-centric pipeline to curate 31,398 high-risk events, including 1,036 near-collision cases. Each event includes synchronized trajectory-metadata-language triplets with scenario descriptions, behavior alerts, and validated causal analyses, enabling interpretable supervision for risk-aware autonomous driving agents.
autonomous drivingrisk-awaretrajectory datasetssemantic annotationsnear-collision
Structural Adversarial Attacks on Relational Deep Learning under Integrity Constraints
The paper investigates adversarial attacks on Relational Deep Learning (RDL) systems under database integrity constraints. The authors propose a white-box attacker that perturbs foreign-key references while preserving schema constraints (e.g., FK validity, functional dependencies), formulating a combinatorial optimization problem. They evaluate seven attack heuristics—including gradient-guided variants using differentiable edge masks—on the RelBench rel-f1 benchmark, finding gradient-based methods outperform random baselines on regression tasks but show limited gains for classification due to label-flip resistance and output stability.
relational deep learningadversarial attacksintegrity constraintsgraph neural networksdifferentiable edge masks
Is Randomness Necessary for Adaptive Data Analysis?
The paper resolves the open question of whether randomness is necessary for Adaptive Data Analysis (ADA) against computationally unbounded analysts. By analyzing deterministic mechanisms in the Random Oracle model, the authors demonstrate that randomness is strictly required to answer more than a trivial number of adaptive queries: any deterministic mechanism fails after $k = \tilde{O}(n)$ queries, contrasting with randomized mechanisms that support $k \approx n^2$ queries. This establishes a fundamental separation between deterministic and randomized approaches to ADA.
adaptive data analysisrandom oracle modeldeterministic mechanismsstatistical queriescomputational unboundedness
Prior-aware and Context-guided Group Sampling for Active Probabilistic Subsampling
The authors propose Prior-aware and Context-guided Group-based Active Deep Probabilistic Subsampling (PGA-DPS), an enhanced version of Active Deep Probabilistic Subsampling (A-DPS) that integrates deterministic prior-informed sampling patterns and group-based top-k sampling to improve optimization robustness. PGA-DPS jointly optimizes subsampling patterns and downstream task models while leveraging dataset priors and group sampling strategies. Theoretical analysis supports the optimization benefits of group sampling. Empirical evaluations on MNIST, CIFAR-10, fastMRI knee, and hyperspectral AeroRIT datasets demonstrate PGA-DPS's superior performance over A-DPS and other methods in classification, image reconstruction, and segmentation tasks.
subsamplingprobabilistic modelinggroup samplingoptimizationprior-aware
An Hybrid Quantum-Classical Diffusion Model for Image Generation
The authors propose a hybrid quantum-classical diffusion model for image generation, combining a classical autoencoder with a mixed-state quantum denoising diffusion probabilistic model (MSQuDDPM). The autoencoder reduces dimensionality to enable quantum processing in a small-qubit Hilbert space, while MSQuDDPM simplifies reverse dynamics by predicting clean state estimates and using analytic backward propagation. Demonstrated on MNIST, this approach shows quantum diffusion's potential for hybrid generative modeling under limited qubit constraints.
quantum diffusionmixed-stateautoencoderdenoisinglatent space
Behavior Leverage Imbalance in Multi-Teacher On-Policy Distillation
The paper identifies behavior leverage imbalance in multi-teacher on-policy distillation (OPD) for tool-using language models, where local token-level signals disproportionately influence global generation modes despite balanced aggregate losses. It proposes Soft Clamp, a method that dynamically compresses extreme token-level Jensen-Shannon divergence while maintaining gradients, to mitigate over-calling tools. Evaluated on APIGen-MT, Soft Clamp reduces over-calling by 4.7 percentage points (13.7% to 9.0%) versus vanilla generalized knowledge distillation (GKD) while preserving decision accuracy, and also improves multi-turn tool-use diagnostics in BFCL.
on-policy distillationbehavior leverage imbalancejensen-shannon divergencetool-usemulti-teacher learning
Gauge-Invariant Learnable Spectral Positional Encodings for Directed Graphs via Hermitian Block Krylov Subspaces
The paper introduces gauge-invariant spectral positional encodings (PEs) for directed graphs using Hermitian block Krylov subspaces. The method constructs PEs as matrix functions of normalized magnetic operators, ensuring gauge invariance via learnable scalar spectral responses and random probes. It achieves computational efficiency with O(log(1/ε)) block steps and demonstrates superior performance on directed SBMs, matching exact eigendecomposition oracles as depth increases. The approach also generalizes to undirected graphs, improving heterophilous benchmarks over baselines.
spectral positional encodingshermitian block krylov subspacesgauge invariancemagnetic laplaciansheterophilous benchmarks
Constrained Decoding for Diffusion Language Models via Efficient Inference over Finite Automata
The paper introduces an exact algorithm for constrained decoding in diffusion language models, enabling sampling from mean-field posteriors under finite automaton constraints. By modeling automata as graphical models, the method guarantees constraint satisfaction while supporting greedy and sampling-based decoding, with depth-reduction techniques achieving logarithmic sampling depth. Evaluations on Dream-7B and LLaDA-8B demonstrate significant accuracy improvements (e.g., 63.9% to 71.5% on BFCL-Live) with minimal overhead (<5%) compared to unconstrained baselines.
constrained decodingdiffusion language modelsfinite automatamean-field posteriordepth-reduction
Online Data Selection Is Implicit Alignment
The paper demonstrates that online data selection during supervised fine-tuning (SFT) acts as an implicit alignment mechanism, inducing behavioral drift without explicit preference optimization. It formalizes online selection as a reweighted SFT objective, where the scorer defines implicit preferences over response styles and safety postures. Experiments with random, loss-based, quality-based, and diversity-based selectors show divergence in refusal rate, verbosity, and sycophancy, predictable from selected data attributes. The authors introduce Alignment Drift Auditing (ADA) for quantifying drift and Alignment-Aware Selection (AAS) to constrain unwanted behavioral shifts while maintaining data efficiency.
supervised fine-tuningonline data selectionbehavioral driftalignment drift auditingimplicit alignment
Tensorized algorithms and scalable filtering methods for hidden Markov and factorial hidden Markov models
The authors propose tensorized algorithms for scalable filtering in factorial hidden Markov models (fHMMs), avoiding the state-space explosion of equivalent HMM representations. By leveraging tensor algebra to exploit fHMMs' multidimensional structure directly, their method accelerates the forward filtering algorithm crucial for evaluation, decoding, and estimation. This approach significantly reduces computational costs, enabling efficient analysis of large-scale systems and datasets that were previously intractable with conventional HMM reformulations.
factorial hidden markov modelstensor algebraforward filteringstate-space explosiontime-series analysis
Dissociating the Internal Representations of Sycophancy in LLMs
This work dissociates sycophancy in Large Language Models (LLMs) into factual and opinion subtypes, addressing whether their internal mechanisms reflect this multi-faceted nature. Using linear probes and steering vectors on model activations, the authors measure representation transfer between subtypes to assess shared mechanisms. Results indicate varying representational structures across LLMs, with some exhibiting unified representations and others showing distinct, causally interfering mechanisms. This dissociation framework provides a methodological advance for analyzing complex behaviors in LLMs.
sycophancylinear probessteering vectorsinternal representationsllms
UP: Unbounded Positive Asymmetric Optimization for Breaking the Exploration-Stability Dilemma
The paper introduces Unbounded Positive Asymmetric Optimization (UP), a universal plug-and-play objective to address the exploration-stability dilemma in reinforcement learning (RL) for large language models. UP restructures optimization by anchoring the policy to its current state via the stop-gradient operator, enabling unclipped gradients for positive advantages to maximize exploration while maintaining clipping safeguards for negative advantages to ensure stability. The method extends across token-level (GRPO, DAPO) and sequence-level (GSPO) frameworks. Experiments demonstrate UP's effectiveness in enhancing exploration capacity and reasoning accuracy across diverse RL algorithms, model architectures, and training modalities.
reinforcement learningexploration-stability dilemmastop-gradient operatorimportance samplingpolicy update
EdgeCompress: Coupling Multidimensional Model Compression and Dynamic Inference for EdgeAI
The paper proposes EdgeCompress, a multidimensional CNN compression framework combining dynamic image cropping (DIC) and compound shrinking (CS) to reduce computational redundancy in both input images and network architectures. DIC employs a lightweight foreground predictor to crop informative regions, while CS compresses depth, width, and resolution dimensions based on their accuracy-computation tradeoffs. The framework further incorporates dynamic inference by cascading models of varying complexity. On ImageNet-1K, EdgeCompress reduces ResNet-50 computation by 48.8% while improving top-1 accuracy by 0.8%, outperforming HRank by 4.1% accuracy at similar computation.
convolutional neural networksmodel compressiondynamic inferenceimage croppingcomputational efficiency
Robust Federated Learning Under Real-World Client Churn
FeLiX introduces a federated learning orchestration framework addressing three challenges in production deployments: transient client availability, dynamic data heterogeneity, and prediction-outcome delays. It employs streaming-aware availability tiers, fresh-utility selection, and informativeness-aware delay-robust aggregation to minimize wall-clock time-to-target accuracy on live interaction streams. FeLiX achieves near-oracular performance without relying on unrealistic client availability knowledge. Evaluations on CIFAR-10, Google Speech, and low-availability traces demonstrate reductions in wall-clock time-to-target accuracy by up to 2.37X and communication bandwidth by 1.30X compared to state-of-the-art synchronous and asynchronous FL baselines.
federated learningclient availabilitydata heterogeneitydelay-robust aggregationwall-clock time
MILES: Modular Instruction Memory with Learnable Selection for Self-Improving LLM Reasoning
The paper introduces MILES, a framework for self-improving LLM reasoning through modular instruction memory with learnable selection. MILES dynamically expands step-wise memory using asymmetric sub-goal embeddings and sub-instructions, each with a learnable selection head, enabling coarse-to-fine retrieval. It trains selection heads from confident samples and applies them to rerank candidates for uncertain samples. Experiments show MILES matches or outperforms prior methods with superior accuracy-efficiency tradeoffs, demonstrating effectiveness, robustness, and transferability.
modular instruction memorylearnable selectionself-improving reasoningcoarse-to-fine retrievalsub-goal embeddings
Rethinking Multimodal Time-Series Forecasting Evaluation
The paper introduces TimesX, a context-enriched multimodal time-series forecasting benchmark addressing limitations of existing benchmarks: poor generalization from small-scale synthetic data, limited textual context types, and evaluation data leakage. TimesX employs an automated pipeline to generate diverse real-world time series with rich textual contexts across multiple domains. Empirical evaluation shows that zero-shot multimodal forecasting methods performing well on existing benchmarks often fail on TimesX, while simple ensemble methods leveraging textual context outperform baselines.
multimodal forecastingtime-series benchmarktextual contextzero-shot learningensemble methods
Flow-ERD: Agent-type Aware Flow Matching with Entropy-Regularized Distillation for Diverse Traffic Simulation
The authors propose Flow-ERD, a multi-agent traffic simulator that jointly optimizes for realism and diversity through two key components: Agent-Type Aware Flow Matching (AFM) preserves type-specific kinematic consistency while enabling multi-modal behavior generation, and Entropy-Regularized Distillation (ERD) fine-tunes rollouts with an entropy-regularized reverse-KL objective to prevent mode collapse. Evaluated on the WOSAC benchmark, Flow-ERD achieves state-of-the-art performance on both realism and a novel diversity metric, dominating the realism-diversity Pareto frontier among reproducible baselines.
traffic simulationflow matchingmulti-agent systemsentropy regularizationreverse-kl divergence
Physical activities enable scalable foundation modelling for broad-spectrum health prediction
The authors propose StepFM, a foundation model for broad-spectrum health prediction using only step counter data, addressing privacy and scalability limitations of high-frequency sensor models. The method employs scalable pre-training on temporal dynamics and behavioral patterns from large-scale step sequences, enabling transfer across 20+ health risk prediction tasks. Experiments show StepFM achieves robust performance across heterogeneous settings while revealing interpretable activity-health relationships, establishing step-based sensing as a practical foundation for health monitoring.
foundation modelstep counterhealth predictiontemporal dynamicstransfer learning
Mathematical methods of reinforcement learning
The survey provides a unified mathematical framework for reinforcement learning (RL), organizing key structures from probability, optimization, and operator theory. It begins with Markov decision processes (MDPs) and Bellman operators, analyzing contraction mappings, monotonicity, and fixed-point theory to derive convergence guarantees for value/policy iteration and temporal-difference methods. The optimization perspective covers stochastic approximation, convex duality, and regularization techniques like mirror/proximal methods. Function approximation is addressed via linear/non-linear settings, with finite-sample bounds and asymptotic results. The work also includes off-policy evaluation, constrained RL, and CMDPs, offering a cohesive entry point for interdisciplinary researchers.
markov decision processesbellman operatorsstochastic approximationconvex dualityfunction approximation
Intrinsic-Noise Consolidation: A Doob-Barrier-Conditioned Diffusion Turns Analog Device Noise into a Continual-Learning Resource
The paper introduces intrinsic-noise consolidation, a method that repurposes analog neuromorphic hardware's intrinsic noise for memory consolidation by conditioning synaptic weight dynamics on a Doob barrier. The approach adds a noise-dependent restoring force (σ² d/dw log h) that prevents weights from crossing critical memory barriers, differing from prior anchored-drift methods. Experiments on Split-MNIST (8 seeds) show a 10.9-point retention improvement at an optimal noise level (p=0.004), with an inverted-U performance curve unique to barrier conditioning. Validation on BrainScaleS-2 hardware demonstrates 15.6-point better prior-task retention without net-accuracy loss, leveraging intrinsic noise as a resource.
doob barrierneuromorphic hardwarememory consolidationdiffusion dynamicssynaptic plasticity
Latency-Constrained DNN Architecture Learning for Edge Systems using Zerorized Batch Normalization
The authors propose a latency-constrained neural architecture learning method for edge devices, introducing zerorized batch normalization and a hardware-customized latency predictor to optimize models via one-shot training. Their approach jointly optimizes accuracy and inference time by dynamically pruning neurons to meet strict latency constraints while minimizing accuracy loss. Experiments on ImageNet-100 demonstrate effectiveness: GoogLeNet's latency reduced from 40.32ms to 34ms (-0.14% accuracy) on Jetson Nano, while VGG-19 achieves 34ms (119.98ms→34ms) with +0.5% accuracy on Jetson TX2. The framework is open-sourced.
edge computingneural architecture searchbatch normalizationlatency predictionmodel compression
Compass: Prostate Cancer Detection Needs Multi-View Context
Compass introduces a transformer-based AI framework for prostate cancer detection using multi-view micro-ultrasound ($μ$US) context, addressing limitations of single-frame analysis. The method aggregates evidence across rotational sweep videos and biopsy-acquired frames, conditioning on probe angle via a transformer, and predicts frame- and study-level risk scores. Evaluated on a multi-center clinical trial dataset, Compass outperforms baseline AI methods and clinical expert risk scores, demonstrating the value of multi-view context in $μ$US-based diagnosis.
prostate cancer detectionmicro-ultrasoundmulti-view contexttransformerevidence aggregation
Smart Scissor: Coupling Spatial Redundancy Reduction and CNN Compression for Embedded Hardware
The paper proposes Smart Scissor, a unified framework combining dynamic image cropping and CNN compression for efficient edge deployment. The method first employs a lightweight foreground predictor to crop salient regions, reducing spatial redundancy while preserving accuracy at low resolutions. It then applies compound shrinking to compress CNN depth, width, and resolution dimensions. On ImageNet-1K, Smart Scissor reduces ResNet50's FLOPs by 41.5% while improving top-1 accuracy by 0.3%, outperforming HRank by 4.1% accuracy at equivalent compute budgets.
spatial redundancyforeground predictorcompound shrinkingcnn compressionedge ai
Converge to Surprise: Evolutionary Self-supervised Image Clustering
The paper introduces a novel self-supervised image clustering framework that eliminates the need for predefined optimization targets. The method employs a 'surprise score' derived from the Principle of Maximum Entropy, measuring deviation from i.i.d. pixel assumptions, and optimizes it via an evolution-strategy outer loop combined with periodic gradient-descent inner loops. This 'converge-to-surprise' approach achieves state-of-the-art results on standard benchmarks in non-parametric self-supervised clustering, where ground-truth class counts are unknown.
self-supervised learningimage clusteringevolutionary strategymaximum entropynon-parametric
Finding a stationary point of a stochastic convex problem
The paper addresses stochastic convex optimization by proposing a stronger convergence criterion requiring the subdifferential to contain a small element, rather than relying on proximity-to-stationarity guarantees. The method leverages dimension theory to decompose the subdifferential graph, demonstrating how stochastic sampling preserves its structural 'pieces' and enables proximal-point-like methods. Theoretical results establish convergence under this non-trivial criterion, overcoming challenges posed by non-uniform subdifferential convergence in convex functions.
stochastic convex optimizationsubdifferential convergencedimension theoryproximal-point methodsstationarity criterion
Best-Arm Identification with Generative Proxy
PROBE introduces a phase-elimination algorithm for fixed-confidence best-arm identification that leverages cheap proxy scores correlated with costly reward observations. The method uses ordinary least squares to maintain an upper certificate on residual variance, ensuring correctness despite unknown proxy-reward correlation. Theoretical analysis shows PROBE is δ-PAC and achieves oracle sample complexity up to constant factors, with extensions to (ε,δ)-PAC settings. Empirical validation on synthetic and auto-loan pricing datasets demonstrates sample savings scaling with proxy-reward correlation strength.
best-arm identificationproxy scorephase-eliminationordinary least squaresresidual variance
Video2Reaction: Mapping Video to Audience Reaction Distribution in the Wild
The paper introduces Video2Reaction, a multimodal dataset mapping 10,000+ short movie segments to audience emotion distributions from social media reactions. A two-stage multi-agent pipeline using open-source LLMs achieves 86% annotation correctness despite task subjectivity. Benchmark results show pretrained video models fail zero-shot, but finetuning yields state-of-the-art predictors (77% Top-3 F1 with LLaVA-Next), though significant gaps remain in modeling collective reactions.
multimodal datasetemotion distributionmulti-agent pipelinezero-shot learningreaction prediction
Gen4U: Unifying Video Generation and Understanding via Diffusion
Gen4U demonstrates that frozen video diffusion models serve as effective video encoders for both generation and understanding tasks. The method analyzes intermediate activations via mutual-kNN alignment, revealing structured latent spaces where semantics emerge at moderate noise levels while fine details require attention mechanisms. Without fine-tuning, Gen4U achieves strong performance on video classification, depth estimation, pose estimation, and captioning while maintaining video generation capability.
video diffusion modelsmutual-knn alignmentlatent space analysisattention mechanismsfrozen encoders
Geometric Self-Distillation for Reasoning Generalization
GeoSD introduces a geometric self-distillation objective to mitigate reasoning generalization degradation in large language models during privileged-context self-distillation. The method combines a Hellinger loss, which scales teacher preferences by student overlap, and a proximal term penalizing prediction drift via Fisher-Rao distance, both operating in the geometry of next-token distributions. Evaluated across mathematical reasoning benchmarks and model families (1.7B to 32B parameters), GeoSD maintains in-distribution performance while improving out-of-distribution accuracy by 5.7-8.6 points over base models. Analysis reveals GeoSD preserves high-entropy alternatives, contrasting with standard matching's tendency toward confident wrong answers.
self-distillationhellinger lossfisher-rao distancenext-token distributionsout-of-distribution
Tensor Train Diffusion: Leveraging Low-Rank Structures for High-Dimensional Score-Based Sampling
The authors propose Tensor Train Diffusion, a novel method for high-dimensional score-based sampling that leverages functional tensor train (FTT) representations to efficiently solve the Hamilton-Jacobi-Bellman (HJB) PDE governing diffusion models. By combining FTT's low-rank function approximation with a backward-in-time iterative scheme from backward stochastic differential equations (BSDEs), the method achieves both model compression and computational efficiency. Results demonstrate improved sampling fidelity and robustness compared to existing techniques like PINNs or trajectory-based methods, addressing key bottlenecks in training time and hyperparameter sensitivity for high-dimensional distributions.
functional tensor trainhamilton-jacobi-bellman pdescore-based samplingbackward stochastic differential equationsdiffusion models
LEMUR 2: Unlocking Neural Network Diversity for AI
LEMUR 2 introduces a large-scale framework for neural architecture diversity, unifying generative, evaluative, and deployment pipelines across 14,000+ architectures and 750,000+ training records. The system employs AST-based code mutation, genetic/RL evolution, fractal architectures, and LLM-guided synthesis (including NN-RAG with 900+ PyTorch modules), with NN-VR/NN-Lite pipelines for cross-platform deployment benchmarking. Results demonstrate multimodal task coverage (image captioning, text-to-image, language modeling) and architectural transferability analysis, providing a foundation for LLM-driven AutoML and cross-hardware validation.
neural architecture searchcross-domain evaluationast-based mutationretrieval-augmented generationdeployment benchmarking
Generative Diffusion Models of Stochastic Graph Signals
The paper proposes a unified denoising diffusion framework for conditional graph signal generation, addressing limitations of application-specific designs that often regress to conditional means. The method employs a novel U-Graph Neural Network (U-GNN) architecture, which generalizes U-Nets to graph-structured signals via multi-resolution encoder-decoder processing with learned node selection and original-graph convolutions. Evaluations on stock price forecasting and wireless resource allocation demonstrate the framework's effectiveness across diverse domains.
denoising diffusiongraph neural networksconditional generationmulti-resolution processinggraph signals
CaLiSym: Learning Symplectic Dynamics of Real-World Systems through Structured Canonical Lifts
CaLiSym extends symplectic learning to non-conservative robotic systems by embedding physical states and ports into a structured canonical phase space, where dynamics evolve through an exactly symplectic map. The method employs an explicit algebraic lift, avoiding recurrent states or implicit optimization, and introduces GRB-SympNet, a B-spline variant combining local approximation with symplectic structure. Experiments on a dissipative double pendulum, quadrotor, and quadruped show improved out-of-distribution autoregressive prediction while preserving the symplectic form numerically. This demonstrates geometry-preserving dynamics modeling for real-world systems.
symplectic learningcanonical phase spaceautoregressive predictionb-splinenon-conservative systems
On Explicit Super-Expressive Approximation for Neural Networks
The paper introduces explicit parameter-error trade-offs for fixed-architecture neural networks approximating Lipschitz and Hölder-smooth functions. Using the Chinese Remainder Theorem as a constructive encoding mechanism, the authors design networks with bounded width and depth: for Lipschitz functions on $[0,1]^D$, a width-$\max\{D,4\}$, depth-$5$ network suffices; for $C^{r,\gamma}_A([0,1]^D)$ functions, width $\max\{2D, D+5N+1\}$ and depth $r + 9$ yield parameter magnitude $\log_2 \mathcal{P} = \mathcal{O}(\varepsilon^{-2D/(r+\gamma)}\log(1/\varepsilon))$. This contrasts with prior work lacking quantitative bounds.
neural networksapproximation theorychinese remainder theoremlipschitz continuityhölder-smooth
Efficient Bayesian Deep Ensembles via Analytic Predictive Inference
The authors propose an efficient Bayesian deep ensemble method for predictive regression, combining Bayesian inference with deep ensembles to yield interpretable, calibrated uncertainty estimates. Key innovations include low-dimensional ensemble representation via a small set of neural predictors, closed-form Bayesian aggregation for interpretable posterior weights, and independent ensemble training for diversity. Evaluated on standard regression benchmarks, the method demonstrates competitive predictive performance while maintaining reliable uncertainty calibration.
bayesian deep ensemblespredictive regressionuncertainty calibrationclosed-form aggregationindependent training
Efficient Long-Horizon Learning for Learned Optimization
The paper introduces Efficient Long-hOrizon (ELO) learning, a meta-training algorithm addressing scalability and performance limitations in learned optimizers (LOs). ELO reallocates redundant meta-training compute to longer failure regimes and employs decoupled progressive expert supervision for stable meta-learning signals. Evaluated on element-wise and matrix-based LOs across GPT-2-124M/350M, ViT-B/16, and ResNet-50 models, ELO enhances long-unroll performance and out-of-distribution generalization. ELO-Celo2 outperforms AdamW across tasks and remains competitive with Muon in language modeling, requiring less than 7 H100 GPU-hours for meta-training.
learned optimizersmeta-traininglong-horizon learningdecoupled progressive expert supervisionout-of-distribution generalization
Deployment Risk Assessment Using Diff-Aware Features: A Case Study at Prime Video
The study introduces a diff-aware feature framework for deployment risk assessment, addressing limitations of existing approaches that rely on developer metadata or historical data. Using LLMs as multi-language feature extractors, the method systematically evaluates quantitative metrics (code-level, change-level) and qualitative indicators (coding style violations, change type classification). Evaluation on Prime Video's production environment and ApacheJIT dataset shows an average recall of 0.83 and F1 score of 0.81, with structural code complexity outperforming change-level volume metrics as a risk predictor.
diff-aware featuresdeployment risk assessmentllm feature extractioncode complexitychange-level metrics
Trees from Marginals: Autoregressive drafting with factorized priors
We introduce Weaver, a lightweight autoregressive adapter that improves speculative decoding efficiency by constructing proposal trees from top-K marginals of factorized draft models. Weaver restores conditional dependencies between proposed tokens while avoiding full-vocabulary projections, addressing the degradation in acceptance rates observed in independent marginal predictions. The method includes a rollback-free tree-verification algorithm optimized for Gated Delta Net layers and CUDA kernel implementations in SGLang. Experiments demonstrate a 4.37× speedup over standard autoregressive decoding and a 24.7% improvement over the DFlash baseline.
speculative decodingfactorized draft modelsautoregressive adaptertree-verificationgated delta net
Optimization Geometrodynamics: A Framework for Dynamic Geometric Optimization
The paper introduces optimization geometrodynamics, a theoretical framework for analyzing gradient-based optimization as a coupled evolution of parameters, particle distributions, and time-varying Riemannian metrics. The approach separates invariant geometric obstructions from improvable conditioning effects, formalizing dynamic geometric complexity as the minimal metric cost to reduce optimization difficulty. Key results include an exact affine-invariant complexity measure for strongly convex quadratic objectives, analyses of Hessian-matching flows and spectral relaxation, and invariant observables for comparing adaptive optimizers. All claims are formally proven without empirical validation.
optimization geometrodynamicsriemannian metricdynamic geometric complexityhessian-matching flowsaffine-invariant distance
Macroeconomic Message Passing for Anticipating Foreign Exchange Regime Changes: A Deep Logical Learning Approach using Graph Tsetlin Machines
The paper contributes a graph-theoretic method for predicting foreign exchange regime changes by incorporating macroeconomic variables through message-passing operations. The proposed Graph Tsetlin Machine (GraphTM) framework represents multivariate drivers and technical indicators as hypervectorized directed multigraphs, using structured message passing to construct interpretable logical clauses for pattern recognition. Empirical results demonstrate the approach's efficacy in anticipating USD/JPY market regimes.
graph tsetlin machinemessage passinghypervectorized multigraphsmarket regime predictionforeign exchange
Heat-Kernel Entropy Profiles and Geometric Effective Sample Size for Weighted Measures on Manifolds
The paper introduces heat-kernel entropy profiles, a multiscale geometric summary for weighted empirical measures on compact manifolds that accounts for both particle weights and their spatial distribution. The method tracks nonuniformity across scales by diffusing weighted atoms via intrinsic heat flow, with order-two Rényi entropy computed from pairwise heat-kernel overlaps. Theoretical results include monotonicity, asymptotics, and consistency properties, while spherical experiments demonstrate the profile's ability to identify antipodal, multimodal, and duplicate-particle structures missed by weight-only summaries. The approach yields a geometric effective sample size that discounts nearby particles while matching standard ESS for well-separated cases.
heat-kernel entropyrényi entropyeffective sample sizecompact manifoldsspherical harmonics
Robust Human-AI Complementarity under Uncertainty
The paper establishes conditions for robust human-AI complementarity under asymmetric information about prediction quality, demonstrating that negative error correlation between human and AI predictions enables decision makers to construct utility-guaranteeing strategies. Through theoretical analysis and empirical validation on real-world forecasting benchmarks, the authors identify error correlation structure as a key determinant of complementary value extraction. Results show that negative error correlation facilitates reliable performance improvements when combining human and AI judgments.
human-ai collaborationerror correlationdecision theorycomplementary valueforecasting benchmarks
Pelican-VLA 0.5: Attending Before Acting Benefits Generalization
Pelican-VLA 0.5 introduces a unified vision-language-action (VLA) model integrating visual understanding, frame generation, and action prediction without task-specific fine-tuning. The model employs learnable Reasoning Slots between perception and action, routing task-relevant visual information through a compact bottleneck to induce manipulation-centric attention. It demonstrates attention-level generalization across unseen scenes and robot embodiments, outperforming open-source VLA baselines in focusing on instruction-relevant objects and contact regions without explicit supervision.
vision-language-actionreasoning slotsattention-level generalizationmanipulation-centric attentionunified architecture
Dual Attention Heads for Personalized Federated Learning in ECG Classification
The paper proposes FedDualAtt, a personalized federated learning method for ECG classification that splits transformer attention heads into global and local branches. Global heads are aggregated across clients via FedAvg to learn shared patterns, while local heads remain client-specific to adapt to institutional data characteristics. Evaluated on the FedCVD benchmark, FedDualAtt outperforms existing FL and personalized FL approaches, with analysis showing optimal performance depends on client-specific global-local head ratios.
federated learningecg classificationattention headspersonalized flfedavg
From Jumps to Signatures: a Generative Method for Temporal Point Processes
The paper introduces sigTPP, the first signature-based generative model for Temporal Point Processes (TPPs), addressing limitations in existing neural TPP models that lack global sequence-level loss and distributional discrepancy measures. The method employs an interarrival embedding to lift jump paths to continuous paths of bounded variation, enabling the application of rough path signatures to discrete event sequences. Theoretical analysis yields three distributional discrepancies for evaluating generative TPP models. Empirical results show sigTPP outperforms baselines, achieving the best average rank across eight metrics, surpassing or matching baselines in 64% of dataset-metric pairs, and improving against every baseline by at least 19% on average.
temporal point processesrough path signaturesinterarrival embeddinggenerative modeldistributional discrepancy
Entropy-Guided Tensor Compression for Multimodal Federated Learning on Edge Devices
MESH-FL introduces an entropy-guided matrix product state (MPS) update-compression framework for multimodal federated learning on edge devices, addressing modality-specific differences in spectral structure and compressibility. The method estimates spectral entropy via truncated singular value decomposition and adaptively allocates MPS compression ranks across layers, modalities, and devices under per-client payload budgets. Theoretical analysis shows the approach preserves monotonicity and achieves convergence with a compression-dependent error term. Experiments on a 15-node Raspberry Pi cluster demonstrate up to 56.8× compression, 2.01% higher accuracy than FedAvg, and 66× reduction in transmitted data to reach convergence.
federated learningmatrix product statespectral entropymodality-heterogeneousedge devices
POPS: Recovering Unlearned Multi-Modality Knowledge in MLLMs with Prompt-Optimized Parameter Shaking
The paper introduces Prompt-Optimized Parameter Shaking (POPS), an adversarial method to recover supposedly unlearned multi-modality knowledge from Multimodal Large Language Models (MLLMs). POPS optimizes prompt suffixes to elicit private examples from victim MLLMs, then fine-tunes the models using these synthesized outputs to disclose sensitive information. Experiments on Multi-modality Machine Unlearning (MMU) benchmarks demonstrate POPS achieves near-complete recovery of erased data, exposing vulnerabilities in existing MMU-based privacy protections.
multimodal large language modelsmachine unlearningadversarial attackprompt optimizationprivacy vulnerability
Final Checkpoints Are Not Enough: Analyzing Latent Reasoning Faithfulness Along Training Trajectories
The paper investigates the evolution of faithfulness in latent reasoning methods during training, contrasting prior work that only examined converged checkpoints. Using counterfactual edits and noise-ablation activation patches across saved checkpoints, the study reveals divergent trajectories: (i) output-level unfaithfulness manifests similarly at convergence but follows different paths; (ii) activation-level causal contributions decay during training, correlating with output flips; (iii) trajectories vary by answer format, decaying for binary choice but rising for open-ended decoding. Findings emphasize the dependence of latent reasoning faithfulness on training stage and answer format.
latent reasoningfaithfulnesscounterfactual editactivation patchtraining trajectory
ORAN-DEFEND: Subspace Detection and Sanitization of Backdoor DRL xApps in Open RAN
ORAN-DEFEND proposes a retraining-free defense against backdoor attacks in DRL-based xApps for O-RAN, using SVD-based subspace projection to sanitize KPI telemetry. The method analytically establishes a recovery condition dependent on trigger energy concentration in the orthogonal complement of the safe subspace (quantified by $\Eperp$). Evaluated on the COLORAN dataset against four DRL backdoor attacks (TrojDRL, SleeperNets, BadRL, Q-Incept), it achieves 100% return recovery and ≥99.5% defense success rate when subspace assumptions hold, while revealing fundamental limits of linear projection defenses.
open ranbackdoor defensesubspace projectiondrl xappskpi telemetry
Fast determinantal sampling on general spaces and diffusion geometry
The work establishes theoretical guarantees for determinantal point process (DPP) sampling in non-Euclidean spaces, including Riemannian manifolds and weighted networks. By analyzing spectral kernels derived from Laplacian operators and Markov diffusions, the authors demonstrate sampling rates that adapt to intrinsic dimensionality $d_{\text{int}}$, achieving $\big(\text{sample size}\big)^{-\frac{1}{2}-\frac{1}{2d_{\text{int}}}}$ convergence matching Euclidean benchmarks. Key techniques involve Weyl's Law for manifold spectra, Dirichlet forms, and pseudodifferential operators, extending DPP sampling theory to geometrically complex domains.
determinantal point processesspectral kernelsriemannian manifoldsmarkov diffusionsintrinsic dimensionality
The Power of Backdoor Absorption in Community Training
The paper proposes a defense against backdoor attacks in decentralized AI training by analyzing the resilience of optimization dynamics under Byzantine perturbations. The authors formalize the attack-defense interaction as a Discrete-Time Markov Chain (DTMC), proving that adversarial success probability asymptotically collapses to zero when combining natural absorption, randomized scheduling, and lazy verification. Empirical results show 10% verification overhead suffices for complete backdoor suppression without utility degradation, offering a computationally efficient solution for safety-critical systems.
backdoor attacksbyzantine robustnessdecentralized trainingmarkov chain analysislazy verification
The Approximation Ratio for the Risk of Myopic Bayesian Active Learning for Linear Regression
This work establishes a tight approximation ratio for the risk of myopic Bayesian active learning in linear regression, addressing the greedy algorithm's performance in optimal experiment design. The authors introduce the maximum initial leverage score (MILS) as a fundamental quantity governing the algorithm's behavior and prove a linear approximation ratio in terms of MILS. Numerical simulations demonstrate the theoretical results, validating the tightness of the bound. This constitutes the first approximation ratio analysis for greedy active learning in this setting.
active learninglinear regressiongreedy algorithmapproximation ratioleverage score
UASPL: Uncertainty-Aware Self-Paced Learning with Evidential Neural Networks
The paper proposes UASPL, an uncertainty-aware self-paced learning method using evidential neural networks to improve sample selection reliability. The approach integrates predictive uncertainty estimation via a Subjective Logic-based loss function, enabling interpretable sample selection while maintaining compatibility with existing SPL variants. Experiments across multiple datasets demonstrate UASPL's superior classification performance (quantitative metrics unspecified), interpretability, and generality compared to standard SPL methods.
self-paced learningevidential neural networksuncertainty estimationsubjective logicsample selection
Optimized Instance Alteration for Explaining and Assessing Robustness of Classifiers
The authors propose a unified optimization framework for diagnosing misclassifications and assessing classifier robustness. The method modifies instances via an objective combining an explainability-aware $L_0$ penalty (XA-$L_0$) for sparse, interpretable changes and a classifier loss term to steer predictions toward target labels. They introduce the Tolerance Region Confusion Matrix (TOR-Confusion Matrix) to quantify robustness via class transition probabilities under bounded perturbations. Experiments on image and tabular datasets demonstrate joint interpretability and robustness assessment capabilities.
explainability-aware penaltytolerance region confusion matrixclassifier robustnessl0 sparsityblack-box diagnostics
When Do Geometric Algebra Layers Beat Scalarization? A Controlled Study on SO(3)-Equivariant Vector Laws
The paper investigates whether geometric algebra (GA) layers in Cl(3,0)-based networks offer advantages beyond exact SO(3)-equivariance for learning 3D vector laws. Through controlled comparisons with scalarization baselines (MLPs on invariant dot products), the study finds that GA provides no benefit for single-stage operations (e.g., cross products) but significantly outperforms baselines (10-30x sample efficiency) on compositional tasks involving nested group operations (e.g., chained rotations). The performance gap persists against strengthened baselines and correlates with rotation chain depth, while scalarization fails entirely on 4-rotation chains. GA’s advantage is specific to compositional group operations, not general polynomial invariants.
geometric algebraso(3)-equivariancescalarizationcompositional learninglow-data regime
Does Demand Response Increase Vulnerability to Cyber Attacks by Adversarial Data Modifications?
The paper investigates adversarial attacks on industrial demand response systems by manipulating electricity price forecasts. Using a generalized process model, the authors design attacks targeting scheduling optimization in energy-intensive production, evaluating vulnerability across varying process flexibility levels. Results show attacks can reduce demand response profits, yet 90% financial advantage persists under stealthy perturbations; impact depends more on perturbation orientation than magnitude, suggesting attack analyses should incorporate scheduling model sensitivities.
adversarial attacksdemand responseprice forecastingscheduling optimizationprocess flexibility
When Certificates Fail: A Unified Safety Framework for Embedded Neural Interface Models
The authors propose a unified empirical audit framework addressing three safety failures in embedded neural-interface models: verification insufficiency, proxy-fidelity divergence, and latent information exfiltration. They demonstrate that formal robustness certificates can remain valid while task accuracy collapses, exemplified by EEGNet's 25.7% accuracy drop under projected-gradient attack despite valid Lipschitz-style certificates. The framework is instantiated on BCI Competition IV 2a and SEED-IV datasets using EEGNet, CSP+LDA, and FBCSP+LDA models, revealing architecture-independent verification gaps. Results show that operational safety auditing, beyond certificate verification, is essential for responsible neural-interface deployment.
robustness certificatesneural interfaceseegnetprojected-gradient attackverification insufficiency
STST-JEPA: Shallow-Target Spatio-Temporal Joint Embedding Prediction Architecture For EEG Self-Supervised Learning
The authors propose STST-JEPA, a self-supervised transformer for EEG-based brain age prediction, addressing challenges of cross-site heterogeneity and limited labeled data. The model combines latent-prediction (masked-token representation prediction with EMA targets) and signal-reconstruction objectives on 30-second multi-channel windows, pretrained on 47,703 sessions (ages 5-81) from brain.space and HBN. Frozen embeddings achieve 3.06 years MAE (r=0.924) on age regression (3,367 sessions), outperforming a 10-year MAE baseline. Fine-tuning yields state-of-the-art performance on NeuralBench tasks: sex classification (0.911 balanced accuracy), age prediction (r=0.749), and psychopathology regression (r=0.215). Age-prediction residuals correlate negatively with cognitive efficiency.
self-supervised learningeeg foundation modelbrain age predictiontransformer architecturespatiotemporal masking
Reward Valuation in Vision Language Models: Causal Mechanisms Underlying Anhedonia
The study identifies and causally manipulates reward-anticipatory units in Vision-Language Models (VLMs), drawing parallels to neural mechanisms of anhedonia in major depressive disorder. Using perturbations inspired by Nucleus Accumbens (NAc) dysregulation, the authors demonstrate that targeted unit disruptions induce anhedonia-like behavior: models favor low-effort, low-reward options in effort-based decision tasks while maintaining baseline performance in non-reward contexts. Results align with clinical scales (DARS, MAP-SR), suggesting VLMs encode human-like reward valuation circuits. Methodologically, the work combines neuroscientific frameworks with mechanistic interpretability techniques.
vision-language modelsanhedonianucleus accumbensmechanistic interpretabilityreward valuation
Creating Power Distribution Network Layouts Using Generative Adversarial Networks and Image-Based Representations
A generative framework leveraging Generative Adversarial Networks (GANs) is proposed for creating power distribution network layouts using image-based representations. The model operates in unconditional and conditional modes, learning layout patterns from rasterised views of distribution systems and incorporating geographical context such as street maps and consumer distributions. Training involves dataset preparation from Geographic Information System (GIS) sources, GAN architecture design, and stability analysis. Results demonstrate the model's ability to reproduce low, medium, and high voltage feeder topologies while aligning layouts with geographical structures. Limitations include training instability, resolution-dependent artefacts, and lack of explicit electrical constraints, suggesting future extensions for power flow validation.
generative adversarial networkspower distribution networksgeographic information systemimage-based representationstraining stability
Fingerprint, Not Blueprint: How Positional Schemes Set the Default Spectral Algebra of Attention
The paper demonstrates that positional encoding schemes determine the default spectral algebra of attention heads, acting as fingerprints rather than blueprints. Analyzing seven pretrained models with three positional schemes (RoPE, learned-absolute, ALiBi), the authors show that rotational spectral signatures emerge only after functional behavior develops, not as a precursor. Key findings include perfect model-level separation of rotational heads under RoPE (permutation p=0.029), elimination of induction by zeroing RoPE's phase component, and rerouting capability in constrained training despite spectral bans (q_BH ≤ 0.016). The study combines static spectral analysis, dynamic checkpoint tracking, and causal experiments to establish positional schemes as post-hoc sculptors of attention head solutions.
attention mechanismspositional encodingspectral analysisnon-hermitian algebrarotary embeddings
HiFuzz: Hierarchical Reinforcement Learning for Semantic-Aware and Adaptive CPU Fuzzing
HiFuzz introduces a hierarchical reinforcement learning framework for CPU fuzzing, replacing traditional mutation with a two-layer generation process: a Program Agent for global layout and a Basic Block Agent for instruction-level precision. The method integrates adaptive coverage rewards and a semantic-aware basic block encoder to address reward sparsity. Evaluations on three RISC-V cores show HiFuzz outperforms state-of-the-art fuzzers in coverage and bug detection.
hierarchical reinforcement learningcpu fuzzingsemantic-aware encodingadaptive coverage rewardrisc-v verification
📰 Industry Media (8)
OpenAI Releases GPT-5.6 (Sol, Terra, Luna): A Three-Tier Model Family With Programmatic Tool Calling in the Responses API
OpenAI released GPT-5.6, a three-tier model family (Sol, Terra, Luna) with programmatic tool calling via JavaScript execution in an isolated V8 runtime. The flagship Sol model achieves 80.0 on the Artificial Analysis Coding Agent Index (2.8 points above Claude Fable 5) and 92.2% on BrowseComp, while maintaining cost efficiency with 85% fewer output tokens than Claude Opus 4.8 on OSWorld 2.0. Performance gaps remain on SWE-Bench Pro (64.6% vs Claude Mythos 5's 80.3%) and Toolathlon (58% vs Fable 5's 61.7%). The system introduces multi-agent parallelism (Ultra mode) and prompt caching with 30-minute minimum cache life.
programmatic tool callingv8 runtimemulti-agent parallelismprompt cachingterminal-bench 2.1
Meet Nemotron Labs 3 Puzzle 75B A9B: A Compressed Hybrid MoE LLM Delivering 2.03x Server Throughput
Nemotron-Labs-3-Puzzle-75B-A9B introduces a compressed hybrid MoE architecture achieving 2.03x server throughput while preserving the parent model's 88-block layout (40 Mamba, 40 MoE, 8 attention). Through iterative neural architecture search (Puzzletron) combining intermediate channel pruning, top-k expert reduction, and Mamba SSM state compression (128→96), it reduces total parameters from 120.7B to 75.3B and active parameters from 12.8B to 9.3B. On 8xB200 nodes, decode-heavy workloads show 2.14x throughput gains, while single-H100 concurrency improves from 1 to 8 requests at 1M context. Benchmarks reveal modest accuracy drops (-4.2 on Arena-Hard-V2, -2.6 on SWE-Bench) but maintain performance on long-context tasks (RULER 1M: -1.7).
mamba-ssmmixture-of-expertskv-cacheneural-architecture-searchparameter-pruning
Datalab Lift vs the Field: How a 9B Schema-First Extractor Compares with NuExtract3, LlamaExtract, Marker, and Docling
Datalab Lift introduces a 9B vision model for schema-first document extraction, directly converting PDFs or images into structured JSON without intermediate Markdown conversion. It employs schema-constrained decoding to produce application-ready fields in a single visual extraction pass. Benchmarks show Lift achieves 90.2% field accuracy, outperforming NuExtract3 (81.5%) and demonstrating lower latency (9.5s) compared to Gemini Flash 3.5 (28.1s). Lift is optimized for scenarios requiring low latency, self-hosting, and large-volume cost control, contrasting with broader cloud platforms and enterprise extraction systems.
schema-first extractionvision modeljson schemastructured outputdocument parsing
Robbyant Releases LingBot-VLA 2.0: An Open-Source 6B Vision-Language-Action (VLA) Model for Cross-Embodiment Robot Manipulation
LingBot-VLA 2.0, a 6B parameter Vision-Language-Action (VLA) model, advances cross-embodiment robot manipulation through improved generalization, an expanded action space, and predictive dynamics modeling. Built on Qwen3-VL-4B-Instruct, it employs Mixture-of-Experts (MoE) layers and dual-query distillation supervised by LingBot-Depth and DINO-Video. Pretrained on 60,000 hours of data spanning 20 robot configurations, it unifies actions via a 55-dimensional canonical vector. Evaluated on the GM-100 benchmark, LingBot-VLA 2.0 achieves 66.2/34.4 progress/success on AgileX Cobot Magic and 34.6/15.6 on Galaxea R1Pro, outperforming prior versions and baselines in both in-domain and out-of-domain tasks.
vision-language-actionmixture-of-expertsdual-query distillationcross-embodimentpredictive dynamics
SpaceXAI Releases Grok 4.5, a Cursor-Trained Model for Coding, Agentic Tasks, and Knowledge Work at $2/M Input
SpaceXAI released Grok 4.5, a general-purpose model optimized for coding, agentic tasks, and knowledge work, achieving state-of-the-art performance on Harvey’s Legal Agent Benchmark. Trained on tens of thousands of NVIDIA GB300 GPUs with Cursor, the model employs large-scale reinforcement learning on multi-step software engineering tasks, emphasizing data curation and per-token efficiency. Benchmarks show competitive results: 62.0% pass@1 on DeepSWE 1.0, 83.3% on Terminal Bench 2.1, and 64.7% resolve rate on SWE Bench Pro, with 4.2× fewer output tokens than Opus 4.8 (max). Priced at $2/M input and $6/M output tokens, it operates at 80 TPS.
agentic taskstoken efficiencyreinforcement learningnvidia gb300swe bench pro
Netflix AI Team Cuts Wide-Partition Read Latency from Seconds to Milliseconds by Splitting Cassandra Partitions Per ID
Netflix engineers introduced dynamic partitioning in Apache Cassandra to address wide-partition latency issues in their TimeSeries Abstraction platform. The method asynchronously splits oversized partitions per TimeSeries ID using a three-stage pipeline (detection via read-path byte counting, planning/splitting with Bloom filters, and metadata routing) while maintaining backward compatibility. This reduced read latency from seconds to low double-digit milliseconds (tail latency ~200ms) and maintained availability for 500MB+ partitions without requiring application changes.
cassandratime-seriespartitioninglatencybloom filter
AWS GraphRAG deployment cuts drug research cycles by 87%
AWS GraphRAG deployment accelerates pharmaceutical research by 87% through knowledge graph integration of fragmented proprietary and public datasets. The system combines Amazon Neptune Analytics for graph storage, Claude 4.5 Sonnet via Amazon Bedrock for NLP processing, and fuzzy entity linking to map natural language queries to structured data. Performance metrics show 85% faster data retrieval, 70% reduced review times, and research cycles shortened from six months to three weeks while maintaining verifiable citation trails and compliance with regulatory requirements.
graphragknowledge graphentity linkingamazon neptuneclaude 4.5
NHS AI blood test could reduce invasive womb cancer checks
The PinPoint AI blood test, developed by PinPoint Data Science, reduces invasive diagnostic procedures for suspected womb cancer by analyzing approximately 30 blood markers to classify patients into low, elevated, or high-risk categories. The machine learning-based test demonstrated 99.1% sensitivity for cancer detection and a 99.8% negative predictive value in a trial involving 16,481 patients. Current NHS adoption aims to spare 18,000 women annually from transvaginal ultrasounds, with deployments planned for gynaecological and gastrointestinal cancer pathways across multiple NHS trusts.
machine learningnegative predictive valuegynaecological cancerdiagnostic triageblood biomarkers
Generated automatically at 2026-07-09 21:08 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.