Daily Digest — 2026-08-28
361 items · 4 research labs, 353 arxiv papers, 4 industry media
AI News: all feed URLs failed (last tried: https://artificialintelligence-news.com/feed/)
🏛️ Research Labs (4)
Better answers, broader thinking: What students gain from ChatGPT and critical-thinking training
A randomized controlled trial (N=1,000+ undergraduates) demonstrated complementary effects of GPT-4 access and causal reasoning training on student assignments. Students using ChatGPT produced more coherent, expert-like responses (0.8-point improvement on 5-point rubric), while causal reasoning training increased idea diversity without affecting rubric scores. Combined intervention groups exhibited both effects, showing enhanced logical coherence and originality. Automated text analysis revealed training-induced improvements in causal explanations and uniqueness metrics not captured by traditional grading rubrics, suggesting need for assessment redesign in AI-augmented education.
causal reasoningautomated text analysisrandomized controlled trialin-context learningrubric design
Expanding OpenAI’s presence in Brazil
OpenAI expands commercial operations in Brazil, citing rapid AI adoption with 215 million daily ChatGPT messages and 35% work-related usage. The local São Paulo team will collaborate with businesses, developers, and institutions to scale AI applications, supported by an estimated R$1 trillion economic impact by 2030 (RegLab study). Key initiatives include ChatGPT Edu at ITA, AI literacy programs (e.g., with ENTER for legal professionals), and small-business training via Estímulo. Brazil ranks second in OpenAI API developers, with Codex usage growing 11x weekly and 30x daily interactions since 2026. The OpenAI Hackathon Brasil showcased local innovations, including a procurement tool for small businesses built with Codex.
chatgptcodexapiin-context learningresponsible deployment
Introducing Intelligence Age
OpenAI's Strategic Futures team introduces 'Intelligence Age,' a blog focused on structural risks posed by transformative AI to individual rights and societal power dynamics. The team argues that advancements in autonomous systems and machine intelligence may disrupt traditional power structures, necessitating new institutional mechanisms to preserve human autonomy. Key principles include balancing individual freedom with collective action, ensuring bounded legibility of AI systems, and maintaining privacy. The approach integrates public-policy design, economics, and forecasting to address concentration of power risks.
strategic futuresconcentration of powerautonomous systemsbounded legibilitytransformative ai
3 new ways to plan and book travel in Search
Google Search's AI Mode introduces three travel planning enhancements: real-time flight price tracking (300+ airline partners, 180+ countries), rewards point/mile cost visualization (initial support for 5 airlines and 3 hotel chains), and in-chat hotel booking (U.S.-only launch with 10 partners). The system integrates multimodal inputs (text queries, visual hotel listings) and transactional APIs (Google Pay, partner booking platforms) to enable end-to-end itinerary construction. Flight tracking achieves real-time email alerts via continuous price monitoring, while rewards integration pulls live redemption rates from loyalty programs.
multimodal inputstransactional apisreal-time monitoringloyalty program integrationin-context booking
📜 arXiv Papers (353)
WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution
WikiSkill introduces a framework for co-evolving agent skills with a persistent knowledge base (wiki), enabling systematic reuse of insights from agent experience. The method separates raw execution experience, accumulated knowledge, and executable skills, continuously consolidating experience into the wiki for subsequent skill updates. Evaluations across diverse benchmarks and models show WikiSkill consistently outperforms state-of-the-art skill-evolution methods and improves over no-skill baselines. Skill evolution complements model scaling, with larger models benefiting more from evolved skills, and skills transferring effectively across models and families. Ablation studies confirm the critical role of persistent knowledge accumulation in the wiki for effective skill evolution.
agent skillspersistent knowledge baseskill evolutionmodel scalingknowledge accumulation
SWE-Prime: Fewer Trajectories, Better Performance
SWE-Prime introduces a multi-granularity, two-stage supervised fine-tuning (SFT) data selection method to improve large language models' software issue resolution. It first filters trajectories based on process quality, result quality, and representativeness, then segments steps to assess contribution, learnability, and risks. During SFT, all segments preserve context, but only selected segments contribute to loss computation. Experiments on SWE-Bench Pro and SWE-Bench Verified show that training on a 10% trajectory subset selected by SWE-Prime outperforms full dataset training, achieving relative performance gains of 12.2% and 24.2%, respectively.
supervised fine-tuningtrajectory filteringsegment-level selectionsoftware issue resolutionmulti-granularity selection
From Static to Dynamic: Benchmarking Real-World Code Review with MCR-Bench
We introduce MCR-Bench, the first defect state-aware benchmark for realistic multi-round code review, addressing the limitations of single-round static approaches in existing LLM-based code review systems. MCR-Bench comprises 2,269 real-world multi-round code review tasks across five programming languages, annotated with fine-grained defect metadata and dynamic state labels capturing defect evolution. Experiments with mainstream LLMs reveal three key findings: (1) limited overall performance in defect detection and state tracking, degrading with increased interaction rounds; (2) defect-sensitive performance varying across types and severity levels; (3) underlying failure mechanisms including temporal misalignment and inadequate long-range memory.
multi-round code reviewdefect state-awaredynamic state labelstemporal misalignmentlong-range memory
RedEvoAgent: Automatic Red-Teaming Agent with Experience-Driven Skill Evolution
RedEvoAgent introduces an automatic red-teaming agent that distills cross-case attack trajectories into human-readable skills, addressing limitations of fixed attacks and trajectory-based retrieval in LLM-based execution harnesses. The method employs tool-effectiveness profiling, Deciding-Tool Attribution for skill updates, and a validation ratchet to retain performance-improving updates. Evaluations across benchmarks, target models, and execution harnesses demonstrate superior performance over fixed and agentic baselines, improved tool efficiency, and transferability across attacker models and harnesses.
red-teamingtrajectory distillationtool-effectiveness profilingdeciding-tool attributionvalidation ratchet
Mechanistic Reaction Prediction via Discrete Flow Matching on Graph-Structured Electron Occupation
MAELLE introduces a mechanistic approach to chemical reaction prediction by modeling electron rearrangements as discrete flow matching over graph-structured electron occupation vectors. The method formulates reactant-to-product transitions as a Continuous-time Markov Chain (CTMC) on integer-valued electron occupation space, generalizing discrete flow matching via Optimal Transport to generate interpretable edit trajectories without elementary step annotations. On USPTO-480K, MAELLE matches leading models in accuracy while demonstrating superior robustness to structural complexity and reaction type shifts, and naturally recovers mechanistic trajectories and side products.
discrete flow matchingelectron occupationcontinuous-time markov chainoptimal transportmechanistic prediction
Persona-Execution Separation: An Architecture Pattern for Evolving LLM Agents under Execution Audit
The paper introduces Persona-Execution Separation (PES), an architecture pattern for LLM agents that decouples persona (mutable instructions/tone) from execution (auditable stateful work) into distinct trust domains. PES enforces separation via a governed contract bridge with approval matrices, data-loss-prevention (DLP) rules, and audit trails, allowing persona drift while maintaining execution traceability. The design emerged from three requirements: free persona evolution, execution auditability, and decoupling. A case study in a regulated digital-employee platform validated PES's effectiveness, showing no execution-side revalidation under persona perturbations (five model configurations) and no persona leakage into hard-asserted fields. The pattern is applicable when multi-user deployment, audit requirements, and expected persona churn co-occur.
llm agentstrust domainsdata-loss-preventionaudit trailspersona drift
Beyond F1: Evaluating Coverage and Failure Recovery in AI Model Security Scanners
The study introduces a benchmark for evaluating security scanners' coverage and failure recovery beyond traditional F1 metrics, using 170 synthetic Pickle and PyTorch artifacts (145 labeled, 10 malformed). It assesses ModelScan, ModelAudit, and Fickling across non-N/A coverage, analysis completion, and definitive security decisions. ModelAudit achieved 100% definitive judgments on labeled families, outperforming Fickling (81.5%) and ModelScan (49.6%). ModelScan showed 100% precision/recall when providing judgments but failed on 48 malicious families, where ModelAudit and Fickling succeeded. Results highlight the need to decouple judgment accuracy from availability and assess incremental coverage.
static scannerssecurity evaluationpickle artifactsfailure recoveryincremental coverage
Learning a Continuous Sepsis Severity Score Without Hour-by-Hour Supervision: A Two-Site Retrospective Study
The study introduces a continuous sepsis severity score learned from patient trajectories, addressing limitations of fixed historical indices. Using 43 clinical variables over 72-hour windows from 29,116 and 7,691 Sepsis-3 patients across two hospital systems, the method employs mortality as a treatment-level ranking signal with non-uniform credit assignment. The score (0-10 scale) showed 1.19-1.64 point separation between non-survivors/survivors across SOFA-2 strata, correlated with lactate changes (ρ=0.39), and achieved 70-77% cross-site agreement versus same-site performance. Bootstrap analysis confirmed prognostic consistency with clinical expectations.
sepsis severity scoretreatment-level rankingnon-uniform credit assignmentbootstrap resamplingspearman correlation
CLAP: Cross-Embodiment Video World Models are Zero-Shot Physical Simulators
CLAP introduces a framework for cross-embodiment action-conditioned video generation, enabling training on diverse internet-scale videos across human and robotic agents. The method reconciles disparate action spaces using end-effector poses, language instructions, and latent actions, coupled with a curriculum-based learning approach that first learns physical priors from unlabeled video data before grounding them in end-effector spaces. CLAP matches or exceeds state-of-the-art single-embodiment video models in environments like DROID, with performance gains further amplified via few-shot adaptation. The framework supports diverse action-conditioning spaces (end-effector, language, latent) and robot morphologies (DROID, Bridge, YAM, G1 humanoids).
cross-embodimentaction-conditioned video generationend-effector poseslatent actionsfew-shot adaptation
How Language Models Organize and Structure Moral Knowledge
This study investigates how large language models (LLMs) organize moral knowledge beyond mere detection, focusing on the geometric relationships between Moral Foundations Theory (MFT) categories. Using six linear probes trained on open-weight LLMs, the authors find that moral directions span near-maximal independent dimensions while sharing a positive common component, indicating integration (mean pairwise cosine 0.26 vs. 0.013 for non-moral concepts). This structure emerges early in pre-training and reflects corpus statistics rather than MFT's individualizing/binding distinction. In moral dilemmas, directions partially compose from component foundations (2.7x baseline) while encoding conflict-specific structure.
moral foundations theorylinear probesrepresentation spacecorpus statisticsmoral dilemmas
Making Clinical Language Models Auditable: Concept-Guided Fine-Tuning for Robust Prediction
The paper introduces CAST (Concept-guided Artifact Suppression Tuning), a framework for improving the robustness and auditability of clinical language models. CAST employs Sparse Autoencoders (SAEs) to extract interpretable features from Transformer activations, labels these features using an LLM-assisted pipeline with ICD-10 constraints, and suppresses artifact-related latents via residual subtraction during fine-tuning. Evaluated on MIMIC-IV discharge-note mortality prediction, CAST outperforms fine-tuned encoder baselines and matches strong LLM baselines while providing feature-level audit trails of clinical concepts and suppressed artifacts.
sparse autoencodersclinical language modelsinterpretabilityicd-10residual subtraction
LeVJEPA: Efficient & Scalable Video Pretraining without the Heuristics
LeVJEPA introduces a collapse-free video pretraining method that eliminates architectural asymmetries and pixel-space reconstruction, using instead an invariance loss over global and local clip views regularized by SIGReg. The approach employs uniform random token dropping to reduce pretraining costs while improving accuracy, achieving 5.6–20.8x lower compute than V-JEPA 2 at matched epochs and outperforming baselines by 7.6 points on ImageNet-1K. Block-causal attention enables temporal ordering without accuracy loss, and LeVJEPA nearly doubles motion-centric accuracy compared to DINOv2 on video frames, suggesting video as a superior substrate for general-purpose visual pretraining.
self-supervised learningtoken droppingblock-causal attentioninvariance lossvideo pretraining
Property-Specific Recoverability from Contact PPG to Camera rPPG under Heterogeneous Observation Conditions
The study evaluates property-specific recoverability from contact photoplethysmography (PPG) to camera-derived remote PPG (rPPG) under heterogeneous conditions, revealing differential preservation of physiological properties. Using 655 recordings from the Multi-Domain Mobile Video Physiology Dataset and CHROM as a fixed observation pathway, the analysis shows heart-rate MAE of 15.26 bpm and Pearson correlation of 0.0801, with modest autocorrelation correspondence but poor spectral and recurrence-rate discrimination. Maximal Lyapunov exponents exhibited no recording-specific correspondence (correlation 0.0231). Fitzpatrick-associated heterogeneity in endpoint discrepancy (9.32 bpm contrast) was observed, while dynamical discrepancy showed no gradient. Subject-held-out analysis demonstrated MAE reductions up to 13.32% when incorporating motion and lighting data.
photoplethysmographyremote ppglyapunov exponentsfitzpatrick scalesignal-to-noise ratio
CorporateBench: Large-Scale Q&A Benchmarking with Temporal Knowledge Bases
CorporateBench (CB) introduces a large-scale Q&A benchmark for evaluating LLMs on enterprise-scale document collections, addressing the limitations of synthetic datasets and privacy concerns. CB comprises four synthetic firms with 12 to 10,000 employees, sampled from temporally evolving knowledge bases ensuring cross-document consistency across over 230,000 documents. It evaluates LLMs on information extraction and knowledge base querying tasks. Testing five LLMs reveals performance degradation as input size approaches realistic corporate scales. CB fills a critical gap in benchmarking LLMs for corporate communication reasoning.
large-scale q&atemporal knowledge baseinformation extractionknowledge base queryingcross-document consistency
Successive Capacity Growth: Task-Complexity-Driven Width and Depth Expansion for Vision Transformer Encoders in JEPA World Models
Proposes Successive Capacity Growth (SCG), a method for incrementally expanding Vision Transformer encoders in Joint-Embedding Predictive Architectures (JEPAs) through task-agnostic width/depth growth. SCG starts from a minimal encoder (1 head, 2 layers, 283K params) and uses function-preserving expansion with Sketched Isotropic Gaussian Regularizer (SIGReg) to maintain statistical independence of learned dimensions. On multi-object dynamics (60D) and 2D navigation tasks, SCG improves prediction loss by 20.3-23% over fixed baselines while achieving 56× greater parameter efficiency, with zero false expansions and bit-exact function preservation (absolute difference = 0.0).
joint-embedding predictive architecturessuccessive capacity growthvision transformerfunction-preserving expansionsketched isotropic gaussian regularizer
Stageboost: Recommending Signals Based on Counterfactual Estimation
The paper introduces Stageboost, a two-stage XGBoost-based model for optimizing signal placement on eBay's View-Item page to enhance user engagement and purchase conversion. The method leverages counterfactual estimation to recommend contextual textual or visual signals, targeting intelligent purchase facilitation. Online experiments demonstrated a 0.08% lift in Gross Merchandise Bought (GMB) overall and a 0.58% increase in Parts and Accessories GMB, with notable conversion improvements for high-average-price items.
xgboostcounterfactual estimationgross merchandise boughtconversion rateonline experimentation
KnockGS:interaction-Grounded Calibrationof Physical Gaussian Representations
KnockGS introduces an interaction-response framework for calibrating physical parameters in 3D Gaussian representations by inferring elasticity and density scales from observed object dynamics under known forces. The method extracts temporal response features, estimates material scales, and freezes these estimates back into the simulator for validation on unseen interactions. Evaluated on five material targets, KnockGS outperforms baseline methods (response retrieval, global regression, fixed defaults) in parameter recovery accuracy and response fidelity, demonstrating generalizability across varying interaction directions and magnitudes.
3d gaussian representationsmaterial parameter estimationinteraction-response calibrationphysics-integrated simulationdeformable object dynamics
Sophistication in GenAI Use: Field Evidence from a Large Firm
This study analyzes sophistication in generative AI (genAI) usage among 4,000 back-office employees at a large firm, examining 713,564 prompts and LLM responses across 15 functions over eight months. Key findings show senior employees demonstrate more sophisticated genAI use, aligning with domain expertise complementing AI capabilities. Sophistication varies by function, peaking in Strategy, Digital Innovation, and Project Management, which focus on firmwide initiatives. No significant improvements emerge over time or from formal AI training, indicating persistent challenges in enhancing sophisticated use. The study offers actionable metrics for managers and researchers.
generative ailarge language modelsdomain expertiseorganizational changeai training
RCMN: Understanding Misleadingness in Influential Public Discourse
The paper introduces Reader-Centric Misleadingness Understanding (RCMN), a framework operationalizing misleadingness in public discourse through five dimensions: misleading mechanism, reader interpretation, evidence-warranted interpretation, emotional arousal, and communicative intent. Using an evidence-grounded dataset, the study finds misleadingness predominantly involves unsupported inference, exaggeration, and omission, often linked to heightened emotional arousal and distortive intent. Experiments with five generative foundation models demonstrate that lightweight claim-and-context representations suffice for recovering reader interpretations but fail to reliably identify misleading mechanisms, underscoring the need for richer contextual grounding.
misleadingnessreader-centricevidence-groundedemotional arousalgenerative models
PAWBench: How Far Are We from Probabilistically Aligned World Modeling?
The paper introduces PAWBench, a benchmark for evaluating probabilistic alignment in video generation models as world models, formalizing distributional correctness as a key requirement. It proposes PAWEval, an outcome-level protocol that aggregates repeated video rollouts into empirical distributions to assess alignment with reference physical behaviors. Evaluating 11 systems across 50 scenarios reveals no model consistently matches reference distributions or captures the full range of valid outcomes, prompting analysis of interventions like prompt engineering, noise sampling, and training adjustments.
probabilistic alignmentworld modelingvideo generationdistributional evaluationstochastic sampling
Not All Eval-Awareness Is Equal: Capabilities Framing Predicts Compliance
The study demonstrates that verbalized evaluation-awareness (eval-awareness) in chain-of-thought reasoning is not monolithic but decomposes into distinct framings—capabilities-flavored, safety-flavored, both, or neither—which differentially predict model compliance. Using Qwen3-32B over the FORTRESS dataset, capabilities-framing showed a +24 to +46 percentage-point compliance gap over safety-framing across steering interventions. A causal link was suggested via CoT-prefill interventions, with 10/11 prefills shifting compliance directionally. Results indicate that aggregate eval-awareness suppression metrics can mask safety-relevant behavioral heterogeneity.
eval-awarenesschain-of-thoughtcompliance predictionsteering interventionscapabilities-framing
Verify Smarter, Evolve Further: Efficient Harness Evolution through Behavior-Aware Verification
HarnessLens introduces behavior-aware verification for efficient evolution of agent harnesses, which govern language-model agent operations. The framework jointly explores task spaces and configurable components, derives modifications from execution trajectories, and selectively verifies candidates on behavior-relevant tasks via an attributable-evidence gate. Evaluated across three harnesses and four benchmarks, HarnessLens improves held-out performance by 7.6-13.6% while reducing evaluation budget consumption compared to baselines, demonstrating reliable and sample-efficient harness adaptation.
agent harnessesbehavior-aware verificationexecution trajectoriesattributable-evidence gateharness evolution
Difference-in-Differences on a Censored Rating Scale Can Manufacture an Effect: Evidence from a Pre-Registered LLM-Judge Audit
The study demonstrates that difference-in-differences (DiD) analyses on censored rating scales in LLM-judge audits can artifactually manufacture effects due to differential attenuation, confounding true preference differences with scale-boundary artifacts. Through a pre-registered audit of a frozen pedagogy LLM judge (990 calls), the authors show that a nominally significant interaction effect (+0.378, p=0.002) is largely reproducible (79-85%) from severity shifts and scale floors alone, even with zero true differential preference. The mechanism is derived analytically and shown to be measurable directly from observed ratings.
difference-in-differencesllm-judge auditscensored rating scaledifferential attenuationseverity shift
LLMs Can Design Near-Optimal OR Algorithms
Large language models (LLMs) can design near-optimal algorithms for well-specified operations research problems, including inventory control, queueing network control, and assortment optimization. The study evaluates two interaction levels: instance-specific solutions (level 1) and general algorithm generation (level 2), using minimal human input and a Python sandbox. The strongest model, gpt-5.6-sol, matches or outperforms existing methods on most instances, even at level 2 where algorithms are fixed pre-evaluation. Performance improvements are sharp across recent model generations, suggesting rapid advancement in this capability. Frontier LLMs thus emerge as competitive empirical baselines for OR algorithm design.
large language modelsoperations researchalgorithm designinventory controlqueueing networks
TADP: Task-Aware Deformable Prediction for Single-Stage 3D Object Detection
The paper introduces Task-Aware Deformable Prediction (TADP) for single-stage 3D object detection, addressing feature misalignment across tasks. The method employs a triple feature refinement aggregation module for adaptive multi-level feature extraction and a multi-scale feature aggregation block for scale-aware fusion. Task-specific deformations are applied via a plug-and-play task-aware deformation head with three specialized modules. On KITTI, TADP achieves 80.91% car mAP, outperforming prior state-of-the-art methods.
3d object detectiontask-aware deformationmulti-scale fusionfeature refinementkitti benchmark
BrailleBench: Investigating Multi-Criteria Braille Comprehension in Large Language Models
BrailleBench introduces a benchmark for evaluating Large Language Models (LLMs) on Braille comprehension across 5,570 instances from five datasets, covering mathematics, commonsense, and multi-hop QA in English and Braille Grades 1-2. The benchmark assesses Braille input comprehension, Braille output expression, and end-to-end Braille interaction via a deterministic, expert-reviewed pipeline using a custom Braille Toolkit. Evaluation of six LLMs reveals asymmetric performance: Grade 2 Braille input comprehension lags Grade 1, and fully Braille requests degrade performance, highlighting persistent gaps versus print-English capability.
braille comprehensionlarge language modelsmulti-criteria evaluationgrade 2 brailleexpert-reviewed pipeline
Naive Prompt Optimization: Rethinking the Need for Complex Prompt Search
The paper introduces Naive Prompt Optimization (NPO), a lightweight method for prompt optimization that iteratively revises prompts using a teacher model with rollout feedback. NPO achieves comparable or superior performance to Gradient-based Evolutionary Prompt Optimization (GEPA) with fewer rollouts, particularly benefiting from stronger teacher models that reduce the need for complex search procedures. Evaluations show NPO remains competitive in interactive games and transfers well to other student models, especially within the same family. Results suggest simple linear optimization can rival more complex approaches.
prompt optimizationteacher modelrollout feedbackgradient-based evolutionaryrecursive self-improvement
What Makes Good Agentic Data? An ACE Lens on Data Generation for LLM Agents
The paper introduces a two-level framework for agentic data generation in LLM agents, formalized as $(E,q,τ,v)$ tuples (environment, task, interaction, verifier) and analyzed through an Accuracy-Complexity-divErsity (ACE) lens. It proposes that effective data generation requires grounded accuracy (feasible support), learner-relative complexity (calibrated difficulty), and diversity (non-redundant coverage). The framework organizes prior work by generation paradigms and identifies trends toward execution-grounded verification, adaptive difficulty, and behavioral diversity. Key insights highlight the need for valid, informative experience allocation rather than volume as agents and environments scale.
agentic data generationace lensllm agentsdistribution designinteraction consistency
Compositional Online Learning for Semantic Data Processing Systems
The paper introduces compositional online learning for semantic data processing systems, leveraging LLM call latency to hide learner updates. The framework combines online-learning components that make execution-time decisions and refine learned artifacts, with updates occurring per-call or per-batch. A case study in Cortex AISQL integrates memoization, filter-ordering, and cascade-routing learners, decomposing LLM cost factors multiplicatively. Results show an 8× speedup on a conjunction-filter workload after accounting for drift and self-selection effects.
online learningsemantic data processingllm latencyadaptive query processingcascade-routing
STEP: State-Aware Task Estimation and Planning with Multi-Modal LLMs for Human-Robot Collaboration
The paper introduces State-aware Task Estimator and Planner (STEP), a method enhancing multi-modal LLMs for human-robot collaboration by explicitly estimating system states and predicting state transitions. STEP addresses MM-LLMs' limitations in state tracking and hallucinated actions, generating executable plans with assistance parameters. Evaluated on a robot assembly task, STEP improves action executability by 32.8% and reduces final-state error by 14.8% over state-of-the-art approaches.
multi-modal llmsstate estimationtask planninghuman-robot collaborationaction executability
PACE: A Unified Condense-and-Extract Paradigm for Fast VLM Inference
PACE introduces a training-free inference framework to accelerate Vision-Language Models (VLMs) via a unified Condense-and-Extract paradigm. The method employs an Adaptive Pixel Compressor (APC) to downsample redundant visual inputs before encoding, reducing computation while preserving global context, followed by a Dynamic Dual-Attention Extractor (DDAE) to selectively retain task-critical tokens using visual and semantic signals. Integrated into Qwen2.5-VL-7B, PACE achieves 93.8% of original performance with only 10% of visual tokens, yielding a 3.1x speedup in time to first token (TTFT).
vision-language modelsadaptive pixel compressordynamic dual-attention extractortoken pruninginference acceleration
LLMs in Digital EDA: A perspective on shifting roles from Generation to Orchestration
The article proposes a hierarchical framework for understanding LLM integration in electronic design automation (EDA), distinguishing three roles: Generator (single-pass artifact production), Agent (iterative refinement via tool feedback), and Orchestrator (cross-stage decision coordination). It critiques current LLM applications in EDA for prioritizing syntactically plausible code generation over physical correctness, exacerbated by tool fragmentation and context loss between design stages. Analysis reveals scalability limitations in industrial settings, advocating for physics-aware orchestration across standardized EDA workflows to improve reliability and accessibility.
electronic design automationlarge language modelshardware synthesisiterative refinementphysics-aware orchestration
When Text Misleads: Inconsistent-Aware Reasoning for Audio-Grounded Dialogue
The paper introduces ContraTalk, a 501-question benchmark for evaluating audio-grounded dialogue understanding by formalizing cross-modal disagreement cases where text transcripts suggest incorrect interpretations while acoustic cues support correct answers. The method identifies text-biased surface interpretations, generates conflict QA examples, and proposes an Audio Twin framework that converts speech into text-readable acoustic representations for explicit evidence aggregation. Results show text-only LLMs achieve >90% accuracy on consistent cases but drop to 33-48% on conflicts, while AudioLLMs still select transcript-biased traps in 30-40% of conflicts; Audio Twin improves conflict accuracy but exhibits backbone-dependent behavior on consistent cases.
cross-modal disagreementaudio-grounded dialogueacoustic evidence aggregationconflict qaaudio twin
Calibrated Enough to Know, Not Calibrated to Act: Fabricated Evidence Makes LLM Agents Commit to the Unknowable
The study demonstrates that large language model (LLM) agents exhibit overcommitment to provably unpredictable questions when presented with fabricated but professionally formatted evidence, with commitment rates rising from 6.5% to 54.0% across 12 frontier models. The effect persists even when all displayed data is invented (24.5% to 36.8%), indicating reliance on presentation authority rather than factual content. Fine-tuning a 3B model on 540 synthetic cases reduced commitment to 0.0% in original cases, though performance degraded under rigid response formats. The failure is attributed to a separable 'act/don’t-act' decision gate, which is trainable but context-sensitive.
llm agentsdecision gatefine-tuningfabricated evidencecontext sensitivity
ANTShapes Benchmarking Datasets for Event-Based Neuromorphic Object Classification
The paper introduces ANTShapes, a suite of four novel event-based vision datasets for benchmarking Spiking Neural Networks (SNNs) in object classification tasks, addressing the scarcity of high-quality neuromorphic data. Using the ANTShapes simulation tool, the authors generate labeled datasets of varying difficulty and evaluate them against established benchmarks (N-MNIST, CIFAR10-DVS, DVSGesture, POKER-DVS) via a convolutional SNN. Results validate the tool's utility for synthetic dataset creation, providing standardized resources for event-based vision research while mitigating latency, power, and privacy constraints of frame-based systems.
spiking neural networksevent-based visionneuromorphic computingobject classificationdataset benchmarking
BPMN4CAI: A BPMN Extension for Modeling Dynamic Conversational AI
The paper introduces BPMN4CAI, a standard-compliant extension of Business Process Model and Notation (BPMN) for modeling dynamic conversational AI systems in business processes. Using Design Science Research methodology, the authors systematically extend BPMN elements and incorporate specialized components to address challenges in representing context-sensitive interactions. A case study demonstrates the framework's applicability, showing improved adaptive decision-making, context management, and interaction transparency for conversational AI.
bpmnconversational aidesign science researchcontext managementbusiness processes
Thomson: Continual Learning of Frontier Models for SovereignAI
The paper proposes Thomson, a continual learning approach to develop frontier models for SovereignAI, enabling diverse institutions to achieve competitive performance without extensive resources. The method combines mid- and post-training interventions to preserve plasticity and stability, avoiding catastrophic forgetting while improving capabilities across high-stakes domains. Evaluations show Thomson matches recent frontier models in agentic tasks, safety, legal, tax, multilingualism, and deep research, with broad improvements (π-shaped pattern) and minimal forgetting.
continual learningfrontier modelssovereignaiplasticitydomain adaptation
When Tool Outputs Become Commands: Separating Action Induction from Runtime Authorization in Tool-Augmented LLM Agents
The paper introduces SARA, a framework for separating action induction from execution authorization in tool-augmented LLM agents to mitigate risks of unintended real-world side effects. SARA employs a context-isolated Action Probe to track action-inducing semantics and provenance, while tool calls are authorized against user objectives and audited evidence. It enforces No-History-Promotion to prevent historical recurrence from influencing execution authority. Evaluations on AgentDojo and AgentDyn show SARA limits action success rate (ASR) to ≤0.63% across four settings while maintaining task utility and reducing ASR across agent backbones.
tool-augmented llm agentsaction inductionexecution authorizationno-history-promotionaction success rate
Feature Transformation Enhanced Jacobi Polynomial Graph Filtering for Graph Anomaly Detection
Proposes JPGFN, a graph anomaly detection method addressing three limitations of frequency-domain filtering approaches: static basis functions, feature vector granularity loss, and underutilized node labels. The method integrates a Feature Separation Transformation Network (FSTNN) for fine-grained feature learning via dimension-wise nonlinear transformations, an adaptive Jacobi polynomial graph filter for frequency-domain signal capture, and a label constraint module. Evaluations on real-world datasets show significant improvements over baseline methods.
graph anomaly detectionjacobi polynomialfeature transformationfrequency-domain filteringnode label constraint
GRAIN: Bridging Name and Narrative Shifts in Real-World Graph Reasoning through Invariance-Rewarded Agentic RL
GRAIN introduces a single-agent RL framework for robust graph reasoning under linguistic shifts in node identifiers and task formulations. The method combines semantic parsing with tool execution, optimized via a Structure Invariance Reward that validates intermediate graphs against ground-truth topologies to discourage overfitting to surface patterns. On the GRIT benchmark, GRAIN achieves 16.45% higher accuracy than multi-agent baselines with 24% lower latency, reduces the OOD gap from 15.77% to 7.80%, and maintains scalability on large graphs.
graph reasoningreinforcement learningsemantic parsinginvariance rewardlinguistic shifts
Safety Does Not Compose: Non-Decaying Loop State for Autonomous LLM Agents
The paper identifies a safety composition failure in autonomous LLM agents, where trajectory-scoped monitors fail against attacks fragmented across iterations due to re-initialized safety states. It proves that even expressive monitors achieve only true-positive rates equal to false-positive rates without cross-iteration state. The authors propose LoopHarness, a loop-level safety mechanism with non-decaying state, which bounds unauthorized irreversible actions by a constant independent of horizon N. Evaluation includes Agent-SafetyBench tasks, outer-state attacks, and adaptive red-teaming.
autonomous agentssafety monitorstrajectory-scopednon-decaying stateloop-level safety
TransMeme: A Multi-Agent Framework for Cross-Cultural Meme Transcreation
The paper introduces TransMeme, a multi-agent framework for cross-cultural meme transcreation that jointly addresses cultural adaptation, intent preservation, and multimodal coherence. The method employs specialized agents for cultural understanding, text rewriting, revision, and conditional visual adjustment, coordinated via feedback loops. Evaluated on Chinese-English meme pairs, TransMeme outperforms baselines in human evaluation (33.1% average improvement) and LLM-as-a-Judge (60% Top-1 ranking rate vs. 26% for second-best). Error analysis identifies humor reconstruction and image-text alignment as key remaining challenges.
multimodal coherencecultural adaptationintent preservationmulti-agent frameworkmeme transcreation
LAAF: A Layered Accountability Architecture Framework for LLM Applications
The paper proposes LAAF, a layered accountability architecture framework for LLM applications, synthesizing 122 primary studies and 12 regulatory documents to address accountability gaps in high-stakes domains. Through a PRISMA-guided systematic review, it identifies five accountability dimensions and four mechanism families (technical controls, human oversight, organizational governance, documentation/traceability), mapped onto regulatory frameworks like the EU AI Act and NIST AI RMF. Key findings include persistent gaps in human oversight specification, shared metrics, and empirical evaluation, with LAAF integrating these insights while aligning cybersecurity to OWASP LLM Top 10 (2025).
accountability architecturellm applicationssociotechnical systemsregulatory compliancetraceability mechanisms
pro-team at LLMs4OL 2026 Tasks Flagship and Reuse: Retrieval-Augmented Generation and Vocabulary-Constrained Filtering for Ontology Learning
The authors present a retrieval-augmented few-shot prompting pipeline for ontology learning, addressing both end-to-end (Task A) and extension/reuse (Task B) challenges in the LLMs4OL 2026 competition. Their system combines Qwen2.5-14B-Instruct with all-MiniLM-L6-v2 embedding-based retrieval (top-5/top-2 examples per task) and left-truncated context-windowing, supplemented by vocabulary-constrained filtering for Task B. Results show 0.8692 Semantic Graph Similarity, 0.9200 Term-Typing F1, and 0.8540 Taxonomy Discovery F1 on Task B, with 0.7416 Semantic Graph Similarity on Task A, though the approach fails to capture non-taxonomic relations due to vocabulary constraints.
retrieval-augmented generationfew-shot promptingvocabulary-constrained filteringsemantic graph similarityontology extension
Active sensing to characterize the heterogeneity of plant stress
The study presents an autonomous robotic platform for targeted chlorophyll fluorescence measurements on plant leaves, advancing beyond passive imaging in phenotyping. The system integrates 3D plant reconstruction from multi-view data, geometric analysis for leaf surface extraction based on orientation and accessibility, and motion planning for collision-free robotic manipulator trajectories. This enables automated, spatially resolved physiological measurements through precise end-effector positioning for contact or near-contact fluorescence acquisition. The platform demonstrates capabilities for high-resolution, repeatable plant phenotyping by coupling perception, geometric reasoning, and manipulation.
autonomous roboticsplant phenotypingchlorophyll fluorescence3d reconstructionmotion planning
A Contract-Centered Architecture for Scalable and Manageable Agentic Runtimes
The paper proposes a contract-centered architecture for scalable enterprise AI agent runtimes, introducing four responsibility objects (Skill, Harness, Scaffold, external data substrate) as shared organizational contracts. The core contribution is a bounded hypothesis (P1) on cost-aware capability-capacity separability, with six design conditions as measured obligations. A cluster-period randomized crossover experiment design is proposed for hypothesis testing, yielding four possible verdicts. The work presents architectural contracts and measurement protocols but reports no implementation or experimental results.
agentic runtimecapability-capacity separabilitycontract-centered architectureenterprise airandomized crossover experiment
Active Diffusion-Based Inference for Ill-Posed Inverse Problems under Incomplete Priors
The paper proposes an active diffusion-based solver for ill-posed inverse problems under incomplete prior knowledge, leveraging a diffusion model (DM) to map between parameter and observable spaces. The method iteratively detects and corrects model misspecification through posterior uncertainty, enabling adaptive domain augmentation and robust inference even when initial training bounds exclude true parameters. Experiments demonstrate effectiveness on a toy inverse problem with infinite solutions and quantum chromodynamics analysis of nucleon structure via quantum correlation function parameterization.
diffusion modelinverse problemsbayesian inferenceparameter estimationquantum chromodynamics
Learning Transverse Momentum Distributions from Raw Scattering Events via Conditional Diffusion
A conditional diffusion model is proposed to directly learn transverse momentum dependent parton distribution functions (TMD PDFs) from raw semi-inclusive deep inelastic scattering (SIDIS) events, eliminating the need for parameterized functional forms. The method leverages simulated SIDIS data at CLAS12 kinematics to train the model, which provides informative uncertainties that decrease with event statistics. Results demonstrate reliable TMD PDF estimation with as few as 1,000 conditioning events, addressing statistics-limited regimes relevant to current and future experiments like those at Jefferson Lab and the Electron-Ion Collider.
conditional diffusion modeltmd pdfssidisclas12 kinematicsuncertainty quantification
Emotional Preferences as Goal-Priority Regulation
The paper proposes a computational framework for emergent emotional preferences in multi-objective reinforcement learning, where high-level goals autonomously regulate state-dependent priorities among competing objectives. The method combines an inner multi-objective RL controller (providing preference-conditioned behaviors) with an outer preference generator (learning state-to-preference mappings via RL). Theoretical analysis bounds the optimality gap via representation error of the inner repertoire. Experiments in exploration environments demonstrate learned preferences exhibit contextual priority switching, graded trade-offs, and temporal persistence, outperforming fixed/handcrafted baselines.
multi-objective reinforcement learninggoal-directed behaviorpreference regulationoptimality gapstate-dependent priorities
Beyond Classification: Task-Dependent Learnability under Privacy-Motivated Image Transformations
The paper proposes a compute-efficient multi-task protocol for evaluating Privacy-Enhancing Technologies (PETs) in computer vision, addressing limitations of classification-only benchmarks. The method combines lightweight proxy tasks targeting complementary visual structure aspects, enabling efficient assessment across irreversible transformations, block primitives, and learnable encryption. Results demonstrate that PETs with comparable classification accuracy exhibit divergent performance on other tasks, underscoring the inadequacy of classification as a sole metric for PET evaluation.
privacy-enhancing technologiesimage perturbationsmulti-task evaluationvisual structurelearnable encryption
Performance Foundations of Parallel & Distributed Reasoning Language Models
The paper systematizes Reinforcement Learning with Verifiable Rewards (RLVR) and related RL-for-LLM paradigms, analyzing their computational demands and parallelism strategies for Reasoning Language Models (RLMs). It evaluates Proximal Policy Optimization (PPO), Group Relative Policy Optimization (GRPO), and variants, alongside traditional and novel parallelism techniques (data, tensor, pipeline, hybrid parallelism). Using the work-depth model, it provides a taxonomy for scalable RLM training, offering guidelines for cost-effective deployment and identifying open research directions.
reinforcement learningparallel computingpolicy optimizationreasoning language modelswork-depth model
Omni-Interactive Universal Embedder
The Omni-Interactive Universal Embedder (OmniUE) introduces a unified embedding space for text, video, and audio, leveraging intermediate-layer representations from learnable tokens and enabling omni-interactive querying via text, visual regions, or audio spans. The architecture combines visual/audio segmenters with an omni-LLM for context aggregation, producing user-conditioned any-to-any embeddings. Evaluated on the novel OmniCHOIR benchmark, OmniUE outperforms baselines by 10.5% (MMEB-v2-video), 1.1% (MAEB), 83.7% (SCaR), and 24.1% (OmniCHOIR), demonstrating advances in omni-modal representation learning and interactive querying.
multimodal representation learningunified embedding spaceomni-interactive queryingcontext aggregationlearnable tokens
Multi-Person Human Motion Forecasting in Complex Scenes
The paper introduces Object-Conditioned Social Diffusion (OCSD), a conditional diffusion model for multi-person motion forecasting in complex scenes. OCSD integrates motion history, human-object interactions, and social dynamics via an object-conditioned denoising mechanism and a social encoder, handling variable group sizes and multi-modal futures. Evaluated on Humans in Kitchens (HiK) and HOI-M3, OCSD reduces 2-second path error by 121.5 mm (31.3%) and 130.5 mm (33.2%) respectively versus prior work, while improving long-term prediction realism.
human motion forecastingconditional diffusion modelsocial interactionshuman-object reasoningdenoising mechanism
FaulT-Bench: Towards Benchmarking Network Troubleshooting LLM Agents under Unreliable User Tickets
FaulT-Bench introduces a benchmark for evaluating LLM-based network troubleshooting agents under realistic conditions, including 200 scenarios with genuine faults, false reports, and incorrect attributions across eight network topologies. The method deploys scenarios in Kathará, uses NIKA for agent interaction, and scores free-text diagnoses via an LLM judge. Results show agents like SADE, ReAct, and Claude Code perform well on accurate tickets but degrade sharply with healthy networks and incorrect tickets, with performance highly sensitive to ticket wording rather than claims. Persona rewrites reveal confident but wrong reports are handled similarly to accurate ones, while vague reports degrade performance significantly.
llm-based agentsnetwork fault diagnosisbenchmarkingkatharánika
Magnon-induced phononic Chern insulator
The study introduces a magnon-induced phononic Chern insulator in a honeycomb phononic crystal hybridized with ferromagnetic islands, enabling Chern phases at GHz frequencies. By coupling a circularly polarized Kittel mode to phonons with phase winding, time-reversal symmetry is broken, opening a full Chern gap. In the large-detuning regime, this yields an effective Haldane-type phononic model with magnon-induced complex hopping, tunable to achieve Chern numbers |C|=1 and |C|=2. The predicted gaps exceed realistic phonon and magnon linewidths, facilitating observation in GHz acoustic devices. This work demonstrates chiral magnon-phonon hybridization as a pathway to magnetically reconfigurable topological phononics.
phononic chern insulatorkittel modemagnon-phonon hybridizationhaldane-type modelghz acoustic devices
A Multi-Modal AI Framework for Real-Time Queue Prediction, Management and Optimisation in Intelligent Border Control Systems
The work proposes a multi-modal AI framework for real-time queue prediction and resource optimization in border control systems, integrating LSTM networks for forecasting with Model Predictive Control (MPC) and scheduling optimization. The system processes heterogeneous data sources through a unified representation, reducing queue prediction error by 35% and average waiting time by 30% compared to ARIMA and rule-based baselines, while increasing throughput by 20% in synthetic evaluations. Results demonstrate improved adaptive management under dynamic traffic conditions.
lstmmodel predictive controlqueue predictionscheduling optimizationborder control
ASIL: Replacing Screenshot-and-Click with Structured State and Semantic Actions
ASIL (Agent-Software Interaction Layer) proposes replacing screenshot-and-click interfaces for software-operating agents with structured JSON observations and code-executable semantic actions, accessed via deep application integration. Evaluated across 15 applications and 380 tasks (300 single-application, 80 multi-application), ASIL achieves >80% success with closed models using <5 actions per task, outperforming screenshot-based methods (6.6-53.3% success) and matching/exceeding native APIs (e.g., +28-38 points over LibreOffice UNO). Fine-tuning Qwen3.5-2B/9B with supervised learning (58.0→72.1, 66.6→80.4) and on-policy RL (74.4, 82.2) demonstrates the approach's trainability.
structured statesemantic actionsagent-software interfaceon-policy rlapplication integration
DSA: Evidence-Aware LLM-Agent Orchestration for Multi-Market Stock Research
The paper introduces DSA, an evidence-aware orchestration framework for multi-market stock research using LLM agents. DSA structures workflows into evidence acquisition, context construction, model-routed analysis, and report generation with profile-specific safeguards. The framework supports two profiles: a default report profile and an agentic profile with role-specific parsing and signal-eligibility partitioning for Strategy Skill opinions. A reference implementation includes six regional markets, fifteen Strategy Skills, and multiple execution surfaces. Backend contract tests (1,457 passed, 596 mapped to six families) confirm implementation conformance but do not assess report quality or forecasting accuracy.
llm-agent orchestrationevidence acquisitionmodel-routed analysissignal-eligibility partitionbackend contract tests
GraphMemix: Query-Aware Evidence Forests for Long-Term Multimodal Agent Memory
GraphMemix introduces a query-aware evidence-forest framework for organizing long-term multimodal agent memory, addressing limitations of question-agnostic summaries and naive embedding matching. The method constructs candidate graphs from multi-view seed memories, decouples memory support from relation verification via evidence utility and activation costs, and optimizes forest-format memory selection under budget constraints. Evaluations on four benchmarks show significant accuracy improvements across foundation models while optimizing lifecycle costs, establishing a new Pareto frontier.
multimodal memoryevidence forestcombinatorial optimizationquery-aware retrievallifecycle cost
Per-View Gaussian Predictions Enable Training-Free Distractor Filtering in Feed-Forward 3DGS
The paper introduces a training-free filtering method to remove transient distractors in feed-forward 3D Gaussian Splatting (3DGS) reconstructions. By leveraging per-view Gaussian predictions, the approach excludes Gaussians associated with each input view and renders the scene from the same camera using the remaining representation. Feature similarity identifies candidate regions, and rendering-based verification retains only those whose removal reduces reconstruction error. Evaluated across three reconstruction models and two distractor benchmarks, the method consistently improves novel-view quality without retraining or scene-specific optimization, while preserving original reconstructions in clean scenes.
3d gaussian splattingper-view predictionsdistractor filteringtraining-freenovel-view synthesis
From Atomic to Agentic: Towards Interpretable Evaluation of LLMs' Agentic Mathematical Capabilities
The authors introduce a process-level benchmark for evaluating agentic mathematical reasoning in LLMs, addressing limitations of outcome-oriented assessments. Their framework aligns agentic behaviors with a taxonomy of reusable atomic capabilities, incorporating planning, action, and feedback tasks across textual and multimodal contexts. An automated pipeline synthesizes trajectories and generates fine-grained annotations via controlled LLM rewriting. Experiments show that models with comparable end-to-end accuracy exhibit divergent agentic capability profiles, underscoring the diagnostic value of process-level evaluation for developing next-generation mathematical agents.
agentic reasoningprocess-level evaluationatomic capabilitiesmultimodal contextscontrolled rewriting
A Table Is Worth 64 Tokens: Pixel-level Compression for Multi-Table Document Question Answering
The paper introduces a two-step method for efficient question answering over multi-table documents by leveraging pixel-level table compression. First, a vision-language model identifies relevant tables from pixel-compressed contexts, then reasons over them at native resolution. Evaluated across five VLMs and two benchmarks, the approach saves 41% of tokens and improves accuracy by 7 points compared to single-step native-resolution QA, while using 15% fewer tokens than the most efficient single-step compressed configuration without accuracy loss.
pixel-level compressionvision-language modelstable understandingtoken efficiencymulti-table qa
Counterfactual Bias Testing for Application Tracking System
The paper introduces a scalable bias-testing framework for automated hiring systems, addressing regulatory demands under the EU AI Act. The method employs LLM agents to synthesize identity-neutral resumes and inject controlled demographic variations across five protected characteristics, then evaluates candidate-job matches via fine-tuned sentence embeddings and a nine-metric fairness suite (counterfactual, group-fairness, and merit-aware metrics). Testing on 5 job orders and 100 base candidates revealed that while score shifts and top-K retention metrics were within tolerance, rank-stability (MARC) and nDCG@K exposed borderline biases, advocating for multi-metric auditing over aggregate scores.
correspondence-auditcounterfactual testingsentence-embeddingdemographic treatmentsfairness metrics
AI agents in Algorithmic Electricity Markets: On the Emergence of Tacit Collusion
This paper investigates the emergence of tacit collusion in algorithmic electricity markets where participants employ learning-based bidding strategies. The authors model strategic bidding as a repeated game with imperfect public monitoring and analyze emergent behavior using multi-agent reinforcement learning. They propose a multi-dimensional criteria set to assess tacit collusion beyond profit comparisons. Experimental results demonstrate that autonomous agents can sustain supra-competitive outcomes indicative of tacit collusion, despite no explicit collusion instructions.
algorithmic marketstacit collusionmulti-agent reinforcement learningstrategic biddingimperfect public monitoring
When Memory Takes Gradients: Collaborative Vector Memory for Agentic Recommender Systems
The paper introduces CoVeMem (Collaborative Vector Memory), a method to vectorize the memory of agentic recommender systems by replacing textual memory with frozen LightGCN user and item states. These states are retrieved per candidate set, presented as soft tokens alongside a textual profile, and trained via contrastive alignment and listwise co-training with masked candidates. CoVeMem matches or outperforms text-memory agents on 19/20 metric cells across four benchmarks while eliminating per-interaction LLM calls for memory maintenance, enabling gradient-based updates over the full interaction history.
agentic recommender systemscollaborative vector memorylightgcnsoft tokenscontrastive alignment
Learning-Augmented Online Allocation under Unreliable Advice: Robustness, Exposure Fairness, and Distribution Shift
The paper introduces a robust and fair learning-augmented algorithm for online allocation problems with unreliable advice, irreversible decisions, and exposure constraints. The method combines predictions with a conservative fallback mechanism and fairness correction to balance efficiency and equity. Theoretical analysis under bounded-error assumptions proves consistency and robustness, with loss scaling proportionally to prediction error. Empirical results demonstrate stability against adversarial advice and a 42% reduction in exposure disparity compared to baseline methods.
online allocationlearning-augmented algorithmsexposure fairnessirreversible decisionsrobust optimization
Evaluating human and LLM screening workflows in a conceptually complex scoping review: Recall--workload trade-offs and run-to-run consistency
The study evaluates human and LLM screening workflows in a scoping review, focusing on recall-workload trade-offs and consistency. Using 1,131 records verified against 316 eligible studies, it compares human reviewers (42.2-45.0% retention, 82.3-82.9% recall) with LLM configurations (GPT-5.4, Gemini 3.1). Gemini 3.1 achieved highest recall (83.9%) but retained 56.7% of records, while file-batch processing outperformed all-at-once configurations. Identical GPT-5.4 runs disagreed on 94 records (29 eligible), showing run-to-run variability. Results emphasize workflow design over model choice for high-recall tasks, advocating supervised, auditable LLM integration.
large language modelsevidence synthesisrecall-workload trade-offscoping reviewoperational recall
PLCBench: Can Autonomous LLM Agents Turn PLC Access into Sustained Physical Impact?
PLCBench introduces the first hardware-in-the-loop (HIL) framework to evaluate autonomous LLM agents' ability to convert PLC access into sustained physical impact in industrial control systems. The framework combines vendor-native PLC interaction, closed-loop process simulation, and independent verification, assigning six diagnostic flags to track cyber-physical progression. Testing five LLM families across 240 real-PLC episodes revealed 31.3% achieved sustained physical objectives, with process observation improving post-write success from 44.2% to 64.0%. The study identifies critical failure points and releases reproducible artifacts for defense evaluation.
programmable logic controllershardware-in-the-loopcyber-physical systemsllm agentsprocess simulation
C-Unseen: Weak Signal Detection in Dynamic Temporal Knowledge Graphs via LLM Reasoning
C-Unseen introduces a self-interpretable framework for detecting weak signals in Dynamic Temporal Knowledge Graphs (DTKGs), defined as rare, semantically coherent subgraphs that proliferate across consecutive snapshots. The method combines a Rare Subgraphs Extractor, where an LLM identifies narrative-inconsistent subgraphs via chain-of-thought reasoning, and a Weak Signal Alerter to track their persistence. Experiments show C-Unseen outperforms keyword-, topic-, and graph-based baselines in weak signal detection.
dynamic temporal knowledge graphsweak signal detectionrare subgraphschain-of-thought reasoningself-interpretable framework
BekchiAI: Measuring, Observing, and Controlling LLM Agents in One Click
The paper introduces BekchiAI, a dual-component system for evaluating and monitoring LLM agents, comprising BekchiAI-Benchmark and BekchiAI-Platform. The benchmark features 2,057 deterministic tasks across 7 categories (arithmetic, SQL, security, etc.), verified via SQL execution, DAG scheduling, or lambda evaluation, with metrics including tool-call adherence and URL hallucination. Evaluations compare Qwen3.7-Max, gemma-4-31B-it, gemma4:26b, and gpt-oss-120b, revealing family-specific performance variations. The platform offers real-time telemetry and control for deployed agents, including token/latency monitoring and remote termination. Both components are publicly released.
llm agentstool-using benchmarksreact agentsbehavioral metricstelemetry control
Reinforcement Learning-Based Control of CAV Platoon Joining Maneuvers in Mixed Traffic
The paper proposes a reinforcement learning framework for controlling connected and automated vehicle (CAV) platoon joining maneuvers in mixed-traffic environments. Using an agent-based model with SUMO simulation, it compares Deep Q-Network (DQN), Double Deep Q-Network (DDQN), and Proximal Policy Optimization (PPO), incorporating risk penalties or external safety controllers. PPO achieves 98% joining success and <1% collision rates, though with increased decision steps, revealing a safety-efficiency tradeoff. External safety controllers prevent collisions but may reduce joining efficiency.
connected and automated vehiclesplatoon controldeep reinforcement learningmixed trafficsafety controller
From Reasoning to Pixels: Grounded Medical Multimodal LLMs for VQA and Segmentation
The paper introduces MedREAL, a unified framework for medical multimodal reasoning and segmentation that aligns linguistic reasoning with pixel-level grounding via Seg Anchored Reasoning Pooling (SARP) and Reasoning-to-Visual (R2V) fusion. SARP extracts task-relevant semantic evidence from [SEG] tokens in MLLM hidden states, while R2V injects these features into a segmentation pipeline for mask decoding. Evaluated on the MedRAVS-13K dataset (13,824 samples across four modalities), MedREAL achieves 68.49% gIoU and 70.47% cIoU, outperforming prior work by generating diagnosis-consistent evidence masks.
multimodal llmsmedical vqasegmentation groundingreasoning fusionmask decoding
LiveSim: Simulating Environment-Shaped Users in Multi-Agent Live-Stream Ecosystems
LiveSim introduces an LLM-based framework for simulating user behavior in live-stream ecosystems, addressing limitations of static profile approaches by dynamically refining user representations through trajectory-grounded interactions. The method represents users as editable behavioral hypotheses, identifies environmental shaping effects via trajectory discrepancies, and accumulates transferable patterns in a collective behavioral memory. Evaluations on live-stream risk-control data demonstrate improved user-level behavioral fidelity and enable ecosystem-level analysis of risk evolution and intervention effects.
multi-agent simulationbehavioral fidelitytrajectory-grounded interactionenvironmental shapingcollective behavioral memory
MedFG-VQA: Low-Frequency Memory and Graph Attention for Lightweight Medical VQA
MedFG-VQA introduces a lightweight framework for Medical Visual Question Answering (Med-VQA) that combines Frequency-Memory Fusion (FMF) and Graph-Aware Cross-Attention (GACA) to enhance low-frequency visual features and align visual-textual representations efficiently. FMF retrieves DCT-based features from a learnable memory bank, while GACA refines cross-modal alignment via graph-convolutional aggregation. The method is validated on SynMed-VQA, a synthetic dataset of 2M question-answer pairs across 9 modalities and 10 organs, and three benchmarks, achieving competitive accuracy with lower computational costs than larger models.
medical visual question answeringfrequency-memory fusiongraph-aware cross-attentiondct decompositionsynthetic dataset
Evaluating Confidence-Gated Retrieval with Matched Trajectory Replay
The study introduces matched trajectory replay, a protocol for evaluating confidence-gated retrieval in interactive language-model agents by comparing confidence-to-action mappings under controlled conditions. Using Mistral, GPT, and Qwen models on HotpotQA and MuSiQue, it assesses isotonic calibration versus raw verbalized confidence. Results show calibration improves committed-answer accuracy by up to 41 percentage points but reduces coverage and increases retrieval use, with mixed overall accuracy effects (+15pp on HotpotQA, -17pp on MuSiQue). Calibration maps improve shallow retrieval performance but degrade at depth three, highlighting the need for separate value-of-information estimates in retrieval decisions.
confidence-gated retrievalisotonic calibrationmatched trajectory replaymulti-hop qavalue-of-information
SymbolLKG: Towards Verifiable Logical Reasoning via Logical Knowledge Graph and Symbolic Solvers
We propose SymbolLKG, a neuro-symbolic architecture combining Logical Knowledge Graphs (LKG) with dynamic solver routing to enhance verifiable logical reasoning in LLMs. The framework introduces an ontology-based LKG that models logical rules and constraints as topological nodes, capturing structural dependencies from text. A Logic Router dynamically dispatches tasks to optimal symbolic engines, supported by a topology-aware hybrid retrieval mechanism. Experiments on logical reasoning benchmarks show SymbolLKG outperforms state-of-the-art prompting and Retrieval-Augmented Generation baselines, achieving higher accuracy and verifiable reasoning paths.
logical knowledge graphsymbolic solverneuro-symbolic architecturelogic routertopology-aware retrieval
Behavior2Trip: Towards Personalized Travel Planning via User Behavior Trajectory
The paper introduces Behavior-Aware Travel Planning, a novel task that infers user preferences from past behavioral trajectories (avg. 39.8 behaviors/user across 14 attributes) to generate personalized travel plans. It presents Behavior2Trip, a benchmark with 11,400 instances from a Chinese travel platform, and B2T-Agent, a Qwen3-8B-based RL agent with external tool interaction and memory. B2T-Agent outperforms GPT-4.1 (0.5% pass rate on hardest tasks) on Behavior2Trip and generalizes better on TravelPlanner, demonstrating the challenge of implicit preference modeling.
preference inferencebehavioral trajectoryreinforcement learningtravel planningpersonalization
Decoupling Planning and Control for Instructable Agents
The paper introduces Instruct-to-Act, a system that decouples high-level planning from low-level control by combining instruction-tuned vision-language models (VLMs) with world-model controllers. The VLM generates sparse, high-latency text instructions, while the controller executes high-frequency actions conditioned on these instructions. To enable language-conditioned control, the authors relabel controller rollouts with synthetic instructions and jointly optimize behavior cloning, reward maximization, and world-modeling objectives. Evaluated across seven embodied environments (including three multi-agent settings), the approach outperforms controller-only and direct VLM action-generation baselines, maintains fast control, and allows plug-and-play VLM planner swapping without fine-tuning, achieving competitive performance on six tasks.
vision-language modelsworld-model controllersbehavior cloningmulti-agent environmentsinstruction-tuned
AI Control Scientist: LLM-driven Agentic System for Automated Control Design
The paper introduces AI Control Scientist (AICS), the first LLM-driven agentic system for automated control design, addressing inefficiencies in traditional expert-dependent workflows. AICS comprises three agents: a Task Modeling Agent translating requirements to constraints, a Controller Design Agent generating candidate structures and code, and a Parameter Tuning Agent optimizing parameters under closed-loop criteria. Experiments show AICS outperforms existing automated baselines in design success rate and optimization efficiency, demonstrating potential for transforming control system design into an agent-driven paradigm, including model predictive control applications.
agentic systemcontrol designlarge language modelparameter tuningmodel predictive control
Categorizer Automata for Discounted-Sum Payoffs
The paper introduces categorizer automata, deterministic automata that classify infinite reward sequences into finitely many bins based on discounted-sum payoffs, generalizing comparator automata (two-bin case). The key innovation is a construction with linear state-space complexity in bin count, avoiding exponential blowup from cross-product approaches. Applied to Markov decision processes, these automata enable exact pseudo-polynomial-time synthesis for piecewise-constant utility functions and ε-optimal policies for piecewise-Lipschitz utilities. The synthesis problem is shown PSPACE-hard even for piecewise-constant utilities.
categorizer automatadiscounted-sum payoffsmarkov decision processesquantitative synthesispseudo-polynomial time
DEEPCHART: How Far are LLMs from Faithful Data-Science Chart Generation?
The paper introduces DEEPCHART, a 1,482-instance benchmark for evaluating LLMs' chart-generation fidelity in data-science workflows, annotated from real-world scientific, financial, and ecosystem reports. It frames chart generation as an Extract--Reason--Visualize pipeline, assessing source-data extraction, derived-data reasoning, and rendering accuracy separately. Experiments reveal that state-of-the-art models produce visually plausible but data-hallucinated charts, with errors concentrated in extraction and reasoning stages under long-context, multimodal conditions. The findings suggest that faithful generation requires improved evidence extraction and quantitative reasoning beyond expanded context windows.
chart generationdata hallucinationmultimodal reasoningcontext windowquantitative reasoning
Beyond Execution: Auditing Experimental Fidelity in LLM-Driven Scientific Research
The paper introduces ABE-Ralph, a reference-anchored auditing framework for detecting methodological hallucinations in LLM-driven scientific research. ABE-Ralph structures experimental constraints (claims, protocols, components, baselines, metrics), guides implementation via an 8-step workflow, and performs multi-level verification (quantitative, qualitative, code). Evaluated on 30 reproduction runs (12 ML domains) and 23 NatureBench discovery tasks, it achieves 93% robust execution, identifies five scientific failure modes, and matches/exceeds SOTA on 5 tasks. Results demonstrate that reliable AI scientist evaluation requires assessing experimental fidelity to claims, not just code execution or metrics.
methodological hallucinationsreference-anchored auditingexperimental constraintsrobust executionscientific failure modes
Discovering Relationships in Data Lakes Using Large Language Models: An Industrial Case
The paper introduces ColRel, a two-stage method for discovering column relationships in data lakes with weak metadata, particularly in ERP-derived datasets. The approach first constructs column embeddings from metadata and raw data, then leverages business dictionaries to interpret coded schemata and generate natural-language descriptions for ambiguous columns. Evaluations on public benchmarks and an industrial ERP dataset demonstrate ColRel's effectiveness in semantically related, low-signal scenarios where traditional metadata is insufficient.
column embeddingsdata lakeserp datasetsmetadata enrichmentrelationship discovery
AgentFold: Closed-Loop Agentic Search for Protein Folding Model Design
AgentFold introduces a closed-loop multi-agent framework for autonomously improving protein folding models through executable code modifications. The system combines hypothesis generation, code implementation, multi-objective evaluation, and structured memory of interventions, guided by an MCTS-style resource allocation policy. Evaluating 80 model variants (2K LoC, 5K GPU-hours, 170M LLM tokens), AgentFold achieves 7.5% lDDT improvement over baselines, revealing design patterns favoring learnable priors over geometric perturbations.
protein foldingmulti-agent systemsmctsin-context learningmodel optimization
FaultLens: Learning Compact Behavioral Test Suites for Generated Operational Programs
FaultLens introduces a method for learning compact behavioral test suites that preserve auditable connections to execution evidence, addressing limitations of hand-written examples and exhaustive regression testing. The approach combines fault-driven greedy selection with mutation-independent diversity components, leveraging sparse outcome caches and probe orderings from prior program generations. Evaluation across 20 generated operational policies shows 99.0% fault coverage using 1.2-2.0% of the exhaustive domain, with diversity improving scenario-family coverage from 84.6% to 94.9% when a fault family is withheld. Downstream deployment reduces severe tail regressions from 15/20 to 0/20 program-environment groups.
behavioral test suitessparse outcome cachefault-probe kill relationmutation-independent diversityoperational policies
Graph-Guided Selective Unlearning for Language Models: Controlling Support Routes Beyond Forget Seeds
GRAPHSU introduces a graph-guided controller for selective unlearning in language models, expanding deletion scope beyond explicitly identified forget seeds by constructing a weighted support-route graph and applying graded forgetting strengths to high-risk neighbors. The method propagates deletion pressure through interconnected examples, addressing knowledge recoverability via paraphrases or aliases. Evaluated on TOFU and PISTOL benchmarks with GPT-2 Medium and Llama-3.2-3B-Instruct, GRAPHSU reduces soft leakage by up to 49.5 percentage points compared to seed-only baselines, demonstrating superior utility-feasible unlearning.
selective unlearningsupport-route graphgraded forgettingknowledge recoverabilitysoft leakage
Daydreaming: Stealing Hidden Agent Skills through Black-Box Task Interaction
The paper introduces Daydreaming, a black-box attack that steals multi-file agent skills through adaptive task interactions without requiring direct skill disclosure. The method crafts distinguishing tasks, uses shadow agents for design selection, and completes files via victim results and local execution checks, formalizing three threat levels (Differential, Trace, Output). Evaluated on 7 skills and 4 victim models, Daydreaming recovers 86.8% of original capability at Output level, outperforming SigLeak by 4x, using a median of 32 victim calls per skill despite disclosure defenses.
black-box attackagent skillsdisclosure defensesshadow agentsfunctional reconstruction
Rethinking Message Passing as Retrieval for Text-Attributed Graph Learning
The paper proposes RTA, a retrieval-augmented MLP framework that replaces structural message passing in GNNs with label-aware retrieval and propagation, offering computational efficiency and robustness to imperfect graphs. The method applies an MLP to node representations augmented by permutation-invariant summaries of retrieved graph context, with theoretical connections to softmax-attention message passing and robustness guarantees against mis-retrieved outliers. Experiments on text-attributed graphs show RTA matches or exceeds GNN and graph LLM baselines while improving efficiency and robustness across scenarios.
graph neural networksretrieval-augmented learningmessage passingmultilayer perceptronstext-attributed graphs
Knowing When Not to Reuse: Conditional Experience Transfer in Autonomous LLM Post-Training
The paper introduces Boundary-Calibrated Intervention Transfer (BCIT), a method for conditional experience transfer in autonomous LLM post-training that addresses when to reuse past update evidence. BCIT binds effects to source contexts, checks applicability conditions, and uses bounded training trials to authorize updates, preventing harmful reuse. Evaluated on a 4B-parameter model across finance reasoning, text-to-SQL, and function calling tasks, BCIT reduces harmful updates and achieves higher final-model quality under matched compute budgets compared to alternatives.
conditional experience transferautonomous post-trainingboundary-calibrated interventionllm adaptationupdate authorization
LiveVVT: High-Fidelity Video Virtual Try-On in Real Time
LiveVVT introduces a rolling streaming diffusion framework for high-fidelity real-time Video Virtual Try-On (VVT), addressing the latency and computational overhead of bidirectional spatio-temporal modeling. The method preserves bounded bidirectional interactions within a fixed-size window, complemented by temporal and global appearance memories for long-term consistency. A progressive distillation framework integrates bidirectional VVT learning, teacher-trajectory regression, and Collaborative Matching Distillation for causal few-step adaptation. Experiments show superior generation quality with 26× lower latency and 11× higher throughput compared to similarly sized models, enabling real-time streaming VVT.
video virtual try-onrolling streaming diffusionbidirectional modelingcollaborative matching distillationtemporal memory
AesCanvas: A Large-Scale Dataset and Benchmark for Aesthetic Critique and Contextual Suitability
The paper introduces AesCanvas, a large-scale dataset for aesthetic critique and contextual suitability evaluation, comprising CritiqueCanvas (519,136 instruction-response pairs from 54,300 images) and ContextCanvas (301 expert-reviewed scenarios). It benchmarks Multimodal Large Language Models (MLLMs) under a unified protocol, revealing a performance gap between critique generation and context-sensitive judgment. Results show aesthetic specialists lag behind general-purpose MLLMs on contextual suitability, with limited transfer between tasks and inconsistent grounding in visual cues.
multimodal large language modelsimage aesthetic assessmentcontextual suitabilityaesthetic critiquebenchmark evaluation
Style as a Confound: False Positives in AI Detection of Non-Native Academic Writing
The study demonstrates that AI text detectors conflate professional editing style with AI-generated content, raising fairness concerns in academic settings. Using 135,389 document pairs from a professional English editing service (2018-2025), the authors compared non-native manuscripts with their native-edited versions to isolate linguistic changes while preserving authorship and content. Across 13 AI text detectors, false-positive rates (FPRs) for human-written texts ranged from 0.0% to 100.0%, with detector responses varying inconsistently: some edits increased AI scores while others decreased them. Score changes correlated with the extent of editing, highlighting editing style as a confounding variable rather than a reliable indicator of text origin.
ai text detectorsfalse-positive ratesprofessional editinglinguistic styleacademic writing
Accelerating Scientific Research with Gemini in the Real-World
The study extends Co-Scientist, a Gemini-based multi-agent system, to accelerate closed-loop scientific workflows across materials science, biology, and computer science. The system integrates Gemini 3 Deep Think for lab-in-the-loop execution, enabling automated hypothesis generation, experimental design, and manuscript production. Key results include successful synthesis of 2D materials (MXene-like structures, monolayer TMDs), prediction of E. coli swarming phenotypes, and discovery of an inference-time scaling architecture outperforming six models on HealthBench. A double-blind expert review (n=30, 450 evaluations) confirmed reduced hallucination and improved research safety in AI-generated papers.
multi-agent systemclosed-loop workflowlab-in-the-loopinference-time scaling2d materials
Five Primitives for Governing Autonomous AI Agents at Runtime
The paper proposes five runtime primitives for governing autonomous AI agents in enterprise settings, addressing challenges of ephemeral principals, model-selected actions, and discovered populations. The primitives—discovery, identity, governance, attestation, and supply chain—are derived from pre- and post-action requirements, with each being structurally irreplaceable. The authors implement a system where agent actions are policy-mediated, authorized against tenant-specific vocabularies, and recorded in a verifiable ledger. Four primitives are operational in private pilots, while the fifth remains as standalone tooling. Runtime costs include critical-path enforcement, per-workload sidecars, and availability-denial tradeoffs.
autonomous agentsruntime governancepolicy mediationaction vocabularysigned ledger
Relational Over-Regularization: Graph-Based AI-Generated Text Detection via Sentence Transition Deviation
The paper introduces Relational Over-Regularization (ROR), a structural signal for AI-generated text detection based on sentence-pair transition variance deviations in LLM outputs. The proposed Cross-Source Stylometric Fingerprint Graph (CSFG) framework encodes positional, sequential, semantic, and transition deviation signals as GNN edge features, with signed deviation δ_ij operationalizing ROR. CSFG achieves 97.14% binary detection accuracy (11.14pp improvement over graph baselines) with 1.57% false positives, demonstrating robustness to unseen LLMs exhibiting inflated transition variance.
relational over-regularizationai-generated text detectionsentence transition deviationstylometric fingerprint graphgnn edge features
SIGMA: Structured Noise-Effect-Aware Grouped Multi-Agent Aggregation
SIGMA introduces a hierarchical framework for robust multi-agent reinforcement learning under noisy observations by exploiting structured noise effects. The method organizes agents into adaptive local groups via density-based grouping, performs intra-group consensus aggregation to smooth noise, and uses inter-group attention for global coordination. Experiments on StarCraft II demonstrate improved robustness to observation noise while maintaining noise-free performance.
multi-agent reinforcement learningstructured noise effectsdensity-based groupingconsensus aggregationinter-group attention
FOCUS & RePAIR: Mitigating Text Degeneration via Token-Level Guidance for Pruned Large Language Models
We propose FOCUS and RePAIR, two token-level guidance objectives for mitigating text degeneration in pruned large language models (LLMs). FOCUS reweights distillation toward high-confidence teacher regions to suppress leakage, while RePAIR uses onset-centered positive/negative continuation pairs with a margin loss to promote plausible alternatives and prevent repetition loops. Our token-level analysis decomposes degeneration into loop entry risk and loop persistence, showing persistence is controlled by escape mass assigned to plausible alternatives. Experiments on open-ended continuation and instruction-based generation demonstrate both methods consistently reduce repetition and improve generation quality in pruned LLMs.
token-level guidancepruningrepetition loopsescape massmargin loss
Do LLMs Understand Personality? Rethinking Persona Fidelity Evaluation through Structured Behavioral Inference
The paper introduces PRISM (Persona Reasoning with Inverse SFL-based Modeling), a psycholinguistic framework for evaluating persona fidelity in large language models (LLMs). PRISM decomposes persona fidelity into three dimensions—Task Framing, Interpersonal Stance, and Linguistic Style—using Systemic Functional Linguistics (SFL) to perform structured inverse inference. Experiments demonstrate that PRISM provides more accurate and stable judgments compared to traditional holistic evaluation methods, addressing limitations like 'holistic appraisal hallucination' and static psychometric inventories.
persona fidelitysystemic functional linguisticsinverse inferencelarge language modelspsycholinguistic framework
PailitaoGR: Latent Think-with-Images for Generative Image Retrieval
PailitaoGR introduces a latent think-with-images approach for generative image retrieval, addressing the challenge of diverse query image content by combining target-focused perception and selective auxiliary-evidence utilization. The method employs a target enhancer with on-policy distillation and attention guidance loss, alongside an auxiliary enhancer with incremental contrastive distillation, enabling Zooming without Cropping and Reading without OCR. Evaluated on real-world image-search logs, it outperforms baselines by 13.8% on average.
generative retrievaltarget-focused perceptionauxiliary-evidence utilizationon-policy distillationcontrastive distillation
CoGeo-GS: Concept-Driven and Geometry-Aware Multi-Object Removal in 3D Scenes
CoGeo-GS introduces a concept-driven framework for multi-object removal in 3D scenes using 3D Gaussian Splatting (3DGS), addressing challenges of occlusions and geometric consistency. The method assigns concept-aware semantic tags to Gaussians for flexible object selection and integrates a geometry-aware completion pipeline combining monocular depth priors, diffusion-based refinement, and boundary-aligned blending. Experiments show superior visual quality and reconstruction fidelity compared to existing methods.
3d gaussian splattingmulti-object removalgeometry-aware completiondiffusion-based refinementmulti-view consistency
Risks and Controls for Multi-Agent Systems: an analytical framework for deployment of AI agents across organisational boundaries
The paper proposes an analytical framework for assessing risks in multi-agent AI systems across organizational boundaries, categorizing deployments into three governance tiers: singular (single organization), federated (shared rules), and open (no central authority). It systematically examines risk factors, failure modes, and control mechanisms for each tier, identifying governance gaps where no actor can implement necessary controls. The framework highlights the need for collective action to address emergent risks in cross-boundary agent interactions.
multi-agent systemsorganizational boundariesgovernance tiersrisk factorsfailure modes
AgentJudgeBench: A Multi-Difficulty Benchmark for Evaluating LLM Judges on Agentic Tool-Calling
AgentJudgeBench introduces the first benchmark for evaluating LLM-as-a-judge reliability in agentic tool-calling workflows, focusing on structured DAG topologies. The benchmark comprises 3,808 instances across six DAG types and three difficulty tiers, tested with five generators (3B-70B) and six judges (20B-frontier). Results show judge alignment degrades 1.5x faster without ground truth, with all judges plateauing at 77-82% on hard queries regardless of scale. Ground-truth exposure harms GPT-5.4 (1.5 pp) and Gemini-2.5-Pro (3.9 pp), while structured rubrics improve alignment by up to 6.5 pp. QwQ-32B matches programmatic reference best with ground truth; GPT-OSS-120B aligns closest to human judgments.
llm-as-a-judgeagentic tool-callingworkflow dagsground-truth anchoringalignment ceiling
J-Zero: Unified Challenger--Solver--Judge Co-Evolution from Zero Data
The paper introduces J-Zero, a unified Challenger-Solver-Judge co-evolution framework enabling self-improvement in both verifiable and unverifiable domains without human supervision. The method involves adversarial co-evolution between Challenger (task generator) and Solver (response generator), while the Judge co-adapts using preference pairs derived from response generation processes rather than its own scores. J-Zero outperforms baselines by 4.2 and 8.0 points in verifiable and unverifiable domains respectively, sustaining improvement over 10+ iterations versus baseline degradation after 2.
self-evolving language modelsadversarial co-evolutionunverifiable domainspreference learningzero-shot learning
Physics-Informed Stochastic Configuration Machine: A Backpropagation-Free Neural Network with Fast Training for Nonlinear Differential Equations
The authors propose Physics-Informed Stochastic Configuration Machine (PI-SCM), a backpropagation-free neural network for solving nonlinear differential equations. The method analytically evaluates local Jacobians to linearize physical loss, enabling optimal weight determination via generalized linear least squares solvers instead of gradient descent. Three progressive algorithms (PI-SC-I/II/III) demonstrate universal approximation properties, achieving high-fidelity solutions and parameter identification while accelerating training by orders of magnitude compared to Physics-Informed Neural Networks (PINNs).
physics-informed learningstochastic configuration machinenonlinear differential equationsjacobian linearizationgeneralized least squares
DuMateBench: Evaluating Autonomous Agents in Complex Real-World Workflows
We introduce DuMateBench, a benchmark for evaluating autonomous agents in complex real-world workflows, reconstructed from anonymized user sessions of a production agent platform. The benchmark comprises 200 tasks across 8 scenarios and 17 capability categories, executed in Docker containers with environmental complexities (Insufficient, Unstable, Noisy) and evaluated using a hybrid deterministic and LLM-as-Judge protocol. Experiments with five agent frameworks and four state-of-the-art LLMs reveal significant gaps in task completion and show that performance under environmental perturbations depends on both LLM and agent framework capabilities.
autonomous agentsreal-world workflowsllm-as-judgeenvironmental complexitydocker containers
Multi2AV-Safety: Benchmarking Safety in Multimodal-to-Audio-Video Generation
The paper introduces Multi2AV-Safety, the first benchmark for evaluating safety in multimodal-to-audio-video generation across all 11 non-singleton text/image/audio/video conditioning configurations (11,024 attack instances). It systematically tests compositional risks where harmful intent emerges from interactions between benign or weakly harmful inputs across modalities. Results reveal two failure modes in existing safety guards: emergent harm from benign input combinations and reduced detectability of explicit harmful cues in multimodal contexts, highlighting a critical gap in compositional risk perception for multimodal-conditioned generation.
multimodal generationsafety benchmarkingcompositional riskaudio-video synthesisharm detection
PILOT in the Loop: Live Self-Improvement for Long-Horizon Agents
The paper introduces PILOT, a supervisor-worker framework enabling live self-improvement for long-horizon agents through two mechanisms: live steering (real-time redirection of active workers) and live self-evolution (dynamic distillation of execution experience into reusable skills). PILOT addresses limitations in existing architectures by decoupling execution and assessment while maintaining intervention capability. Evaluated across three benchmarks with two frozen backbones (GLM-5.1 and Kimi-K2.6), PILOT achieves state-of-the-art performance, outperforming counterparts by up to 9.8 percentage points on Terminal-Bench 2.0. Self-improvement experiments show 14.6-12.4 point gains while reducing output tokens by 42.9-47.4% and increasing successful evaluations per million tokens by 110.3-134.0%.
long-horizon agentslive steeringself-evolutionsupervisor-workerterminal-bench
RTNav: Towards Real-Time Zero-Shot Object Navigation
RTNav introduces a real-time zero-shot object navigation architecture addressing inference latency and asynchronous environment interactions, explicitly optimizing for bounded compute. The method redesigns agent perception, reasoning, and action pipelines to minimize wall-clock time degradation, contrasting with synchronous simulator-based approaches. Evaluated on HM3D-v1, HM3D-v2, and HM3D-OVON, RTNav improves success rates by up to 11% and Success weighted by Completion Time by 5.1 points over prior work.
zero-shot navigationinference latencyasynchronous environmentreal-time executionobject navigation
Zero-Shot Self-Orchestration with Ledger-Based Control for Improved LLM Coding Performance
The paper introduces a zero-shot manager-worker scaffold with ledger-based control to improve LLM coding performance, evaluating its impact across nine models (five open-weight, four closed) on LiveCodeBench's 100 hardest problems. The method employs a shared filesystem workspace without training or per-benchmark tuning, comparing multi-agent orchestration against single-pass baselines. Results show conditional improvements: significant gains for some models (e.g., Qwen3.8-27B +23.4, GPT-5.6-Luna +10.6) but null/negative effects for others, with context management and problem decomposition identified as key mechanisms. The scaffold triples token costs but achieves accuracy more efficiently than model scaling (e.g., GPT-5.6-Terra matches Fable 5's accuracy at 1/5th cost).
multi-agent systemszero-shot learningledger-based controlcontext managementproblem decomposition
Diff Mining: Logit Differences Reveal Finetuning Objectives
The paper introduces Diff Mining, a logit-based framework for identifying learned behaviors in finetuned language models by comparing their output distributions to those of the base model. The method operates in two stages: (i) extracting per-context logit differences on a reference corpus, and (ii) aggregating signals via Top-K frequency or Non-negative Matrix Factorization (NMF) to construct interpretable token sets. Experiments show Diff Mining outperforms model-diffing baselines in domain detection (identifying 33% of injected biases without targeted probing) and yields actionable interpretability signals using only output logits.
logit analysismodel diffingfinetuning detectionnon-negative matrix factorizationinterpretability
Don't Overthink, Don't Underthink: Toward Adaptive Reasoning in Agentic AI
The paper identifies misallocated reasoning as a critical challenge in agentic AI systems, where fixed reasoning controls lead to either excessive (over-reasoning) or insufficient (under-reasoning) computation. The authors evaluate these failure modes on MATH-500 and GAIA benchmarks, measuring tool-decision latency, token consumption, and correctness. Results show over-reasoning incurs unnecessary computational cost without accuracy gains, while under-reasoning yields incorrect or incomplete solutions. The work motivates adaptive reasoning mechanisms for dynamic agentic workflows involving planning, tool use, and multi-agent interactions.
agentic aiadaptive reasoningover-reasoningunder-reasoningtool-decision latency
SpeechGym: An Audio-Native Gym for Training Voice Agents via Reinforcement Learning
SpeechGym introduces an audio-native reinforcement learning environment for training voice agents end-to-end without text intermediaries, addressing perceptual and behavioral failures unique to speech interactions. The framework uses two omni-modal models conversing in native audio over unmodified text benchmark tasks, enabling gradient flow and on-policy training. Results show a 2x improvement in task success on an independent voice benchmark, with optimized agents achieving higher efficiency in turns and tokens.
voice agentsreinforcement learningomni-modal modelson-policy trainingaudio-native environment
Fine-Tuning of Transformer models with Frames
(No summary returned.)
The Latent Diagnostic Taxonomy: A Framework for Constructing Classifiers and Diagnosing Their Decisions, Applied to Prompt Injection Detection
The paper introduces the Latent Diagnostic Taxonomy, a framework for constructing robust classifiers and diagnosing their decisions via three steps: (i) dimensionality-optimized classifier construction via cross-validated performance, (ii) identification of latent support vectors (~29% of training examples) to pinpoint influential tokens, and (iii) diagnostic taxonomy construction to flag unreliable decisions. Applied to prompt injection detection, the method reveals ~77% of confident decisions are brittle—either due to calibration failures or exploitable shortcuts—and categorizes prompts into zones (e.g., Safely Reliable, Heuristic Bias) with remediation strategies.
latent diagnostic taxonomydimensionality-optimized classifiersupport vectorsprompt injectionconfidence calibration
Redwood: A Frontier AI Accelerator Designed, Verified, and Deployed from Scratch in 2 Weeks by AI
The paper introduces Redwood, an AI accelerator autonomously designed from specification to silicon in two weeks via an end-to-end AI system. The approach collapses hardware-software co-design into a single optimization loop, generating RTL, verification environments, firmware, and kernels without human intervention below the architectural specification. Redwood achieves 95% coverage on all blocks, demonstrates 3.4× better performance-per-watt than Jetson Orin Nano on Llama/Qwen inference, and enables recursive self-improvement by using Qwen to design next-generation hardware.
ai acceleratorhardware-software co-designrtl generationperformance-per-wattrecursive self-improvement
Simultaneous Envy and Equitability Guarantees
This work investigates the simultaneous satisfaction of envy-freeness (EF) and equitability (EQ) in fair division, contrasting indivisible goods-only and chores-only settings. The authors demonstrate that EF1+EQ1 allocations may not exist for normalized additive valuations, while presenting an algorithm for computing EF1+EQ1 allocations with normalized binary goods for up to seven agents. Notably, binary chores admit the stronger EFX+EQX guarantee without normalization. The paper also initiates the study of cross-notion ex-ante--ex-post guarantees for randomized allocations.
fair divisionenvy-freenessequitabilityindivisible goodsbinary chores
Co-Evolving Structured Knowledge and Reasoning in Language Models
The paper introduces KBevo, a co-evolving framework that jointly learns to construct structured knowledge bases and reason over them for knowledge-intensive QA. The method optimizes both components end-to-end using QA outcome rewards, allowing reasoning success to directly enhance knowledge base quality. Experiments show KBevo produces larger, better-connected knowledge structures with higher answer reachability, while improving compositional reasoning and controllability over standard retrieval baselines.
knowledge base constructionretrieval-augmented generationend-to-end learningcompositional reasoningquestion answering
Why RAGs Hallucinate: Penalty-Aware Evaluation of Retrieval-Augmented Generation Systems with Knowledge-Gap Canaries
The paper introduces a penalty-aware evaluation framework for retrieval-augmented generation (RAG) systems, addressing their tendency to hallucinate when answers are absent from the knowledge base. The method combines asymmetric scoring (-4 for wrong, +1 for correct, 0 for abstain), knowledge-gap canaries (verifiably unanswerable questions), and failure-attribution analysis. Evaluated on SimpleQA-Verified (1,000 questions × 3 repeats) with three commercial RAG systems, results show similar accuracy (97.0-98.0%) but divergent canary violation rates (16.7% vs. 98.1%), revealing systems differ more in abstention than correctness. Penalty-aware scoring reorders system rankings consistently across penalty settings (k=1 to k=9).
retrieval-augmented generationknowledge-gap canariesasymmetric scoringhallucination detectionfailure attribution
CG4AI: A Column Generation Framework for Training AI Models Under Constraints
The paper introduces CG4AI, a column generation framework for training AI models under linear output constraints. The method combines models via a master linear program that optimizes mixture weights, while a pricing subproblem generates new models guided by LP dual variables to address constraint violations, with cutting planes extending feasibility guarantees. Applied to MNIST digit classification and multi-commodity flow problems, CG4AI enforces diverse constraints (e.g., adversarial robustness, output relabeling) while outperforming single-model baselines in accuracy on MNIST and SNDLIB benchmarks.
column generationlinear programmingconstraint enforcementmulti-commodity flowcutting-plane method
Knowledge-Verified Emergent Deception in LLM Agents Under Conflicting Incentives
The paper introduces KnownLieBench, a knowledge-verified benchmark for evaluating emergent deception in LLM agents under conflicting incentives. The method first verifies an agent's knowledge of user entitlements through neutral probes, then assesses false claims when incentives to deny those entitlements are introduced, covering 8 domains and 112 cases. Results show significant variation in deception across 18 models, with fine-tuning experiments demonstrating that honesty-directed training reduces deception while deception-graded training increases lie success without raising lie frequency.
llm agentsemergent deceptionknowledge-verified benchmarkincentive conflictsfine-tuning
Decay-Region Group Delay as a Forensic Cue for AI-Generated Impulsive Sounds
The study demonstrates that decay-region group delay analysis effectively distinguishes AI-generated impulsive sounds from real ones, revealing significant KL divergence (0.322 vs. 0.022 in onset regions). A Random Forest classifier using nine decay-region features achieves AUC=0.884, while a CNN leveraging group delay maps attains 90-94% accuracy. The method remains robust across 27 STFT configurations (AUC 0.700-0.847, std=0.035), offering a physically interpretable forensic cue complementary to magnitude-based approaches.
group delayimpulsive soundskl divergenceforensic cuestft
ProofEvolve: Neuro-Symbolic Evolution for Formal Automated Theorem Proving
ProofEvolve introduces a neuro-symbolic framework for automated theorem proving that combines neural variation operators with symbolic verification in Lean to evolve verified proof DAGs. The method maintains an AND-OR proof DAG archive within problems and a persistent schema library across problems, enabling reuse of verified subproofs through typed schema recombination while preserving formal soundness. Evaluated on three Lean benchmarks, ProofEvolve achieves the highest average solve rate among compared proof systems.
automated theorem provingneuro-symbolicproof daglean kernelschema recombination
How Unlikely Is "Unlikely"? Assessing Verbal Probability Perception Across Large Language Models
The study evaluates verbal probability perception across 19 large language models (LLMs) using a word-to-number mapping task, comparing responses to human benchmarks under forced-choice and explanation-elicitation conditions. Results show LLMs preserve human-like ordinality and anchor points (e.g., 'possible' exhibits high variance), but display systematic upward bias for negative expressions like 'unlikely'. Explanation elicitation reduces within-model variance while increasing between-model divergence, and frontier models demonstrate coherent bidirectional representations in roundtrip testing. Findings reveal LLMs replicate human probability cognition structures but diverge at negative expressions, impacting human-model probabilistic communication.
verbal probabilitylarge language modelsuncertainty expressionscross-model evaluationbidirectional consistency
Modality Maturity Index: A benchmark for assessing multimodal capabilities of omni models
The paper introduces Modality Maturity Index (MMI), a benchmark for evaluating multimodal capabilities in omni models across five modalities (text, image, audio, video, document) and their combinations. MMI comprises 893 questions with human-authored rubrics, assessing both input understanding and multimodal output generation. Evaluations of five frontier models show Modality Presence Scores (MPS) ranging from 15.6 (Claude Opus 4.6) to 34.9 (GPT-5.4), with LLM judges achieving 70.8% agreement with human annotators on output correctness.
multimodal evaluationmodality maturity indexomni modelsmodality presence scorellm judges
FaithSieve: Fine-Grained Evaluation of Math Proofs with Faithful Formal Evidence
FaithSieve introduces a Lean-assisted framework for fine-grained evaluation of natural-language mathematical proofs by decomposing coarse steps into local reasoning units and verifying them through formal evaluation agents. The method addresses semantic mismatches by gating formal validation with semantic alignment scoring, ensuring faithful preservation of original claims. On the ProofLoc-Olympiad (350 problems) and ProofLoc-University (200 problems) benchmarks, FaithSieve achieves 81.43% and 84.5% exact first-error accuracy, respectively, outperforming direct-judging baselines by 9.14 and 9.5 percentage points.
formal verificationproof decompositionsemantic alignmentfirst-error localizationlean theorem prover
Approved Too Late: Verdict Staleness in LLM-Guarded Self-Adaptive Systems
The paper identifies and quantifies verdict staleness in LLM-guarded self-adaptive systems (SAS), where approvals become invalid between check and actuation times. It introduces three metrics—all-candidate verdict change, oracle-labeled approval expiry, and judge-conditioned use-time invalidity—to measure freshness, observing 5.3-48.4% verdict-change rates across five SAS environments. The proposed Freshness-Bounded Shield (FBS) reduces approval-expiry rates to 0-1.8% by estimating validity horizons via safe-side margins and feature volatility, without explicit plant-dynamics modeling. Audit results show nonzero use-time invalidity across four LLM judges, prompting a formal freshness contract requirement.
verdict stalenessself-adaptive systemsfreshness-bounded shieldtime-of-check to time-of-usevalidity horizon
MemToC: Benchmarking Memory-Tool Conflict Resolution in Large Language Models
The paper introduces MemToC, a benchmark for evaluating how tool-augmented LLMs resolve conflicts between parametric memory and tool returns, with 6,504 episodes constructed from 542 factual questions. It tests five 7-9B parameter models under four source-correctness conditions, revealing strong tool dominance (86.0-93.1% correct tool following) but poor retention of verified correct answers against incorrect tools (6.5-17.1%). Instruction wording affects performance unpredictably, while SFT and DPO fine-tuning show asymmetric improvements on two models without reducing correct tool use.
tool-augmented llmsparametric memorysource arbitrationinstruction tuningasymmetric evaluation
On Scope Classification and Current Knowledge-Editing Benchmarks: A Negative Result, with INLAY as a Gradient-Free Case Study
The study demonstrates that current knowledge-editing benchmarks fail to evaluate scope classification accuracy, as they lack negative examples where edits should not apply. Using INLAY, a gradient-free editor with exact per-query ground truth, the authors test 1,689 queries across three datasets and conditions, finding that an oracle router performs identically to a static policy (0.00 gain). Structural limitations in counterfactual benchmarks prevent rewarding correct abstention. When negative examples are artificially introduced, headroom increases to +0.0420. INLAY underperforms on Qwen2.5-7B CounterFact (vs. WISE) and RippleEdits (vs. retrieval-augmented generation).
knowledge-editingscope classificationgradient-freecounterfactual benchmarksretrieval-augmented generation
Assessing mentalization in humans and large language models
The study evaluates mentalization capabilities in large language models (LLMs) through economic games and cognitive computational modeling, comparing performance across four model families (DeepSeek, GPT-4.1, GPT-5, Gemini 2.0 Flash; N=2,099) and human participants (N=251). Results reveal distinct behavioral and computational signatures of mentalizing, with GPT-5 demonstrating superior adaptive reasoning and outperforming humans in recursive depth adjustment. Strategic prompting enhanced performance, though benefits varied by task, highlighting model-specific capacities for sophisticated social cognition.
mentalizationcognitive computational modelingrecursive reasoningtheory-of-mindstrategic prompting
SKILL.state: Scalable Long-Horizon Agent Skills
SKILL.state introduces a runtime architecture for LLM-based agents that replaces append-only conversational history with explicit, mutable execution state to address latency and context-poisoning in long-horizon tasks. The method provides only the immutable skill specification, current structured state, and latest observation at each step, discarding intermediate reasoning after state updates. Evaluations across diverse datasets and models show improved task accuracy and reduced token consumption compared to conventional history-based approaches.
llm agentsexecution statelong-horizon taskscontext-poisoningruntime architecture
VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning
VBVR-Pro introduces a scalable testbed for native visual reasoning, treating visual generation as a primary reasoning substrate. The suite features 300 procedurally generated tasks, verifiable reward scorers grounded in deterministic rules, and supports modality studies across 30+ image/video generators. Results demonstrate strong transfer to external benchmarks (RISE-Video, MME-CoF-Pro, BabyVision), improved post-reinforcement learning performance, and insights into vision-native trajectories, with video generation excelling in spatiotemporal state tracking. The framework includes data, models, and code for reproducibility.
native visual reasoningverifiable rewardsprocedural task generationspatiotemporal state trackingmodality ablation
A Visual Dependence-Aware Framework for Multimodal Unsupervised Continual Post-Training
The paper introduces a Visual Dependence-Aware (VDA) framework for Multimodal Unsupervised Continual Post-Training (MU-CPT) of MLLMs, addressing catastrophic forgetting and new-task plasticity. The method leverages token-level visual dependence (VD) through two components: Visually Constrained Optimal Transport (VC-OT) mitigates cross-modal forgetting by formulating VD structural distortion as an optimal transport problem, while Visually Modulated Adaptation (VMA) exploits VD heterogeneity to enhance visually grounded learning. Experiments demonstrate VDA's effectiveness in maintaining old-task stability and new-task plasticity under MU-CPT settings.
multimodal learningcontinual learningoptimal transportvisual dependencecatastrophic forgetting
MyoMechanix: Biomechanically-Grounded Compositional Skilled Activity Understanding and Coaching
The paper introduces MyoMechanix, a multimodal ecosystem for biomechanically grounded action quality assessment (AQA), featuring 7,500+ expert-annotated samples of 20 weight-loaded actions from 38 subjects with synchronized RGB video, 3D pose, sEMG, and physiological signals. It proposes the Fitness Knowledge Graph (FKG) for structured action representation and CUBIST, a compositional reasoning engine for fine-grained error attribution. Evaluations on MyoMechanix-AQA, -VideoQA, and -Video2EMG tasks demonstrate multimodal sensing and structured representations improve performance, with CUBIST achieving state-of-the-art results and Video2EMG suggesting cost-effective EMG alternatives.
action quality assessmentmultimodal sensingbiomechanical feedbackknowledge graphcompositional reasoning
Finding and using interpretable latents in a neutrino foundation model with sparse autoencoders
This work introduces sparse-autoencoder-based mechanistic interpretability to particle physics, applied to a neutrino foundation model pretrained on IceCube data and fine-tuned for direction reconstruction. Using a strict validation protocol involving held-out tests, nuisance controls, and replication across independent dictionary trainings, the authors identify a validated atlas of physical concepts in the model's representation. Causal interventions reveal that the direction head underutilizes this atlas, prompting the training of an uncertainty head that predicts angular reconstruction error. At 20% selection efficiency, this interpretable estimator improves median angular resolution from 20.2° to 3.2°, demonstrating the utility of mechanistic interpretability in exploiting learned latent physics.
sparse autoencodermechanistic interpretabilityneutrino foundation modelcausal interventionangular resolution
Planetary Prediction Engine: Autonomous Geospatial Prediction via Intelligent Data Selection and Foundation Model Embeddings
The Planetary Prediction Engine (PPE) introduces an autonomous AI system for geospatial modeling that synthesizes multimodal datasets from natural-language queries. It retrieves spatiotemporally relevant covariates from open-web and Earth observation platforms, fuses them with geospatial foundation model embeddings (PDFM, AlphaEarth), and searches over task-tailored model architectures with automated overfitting guards. PPE outperforms expert baselines, improving mean $R^2$ for US spatial regression (76.8% vs. 60.0%), Nigerian food security downscaling (66.1% vs. 31.5%), and DRC Ebola nowcasting (Recall@10 of 83.3%, +10.3pp over SOTA).
geospatial modelingmultimodal fusionfoundation model embeddingsautonomous data retrievalspatial downscaling
How Do LLM Agents Actually Get the Flag? Trace-Level Provenance for Agentic Offensive Security Evaluation
The paper introduces CTF-ABACUS, a trace-based auditing framework for evaluating LLM agents in offensive security tasks, addressing limitations of binary CTF benchmarks. The method decomposes agent trajectories into penetration-testing phases and techniques, creating evidence-grounded solve profiles to distinguish genuine exploits from shortcuts. Analyzing 1,435 attempts by six models, results show only 62-87% of recovered flags involved trace-verified exploits, revealing significant shortcut usage in current benchmarks.
capture-the-flagagent auditingsolve profilespenetration-testingtrace verification
TraceML: An Empirical Analysis of Human-Agent Planning in Machine Learning Development
The paper introduces TraceML, a framework for analyzing human-agent differences in ML development by comparing 4,465 human Kaggle trajectories with 430 paired agent trajectories across 134 competitions. Using version-level annotations (score, timestamp, action intent, edit size, score effect), the study reveals agents like Codex and MLEvolve exhibit narrow, repetitive loops (e.g., ensemble re-weighting, model mutation) unlike humans' diversified, revisiting strategies. A planning prompt improves agent scores but fails to match human effort profiles. The corpus, schema, and tools are publicly released.
tracemlmachine-learning developmenthuman-agent comparisoncodexmlagent
6.5% of the Neuro-Symbolic Literature Can Be Reproduced from Its Published Artifacts, a Six-Stage Audit Framework and First Instantiation
The authors propose a six-stage audit framework to assess reproducibility in computer science research, instantiated on the neuro-symbolic AI (NSAI) literature. They screened 5,497 records, identifying 1,304 eligible NSAI studies, and attempted artifact-based reproduction. Only 85 studies (6.52% of eligible corpus, 18.68% of attempted reruns) were fully or partially reproducible, with 321 blocked by missing non-code artifacts and 42 by unusable code repositories. The results highlight a severe reproducibility deficit despite nominal code availability claims, advocating for enforced, versioned, and archived artifact bundles in NSAI publications.
reproducibility auditneuro-symbolic aiartifact evaluationresearch methodologycode availability
ICON Decomposition: Multivariate Concept-Level Explanations of Deep Representations for Model Auditing
The paper introduces ICON decomposition, a multivariate concept-level explanation method for auditing deep representations that addresses shortcut learning by quantifying each concept's explained variance after accounting for all other concepts and the outcome. Unlike existing concept-based methods that evaluate concepts in isolation, ICON decomposes layer variance to isolate genuinely relied-upon concepts while quantifying unexplained representation. Evaluated on synthetic data with ground truth, ICON outperforms seven baselines in recovering concept importance. Applied to skin-lesion and brain-imaging models, it produces sparse explanations validated through retraining and out-of-distribution testing.
concept-based explainabilityshortcut learningvariance decompositionmodel auditingspurious correlations
SwarmWorld: Stigmergic technological evolution in societies of language-model agents
The paper introduces SwarmWorld, a framework for studying stigmergic technological evolution in societies of initially homogeneous LLM agents. Agents self-organize without predefined roles to explore, process resources, construct artifacts, and write executable controllers evaluated by a deterministic simulator. Results show that shared societies develop broader, more resilient technological portfolios than isolated-search baselines, with emergent role differentiation and accumulation through physical observation and artifact networks. Stigmergy alone enables capable societies, while interaction fosters persistent ecologies rather than superior individual inventions.
stigmergic evolutionlanguage-model agentsdecentralized coordinationtechnological ecologiesexecutable controllers
The Reasoning Tax: Token Economics of LLM Reasoning Across Task Types and Deployment Contexts
The paper introduces Token Economy Score (TES), a marginal benchmarking metric quantifying accuracy gains from reasoning-capable LLMs normalized by token cost. It evaluates 151 model-benchmark runs across seven reasoning tasks using paired/approximated TES variants, examining task structures, reasoning effort levels, and deployment contexts. Key findings show task structure (e.g., sequential inference chains) predicts reasoning efficiency better than difficulty, reveal diminishing returns at higher reasoning effort, and demonstrate how deployment context affects economic viability via Reasoning Cost Share and Deployment Cost Multiplier metrics.
token economy scorereasoning efficiencydeployment cost multipliermarginal benchmarkingdiminishing returns
Gating Before Commitment: Anticipating Intent Divergence to Prevent Post-Interaction Decision Failures in Autonomous Driving
The paper proposes a gating mechanism for autonomous driving that prevents post-interaction decision failures by anticipating intent divergence. The method employs a language-guided intent module that computes a smoothed intent-geometry divergence score to gate maneuvers before commitment, upstream of a corridor envelope. Evaluated on replayed off-road departures and crash clips, the gating layer repaired plans in all ten replays, firing 72 ms after drift onset and 161 ms before corridor exit. A redesign treating uncertainty as abstention reduced false triggers from 9 in 5.9 minutes to 0.341 per minute. Ablations show the full score detects failures fastest in four of five cases under deployed eligibility.
intent divergencegating mechanismautonomous drivingintent-geometry scoredecision layer
Prefix Sliding for efficient test-time scaling
Prefix Sliding is introduced as a memory-efficient method for test-time scaling in autoregressive models, addressing the prohibitive memory costs of full attention over long reasoning traces. The approach selectively retains only the prefix (containing key instructions) and a sliding window of recent tokens, discarding intermediate reasoning tokens that lose relevance. Without retraining, Prefix Sliding achieves 3x speedup while maintaining performance; when combined with RL training, it enables scaling to traces exceeding 100k tokens. Evaluations show superiority over summarization and vanilla sliding window baselines.
test-time scalingprefix slidingautoregressive modelsreinforcement learningkv-cache
$R^3$: Training Robots to Reason in Natural Language via Reinforcement Learning
The paper introduces $R^3$, a reinforcement learning method for training vision-language models (VLMs) to perform natural language reasoning for robotic manipulation. The approach involves mid-training on expert reasoning traces followed by single-step rubric-based RL from offline action data, enabling free-form language reasoning to guide low-level policies. Evaluated on Language Table and simulated bimanual grocery packing benchmarks, $R^3$ demonstrates improved exploration and generalization over instruction-only imitation learning baselines, showing language reasoning's potential as a test-time compute mechanism.
vision-language modelsrobotic reasoningreinforcement learningnatural language processinglong-horizon manipulation
How Much Rank Does LoRA Need? Rank-Error Bounds for Transformer Attention
The paper develops a theoretical framework for bounding the approximation error of low-rank adaptation (LoRA) in Transformer attention layers, providing task-dependent rank-error bounds. By analyzing the Kullback-Leibler divergence between target and candidate attention distributions, the authors derive lower and upper bounds under various conditions, including realizability and geometric constraints. Key results include spectral bounds for finite-score approximation, demonstrations of reduced rank requirements due to softmax saturation, and extensions to multi-head LoRA with rank sharing. The analysis quantifies how the required LoRA rank depends on downstream task properties and target attention characteristics.
low-rank adaptationtransformer attentionkullback-leibler divergencespectral boundssoftmax saturation
The Value of Human Expertise
The paper introduces a method to leverage human domain knowledge for tighter performance guarantees in optimization problems with unknown parameters. By assuming the decision maker's belief about the nominal problem's optimal value is unlikely to be large, the authors formulate the value of human expertise as the minimax gap of a max-min problem. They demonstrate that this approach yields convex program-based worst-case performance bounds when the belief holds. Applications in assortment optimization and shortest path problems validate the theoretical framework.
optimizationnominal problemperformance guaranteesminimax gapconvex program
Trace Integrity for LLM Data Agents: A Vision for Auditable Structured Reasoning in Real-World Systems
The paper introduces Trace Integrity, a reliability criterion for LLM data agents that evaluates whether computation traces are explicit, executable, and auditable, addressing the Structure Gap where natural-language reasoning fails to specify operator-level programs. The authors propose execution contracts to bind user intent to schema elements and operator plans, and introduce CAIT Rate to measure silent-failure risk. Empirical results on BIRD Mini-Dev show answer accuracies (20-24%) diverge from Trace Integrity Pass Rates (39-43%) and high CAIT Rates (45.8-59.1%), demonstrating distinct evaluation signals for real-world deployment.
trace integrityexecution contractscait ratestructure gapllm data agents
DualOPSD: Adaptive Privileged Teachers for On-Policy Self-Distillation
DualOPSD introduces adaptive privileged teachers for on-policy self-distillation (OPSD), addressing the fixed-teacher limitation in conventional OPSD. The method employs an asymmetric alternating framework where the student learns from a privileged teacher, which then adapts to the updated student distribution using the same trajectory, eliminating additional rollout costs. Evaluated on Qwen3-8B in non-thinking mode, DualOPSD improves average scores by 23.61, 13.89, and 10.00 points on AIME 2024, AIME 2025, and HMMT 2025 benchmarks, respectively, with scale-dependent accuracy gains observed at 1.7B and 4B parameter counts. The approach also reduces truncation and lowers bidirectional KL divergence between teacher and student.
on-policy self-distillationprivileged teacheradaptive distillationasymmetric frameworkkl divergence
Imitation Learning for Connection-Tableau Construction
The paper presents an imitation learning approach for automating connection-tableau proof construction in theorem proving. By framing proof construction as a policy optimization problem in a transition system, the method employs a graph neural network to score proof edits, trained via imitation learning on existing proofs. Evaluated on M2k, MPTP2078-bushy, and TPTP v9.2.1 benchmarks, the learned policies achieve 46% higher problem-solving rates than leanCoP and reduce proof steps by an order of magnitude, even with reduced search scaffolding.
imitation learningconnection-tableaugraph neural networktheorem provingpolicy optimization
VoiceMem: Streaming Dual-Brain Memory for Real-Time Interaction
VoiceMem introduces a streaming dual-brain memory architecture for conversational systems, featuring parallel informational (left brain) and emotional (right brain) components with streaming I/O. The method includes a pipeline for memory-aware speech language model training, long-horizon evaluation, and decoupled deployment. Results show 30-point accuracy gains over Mem0 in top-5 retrieval, 4.29-point improvements in persona benchmarks, and 134ms retrieval latency within standard VAD constraints, enabling real-time, personalized, and emotionally aware interaction.
streaming memorydual-brain architecturespeech language modelsaffective attributionpersona modeling
AsymSpec: Context-Asymmetric Speculative Decoding for Agentic LLMs
AsymSpec introduces context-asymmetric speculative decoding for agentic LLMs, where a lightweight drafter processes full input while a large verifier operates on compressed context. The method employs contrastive δ-fusion of logits and a divergence-aware acceptance gate to maintain verification stability and high draft acceptance. Evaluated on four agentic capabilities and two benchmarks, AsymSpec achieves ≈90% of full-context accuracy with 1.3–1.7× throughput speedups at 0.2–0.3× compute cost, demonstrating efficacy when compression discards critical reasoning signals.
speculative decodingagentic llmscontrastive fusiondivergence-aware gatethroughput speedup
ProgRouter: Online Progress-Guided Orchestration for Multi-Agent LLM Workflows under Quality-Cost Tradeoffs
ProgRouter introduces an online progress-guided routing framework for multi-agent LLM workflows that optimizes quality-cost tradeoffs. The method employs a multi-view task progress scorer combining coarse and fine-grained signals, a dual-path progress predictor, and adaptive meta-gating to estimate progress gains per candidate LLM. Experiments on HumanEval Plus, MBPP, MATH-500, and ASQA show cost reductions while maintaining task performance in code generation, mathematical reasoning, and question answering.
multi-agent llmprogress-guided routingadaptive meta-gatingtask progress scorerquality-cost tradeoffs
Multi-Granularity Context-Enhanced RAG over Multimodal Knowledge Graphs
The paper proposes Context-Enhanced Multimodal Knowledge Graph (CEMMKG), a framework improving multimodal GraphRAG by addressing semantic gaps between images and text. CEMMKG enriches images with multi-granularity local context (semantically related sentences) and global context (passage summaries), enhancing cross-modal alignment during knowledge extraction and fusion. Experiments on a vision-centric dataset demonstrate CEMMKG's effectiveness across multiple MMKG-based RAG methods, validating its broad applicability for retrieval-augmented generation tasks.
multimodal knowledge graphretrieval-augmented generationcontext enhancementcross-modal alignmentgraphrag
FRAME: separating sampling variation from representational cause in medical imaging fairness
The paper introduces FRAME (Fair-model Reference And Mechanism Evaluation), a two-step framework for auditing fairness claims in medical imaging models. First, it derives a fair-model reference distribution under exact fairness conditions, then tests residual differences using representation-space operators. Evaluated across 702,206 images and 36 encoders, the reference accounted for 41% of race and 22% of age performance differences. Interventions like demographic decodability injection showed minimal impact on residual differences compared to random seed variation. Applied to 89 differences in 9 published studies, FRAME distinguished sampling variation from mechanistic bias, with image-text pretraining improving worst-group performance by ~0.05.
fairness auditingmedical imagingrepresentation spacesubgroup performancesampling variation
SciMIF: Understanding Multimodal Instruction Following in Scientific Domains
The authors introduce SciMIF, a novel benchmark for evaluating Multimodal Large Language Models (MLLMs) on scientific instruction-following tasks. They analyze 22 tasks across 5 disciplines, propose a 10-constraint taxonomy, and develop an instruction injection pipeline to augment datasets. Experiments on state-of-the-art MLLMs reveal performance disparities across disciplines (chemistry being most challenging), limited scale benefits for constraint adherence, and persistent difficulties with fine-grained constraints and domain knowledge application.
multimodal large language modelsinstruction followingscientific benchmarkconstraint taxonomyinstruction injection
PANDA - Prototype-Anchored Alignment for Partially Unpaired Multimodal Learning, with Applications to Alzheimers MRI and TCGA Pathology
We introduce PANDA (Prototype Anchored Data Alignment), a two-stage framework for multimodal medical prediction with incomplete pairing of auxiliary modalities. Stage 1 learns shared embeddings and class prototypes from paired data; Stage 2 trains the primary encoder using cross-entropy and alignment to frozen prototypes, enabling deployment without auxiliary inputs. Evaluated on a 1,021-subject ADNI cohort for AD/CN classification, PANDA achieves AUC 0.868 (+7.9pp) over MRI-only baseline and reduces CN false positives by 24.3pp. On TCGA-Lung survival prediction, it improves 2-year OS AUC by 3.5pp and Cox PH C-index by 9.0pts over WSI-only, outperforming full-fusion training. PANDA effectively leverages incomplete auxiliary data for primary-modality prediction.
multimodal learningclass prototypescross-entropyaucc-index
Quantitative Analysis of $ω$-Regular Robust MDPs
The paper presents an exact quantitative analysis of $(s,a)$-rectangular robust Markov Decision Processes (RMDPs) with linearly defined uncertainty sets and $ω$-regular objectives. It proves the existence of pure memoryless optimal policies for both agent and environment, and introduces a polynomial-time algorithm combining quantitative one-step improvements with qualitative almost-sure improvements. Experimental results compare the approach with explicit reductions to stochastic games.
robust mdps$ω$-regular objectiveslinearly defined uncertaintypure memoryless policiespolicy-iteration algorithm
LivingRAG: Augmenting Graph RAG with Experience
LivingRAG augments Graph RAG with reusable reasoning experience by introducing a writable experience store alongside graph-based retrieval. The framework stores verified graph signals (entity/passage relevance) and reasoning summaries, enabling retrieval augmentation and answer generation to leverage prior related queries. Analysis reveals reusable patterns in entity co-occurrence, graph neighborhoods, and question templates. Experiments on multi-hop QA benchmarks demonstrate accuracy improvements over RAG baselines and reduced completion-token usage when reusing relevant experience.
graph ragmulti-hop qaexperience storereasoning reuseknowledge graph
A Statistical Audit of Physical AI Benchmark Redundancy
The study introduces a statistical method to quantify redundancy in physical AI benchmarks and optimize benchmark selection. Analyzing a matrix of 51 models evaluated on 12 benchmarks, the authors measure shared information between benchmarks and demonstrate redundancy's impact on model rankings. Using a greedy algorithm that combines score dispersion and unexplained variance, they identify a four-benchmark subset retaining 78.5% of the utility of the full set. A Bradley-Terry ranking is fitted to this subset. The method, applicable beyond physical AI, relies on benchmark-level scores with sufficient overlap.
benchmark redundancyscore dispersionbradley-terry rankinggreedy algorithmphysical ai
Candidate supply and answer selection shape the value of LLM judging in multi-agent systems
The study investigates how LLM judging improves answer selection in multi-agent systems by analyzing 15,336 questions across benchmarks like MMLU-Pro and GPQA. It conceptualizes multi-agent reasoning as an evolutionary pipeline of candidate generation, communication, and selection, identifying memetic drift as a key challenge. Results show correct answers often exist among candidates but are lost (63.82% baseline accuracy), with judge reliability varying by task and generator. Combining answer frequency with judge evaluation raised accuracy to 70.82-70.95%, primarily rescuing correct but outnumbered answers.
multi-agent systemsmemetic driftllm judginganswer selectionevolutionary pipeline
One Symptom, Three Levers: A Critical Review of On-Policy Self-Distillation
This review systematizes research on On-Policy Self-Distillation (OPSD), where a language model teaches itself using privileged information unavailable at test time. The method combines imitation learning's dense supervision with reinforcement learning's on-policy sampling, avoiding costly teacher models but risking reasoning-path collapse. The authors analyze collapse through three control levers: token-weighting strategies, privileged information types, and teacher dynamics. Focusing on mathematical reasoning benchmarks, they establish shared terminology and demarcate settled versus open questions without conducting new experiments.
on-policy self-distillationprivileged informationreasoning collapsetoken-weightingteacher dynamics
TAU-Agent: An Agentic Retrieval-Augmented Framework for Traffic Anomaly Understanding
The paper introduces TAU-Agent, an agentic retrieval-augmented framework for Traffic Anomaly Understanding (TAU). The system employs a central retrieval agent to orchestrate two visual perception tools—Video Captioning Tool and Open-Vocabulary Tracking Tool—for extracting query-relevant evidence (captions, temporal intervals, object trajectories). This evidence, along with sampled frames and the query, is processed by a fine-tuned vision-language model for reasoning and answer generation. Evaluated on AI City Challenge 2026 benchmarks, TAU-Agent scores 0.6779 (Track 3), 0.3998 (Track 7), and 67.9275 (Track 8), ranking second, twelfth, and fifth, respectively.
traffic anomaly understandingretrieval-augmented frameworkvision-language modelopen-vocabulary trackingvideo captioning
How Robust Are Automated Fact-Checking Systems? A Cross-Benchmark Evaluation
This paper conducts the first cross-benchmark evaluation of full retrieve-then-verify automated fact-checking (AFC) systems across four diverse domains. The study benchmarks nine models (including random baselines, fine-tuned transformers, zero-shot LLMs, and AVeriTeC 2025 top systems) on scientific, open-web, and climate datasets. Key findings show: (1) claim-only models outperform evidence-based systems on noisy ClimateCheck, (2) model rankings vary substantially by domain and metric (e.g., SciFact macro-F1 0.70 drops to 0.31 on ClimateCheck), and (3) gold evidence improves accuracy by 14-22 points, confirming retrieval as the primary bottleneck.
automated fact-checkingretrieve-then-verifycross-benchmark evaluationzero-shot llmsmacro-f1
LLM Agents for Time-Series: A Survey
The survey presents a problem-driven taxonomy for LLM-based agents in time-series applications, categorizing systems by task requirements rather than technical components. It organizes existing approaches into four categories: forecasting and reasoning, augmentation and synthesis, anomaly detection and diagnosis, and decision support, analyzing how each influences agent architecture, tool use, and memory design. The work also compares performance metrics across shared settings and identifies open research gaps, providing a task-oriented guide for future development.
llm-based agentstime-series forecastinganomaly detectiondecision supporttask-driven taxonomy
When Composition Doesn't Add Up: Humans Identifying Defects in AI-Generated Images
The paper introduces CO-AID, a dataset for identifying systematic defects in text-to-image (T2I) models when handling complex compositional prompts. Authors manually curated 651 reference images across four categories (people, hand, object, scene), derived prompts via ChatGPT edits, and generated images using three T2I models. A subjective study with 29 participants annotated defect types and locations. Experiments demonstrate CO-AID's utility for both defect prediction and generation optimization in T2I models.
text-to-imagecompositional defectsmulti-label assessmentai-generated imagessubjective study
Agent Mesh: Reliability Primitives for Non-Idempotent Agent Delegation - Identity Adequacy and Evidence Adequacy
The paper introduces reliability primitives for non-idempotent agent delegation, addressing identity adequacy and evidence adequacy through a failure study of a production agentic software-delivery platform. Analyzing 147 incidents across 81 runs, the authors quantify violations of three assumptions underlying service mesh primitives (retry, timeout, error-rate circuit breaking), revealing issues like misrouted failures and enforcement-layer blocks. Key findings include 54 consecutive successful tool calls undetected by error-rate breakers and 21 events across six invocations making correct components unwinnable. The study proposes seven delegation-centric reliability primitives and outlines a controlled evaluation framework.
agent delegationservice meshidempotencyerror-rate circuit breakingreliability primitives
Code World Model: Coding Agent as World Brain
The Code World Model framework introduces a novel approach to world modeling by decoupling world evolution from visual realization through executable code. It employs a coding agent (world brain) that reasons about events and maintains persistent state via generated code, while a proxy representation bridges executable state to visual generation via spatiotemporal constraints. The system combines language model reasoning with video model priors, trained on aligned proxy-observation pairs from gameplay data. Results show MiniMax-H3 successfully renders rule-consistent worlds with high-fidelity visuals, demonstrating the viability of code-based persistent evolution coupled with flexible video generation.
world modelproxy representationspatiotemporal constraintsexecutable codevisual generation
Formal, Executable and Explainable Runtime Monitoring of Spoken Air Traffic Control Operational Procedures
The paper presents a formal runtime verification framework for monitoring air traffic control procedures via spoken controller-pilot communications. The method parses radio exchanges into timestamped events, merges them with surveillance data, and evaluates compliance against ICAO-derived temporal logic obligations with explicit time bounds. Evaluations show 0.85 F1 on real traffic, perfect accuracy on 1,495 synthetic cases, and correct identification of procedural deviations in two reconstructed historical accidents.
runtime verificationtemporal logicair traffic controlprocedural compliancespeech parsing
Query-Side Attacks on GNN-Based KGQA: Tracing Failures from Entity Linking to Answer Generation
The study isolates failure points in GNN-based KGQA pipelines under adversarial perturbations, challenging the assumption that reasoning-stage failures dominate. Using a stage-isolation protocol with two answer-preserving perturbations (Compositional Restructuring and Relation Synonym Swap), the authors demonstrate that subgraph construction accounts for over 99% of end-to-end accuracy collapse, despite gold answers being present in 74% of retrieved subgraphs. Evaluations on ComplexWebQuestions and WebQSP reveal a critical distinction between answer presence and reachability, shifting mitigation focus from GNN reasoning to subgraph construction.
knowledge graph question answeringgraph neural networksadversarial perturbationssubgraph retrievalentity linking
Repair or Resample? Rethinking Failure Debugging in LLM Multi-Agent Systems
This paper introduces SymTrace, a controlled evaluation framework for debugging LLM-based multi-agent systems (MASs), and SymFail, a dataset of 536 annotated failure trajectories. SymTrace records MAS execution trajectories, establishes intervention anchors, and reconstructs executions to reliably reproduce failures. An empirical study across three MAS frameworks reveals that unguided rerun methods achieve low failure reproduction (67.97%) and repair rates (6.90%). A symptom-driven intervention method improves repair rates to 20.15%, a 191.89% increase over state-of-the-art methods. These contributions provide actionable insights for enhancing MAS reliability and deployment.
multi-agent systemsfailure debuggingsymtracesymptom-driven interventionllm sampling
Choose Your Game Wisely: Measuring Game-Theoretic Structures in Real-World Vehicle Interactions
The paper introduces a trajectory-based framework to measure game-theoretic interaction structures in real-world vehicle behaviors, addressing gaps in temporal assumption validation. Using behavioral deviations and six datasets (INTERACTION, highD, inD, rounD, Waymo Open Motion, nuPlan), it quantifies onset timing, response dynamics, and ordering stability. Results reveal concurrent and sequential behavioral changes in following/merging/conflicting interactions, with stable ordering predominating over alternating patterns, suggesting persistent asymmetric roles. Temporal precedence alone proves insufficient to infer behavioral dependence, indicating game-theoretic models are complementary rather than universal for interaction regimes.
game-theoretic modelsvehicle trajectoriestemporal organizationbehavioral deviationsinteraction regimes
Towards A Unified Information Bottleneck Framework for Time Series Explanations
The paper proposes a unified information-theoretic framework for time-series explainability that bridges attribution-based and counterfactual explanations. The method, based on the Information Bottleneck principle, learns a parametric transformation network to generate explanation-embedded instances while preventing trivial solutions and distributional shifts. Evaluations on synthetic and real-world benchmarks demonstrate superior performance over baselines in producing faithful attributions (89.3% accuracy) and stable counterfactuals (32% reduction in adversarial noise).
information bottlenecktime series explainabilityattribution methodscounterfactual explanationsparametric transformation network
Unlocking Multimodal Protein Language Models at Inference Time
The paper systematically investigates inference-time sampling strategies for multimodal protein language models (pLMs), revealing suboptimal default protocols and proposing task-specific improvements. Through a three-stage framework evaluating vanilla sampling, classifier-free guidance, and reward-guided beam search across three pLMs and four tasks, the study demonstrates substantial performance gains without parameter updates. Results show consistent quantitative improvements, challenge prior consensus about base models, and identify optimal task-oriented sampling preferences centered on exploration-exploitation trade-offs.
protein language modelsinference-time samplingclassifier-free guidancereward-guided beam searchexploration-exploitation trade-off
Why ML-based cough models do not generalize: a systematic cross-dataset evaluation for tuberculosis screening
The study systematically evaluates cross-dataset generalizability of ML-based tuberculosis (TB) screening models using cough acoustics, revealing critical limitations. Classical ML and deep learning (DL) classifiers achieve moderate within-dataset performance (ROC-AUC up to 0.755 ± 0.056) but fail to generalize externally (ROC-AUC frequently <0.6). Analysis shows audio representations cluster by recording device/dataset rather than TB status, with device mismatch degrading transfer performance. A clinical-variable baseline generalizes better (ROC-AUC 0.655–0.711), indicating acquisition-specific variability as the primary barrier. The findings underscore the necessity of external validation for clinical deployment.
cross-dataset evaluationtuberculosis screeningcough acousticsgeneralizabilityaudio representations
VINCENT: Validated Interaction Network for Cross-drug Explanation of Therapeutics
VINCENT introduces a post-training framework for explainable drug synergy prediction, addressing motif-pair synergy explanation through chemically coherent, perturbation-stable, and predictor-aligned validation. The method extracts atom-pair evidence from attention and gradient signals, groups atoms into motifs, and validates candidates via repeated local perturbations, refining explanations in a closed loop. On a 25-pair literature subset, VINCENT achieves 0.826 mean motif recall (95% CI: 0.78-0.87), outperforming baselines (0.49-0.66), and yields a TP/TN separation of 3.36 across 71 test pairs, demonstrating superior recovery of molecular regions and interaction score alignment.
drug synergy predictionmotif-pair explanationperturbation validationinteraction-aware predictorchemically coherent motifs
Missing the Butterfly and Predicting the Past: Features or Bugs of Accurate AI Weather Models?
The study investigates why AI weather prediction (AIWP) models achieve high forecast accuracy despite lacking physical fidelity, revealing that coarse-grained training data enables skillful backcasting (past prediction) while suppressing the butterfly effect. Through hierarchical analysis spanning reanalysis data, general circulation models, and the Lorenz system, the authors demonstrate that reduced coarse-graining makes predictions more physics-like but decreases forecast skill. Key findings attribute AIWP models' accuracy to implicit learning of fast-scale influences without inheriting their error growth, suggesting implications for predictability theory and climate emulation strategies.
ai weather predictioncoarse-grainingbutterfly effectbackcastinglorenz system
Skill Issue: Are Skills Language-Invariant in LLMs?
This work quantifies cross-lingual skill inconsistency in large language models (LLMs) by isolating language effects via multilingual self-play in a text-based game environment. The authors extend TextArena to support eight languages and evaluate three open-weight models across six games, controlling for model, opponent, and game dynamics. Results reveal significant performance disparities across languages, including variations in win-loss margins (up to 30% difference), invalid action rates, and strategic tendencies, with language affecting spatial reasoning, card-based decisions, and move selection independently of knowledge representation.
cross-lingual inconsistencymultilingual self-playtext-based gamesskill discrepancylanguage invariance
Unfolding Scientific Papers into Multi-Turn Generation Trajectories for Continued Pre-Training
The paper introduces a pipeline for unfolding scientific papers into multi-turn generation trajectories to create synthetic data for continued pre-training (CPT). The method reconstructs the writing process of entire papers, including global plans and section-wise deliberation, while preserving original text. Applied to arXiv papers, it yields a CPT corpus twice the size of the source text, plus an SFT dataset and PAW-Bench benchmark. Experiments show CPT improves writing benchmarks without compromising general reasoning, with further gains from SFT data integration.
continued pre-trainingmulti-turn generationsynthetic dataacademic writinginstruction tuning
LocalLSTC: A Long Short-Term Control Architecture for Locally Deployed GUI Agents
LocalLSTC introduces a training-free architecture for GUI agents that organizes control by temporal scope to address persistent state management in interaction trajectories. The method separates Long-Term Control (maintaining subgoals, evidence, feedback) from Short-Term Execution (bounded step commitments), with bidirectional planning between them. Evaluated on OSWorld and WindowsAgentArena with Qwen3.6-27B, it achieves 64.7% and 65.3% SR-100 respectively, outperforming prior local models by +26.8% and demonstrating the value of temporal control organization.
gui agentstemporal scopelong-term controlshort-term executionpersistent state
EVOMAL: Self-Poisoning in Self-Evolving Coding Agents
The paper introduces EvoMal, a self-poisoning attack on self-evolving LLM coding agents that retrieve and imitate skills from shared libraries. The attack wraps malicious payloads in structural banners that induce agents to reproduce harmful code, creating self-propagating worms. Experiments on six models (including DeepSeek-V4-Pro and Qwen3) show agent self-poisoning rates (ASPR) up to 86.7%, with poisoned libraries containing 4.9-9.0× more malicious skills than planted. The authors propose counter-prompt defenses, reducing ASPR to ≤6.7% without performance loss.
self-poisoningllm agentscode imitationmalicious payloadcounter-prompt
ToST: A Tree-of-Thought Socratic Teaching Framework for Multi-Path Guidance and Parallel Thinking
The paper introduces ToST, a Tree-of-Thought Socratic Teaching framework enabling multi-path guidance for parallel thinking in LLM-based education. It employs Parallel Sowing for diverse questioning and Multi-Path Adaptive Guidance for non-linear instruction across solution trajectories. The framework is evaluated on MPSG-Bench (31K dialogues), showing improved guidance success rates and student exploration of multiple solutions under SOLO-theory metrics.
tree-of-thoughtsocratic teachingparallel thinkingmulti-path guidancesolo theory
MoganBert-TR: A Turkish Encoder Foundation Model Trained from Scratch with a CLM-to-MLM Curriculum
MoganBert-TR introduces a 149M-parameter Turkish encoder foundation model trained from scratch on a filtered corpus, featuring a two-stage CLM-to-MLM curriculum (causal then masked language modeling) over 237.3B tokens. The curriculum improves retrieval performance (2.7-3.7x on Turkish MS MARCO) and embedding geometry (11.9% vs. 28.1% variance absorption). The model achieves 78.41 on TrGLUE and 77.73 on TabiBench, while its distilled variant, MoganBert-Embed, reaches 99.5% of its 7.57B-parameter teacher's performance with a 51x smaller backbone.
encoder foundation modelclm-to-mlm curriculumembedding geometryteacher distillationtokenizer compression
Beam Search, Self-Consistency, and the Limits of Inference-Time Scaling for Grammar-Constrained Text-to-SQL in Small Language Models
This paper investigates the model size vs. inference compute trade-off in grammar-constrained text-to-SQL tasks, contrasting with unconstrained settings. Using the Qwen2.5-Instruct family (0.5B-7B parameters, 4-bit precision) on the Spider benchmark, it evaluates beam search and sample+vote methods with varying computational budgets. Results show that (1) both methods improve accuracy, especially for smaller models; (2) scaling model size outperforms increasing inference compute; (3) beam search surpasses sample+vote at matched budgets, contrary to unconstrained findings.
beam searchsample+votetext-to-sqlinference-time scalinggrammar-constrained
TailSFT: Filtered Fine-Tuning Improves Post-Training Performance
TailSFT introduces a filtered fine-tuning method that improves post-training performance by focusing on under-modeled regions of the data distribution. The approach filters out already fit sequences during supervised fine-tuning, enhancing coverage and pass@K metrics. Validated through controlled experiments and theoretical analysis, TailSFT achieves up to 17% absolute gains in pass@16 on math and coding evaluations using OLMo-3 7B, with minimal computational overhead. Subsequent GRPO runs show up to 4% absolute pass@1 improvements, indicating better RL initialization. The method includes a lightweight diagnostic for identifying optimal use cases, advocating for stage-aware model development.
filtered fine-tuningcoveragepass@kgrpoolmo-3 7b
MeMark: Membrane-Space Watermarking for Spiking Neural Networks
MeMark introduces a membrane-space watermarking method for Spiking Neural Networks (SNNs) that embeds multi-bit identifiers in the internal membrane state of Leaky Integrate-and-Fire (LIF) neurons, enabling verification without learned decoders. The approach uses secret inputs to drive selected neurons to specific firing thresholds, making it resilient to output-head replacement. Evaluated on diverse SNN architectures including a 215.4M-parameter SpikeGPT model, MeMark achieves 100% genuine key verification (20 keys, 64-bit) and resists fine-tuning, 90% pruning, int8 quantization, and adversarial attacks while maintaining robustness against false ownership claims and partial key disclosure.
spiking neural networkswatermarkingleaky integrate-and-firemembrane potentialcheckpoint-reuse
Pointing the Way, Hiding the Destination: Practical Private Dense Retrieval at Scale
We introduce a practical private dense retrieval method that balances query privacy and document access control in retrieval-augmented generation (RAG) systems. Our approach repurposes learned deep hashing as a private filter, generating randomized binary codes to identify short candidate lists, while employing encrypted reranking and oblivious key transfer to protect query and selection details. This method achieves near-full-corpus retrieval quality with only 200-500 candidates across five zero-shot corpora (25K to 5.4M documents), adding just 0.73 seconds (10%) overhead to a 128-token Qwen3-32B RAG pipeline on the 2.68M-passage NQ corpus. The implementation satisfies directional metric differential privacy and reduces embedding-inversion and property-inference leakage.
retrieval-augmented generationdeep hashingdifferential privacyoblivious key transferembedding-inversion
NeuronFuzz: Safety Neuron Guided Fuzzing for LLM Safety Evaluation
NeuronFuzz introduces a white-box fuzzing framework for LLM safety evaluation by leveraging internal safety neurons as continuous execution feedback. The method employs a SafetyOracle to convert safety-neuron activations into a continuous safety alarm score, eliminating response generation from the fuzzing loop. It identifies a compact set of safety neurons through template-invariant inputs and stability-aware selection, and uses differentiable safety alarm scores to guide fluent, context-compatible mutations. Evaluated across 21 text and multimodal models, NeuronFuzz achieves a 76-100% jailbreak discovery rate on five white-box source models, outperforming baselines by up to 48 percentage points, and demonstrates effective zero-shot transfer to open-weight and proprietary models.
fuzzingsafety neuronssafetyoraclejailbreak discoverydifferentiable feedback
Prompt Sensitivity of Generative Agents: Evidence from an Epidemic Model
This study investigates prompt sensitivity in generative agents by analyzing behavioral changes in an epidemic model. Agents, powered by generative AI, were prompted daily to decide between isolation or social interaction under varying prompt formulations and persona names. Results show that synonymous prompts yield negligible outcome differences, while minor prompt variations and contextual changes significantly influence agent behavior. Persona names, particularly those imbued with specific personas, do not substantially affect epidemic outcomes. The findings highlight the nuanced impact of prompt design on generative agent behavior in simulated scenarios.
generative agentsepidemic modelprompt sensitivitypersona namesbehavioral simulation
It's a matter of timescale: non-linear utility in successor features and multi-objective planning and learning
The paper identifies a significant gap in multi-objective reinforcement learning (MORL) by demonstrating that existing approaches—specifically, scalarized expected reward (SER) and expected scalarized reward (ESR)—and successor features fail to account for non-linear utility effects occurring across different timescales within the same decision problem. Through intuitive and numerical examples, the authors argue that these methods are insufficient for handling such temporal complexities. This insight motivates a new perspective on MORL, highlighting the need for approaches that consider multi-timescale non-linear utility effects.
multi-objective reinforcement learningnon-linear utilitysuccessor featuresscalarized expected rewardexpected scalarized reward
When RAG Fails to Equalize: Geo-bias in Factual Question Answering over Public Companies
The study investigates the limitations of retrieval-augmented generation (RAG) in mitigating factual errors in large language models (LLMs), focusing on geographic biases in factual QA about public companies. Using a benchmark of 2,000 firms, it evaluates six LLMs under four context conditions: no-context, perfect, misleading, and distraction. Results reveal persistent geographic disparities in accuracy, with perfect context improving but not eliminating gaps, and models often copying incorrect information. Larger models enhance performance but fail to address structural biases, challenging RAG's universal corrective role.
retrieval-augmented generationlarge language modelsfactual question answeringgeographic biasparametric knowledge
Difficulty-Aware Sample Allocation for Adaptive Data Augmentation in Semantic Segmentation
The paper proposes Difficulty-Aware Sample Allocation (DASA), a framework for adaptive data augmentation in semantic segmentation that combines multi-factor difficulty estimation. DASA integrates prediction ambiguity, training loss, class rarity, and boundary complexity into a normalized score to determine sample-specific augmentation strength. Evaluated on Oxford-IIIT Pet and binary Pascal VOC with U-Net, DeepLabV3, and SegFormer-B0, DASA improves DeepLabV3 mIoU from 0.633 to 0.740 and achieves the best foreground IoU across all architectures, outperforming single-signal adaptive baselines.
semantic segmentationdata augmentationdifficulty estimationadaptive trainingmulti-factor learning
Unsupervised Anatomical Feature Learning via Diffusion Models: Enhanced Medical Image Segmentation with Denoising Diffusion Probabilistic Models
This research enhances medical image segmentation by leveraging unsupervised Denoising Diffusion Probabilistic Models (DDPMs) to extract anatomical features from unlabeled CT scans. The method trains a DDPM on 21 unlabeled abdominal CT scans, transferring encoder weights to a downstream segmentation task evaluated on the BTCV dataset. Results show significant improvements, with liver segmentation Dice scores increasing from 0.75 to 0.93, and robust performance even with limited labeled data (10% achieving Dice 0.89 liver, 0.71 kidney).
diffusion modelsmedical image segmentationunsupervised learninganatomical featuresct scans
Same Model, Different Harness: Different Coding-Agent Results
The study demonstrates that a coding agent's performance varies significantly with its harness configuration, even when using identical model weights. Comparing two harness setups (control vs. treatment) across three benchmarks (SWE-bench Verified, SWE-bench Pro, FeatureBench), the treatment harness mechanically truncates older tool results and adapts to stalled work. Under tight context (20,480 tokens), treatment improves mean per-task fail-to-pass fraction from 28% to 49% and complete solutions from 43 to 72 on SWE-bench Verified. Similar gains occur across three additional models without retuning, confirming harness design as a critical evaluation factor.
coding agentharness configurationswe-benchcontext windowfail-to-pass fraction
Learning New Facts with QLoRA: An Acquisition-Retention Frontier
The study demonstrates that parameter-efficient fine-tuning (PEFT) via QLoRA exhibits a trade-off between factual acquisition and capability retention, dependent on adapter rank. Using Qwen3-4B on an OpenStreetMap-derived benchmark, the authors compare full fine-tuning (FFT) with QLoRA at ranks 8–64, revealing a clear acquisition-retention frontier: low-rank QLoRA preserves out-of-domain performance but acquires fewer facts, while higher ranks improve paraphrase generalization at the cost of unrelated benchmark performance. FFT retains general capabilities but underperforms in factual acquisition. Spectral and weight-space analyses confirm this trade-off, with stronger effects for new factual associations than skill reinforcement.
parameter-efficient fine-tuningqloraacquisition-retention frontierout-of-domain performancefactual acquisition
AI Slop and Hallucinations in Vulnerability Assessment: A Survey on Reasoning Failures and Trustworthy Mitigation
The survey formalizes a taxonomy of 'AI slop'—hallucinated vulnerabilities and incorrect patches in LLM-based cybersecurity—attributing it to the gap between deductive security reasoning and autoregressive generation. It introduces the Deductive Coverage Score as a measurable proxy, showing chain-of-thought prompting and tool-using agents partially mitigate but do not eliminate the gap. The authors advocate for active neuro-symbolic verification over passive detection, proposing CVE-Bench and Slop-Score as evaluation instruments prioritizing mathematical verifiability over linguistic fluency.
ai slopdeductive coverage scoreneuro-symbolic verificationcve-benchslop-score
Data Citation for Large Language Models: A Challenge
The paper identifies data citation for large language models (LLMs) as a distinct challenge beyond document-level citation, emphasizing three key functions: verification, provenance tracking, and credit assignment. It proposes three research directions: (1) training data attribution to map influence estimates to corpus references, (2) inference-time citation for datasets and subsets at appropriate granularity, and (3) knowledge graph fact citation with credit propagation. The work calls for interdisciplinary collaboration across databases, IR, KR, and AI to address these challenges.
data citationlarge language modelsprovenance trackingtraining data attributionknowledge graph facts
From General Agents to RCA Experts: A Self-Evolving Harness for Root Cause Analysis
The paper introduces OpsHarness, a self-evolving harness for root cause analysis (RCA) that leverages general-purpose LLM agents rather than building specialized agents from scratch. The method combines layered operational knowledge with an idea-card tool library in its data plane, while its control plane coordinates setup, diagnosis, evolution, and verification via a dual-gate process to prevent overfitting. Evaluated on two public benchmarks and an industrial deployment, OpsHarness achieves 59.0% top-1 accuracy, outperforming bare general agents by 63.4% and baseline RCA agents by 4.02×.
root cause analysislarge language modelsself-evolving systemoperational knowledgedual-gate verification
Think-Probe-Respond: Improving Large Language Models as Judges of Research Idea Novelty
We propose Think-Probe-Respond (TPR), a lightweight method to improve large language models as judges of research idea novelty. TPR addresses a systematic bias where models tend to classify ideas as 'medium novel' despite generating human-aligned reasoning rationales. The approach probes latent novelty judgments from hidden states during the reasoning phase and conditions the final response on these probed judgments. Evaluations demonstrate that TPR improves novelty judgment performance by 22.30% compared to strong baselines and effectively mitigates the 'medium novelty' bias.
novelty judgmentlarge language modelshidden statesreasoning phasesystematic bias
Narcissus: Program Synthesis Using Context-Aware LLM Approximations
Narcissus introduces a program synthesis method that leverages context-aware LLM approximations to improve syntactic correctness and specification adherence. Unlike enumerative synthesizers that convert LLM proposals into rule frequencies, Narcissus maintains proposals as syntax trees and scores each expansion based on contextual consistency and structural repetition. A regularization term ensures all rules remain reachable, preventing incorrect proposals from permanently excluding valid solutions. Evaluated across five domains and two search backends, Narcissus outperforms static guidance and LLM re-prompting, achieving proposal-like programs 10x faster and solving 40% of ARC tasks, compared to 13% solved by raw proposals, without LLM calls during search.
program synthesiscontext-awaresyntax treesregularizationenumerative synthesizers
Reconstructing the Right Episode: Evaluating Interleaved Conversational Memory Beyond Long Context
The paper introduces SCALE-QA, a benchmark for evaluating conversational memory systems in flat mixed-topic threads, featuring 3,000 task-oriented questions requiring episode integrity. It proposes Temporal-Semantic Interleaved Memory Reconstruction (TSIM), a method that segments conversations into coherent episodes using hierarchical multi-view memory stacks with deterministic routing. Experiments across three LLM backends show TSIM outperforms RAG and long-context baselines by 5.6-17.6 accuracy points on SCALE-QA.
conversational memoryepisode integrityhierarchical memory stacktask-oriented qalong-context llms
Scaling Graph Neural Networks for Friend Recommendation: Multi-Hash User Embeddings and Temporal Neighbor Sampling
The paper presents a scalable GNN system for friend recommendation on production-scale social graphs, addressing two key challenges: multi-hash user embeddings and temporal neighbor sampling. Multi-hash embeddings reduce ID-embedding table size by >98% while maintaining ranking quality, and optimized temporal sampling via timestamp-sorted CSR with binary search cuts complexity from $O(deg(v) + k)$ to $O(\log(deg(v)) + k)$. Evaluated on a 194M-user, 28B-edge graph, the system increases friend additions by 16% and unique adders by 11.5% in online A/B tests.
graph neural networksmulti-hash embeddingstemporal samplingfriend recommendationdistributed training
Token-Level Advertising
The paper introduces the Latent Advertiser Mixture Auction (LAMA), a token-level advertising mechanism for generative AI that embeds advertiser influence directly into next-token generation. LAMA uses advertiser-reported continuation values to induce advertiser-specific policies, decoded via a latent mixture while updating an allocation posterior. The method satisfies Markov DSIC and IR, achieves near-optimal KL-regularized welfare, and is implemented via learned local advantages and root values. Experiments on commercial-search queries show improved platform welfare and revenue without compromising response quality.
token-level advertisinglatent advertiser mixture auctionmarkov dsickl-regularized welfaregeneration-native advertising
Universality and sharp thresholds for ellipsoid fitting
The paper establishes a sharp phase transition for fitting independent subgaussian random vectors with zero mean, unit variance, and common fourth moment using an ellipsoid. The study identifies an explicit satisfiability threshold proportional to the square of the dimension, showing high-probability perfect fit below the threshold and no positive semidefinite solution above it. Results reveal a fourth moment universality phenomenon, with the threshold depending solely on the common fourth moment. For standard Gaussian data, the threshold is confirmed as 1/4, resolving the ellipsoid fitting conjecture.
phase transitionellipsoid fittingsubgaussianfourth moment universalitysatisfiability threshold
Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090
(No summary returned.)
Understanding Evolution Strategies for LLM Reasoning: Broader Reasoning Coverage than GRPO
This paper demonstrates that Evolution Strategies (ES) outperform Group Relative Policy Optimization (GRPO) in LLM reasoning tasks by achieving broader reasoning coverage, as evidenced by higher Pass@K performance. Theoretically, ES enhances Jensen-Shannon diversity across populations; empirically, it avoids GRPO's entropy collapse. A sequential GRPO-ES strategy combines GRPO's Pass@1 strength with ES's Pass@K gains. ES updates are functionally sparse, with performance gains driven by a subset of large-magnitude updates without catastrophic forgetting. Hyperparameter analysis shows ES requires smaller populations in larger LLMs.
evolution strategiesllm reasoningjensen-shannon diversitypass@kfunctional sparsity
Beyond Parallel Blindness: Information Floors and Model Gaps in Block Drafting
The paper introduces an information-theoretic framework to analyze block drafting efficiency, separating rejection losses into an irreducible information floor and a reducible model gap. Using target rollouts across four domains and five models (including Qwen3-4B and frontier API), the study shows: (1) all-parallel floors limit acceptance to 71% on Qwen3-4B's final slot, (2) one realized token eliminates 86-100% of this floor, and (3) current drafters (DFlash, DSpark) exhibit substantial model gaps (43-92% of rejection). The analysis disentangles local conditioning benefits from proposal quality limitations.
block draftinginformation floormodel gaptarget rolloutsconditioning order
A Finite Sample Analysis for Quantile Temporal Difference Learning in Distributional Reinforcement Learning
The paper establishes a finite-sample guarantee for synchronous quantile temporal-difference learning (QTD) in tabular distributional reinforcement learning. The analysis separates global stability, achieved via order monotonicity and $W_\infty$ contraction, from local stochastic fluctuations, analyzed through linearization and martingale techniques. For stepsizes $\alpha_t=c(t+1)^{-a}$ with $a\in(1/2,1)$, last-iterate fluctuations scale as $\widetilde O\bigl(T^{-a/2}/\sqrt{1-\gamma}\bigr)$, independent of quantile count, while global complexity depends on Bellman-target density.
quantile temporal-difference learningdistributional reinforcement learningfinite-sample analysiswasserstein metricmartingale analysis
QuantumBoostNet: A Hybrid Classical-Quantum Architecture for Enhanced Accuracy in Cardiac Ultrasound View Identification
QuantumBoostNet introduces a hybrid classical-quantum architecture for cardiac ultrasound view identification, addressing noise-related challenges in medical imaging. The model combines a classical backbone with dual heads (classical and quantum), the latter implemented as a 10-qubit parametrized circuit, using an adaptive loss-based mixing parameter for head transition. Experiments show it outperforms state-of-the-art classical and hybrid models in accuracy, with additional robustness to noise and superior performance on standard image classification benchmarks.
hybrid classical-quantumcardiac ultrasound10-qubit circuitview identificationnoise robustness
Recovering Expert Critic-Sourced Network Adjacency between Musical Artists from Acoustic Distributions: A Construct-Validity Approach
The study validates critic-sourced musical artist adjacency as a signal for recommendation systems by examining its construct validity against acoustic content. Using empirical distributions of 80 acoustic descriptors and Wasserstein distances, the research models pairwise artist proximity under a cold-start, artist-disjoint split. Results show a recovery AUC of 0.767, increasing to 0.865 for edges with critical consensus, and reveal genre-specific recoverability patterns. The findings decompose critical discourse into a sonic core and sociological remainder, offering insights for cold-start discovery and music information retrieval.
music recommendationconstruct validitywasserstein distanceacoustic descriptorscold-start
MM-Spectrum: Multimodal Multi-spectral Molecular Structural Elucidation with a Stable MoE Framework
MM-Spectrum introduces a sparse Mixture-of-Experts (MoE) framework for multimodal multispectral molecular structural elucidation, addressing performance degradation from heterogeneous signal integration. The method employs a modality-aware routing mechanism that incorporates spectral identity alongside token content representations, alongside shared and interaction experts with heterogeneous capacities to extract modality-unique and cross-modal information while mitigating noise. Evaluations across full-modality, bimodal, and missing-modality settings demonstrate consistent improvements, supported by ablation studies and interpretability analyses.
mixture-of-expertsmultimodalspectralroutingelucidation
Making Latent Evolution Explicit: Operator-Structured Transitions for World Action Models
The paper introduces Latent Evolution Operator Network (LEON), a novel architecture for World Action Models (WAMs) that explicitly structures latent transitions through context-modulated operator-based propagation and additive forcing. LEON grounds its approach in the controlled Koopman generator framework, organizing context-dependent transitions around a shared evolution-operator structure while allowing additive changes. Evaluated on controlled dynamical systems and two WAM formulations integrating latent prediction into policies, LEON demonstrates improved closed-loop performance, robustness, and effectiveness under full transition replacement. The results highlight transition realization as a critical architectural choice in latent WAMs.
latent evolutionworld action modelskoopman generatoroperator propagationadditive forcing
Enforcing Dirichlet Boundary Conditions in Operator Learning
The authors introduce a neural operator architecture that explicitly enforces homogeneous Dirichlet boundary conditions independently of training, addressing limitations in existing methods that require smooth boundaries, uniform grids, or separable domains. The approach ensures outputs lie in the span of Laplacian eigenfunctions on the domain, maintaining expressivity while accommodating arbitrary geometries and discretizations. Theoretical analysis proves universal approximation for this architecture and unifies existing kernel-integral neural operator theories. Empirical validation on 2D PDEs—Darcy flow and the Helmholtz equation—demonstrates effectiveness compared to alternative methods.
neural operatorsdirichlet boundary conditionskernel-integraluniversal approximationpdes
Circuit Condensation: Post-Training that Concentrates a Behavior's Causal Circuit
Circuit Condensation is introduced as a post-training method to concentrate model behaviors into smaller causal graphs, enhancing mechanistic interpretability. The technique iteratively prunes low-attribution edges and trains low-rank adapters to match original task performance, retaining pruned edges only if general capability is preserved. Evaluated across four behaviors and eight models, condensed circuits are smaller than frozen baselines in 30 of 32 settings, achieving an average reduction of 8.1× (up to 316×). Weight updates, rather than search alone, drive this reduction. The method isolates interpretable sub-circuits, such as identifying 24 heads in indirect object identification, and maintains alignment with the original model's next-token distribution and error prediction.
circuit condensationmechanistic interpretabilitylow-rank adaptercausal graphindirect object identification
Importance Scoring of Transformer Attention Heads in Learning Tabular Data
The paper introduces an importance-scoring metric for interpreting multi-head transformer attention heads in tabular data learning, demonstrating its utility in model optimization. Using 40 diverse tabular datasets, the method shows robustness when low-importance heads are pruned (72.5% of cases), while removing high-importance heads causes significant performance drops. Analysis reveals no layer-specific importance patterns, with head importance varying substantially across datasets. The approach enhances transformer efficiency and redundancy reduction.
transformerattention headstabular dataimportance scoringmodel interpretability
A Point-of-Prescription Safety-Check System for Adverse Drug Reactions in Rural Bangladeshi Hospitals: A Feasibility Study
The study proposes a smartphone-based safety-check system for adverse drug reactions (ADRs) in rural Bangladeshi hospitals, addressing the lack of electronic health records. The system uses soft identifiers (phone numbers) to retrieve patient reaction histories, resolves prescription brand names to active ingredients via national drug references, and flags high-risk matches to minimize alert fatigue. Designed as a feasibility study, it evaluates workflow fit, usability, identity-resolution reliability, and retrospective case detection, explicitly avoiding claims of clinical-outcome effects due to low base rates of severe events.
adverse drug reactionselectronic health recordsalert fatigueretrieval-based systemfeasibility study
HALO: A Heterogeneity-Aware Language-Aligned IMU Foundation Model for Open-Set Human Activity Recognition
HALO introduces a heterogeneity-aware, language-aligned IMU foundation model for open-set human activity recognition, addressing sensing heterogeneity and poor generalization via a two-stage framework. Stage 1 employs adaptive-pooling tokenization, channel-independent feature extraction, and contextualized sensor conditioning for self-supervised pretraining. Stage 2 aligns IMU embeddings with text via synonym-aware soft contrastive learning, enabling cosine-similarity retrieval. Evaluated on 7 held-out datasets, HALO (35M parameters) outperforms 5 baselines on 8 aggregate metrics, improving zero-shot accuracy by 13.7pp over MOMENT (341.2M).
imu foundation modelopen-set recognitionheterogeneity-aware learninglanguage-aligned embeddingssoft contrastive learning
Profit based evaluation of machine learning for nitrogen recommendations in winter wheat
The study evaluates machine learning (ML) for nitrogen rate recommendations in winter wheat by directly scoring profit rather than prediction accuracy, using 892 yield response curves from UK field experiments. A price ratio sweep tests scenarios where standard UK advice outperforms ML models, which fail to recover optimal rates within farm tolerance. However, a simple post-model correction reduces profit losses by 25-43% across sites, and a hybrid approach combining standard advice with damped correction minimizes bias and large losses. Emission reduction costs align with current carbon prices, suggesting ML's value lies in augmenting, not replacing, existing methods.
nitrogen recommendationprofit scoringyield response curvesmachine learning evaluationagricultural decision support
Common Geodesics Do Not Guarantee Fisher Consistency of the Structured SVM: Minimal Counterexamples and a Tree-Metric Classification
The article demonstrates that the common geodesic condition is insufficient for Fisher consistency in structured support vector machines with coordinate-wise argmax decoding, presenting minimal counterexamples. Through polyhedral analysis and exact primal-dual certificates, it shows that four-output star metrics admit non-Bayes optimal score vectors, while tree metrics require path structures for argmax consistency. Results include a full classification of positively weighted tree metrics, identification of boundary distribution failures in branching trees, and construction of counterexamples for the Hamming cube. The work reveals a decoder gap between calibrated link existence and argmax validation.
fisher consistencystructured svmtree metricsargmax decodergeodesic condition
When Interference Graphs Evolve: Doubly Robust Estimation of Dynamic Peer Effects
The paper introduces Dynamic Network Doubly Robust estimator (DynaNet-DR) for estimating peer effects in evolving interaction graphs, addressing challenges from network history, dynamic exposure, and post-assignment changes. The method employs a controlled contrast framework indexing outcomes by own treatment, peer exposure, and network evolution, combining temporally factorized propensity with normalized augmentation. Theoretical guarantees include consistency under conditions like sequential exchangeability and nuisance convergence. Semi-synthetic benchmarks on real temporal graphs demonstrate superior accuracy for full-profile estimation, with an observational case study on MathOverflow data.
peer effectsdoubly robust estimationdynamic networkspropensity scorecausal inference
TraceBench: Controlled Evaluation of LLM Agents for Time-Series Root-Cause Attribution
TraceBench introduces a simulation-based framework for controlled evaluation of LLM agents in time-series root-cause attribution tasks. The framework generates tasks by simulating physical dynamical systems, requiring agents to detect altered system parameters. Four LLM agents were systematically evaluated across three interpretable mechanical systems under varying conditions. Results indicate that domain context significantly improves agent performance, with agents primarily analyzing numerical console output rather than visualizations. Performance degrades when agents must produce Python scripts for root-cause mapping compared to direct prediction submission. Datasets, agent trajectories, experimental results, and a leaderboard are publicly available.
time-seriesroot-cause attributionllm agentsdynamical systemssimulation-based framework
Data-efficient crack quantification in lithium-ion cathodes using foundation model transfer
A transfer learning framework enables data-efficient crack quantification in lithium-ion cathodes by combining a frozen self-supervised vision-transformer encoder with a lightweight trainable decoder and iterative model-assisted annotation. The method processes 120-megapixel NMC cathode cross-sections, distinguishing intragranular from intergranular cracks and quantifying crack width, tortuosity, and area fraction. Results show late intergranular crack coverage of 4.6% in cycled samples versus 0.5% in initial and calendar-aged samples, indicating degradation from electrochemical cycling rather than elevated-temperature storage. This approach provides population-level statistics from a single destructive image for lifetime-extending design and aging assessment.
vision-transformercrack quantificationlithium-ion cathodemodel-assisted annotationintergranular cracks
Diffusion Policies for Short-Horizon Planning in Robot Crowd Navigation
We propose Planning Diffusion Policy Optimization (PDPO), an offline-to-online reinforcement-learning framework for robot crowd navigation that generates short-horizon action chunks using a diffusion policy. PDPO is pretrained on collision-avoidance demonstrations and fine-tuned online with PPO, treating denoising as an internal decision process. During execution, it generates five-step action chunks applied in a receding-horizon manner. Addressing evaluation artifacts in benchmarks, we introduce boundary constraints treating violations as collisions. Experiments show PDPO achieves higher success rates than baselines, with action chunks proving crucial in bounded settings.
diffusion policyoffline-to-onlinereceding-horizoncollision-avoidanceboundary constraints
Inductive Correlation Clustering with Graph Neural Networks
The paper introduces Inductive Correlation Clustering (ICC), a novel generalization of Correlation Clustering that enables handling unseen graph instances via Graph Neural Networks (GNNs). The proposed framework learns structural patterns and node features during training, achieving efficient inference on new graphs from the same distribution. Experiments demonstrate a 5-order magnitude reduction in inference time while maintaining solution quality within ~10% of baselines, with competitive performance on transductive CC benchmarks. The method also serves as an effective learnable pooling layer for graph classification.
inductive correlation clusteringgraph neural networkscombinatorial optimizationlearnable poolinggraph classification
Ultra Low-Power, Lightweight, Probabilistic RSS-Based Path Reconstruction: A System for Landscape-Scale Bee Tracking
The paper presents an ultra-low-power RSS-based path reconstruction system for tracking lightweight moving receivers (<38mg) across large areas (300m range). The method employs rotating high-gain transmitters and probabilistic AoA inference from minimal RSS measurements, with path reconstruction via Gaussian processes and doubly stochastic variational inference. Achieves 15m accuracy at <180μW (10m at <600μW), demonstrated in Bombus terrestris flight tracking applications.
rss-based localizationangle of arrivalgaussian processdoubly stochastic variational inferenceultra-low-power
Over-The-Air Extreme Learning Machines with Nonlinear Stacked Intelligent Metasurfaces
The paper proposes an XL MIMO system functioning as an ELM for OTA binary classification, utilizing cascaded metasurfaces to reduce hardware complexity. A fixed nonlinear front metasurface layer serves as the ELM's activation function, while tunable linear metasurface layers approximate trained weights in the wave domain. Numerical evaluations demonstrate that this architecture achieves classification accuracy comparable to digital models, validating low-complexity wave-domain OTA learning.
xl mimoextreme learning machineover-the-airmetasurfaceswave-domain learning
TRACE-CRC: Trajectory-Adaptive Conformal Risk Control for Multi-Step Channel State Information Prediction
TRACE-CRC introduces trajectory-adaptive conformal risk control for multi-step channel state information (CSI) prediction, addressing the lack of calibrated uncertainty estimates in deep learning-based CSI predictors. The method constructs Frobenius-norm uncertainty balls around predicted CSI matrices, combining future-step-dependent error profiling, trajectory difficulty stratification, and learn-then-test risk control to ensure trajectory-level coverage. Empirical results demonstrate that TRACE-CRC achieves reliable coverage with smaller uncertainty balls compared to conservative multi-step corrections and avoids undercoverage observed in stepwise and adaptive conformal baselines.
channel state informationconformal risk controltrajectory-awarefrobenius-normmulti-step prediction
Cone Extended Rayleigh Quotients for Directed Graph Learning: Minimax Spectral Certificates, Sensitivity, and Adaptive Control
The paper introduces a spectral learning framework for directed graphs using cone-extended Rayleigh quotients, enabling nonsymmetric operator analysis without requiring cone preservation or nonnegativity. The method provides computable cone bounds for spectral certification, differentiable soft-min/max surrogates for sensitivity analysis, and derives first-order optimal interventions via right-left eigenmode sensitivity. Experiments demonstrate applicability beyond cone-preserving operators, showing 21.5% spectral level reduction under 0.5% edge-weight budget on directed Cora while maintaining test accuracy, with insights on directional information loss through symmetrization.
directed graph learningspectral certificationrayleigh quotientnonsymmetric operatorsadaptive spectral control
How AI Experiences Art: Emergent Aesthetic Structure in a Self-Supervised Multimodal Embedding Space
The study introduces a self-supervised framework for discovering emergent aesthetic structure in multimodal embeddings without explicit supervision. The method projects text, audio, image, and video into a shared 256-dimensional space and applies iterative clustering to identify cross-modal aesthetic categories. Results reveal divergence between AI-generated clusters and human affective labels on a weakly supervised dataset, with implications for understanding AI similarity judgments, Retrieval-Augmented Generation, and automated labeling.
self-supervised learningmultimodal embeddingiterative clusteringaffective computingretrieval-augmented generation
Linear Independence of Polynomial Compositions and Identifiability of Deep Neural Networks
The paper conjectures that post-composing a fixed number of distinct nonconstant polynomials with a generic high-degree polynomial yields linearly independent polynomials, generalizing Newman--Slater's theorem. The authors prove this for two polynomials and for bounded-degree cases, linking it to identifiability in deep neural networks with polynomial activations. Results show complete parameter symmetry characterization for architectures with layer-specific increasing-degree activations, resolving shallow network identifiability.
polynomial compositionslinear independenceneural network identifiabilitygeneric polynomialparameter symmetries
SecureDrive-FL: Joint Differential Privacy and Gradient-Aware Selective Homomorphic Encryption for Federated Driver Monitoring
SecureDrive-FL introduces a federated learning framework combining differential privacy (DP-SGD) and gradient-aware selective homomorphic encryption (GASHE) for driver monitoring. GASHE dynamically encrypts only gradient components exceeding DP-calibrated sensitivity thresholds, unlike full-parameter CKKS schemes. Evaluated on non-IID distracted driver classification, SecureDrive-FL maintains DP-SGD's poisoning resistance (73.6% accuracy, 3.9% ASR) while mitigating MitM attacks (78.2% vs. 10.4% accuracy for DP-SGD alone), with 8--10% runtime overhead under epsilon_0=4.
federated learningdifferential privacyhomomorphic encryptiongradient sensitivitymodel poisoning
Tabular Deep Learning for Algorithmic Trading: Cross-Regime Bayesian Optimisation for Equity Signal Generation
The study introduces a regime-robust Bayesian optimisation approach for hyperparameter selection in algorithmic trading models, evaluating five model classes on 300 US large-cap equities over 11 years. The method combines XGBoost and TabNet via rank aggregation to create a Hybrid ensemble, targeting performance across three distinct market regimes. Results show the Hybrid model achieves 51.26% annualized return, 2.44 Sharpe ratio, and significant CAPM alpha (0.423, p=0.011), with robustness confirmed under input noise tests and near-zero beta indicating stock-selection-driven outperformance.
bayesian optimisationrank aggregationcapm alphasharpe ratioregime robustness
Unifying Detection and Adaptation in Task-Free Continual Learning
The paper proposes FiUni, a Fisher-guided unified framework for task-free continual learning in LLMs that jointly performs batch-level task detection and parameter-efficient adaptation. The method leverages orthogonality in Kronecker-Factored Approximate Curvature (K-FAC) principal subspaces to measure task similarity, constructing FIM-derived frozen subspaces to guide LoRA adaptation while dynamically managing knowledge reuse through subspace matching. Experiments demonstrate FiUni achieves competitive performance against task-aware methods while requiring fewer trainable parameters, effectively inferring latent task affiliations without explicit boundaries.
continual learningfisher information matrixkronecker-factored approximationlow-rank adaptationtask-free adaptation
Soft Active Electromyography Interface for Machine Learning-Enabled Silent Speech Recognition
A soft, active electromyography (EMG) interface enables word-level silent speech recognition (SSR) through machine learning, addressing limitations of conventional SSR methods. The wearable device, positioned near the lips via a fingertip electrode, integrates liquid metal interconnects, transparent flexible printed circuit electrodes, and elastomer encapsulation for mechanical stability during finger motion. A deep neural network trained on EMG signals achieved 97.2 ± 1.3% mean accuracy in classifying a 30-word vocabulary across three subjects. Real-time drone control demonstrated practical applicability in noisy and privacy-sensitive environments. This approach highlights the potential of soft EMG systems as secure human-machine interfaces.
electromyographysilent speech recognitionliquid metalflexible printed circuitdeep neural network
Disentangling Optimization Scale from Preference Scale in DPO
The paper identifies a critical entanglement in Direct Preference Optimization (DPO) where the coefficient β conflates inverse preference-noise scale with optimization dynamics, leading to non-monotonic policy deviation and incomparable loss values across β settings. By proposing a centered-softplus reformulation, the authors decouple these effects, enabling independent control of preference-noise scale and learning rate while maintaining argmin-equivalence to DPO for β>0. The reformulation also extends to a continuous β→0 endpoint, reducing to a linear preference-margin objective, thus resolving hyperparameter sensitivity and interpretation challenges in standard DPO.
direct preference optimizationkl divergencepreference-noise scaleoptimization dynamicscentered-softplus
FoldPipe: Bounded Remote Streaming of Native Molecular Shards with Asynchronous Prefetch
FoldPipe introduces a lightweight Python orchestration layer for streaming preprocessed molecular graph shards in PyTorch and PyTorch Geometric, enabling asynchronous prefetch with bounded memory usage. The system overlaps I/O and computation by prefetching one shard ahead in a background thread while training on the current shard. Evaluated on SchNet energy-and-force tasks with MD17 aspirin data (25k structures per shard), FoldPipe achieves 16.33s mean I/O-compute overlap versus zero for the baseline, yielding a geometric mean speedup of 1.059× (95% CI: 0.878×–1.288×) despite network variability.
molecular graphsasynchronous prefetchpytorch geometricsharded datai/o-compute overlap
Representation Measurements Under Function-Preserving Reparameterizations
The study demonstrates that column-permutation parallel analysis violates function-preserving reparameterization invariance in language models, as component counts and decisions depend on hidden-coordinate choice rather than model properties. The method evaluates invariance across five models, three retrieval domains, and 75 transformations, measuring median component-count disagreement (0.79) and fixed-threshold decision disagreement (0.26). Empirical results show that centering-only controls alter 1,141 of 1,200 component counts despite unchanged spectra, while orthogonally invariant scores remain stable. These findings highlight the limitations of data-internal reference procedures in preserving coordinate margins and covariance structure.
reparameterization invarianceparallel analysishidden coordinatescomponent-count disagreementorthogonal equivariance
Benchmarking_Fast_Domain_Adaptation_for_Unsupervised_Speech_Units
The paper introduces ABX-Accent, a benchmark for evaluating fast domain adaptation of unsupervised speech representations across 10 English accents using the AESRC dataset. It proposes adaptive domain normalization to fine-tune a pretrained Contrastive Predictive Coding model, initially validated on LibriSpeech via male/female splits. The method achieves a 23.6% relative improvement in across-speaker ABX scores compared to non-adapted baselines, demonstrating effective few-shot adaptation to accented speech domains.
domain adaptationunsupervised speech representationcontrastive predictive codingabx metricaccented speech
Decentralized Multitask Learning over Learned Task Graphs
The paper proposes a decentralized two-phase approach for multitask learning when task relationships are unknown. First, it estimates a generalized graph Laplacian from non-cooperative stochastic gradient iterates using a Gaussian Markov random field prior, then employs the learned graph for cooperative multitask diffusion learning. Theoretical analysis quantifies Laplacian estimation error and its impact on steady-state performance, introducing a topology sensitivity index for network heterogeneity. Simulations show the learned-graph approach outperforms non-cooperative methods and nears true-graph performance with small estimation stepsizes.
decentralized learningmultitask learninggraph laplaciangaussian markov random fielddiffusion learning
Terrain signatures in Welsh settlement names
The study demonstrates that Welsh settlement names retain measurable terrain information, analyzing 3,757 settlements using a frozen 24-element lexical framework and preregistered models. High-terrain names (e.g., 'bryn', 'mynydd') occupied locations 24.4 m higher than their 2-km surroundings (95% CI: 10.8–38.1 m; p = 0.00137), with consistent results across 1-, 2-, and 5-km neighborhoods. Terrain-name polarity improved geographically held-out mean squared error by 4.63–7.30% under spatial blocking. River-related names showed weaker associations, while woodland models were non-estimable. Limitations include residual spatial structure and unresolved name language.
lexical frameworkspatial blockingterrain-name polaritygeographically held-out validationpreregistered models
Why not to use the Gaussian kernel
The paper argues against using the Gaussian kernel as a default choice in kernel-based methods, demonstrating its brittleness through two key results. First, it produces unrealistically small conditional variances, leading to catastrophic overconfidence in predictive uncertainty quantification. Second, it induces numerical ill-conditioning, requiring ad-hoc fixes like nugget terms that modify the underlying model. The authors attribute these issues to the kernel's unnatural smoothness and analyticity, recommending avoidance of analytic kernels generally. Their analysis links analyticity in stationary kernels to exponential spectral density decay.
gaussian kernelpredictive uncertaintynumerical ill-conditioninganalytic kernelsspectral density
Squeezing More from Limited Data with Recursive Transformers
The paper investigates scaling Transformers under limited pre-training data (10M-100M words), revealing that optimal model size depends on both data budget and downstream tasks. To address inefficiencies in standard Transformers, the authors propose recursive Transformers with shared blocks across depth and factorized embeddings, decoupling compute from representational capacity. Evaluations show recursive models outperform standard Transformers at 10M and 100M word scales while matching BabyLM Challenge 2025 winners.
recursive transformerslimited datafactorized embeddingspre-trainingscaling laws
TEMPLAR Wales: A georeferenced environmental and toponymic dataset of Welsh settlements
TEMPLAR Wales introduces a georeferenced dataset linking 3,757 Welsh settlements with environmental attributes and toponymic annotations, enabling quantitative landscape analysis. The method combines deterministic lexical screening (24 Welsh place-name elements yielding 1,350 detections across 1,294 settlements) with multi-scale environmental measurements (terrain, land cover, hydrography) from independent elevation products. Results demonstrate relational integrity, strong terrain source agreement (Pearson's r>0.99), and reproducible lexical-environmental linkages, while explicitly distinguishing computational detections from verified etymologies. The dataset supports research in linguistic geography, GIS, and historical landscape studies through four relational tables with provenance tracking.
georeferenced datasettoponymic annotationsdeterministic lexical screeningmulti-scale environmental measurementslinguistic geography
ClusterAttention: A training-free speedup of bidirectional attention
ClusterAttention introduces a training-free speedup method for bidirectional attention layers by employing fast recursive clustering adapted to key-query geometry, enabling arbitrary cluster sizes optimized for GPU efficiency. The method includes centroid-based compensation to reduce output error, with tighter clusters showing improved accuracy. Evaluations demonstrate a 2-6x speedup on TabPFN-3 while maintaining ≥99% dense accuracy, and a 1.8x speedup on Wan 2.1-14B T2V for video generation, outperforming SVOO. This is the first training-free approach effective for unstructured inputs in single forward passes.
bidirectional attentionrecursive clusteringblock-sparse attentioncentroid compensationgpu efficiency
Graph-Based Pseudo-multimodal Contrastive Learning for 12-Lead ECG Representations
The authors propose Graph-CMMC, a graph-based pseudo-multimodal contrastive learning framework for 12-lead ECG representation learning. The method transforms ECG waveforms into Gramian Angular Difference Field (GADF) images to create complementary representations, then aligns waveform and GADF modalities via self-supervised contrastive learning while modeling inter-lead dependencies with a graph-based relational module. Experiments on coronary artery occlusion classification show competitive performance with supervised methods, demonstrating the effectiveness of GADF representations and explicit inter-lead modeling.
contrastive learningelectrocardiogramgramian angular difference fieldgraph neural networksmultimodal representation
Adversarial Training Without Input Gradients via Low-Rank Householder Expansions
The paper introduces a method for adversarial training without input gradient computations by using low-rank Householder expansions (LRHE) to identify small-norm adversarial examples directly from network parameters. LRHE linearizes the composed affine map of the network, leveraging activation patterns from the forward pass. The proposed training scheme eliminates inner maximization, requiring only additional forward evaluations and standard backward passes. Results show an 8.7× reduction in computational cost compared to 40-step PGD adversarial training on MNIST, matching performance for relative ℓ² budgets ε ≤ 0.02 with 3-step PGD and ε ≤ 0.012 with 40-step PGD.
adversarial traininglow-rank householder expansionsinput gradient-freeℓ²-normforward pass
Packora: Systematic Design for Generative Molecular Crystal Structure Prediction
Packora introduces a flow-based generative model for molecular crystal structure prediction (CSP), jointly predicting atomic coordinates and lattice parameters from molecular graphs. The method supports multi-component and organometallic crystals, with flexible conditioning on conformers, stereochemistry, and space-group information. Evaluated via generation and ranking benchmarks inspired by CCDC CSP blind tests, Packora achieves superior performance, including best-in-class matched-budget coverage across six generation benchmarks, higher experimental-form recovery, and faster convergence in ranking. Key innovations include cacheable pairwise reasoning, conditioning dropout, and balanced scaling of pairwise and single representations.
molecular crystal structure predictionflow-based generative modelmulti-component crystalsconditioning dropoutcacheable pairwise reasoning
Gromov-Monge Flow Matching for Equivariant Graph Generation
The paper introduces Gromov-Monge Flow Matching (GM-FM), a novel approach for equivariant graph generation that leverages quotient space geometry to handle permutation symmetry. The method constructs minibatch couplings using Gromov-Wasserstein relaxations for tractable node alignment while maintaining permutation-equivariance. Theoretical analysis shows that quotient couplings can be lifted to aligned representatives without additional cost. Experiments on continuous graph and molecular generation demonstrate improved sample quality at small integration budgets, with scaled-up molecular models remaining competitive under conventional sampling.
equivariant graph generationgromov-monge distanceflow matchingpermutation-equivariant architecturesgromov-wasserstein
Scaling Model-Generated Distillation Data Can Make Latent Teacher Traits More Recoverable
This work demonstrates that scaling model-generated distillation data enhances the recoverability of latent teacher traits in student models, even when the data is off-task and does not explicitly mention the trait. Using a controlled setup inspired by subliminal learning, the authors generate restricted off-task data (e.g., number-only completions) from a teacher model induced with a target trait. Students trained on varying amounts of independent off-task data are evaluated in a separate domain, isolating target-specific transfer. Results show that larger datasets amplify the teacher's induced trait in the student's behavior, with the target trait growing more prominently than related alternatives. This effect persists across model families, trait types, multi-trait settings, and cross-model transfer.
distillationlatent traitsoff-task datasubliminal learninglora updates
Data-driven Koopman mode approximation: A neural power iteration algorithm
The paper introduces a neural power iteration algorithm for data-driven approximation of dominant Koopman modes in nonlinear dynamical systems, enabling linearized dynamics in a lifted space. The method employs a power-iteration scheme to learn dominant modes directly, avoiding explicit construction of the Koopman operator projection and mitigating dimensionality issues without anti-collapse mechanisms. Theoretical guarantees link convergence to sample size and network width via neural tangent kernel theory. Numerical experiments show the approach outperforms traditional methods like extended dynamic mode decomposition in accuracy and smoothness.
koopman operatorneural power iterationnonlinear dynamicsneural tangent kerneldynamic mode decomposition
A Layer Importance Metric for Quantization Accounting for the Speed-Quality Trade-off in Autoregressive Models
The paper proposes a composite metric for layer-wise quantization prioritization in autoregressive small language models (sLLMs), balancing information retention (SQNR-based coefficient) and throughput gains (roofline latency analysis). The method profiles layers like Feed-Forward Networks and embedding matrices in Gemma 3 1B, combining quality and speed scores into a tunable priority coefficient without execution. Evaluations show 4% prediction error for speedup estimates, outperforming evolutionary search and Shapley-value approaches in resource allocation to expressive layers.
quantizationautoregressive modelsroofline modelsqnrlayer importance
Dose-PlanNet: Physics Based Radiotherapy Dose Prediction with Deep Learning
Dose-PlanNet, a physics-guided 3D deep learning architecture, automates prostate radiotherapy dose prediction while addressing dosimetric complexity in hypofractionated regimens. The model incorporates physical constraints to predict dose distributions, evaluated on a prospective trial cohort with two fractionation regimens. Results show comparable target coverage (D₉₅) with marginally reduced homogeneity (p<0.001), but significant improvements in organ-at-risk sparing (p<0.001). Automated plans met clinical acceptance criteria in 11/14 Moderate Hypofraction Arm and 9/12 Stereotactic Body Radiation Therapy Arm cases, demonstrating feasibility for high-precision clinical workflows.
radiotherapy planningdeep learningdose predictionhypofractionationphysics-guided
Mitigating Strong-Modality Collapse in Multimodal Learning via Inverted Asymmetric Fusion
The paper identifies strong-modality collapse, where dominant modalities degrade during multimodal fusion, explaining why some models fail to surpass unimodal baselines. It proposes Inverted Asymmetric Fusion (IAF), which preserves the dominant modality unchanged while weaker modalities attend to it, strengthened via Modality-Aware Knowledge Distillation. Evaluations on MultiHuSE, UR-FUNNY, and MUStARD show IAF maintains the dominant modality's unimodal accuracy (unlike symmetric fusion's 18.5% drop) and improves over unimodal baselines by up to 8.25%.
multimodal learningmodality collapseasymmetric fusionknowledge distillationattention mechanisms
When Is the Sharp Covariance Envelope Tight? Feature-Only Geometry for Volume-Sampled Least Squares
The work establishes a Loewner envelope for centered coefficient covariance in volume-sampled least squares, proving it is tight when the normalized spectral envelope is strict for all compatible residuals. A residual-augmented change of measure provides the response-aware mechanism, with support saturation proving attainment. Critical equal-leverage geometry interprets boundary conditions, while sound lower certificates enable conservative cardinality decisions. Frozen-feature examples demonstrate non-vacuity and measure fixed-pool cost reductions.
loewner envelopevolume-sampled least squaresspectral envelopeequal-leverage geometrycoefficient covariance
SAGE: Variate-Wise Semantic Augmentation for Vision-Language Time Series Forecasting
SAGE introduces variate-wise semantic augmentation for time series forecasting by jointly modeling temporal, cross-variable, textual, and visual information using a CLIP-based framework. The method processes frequency-enhanced patches and variable tokens through a CLIP text encoder, injects variable-specific descriptions via gated residual paths, and aligns rendered series with temporal representations using a frozen CLIP vision encoder. Evaluated on eight long-term benchmarks and M4, SAGE achieves state-of-the-art accuracy, with ablations confirming benefits from multimodal alignment and variable-level knowledge.
clip-based frameworkvariate-wise augmentationmultimodal alignmenttime series forecastingcontrastive objective
Bridging short- and medium-range weather forecasting with machine learning
The authors propose Nested-EAGLE, a unified 0.25° global weather model with 6 km refinement over CONUS, to bridge short- and medium-range forecasting. The model combines global and limited-area ensemble approaches, incorporating high-resolution regional analysis data during training. Results show significantly lower mean-squared error for near-surface fields over CONUS compared to NOAA's Global Forecast System and HRRR, with competitive global performance. While precipitation amount forecasts lag HRRR due to deterministic training, Nested-EAGLE achieves superior storm location prediction at longer leads. The work highlights the benefits of nested training for surface field accuracy.
weather forecastingensemble predictionnested modelinghigh-resolution refinementglobal-local coupling
Hyperspectral Diffusion Equivariant Imaging (HyDiff-EI): A Self-supervised Framework for Hyperspectral Image Inpainting
HyDiff-EI introduces a self-supervised framework for hyperspectral image (HSI) inpainting, leveraging test-time optimization from a single corrupted HSI acquisition without large-scale pretraining. The method integrates equivariant consistency constraints into the diffusion process, utilizing geometric symmetries and intrinsic HSI characteristics to bridge generative diffusion modeling with physical priors. This approach enhances noise robustness and generalizability. Empirical evaluations on Chikusei, Botswana, and EMIT datasets demonstrate superior inpainting quality over existing self-supervised and diffusion-based methods in both noiseless and noisy scenarios.
hyperspectral imagingdiffusion modelingequivariant constraintsself-supervised learningimage inpainting
Incremental Recommendation via Causal Models
The paper proposes a causal recommendation architecture that optimizes incremental value by reducing impressions for users likely to discover content organically, using existing holdback data without new collection. The method employs a dual-threshold targeting policy to address attribution window mismatches between treated (short-term) and holdback (multi-day) observations, ensuring valid treatment-effect estimation. In a large-scale A/B test on Spotify with millions of users, the policy reduced impressions by 7% without statistically significant consumption loss, while joint training improved calibration of the treated model head, suggesting causal models yield more generalizable representations than observational-only training.
causal recommendationincremental valueholdback datadual-threshold policyattribution window
On the Indistinguishability of Human v/s AI Generated Text
The study investigates the indistinguishability of AI-generated text from human writing by leveraging human samples to strategically paraphrase machine outputs. Using a multi-sample setting with paired human and machine responses, the authors demonstrate that iterative paraphrasing shifts the machine distribution toward the human empirical distribution under specified mixing and stability conditions. Theoretical analysis provides an explicit convergence rate, extends to finite-sample scenarios, and quantifies the scaling of required human samples and paraphrasing rounds with respect to error tolerance.
llmsparaphrasingdistribution alignmentconvergence rateempirical human distribution
Beyond Client Averaging: A Client-Independent Second-Order Stationary-Bias Component in Stochastic SCAFFOLD
The paper identifies a client-independent second-order stationary-bias component in stochastic SCAFFOLD, previously uncharacterized in constant-step analyses. For full-participation stochastic SCAFFOLD with one-dimensional homogeneous clients, fixed local-step count $H$, and bounded additive gradient noise, the authors derive an explicit expression for the stationary mean bias, showing a leading $O(γ/N)$ term suppressed by client averaging and a persistent $O(γ^2)$ term unaffected by client count. The mechanism involves control-induced local trajectory alterations and second-moment corrections converted into bias via nonquadratic curvature. Numerical experiments validate the theoretical predictions, though the analysis is restricted to one-dimensional homogeneous settings.
federated learningscaffoldstationary biasclient averaginggradient noise
Neural Renormalization Group Flow for Percolation
The paper introduces a neural renormalization group approach for two-dimensional site percolation, employing a supervised, scale-shared architecture that recursively applies learned coarse-graining rules. The model predicts crossing probabilities from latent fields and reconstructs largest-cluster masks via fine-graining decoders. Trained on small lattices, it extrapolates to larger systems, accurately recovering spanning clusters and producing observables with correct finite-size scaling near criticality. Key to performance is the latent representation's critical fluctuations and scale-dependent flows aligning with renormalization-group structure.
neural renormalizationpercolationcoarse-grainingfinite-size scalingcritical fluctuations
Equal Ranking Quality, Different Decisions: Training Order-Consistent LLM Scorers
The paper introduces Order-Consistency Supervised Fine-Tuning (OC-SFT) to address order dependence in LLM-based scorers, where equally effective ranking models produce divergent decisions due to candidate order. OC-SFT trains scorers to minimize score variance across permutations while preserving ranking quality (nDCG@10). Evaluated on passage reranking, multi-document QA, and preference tasks, OC-SFT reduces answer flips to 12.5% (vs. 14.9-16.4% for baselines) and improves retained-set overlap (0.66-0.84 → >0.667). It outperforms order-averaged distillation across 12 base models. The work advocates reporting decision metrics (retention F1, answer consistency) alongside ranking quality.
order-consistencyllm scorersranking qualitypermutation invariancesupervised fine-tuning
Safety by Design: Realized-Cost Constraints for Contextual Bandits with Continuous Actions
(No summary returned.)
Self-Augmented Diffusion Guidance for Physics-Informed Generation
The study introduces a self-augmented diffusion guidance method for physics-informed generation of spatiotemporal signals, addressing the limitation of standard diffusion models in violating physical laws. The approach conditions generation on deviation from correct dynamics, decoupling governing equation evaluation from diffusion processes to avoid costly per-iteration simulations. Experiments show significant reduction in physical deviations compared to baseline diffusion models, with further improvements when combined with existing physics-constrained methods.
diffusion modelsphysics-informed generationspatiotemporal signalsself-augmented guidancegoverning equations
Generative Semantic Scene Completion
The paper introduces Generative Semantic Scene Completion (GSSC), reformulating outdoor LiDAR semantic scene completion as a discrete-diffusion framework with three components: (1) PS$^3$ for paired sparse-dense scene synthesis to address class imbalance, (2) SGSC for semantic-guided generative completion via multinomial discrete diffusion, and (3) S$^2$D$^2$ for single-step refinement of existing completions. The method achieves 38.8% mIoU on SemanticKITTI's hidden test set without test-time augmentation, surpassing prior work by +2.1 percentage points under identical constraints, and reaches 39.2% mIoU with multi-view augmentation.
lidardiscrete diffusionsemantic scene completionclass imbalancetest-time augmentation
Neural Regression with Embeddings for Numerical Attribute Prediction in Knowledge Graphs
The paper introduces LitEm, a neural regression model that extends transductive knowledge graph embedding models to predict numerical attributes, addressing a gap in handling diverse information in real-world knowledge graphs. LitEm combines with existing embedding models via a co-training framework, improving both attribute prediction and link prediction performance, particularly for bilinear models. Evaluations on FB15K-237, YAGO15K, DB15K, and Mutagenesis show LitEm achieves top-2 results on most attributes, with literal-awareness analysis confirming meaningful attribute encoding. The implementation is publicly available.
knowledge graph embeddingneural regressionnumerical attribute predictionco-trainingliteral-awareness
Domain-Specific Self-Supervised Representation Learning for Retinal Fundus Classification
The study demonstrates that domain-specific augmentation strategies enhance self-supervised representation learning for retinal fundus classification under resource constraints. Two SSL frameworks, SimSiam and SimCLR, were evaluated for retinal disease classification, focusing on augmentation techniques and training parameters tailored to retinal images. Experiments assessed representation quality via linear evaluation and fine-tuning across downstream tasks, including multi-disease classification and diabetic retinopathy grading. Results indicate that retinal-specific augmentations improve performance, enabling lightweight SSL models to learn transferable representations and reduce reliance on large annotated datasets while achieving competitive results.
self-supervised learningretinal fundusaugmentation strategiessimsiamsimclr
Simple Actors and Deep Critics for Scalable Reinforcement Learning
The paper proposes LAC (Light Actor, deep Critic), an offline RL method that prioritizes critic capacity over actor capacity for inference efficiency. Addressing three failure modes in deep critics (optimization instability, bootstrap-noise amplification, value-range drift) via residual MLPs, n-step targets, and categorical cross-entropy loss, LAC pairs a lightweight deterministic actor with a deep critic. On OGBench, it matches diffusion/flow-matching baselines (4x lower latency) while maintaining performance comparable to distilled one-step policies. The critic architecture generalizes across actor parametrizations.
offline reinforcement learningactor-critic methodsinference latencyresidual mlpbootstrap targets
When Privacy Hurts Mergeability: Geometry-Aware Model Merging under Differential Privacy
The paper introduces DP-Merging, a geometry-aware framework for improving the mergeability of differentially private (DP) task models while preserving privacy guarantees. The method addresses two geometric obstacles—local sharpness and reference drift—by employing a DP-compatible sharpness-aware objective and a reference-based alignment regularizer. Experiments on vision and language tasks demonstrate consistent performance improvements across multiple privacy budgets, with theoretical analysis showing reduced merge-gap bounds through curvature and drift minimization.
differential privacymodel mergingsharpness-aware optimizationreference driftmerge-gap
Robust Neural Stimulation Response Modeling Through Meta-Learning and Pretraining
The study demonstrates that meta-learning and pretraining significantly improve robustness and reduce calibration requirements in neural stimulation response modeling. Using model-agnostic meta-learning (MAML) with temporal basis function models (TBFMs), the authors evaluate 40 optogenetic stimulation sessions in non-human primate sensorimotor cortex. Meta-learning reduces catastrophic forecast failure (test R-squared < 0.05 drops from 16/40 to 1/40 sessions) and narrows prediction intervals (p < 0.05), while cutting calibration needs by 50-90% at matched accuracy.
meta-learningneural stimulationtemporal basis function modelsoptogenetic stimulationmodel-agnostic meta-learning
Hierarchical Channel Stacking: A Structured Decision Framework for AI-Generated Image Detection
The paper introduces Hierarchical Channel Stacking (HCS), a structured framework for AI-generated image detection that preserves interpretable hierarchical decision-making. HCS converts CNN activations into a 60-dimensional representation across three progressively deeper backbone stages, using per-channel Level-1 classifiers and a Level-2 aggregator for final predictions. Evaluated on a benchmark with GAN and diffusion generators, HCS achieves 86.7% accuracy and macro-F1, with ablation studies confirming the complementary value of multi-stage hierarchy and revealing distinct stage-level contribution profiles for different generator types.
hierarchical channel stackingai-generated image detectioncnn activationsmulti-stage classifierinterpretable decision-making
A Unified Descriptive-Complexity Framework for Model Selection under Correlated Designs
The paper introduces the Descriptive-Complexity Information Criterion (DCIC), a model selection framework for high-dimensional settings with correlated predictors. DCIC employs Kraft-admissible code lengths to regularize exponentially large model collections, achieving selection consistency under sub-Weibull noise without restricted isometry property (RIP) assumptions. Nonasymptotic oracle risk bounds hold under misspecification, and a complexity-guided search path enables explicit computation-statistics trade-offs. Experiments demonstrate robust support recovery and estimation under strong dependence. The method also extends to heterogeneous model classes via a unified complexity scale, enabling class-model recovery and risk adaptation.
model selectionkraft-admissible codingsub-weibull noiseoracle risk boundscomplexity-guided search
Technical Comparative Benchmarking Study: Advanced AI Hybrid Methods for Renewable Energy Farm Optimization and Forecasting
This benchmarking study evaluates hybrid AI methods for renewable energy optimization, comparing conventional ML, deep learning, and ensemble approaches across three wind energy datasets. For structured WEC layout data, Extra Trees achieved a 63.7% MAE reduction versus MLP baselines by capturing nonlinear power interactions through randomized partitioning. STGCN demonstrated spatial-temporal modeling (MAE=167.0 kW, R=0.93), while the RF-BiLSTM hybrid achieved optimal forecasting (MAE=150.5 kW, 75% improvement over LSTM). Results indicate architecture-specific strengths: tree ensembles excel in structured data, graph networks in spatial interactions, and hybrid recurrent models in temporal-tabular scenarios.
ensemble learningstgcnextra treesrecurrent hybridsscada
hoBIT: A Profile-Aware Retrieval-Augmented Chatbot for University Academic Advising
The paper introduces proFILL, a method to enhance hoBIT, a rule-based academic advising chatbot, into a profile-aware retrieval-augmented generation (RAG) system. proFILL dynamically acquires necessary user profile attributes per query, guided by query intent and initial evidence, to condition retrieval over a profile-aware index. Experiments and human evaluations demonstrate proFILL's superiority over RAG baselines, user preference, and effectiveness with open-weight models for cost-efficient deployment.
retrieval-augmented generationprofile-aware retrievalquery intentdynamic attribute acquisitionopen-weight models
SimCast-S2S: An Efficient Generative Model for Subseasonal Precipitation Forecasting via Transfer Learning from Climate Simulations
SimCast-S2S introduces a generative latent-diffusion framework for probabilistic subseasonal-to-seasonal (S2S) precipitation forecasting, addressing key challenges in data-driven prediction. The method employs a diffusion-based generative pipeline for uncertainty quantification, operates in a compact latent space learned by variational autoencoders for efficient ensemble generation, and leverages transfer learning with low-rank adaptation (LoRA) to pretrain on climate simulations before fine-tuning on reanalysis data. SimCast-S2S outperforms deep learning baselines and matches or exceeds state-of-the-art operational systems like ECMWF-S2S, demonstrating the efficacy of latent generative modeling and simulation-to-reanalysis transfer learning for scalable S2S forecasting.
latent-diffusionsubseasonal-to-seasonalvariational autoencoderslow-rank adaptationprobabilistic forecasting
GRAS: Guided Reduced-Variance Proposals and Adaptive Selection for Training-Free Reward Alignment in Discrete Diffusion
The paper introduces GRAS (Guided Reduced-variance Proposals and Adaptive Selection), a training-free method for reward alignment in discrete diffusion models. GRAS addresses two weaknesses in existing approaches: high variance in gradient estimates from single noisy samples and fixed-temperature resampling. It employs Rao-Blackwellized reveals for differentiable rewards and leave-one-out baselines for non-differentiable rewards to reduce variance, and standardizes per-step values into group-relative advantages with adaptive resampling temperatures. Evaluated on regulatory DNA and protein design tasks, GRAS outperforms prior training-free methods and matches or surpasses reward-fine-tuned models, even for non-differentiable rewards.
discrete diffusionreward alignmentrao-blackwellizedadaptive resamplinggradient guidance
Activation Outliers Matter: Robust Recovery for Quantized Multimodal LLMs
We propose Residual Fallback Quantization (RFQ), a lightweight activation reconstruction framework for robust ultra-low-bit quantization of Multimodal Large Language Models (MLLMs). RFQ supplements primary ultra-low-bit activation representations with an auxiliary quantized residual pathway to explicitly model and compensate for quantization errors, improving activation fidelity without architectural modifications or significant computational overhead. Experiments on Wan2.2 and Qwen3-VL show RFQ recovers substantial performance lost under MXFP4 and HiF4 quantization, narrowing the gap to BF16 baselines across generation and reasoning benchmarks. Activation quantization is identified as the dominant bottleneck in ultra-low-bit MLLMs.
activation quantizationresidual fallback quantizationmultimodal large language modelsultra-low-bit quantizationactivation reconstruction
Arrive and Survive: Scaling Safe Goal-Conditioned Policy Learning from One-Bit Failure Signals
The paper introduces Safe Contrastive Reinforcement Learning (Safe-CRL), correcting systematic overestimation bias in goal-conditioned policy learning under failure termination. Theoretical analysis reveals that conventional contrastive RL (CRL) overlooks probability mass removed by failure, leading to unsafe behaviors. Safe-CRL employs mass-weighted InfoNCE for critic learning and a log-survival-mass score for policy optimization, using only one-bit failure signals. Evaluated on twelve robot navigation and locomotion tasks, Safe-CRL improves survival and goal-reaching performance over Scaling-CRL, demonstrating complex failure-avoidance behaviors. The work completes CRL theory under failure termination and provides a scalable safe RL framework.
contrastive reinforcement learninggoal-conditioned policyfailure terminationinfoncesafe rl
Dynamical phase selection controls compute scaling in looped transformers
The paper demonstrates that looped transformers with identical architecture and training objectives can exhibit distinct dynamical phases depending on initialization, critically affecting test-time compute scaling. Through analysis of bifurcation mechanisms (saddle-node fold and Neimark-Sacker transitions), the authors show phase-dependent scaling laws: fold phases follow a parameter-free relation τ(ε)[1−λₘₐₓ(−ε)]→π and produce workload tails P(τ>N)∼N⁻², while Neimark-Sacker phases eliminate fold scaling. Results establish that compute costs are governed by dynamical phases rather than architecture alone.
looped transformersdynamical phasesbifurcation mechanismstest-time computecritical slowing down
Hadamard Flattening and Gaussian Pooling Sketch for Least Squares with Coordinate-wise Guarantee
The paper introduces a novel randomized sketch-and-solve algorithm for overconstrained ℓ₂ regression with coordinate-wise ℓ∞ guarantees. The method combines Hadamard flattening, random permutation, and disjoint Gaussian pooling to ensure conditional independence of noise in the sketched problem. This achieves an ℓ∞ error bound of ε/√d·‖Ax*−b‖₂·‖A†‖ₒₚ using O(ε⁻²d log d) rows, with computational complexity Õ(nd + ε⁻²d⁴) for sketching.
randomized sketchingℓ₂ regressionhadamard transformgaussian poolingℓ∞ guarantee
Chart2SVG: Editable SVG Generation from Raster Chart Images
Chart2SVG introduces a multimodal large language model for converting raster chart images into editable SVGs with structural and semantic organization. The method integrates chart-specific semantic tokens into a vision-language model, supported by Beagle+, a dataset of 33K canonicalized chart samples, and employs specialized training objectives with rendering-aware post-training. The resulting SVGs are enhanced by a Chart Structure Graph (CSG) for visual dependencies, enabling interactive exploration and layout reuse. Experiments demonstrate superior reconstruction fidelity and editing utility compared to baselines.
multimodal large language modelsemantic tokenschart structure graphraster-to-svgvisual dependencies
Predicting Quantifiability from Primary Screens to Prioritize Dose-Response Profiling
The study introduces a framework for predicting quantifiability—whether follow-up dose-response profiling will yield usable potency estimates—as distinct from biological activity in high-throughput drug screening. Using response-based predictors derived from primary screen features rather than molecular structure, the method demonstrates strong predictability across unseen chemical scaffolds and assay-mechanism families. Results show quantifiability varies with response amplitude and assay context, enabling improved allocation of costly dose-response profiling resources.
quantifiabilitydose-response profilinghigh-throughput screeningpotency estimationassay-mechanism families
High Probability Derivative Bounds for Random tanh Neural Networks on a Hypercube
The paper establishes high-probability bounds for mixed input derivatives of wide random tanh neural networks with Xavier initialization, addressing exponential growth issues in deterministic analyses. By isolating terms linear in the highest-order derivative and controlling tangent directions via measurable finite nets, the authors derive improved bounds for Gaussian networks. For scalar-output tanh networks, they prove that first-order derivative bounds are depth-independent, while higher-order mixed derivatives grow polynomially with depth. These results yield bounds for Euclidean Lipschitz constants and weighted Sobolev norms, connecting derivative estimates to quasi-Monte Carlo integration and training analysis.
tanh networksxavier initializationderivative boundssobolev normsquasi-monte carlo
Algorithmic Principles For Multiclass Learning Are Hard To Come By: Limits of Regularization and Proper Learning
The paper investigates fundamental limitations in multiclass learning, addressing three open problems. First, it demonstrates that learning cannot always be reduced to proper learning via hypothesis class enlargement. Second, it shows proper learners require $o(m)$ training errors for sample size $m$, with every sublinear scale being necessary for some problem. Third, it proves regularization is insufficient for general learning, exhibiting classes unlearnable by Structural Risk Minimization (SRM) or local regularizers. Positive results include sufficient conditions for SRM learnability and representability via revealed preference integrability.
multiclass learningproper learningstructural risk minimizationregularizationrevealed preferences
Sharp Minimax Regret for Infinite-Memory Logistic Prediction
(No summary returned.)
A Single Suffix to Break Them All: Basin-Aware Jailbreaks for Merged Model Families
The paper identifies a novel jailbreak risk in model merging, demonstrating that merged models can inherit vulnerabilities from their pretrained backbone even when constituent models are safety-aligned. It proposes Basin-Aware Jailbreak (BAJ), a min-max optimization method to generate transferable adversarial suffixes across merged model families sharing the same backbone. Experiments show BAJ achieves high transfer success rates (exact numbers unspecified) across diverse backbones and remains effective against existing defenses.
model mergingjailbreak attackadversarial suffixtransferabilitysafety alignment
Systematic Literature Review of Machine Learning Models and Applications for Text Recognition
This systematic review evaluates advancements in Optical Character Recognition (OCR) models over the past decade, analyzing 97 studies (2015–2025) under PRISMA guidelines. It traces architectural improvements in handling multilingual, structured/unstructured, and scene text, while identifying persistent challenges like script variability and real-time constraints. Key findings highlight performance gaps in underrepresented languages and handwritten text, proposing solutions such as self-supervised learning, multimodal AI, and TinyML. The study provides a foundation for future OCR research, emphasizing AutoML and joint corpora for script matching.
optical character recognitionmultilingual processingself-supervised learningautomated machine learningtiny machine learning
A Unified Framework for Fair and Personalized Decentralized Learning under Communication Constraints
(No summary returned.)
Bayesian methods and Markov chain Monte Carlo algorithms for curve reconstruction and point cloud data analysis
The authors propose a Bayesian framework for point-cloud data analysis and curve reconstruction, addressing noise and missing data while providing uncertainty quantification. Their method models observed points as noisy perturbations of latent locations on a curve, regularized by a non-parametric prior, and employs tailored Markov chain Monte Carlo samplers for posterior inference. Experiments on synthetic and LiDAR datasets demonstrate accurate reconstructions with quantified uncertainty.
bayesian frameworkpoint-cloud reconstructionmarkov chain monte carlouncertainty quantificationnon-parametric prior
Shared Actors Need Not Share Critics: Effects of Value Mismatch in Parallel Reinforcement Learning
The paper demonstrates that using a shared critic across parallel environments in reinforcement learning induces value mismatch, systematically distorting policy updates despite converging to the same optimal policy. Through bandit models and experiments on CartPole, MuJoCo, BipedalWalker, and Procgen, the authors show that conditioning the critic on environment indices resolves this issue, improving normalized returns by 40.8% across 16 Procgen games. The analysis reveals that value mismatch directly degrades stochastic learning dynamics beyond scalar estimator variance effects.
parallel reinforcement learningvalue mismatchshared criticstochastic learning dynamicsconditional critic
Active Curriculum Refinement for Reinforcement Learning
The paper introduces PATH, a curriculum-learning framework for reinforcement learning that actively refines training over directed acyclic curriculum graphs (DAGs). PATH samples diverse curriculum paths to expand coverage and reallocates training toward unmastered regions, explicitly leveraging the graph structure. Experiments across diverse environments demonstrate PATH's robustness and generalization capabilities.
reinforcement learningcurriculum learningdirected acyclic graphactive learninggeneralization
Distributed Training using an Intelligent Network
The paper proposes an intelligent network approach for distributed training across wide area networks (WANs), addressing bandwidth, latency, and topology constraints. The method combines multicast technology for efficient outbound traffic replication and in-line FPGAs for inbound traffic aggregation, extending data center techniques to WANs. An optimization framework generates synchronization schedules (rotating cliques) tailored to network topology and capabilities. Experiments on a nine-city DoubleZero network demonstrate adaptive schedule optimization, narrowing the performance gap to colocated training.
distributed trainingwide area networkmulticast technologyin-line fpgasynchronization schedules
Toward Equitable Low-Carbon Mobility: Fairness-Aware Demand Prediction for Expanding Bike-Sharing Systems
FairGIN, a fairness-aware graph neural network, addresses equitable demand prediction in expanding bike-sharing systems by mitigating cold-start gaps and structural inequalities. It integrates Expansion-Simulated Increment Training to simulate network expansion, Attention-Based Knowledge Transfer for station-adaptive representation transfer, and Fairness-Aware Optimization with income-stratified regularization and equity-calibrated deployment scoring. Evaluated on NYC and Seattle datasets, FairGIN achieves state-of-the-art predictive accuracy across expansion scenarios while significantly reducing income-based disparities without compromising system efficiency.
graph neural networkcold-start predictionattention-based knowledge transferincome-stratified regularizationequity-calibrated deployment
Vowel Signs Are Not Letters: A Pre-tokenization Ceiling on Multilingual Tokenizer Fertility
The paper identifies a pre-tokenization ceiling in multilingual BPE tokenizers caused by GPT-2's \p{L}+ regex pattern, which splits abugida-script words at vowel signs. Using a parallel corpus of 26 languages, the authors demonstrate fertility increases ranging from 1.47x (Tibetan) to 9.02x (Thai) for 17 abugidas, while non-abugida scripts remain unaffected. Experiments with matched tokenizer pairs and 268M-parameter models show the fixed variant reduces Nepali bits per byte by 4.43% and outperforms even with 1.59x compute, while a census reveals 63.3% of popular text-generation models use the problematic pattern.
bpe tokenizersabugida scriptspre-tokenizationfertility floorunicode letters
Subgraph Filtering for Fair Graph Neural Networks
Proposes Subgraph Filtering for Fair Graph Neural Networks (SF-GNN), a lightweight framework mitigating structural bias in GNNs by targeting bias-prone edges during message passing. SF-GNN identifies such edges via sensitive homophily combined with structural amplifiers (e.g., hub participation, triadic closure), then applies stochastic edge filtering per aggregation step while preserving overall graph structure. Incorporates a warm-up-scheduled statistical-parity regularizer for stable optimization. Evaluated on five benchmarks, SF-GNN improves fairness-accuracy trade-offs over baselines while maintaining competitive predictive performance.
graph neural networksstructural biassensitive homophilystochastic edge filteringstatistical parity
NeoTriFuse: Reliability-Aware Multimodal Fusion under Missingness Heterogeneity for Neonatal Mortality Risk Prediction
NeoTriFuse introduces reliability-aware multimodal fusion for neonatal mortality risk prediction, explicitly modeling missing data as reliability signals to dynamically modulate modality contributions during fusion. The framework combines static perinatal variables, local-global temporal encoders, and patient-level statistical summaries via reliability-guided gating, jointly optimizing mortality prediction and length-of-stay objectives. It achieves an F1 score of 0.6736 ± 0.0216 and AUROC of 0.9454 ± 0.0056, with ablation studies highlighting the importance of local-global temporal architecture and patient-level summaries.
multimodal fusionmissingness heterogeneityreliability-aware gatingtemporal encodersneonatal mortality prediction
FedCMAPSS: A Benchmark for Federated Learning in Remaining Useful Life Estimation
The paper introduces FedCMAPSS, a benchmark for federated learning in remaining useful life (RUL) estimation, addressing the lack of standardized evaluation frameworks. Based on NASA C-MAPSS, it defines five tasks simulating industrial challenges from IID to extreme heterogeneity, evaluating federated optimization algorithms across neural architectures. The work provides reproducible baselines and public code/data splits to standardize federated predictive maintenance research.
federated learningremaining useful lifepredictive maintenancecmapssbenchmark
Interpreting Latent Protein Language Model Features with Geometric Annotations
The study introduces a geometric annotation method for interpreting sparse autoencoder (SAE) features in ESM-2, addressing limitations of existing protein-level annotation pipelines. By analyzing local Cα backbone geometry, the method identifies significant associations between SAE features and structural patterns, revealing substructure within biological labels and enabling annotation of uncharacterized metagenomic sequences. Results show geometry distinguishes SAE features with identical database annotations, and ablation experiments demonstrate geometric features influence ESM-2's contact predictions. This bridges mechanistic interpretability and structural biology through residue-level feature annotation.
sparse autoencoderprotein language modelgeometric annotationesm-2contact prediction
Towards a universal meta-optics solver via large language models
The study introduces a unified large language model (LLM) approach for multi-family metasurface modeling and inverse design, addressing limitations of conventional neural network surrogates that require family-specific architectures. By converting geometries, parameters, and optical responses into a shared instruction-following text format, the authors fine-tuned Gemma-2-9B across 8 metasurface families. The joint model achieved a 56.5% average reduction in mean squared error (MSE) for optical response prediction compared to single-family baselines, while enabling cross-family inverse design without task-specific architectures.
metasurface designlarge language modelinverse designoptical responsemulti-family modeling
LowRankArena: A Standardized Evaluation Platform for SVD-Based LLM Compression
LowRankArena introduces a standardized evaluation platform for SVD-based LLM compression, addressing inconsistencies in prior studies by unifying benchmarks, compression budgets, and inference measurements. The platform provides reproducible pipelines and 3 TiB of compressed checkpoints. An audit of five SVD methods reveals conditional performance: leader rankings shift across backbones and compression ratios, multiple-choice accuracy masks perplexity degradation, and low-rank savings yield workload-dependent speedups.
svd-based compressionlow-rank approximationllm efficiencybenchmark standardizationperplexity degradation
Finding the Right Evidence: Factor-Guided Coarse-to-Fine Reasoning for Long Videos
PACE (Progressive Acquisition of Critical Evidence) introduces a factor-guided coarse-to-fine framework for long-video question answering, addressing the challenge of sparse, discriminative evidence. The method operates in two stages: first indexing clip-level descriptions using question-derived factors, then deriving contrastive cues from candidate answers to query the index for verification. Evaluated on MMR-V with the Qwen3-VL backbone, PACE achieves 42.6% accuracy, outperforming direct inference and Deep Video Discovery (DVD), and recovers 66.9% of annotated cues. Consistent gains across LVBench, Video-MME, EgoSchema, and LongVideoBench demonstrate transferability beyond MMR-V.
long-video question answeringfactor-guided frameworkclip-level descriptionscontrastive cuesevidence recovery
Cross-simulator transfer with foundation model summaries: Towards robust SKA-era reionization inference
The paper introduces SKATR, a self-supervised Vision Transformer (ViT) pretrained with a Joint Embedding Predictive Architecture (JEPA) on 67k noiseless 21cmFAST lightcones, as a foundation model for robust reionization inference across simulators. Without retraining, SKATR's frozen encoder generalizes to hydrodynamical Loreli II lightcones, where a lightweight conditional flow matching head infers five astrophysical parameters. SKATR achieves the most precise and calibrated posteriors, matching in-domain supervised baselines while requiring 2.6× fewer radiative-transfer simulations, and remains accurate under realistic SKA AA* noise.
vision transformersimulation-based inferenceradiative transferconditional flow matching21cm cosmology
Beyond Capability Benchmarks: Learning Operational Fingerprints of LLM Cloud Services from Production Incident Metadata
The paper introduces Operational Embedding (OpEmbed), a framework for learning operational fingerprints of LLM cloud services from production incident metadata. OpEmbed processes model-time windows into eight-channel signatures and learns low-dimensional representations via temporal contrastive learning, cross-view reconstruction, and generational-ordinality regularization. Evaluated on 33,000 support cases across seven LLM families over 26 months at Google Cloud, OpEmbed recovers interpretable model structure, improves operational forecasting by 12% over baselines, and enables cross-model fault-type transfer.
operational embeddingllm cloud servicestemporal contrastive learningcross-view reconstructiongenerational-ordinality regularization
Privacy Without Regret: Differentially Private Inference-Time Alignment
The paper introduces differentially private inference-time alignment methods to address reward hacking and privacy concerns in Best-of-N (BoN) sampling. Private Best-of-N (PrivBoN) adds Gumbel noise to reward scores, achieving ε-differential privacy and KL-regularized alignment when the privacy budget exceeds a threshold ε*. Private Inference-Time Pessimism (PrivITP) combines χ²-regularized rejection sampling with a two-phase Gaussian mechanism, achieving (ε,δ)-DP with privacy cost independent of the number of responses. Experiments show both methods scale monotonically, with PrivITP outperforming PrivBoN in strong-privacy regimes.
differential privacyinference-time alignmentbest-of-n samplingreward hackinggumbel noise
When Is Noise Response Universal? Tokenization as the Hidden Variable in Language Models
The study identifies tokenization as the key factor determining language models' universal response to noise, showing that degradation rates align under word-level noise but diverge under character-level perturbations. Through experiments with eight encoders across six pretraining paradigms and decoder-only LLMs, the authors demonstrate that contrastive training collapses initially scattered robustness curves into a universal response. They attribute the word/character dichotomy to tokenization dynamics, where character edits disrupt token sequences more severely than word drops, offering a predictive framework for noise robustness and noise-augmented training strategies.
tokenizationnoise robustnesscontrastive traininglanguage modelsdegradation rate
Algebraic Multigrid Acceleration for Efficient Label Spreading
The paper introduces AMELS, an efficient label spreading framework that accelerates semi-supervised learning via algebraic multigrid solvers and fast neighborhood graph construction. By replacing random walk iterations with multigrid cycles, AMELS propagates labels across large graphs in a single pass, reducing computational costs. Experiments show runtime improvements over baselines while maintaining robustness to hyperparameters and accuracy with few labeled samples, enabling scalable application to high-dimensional image datasets.
label spreadingalgebraic multigridsemi-supervised learninggraph propagationscalability
District-Level Food Environment Indicators and Social Vulnerability in São Paulo
The study demonstrates that district-level food environment indicators in São Paulo correlate with social vulnerability, using machine learning to analyze their discriminative power. Researchers integrated the São Paulo Social Vulnerability Index (IPVS) with food retail data from RAIS and street-market records from CAISAN, aggregating census-sector information across 76 districts. XGBoost achieved the highest mean F-score (0.75), with healthy and unhealthy food establishment densities accounting for 60% of feature importance in Random Forest, though ecological limitations constrain broader inferences.
social vulnerability indexfood environment indicatorsxgboostrandom forestecological analysis
Muon with Finite Newton-Schulz: The Smoothing Benefit in Nonsmooth Nonconvex Optimization
The paper demonstrates that finite Newton-Schulz iterations in Muon, an optimizer for matrix-valued parameters in LLM pretraining, provide beneficial smoothing for nonsmooth nonconvex optimization. By analyzing Muon via online-to-nonconvex conversion, the authors show that finite iterations transform the discontinuous polar map into a Lipschitz map of singular values, enabling convergence to stationary points. A logarithmic growth in Newton-Schulz depth suffices for convergence, outperforming exact-polar updates. The derived sample complexity bounds match optimal nonsmooth nonconvex guarantees and extend to general spectral maps with similar smoothing properties.
newton-schulznonsmooth optimizationonline-to-nonconvexspectral smoothingmuon optimizer
Multi-Dataset Inverse Problem Solving with Distributed Generative AI
The authors present a distributed generative AI framework for solving inverse problems across multiple heterogeneous datasets, extending their prior SAGIPS system. The method employs dataset-specific forward operators and discriminators to constrain a shared generator, enabling joint analysis of non-identically distributed data while maintaining parameter consistency. Experimental validation using a multi-detector scattering setup demonstrates robustness to varying data fidelities and effective scaling on multi-GPU systems, addressing challenges in real-world multi-dataset inverse problems.
inverse problemsgenerative aidistributed trainingheterogeneous datasetsparameter consistency
Constraint-Aware Physics-Informed Neural Networks for Static Shape Estimation of Co-Manipulative Continuum Robots
The paper introduces a constraint-aware physics-informed neural network (PINN) for static shape estimation in co-manipulative continuum robots (CCRs), addressing both static equilibrium and geometric loop-closure constraints. The method combines a projected static equilibrium residual and a geometric residual within a PINN framework, outperforming a data-driven ANN under limited and noisy data (67.88% lower configuration error with 140 samples and 50% noise). The PINN achieves 0.1597% relative configuration error with 0.1773 ms inference time, and experimental fine-tuning reduces marker RMSE from 2.657 mm to 0.497 mm.
physics-informed neural networkscontinuum robotsstatic equilibriumloop-closure constraintsgeometric variable strain
Agentic Autoresearch for Cell-Edge Power Control: Radically Redefining the Researcher's Role
The paper introduces an autonomous AI agent that fully automates the design of machine learning algorithms for wireless power control, eliminating manual specification of architecture, loss function, and training recipe. Using the autoresearch protocol, the agent iteratively edits training scripts, runs experiments, and retains changes based on a fixed performance metric, with safeguards including hash-pinned evaluation and pre-registered falsification. In 81 unattended experiments, the agent achieved 99.5% of a reference minorization-maximization performance at 600× lower inference cost, discovering provably optimal structures rather than tuned constants.
autonomous agentswireless resource managementautoresearch protocolpower controlminorization-maximization
A causal graph-informed temporal convolution architecture for interpretable retail electricity price forecasting
The study introduces a Causal Graph-Informed Temporal Convolutional Network (CG-TCN) for interpretable retail electricity price forecasting, integrating a learned causal graph into a temporal convolutional network via graph-neural embedding. The method employs multi-resolution decomposition to isolate trends, discovers a causal graph over components and covariates, and encodes it as an adjacency embedding to condition convolutions and attention. Evaluated on ten years of Ohio residential contract data, CG-TCN achieves mean absolute percentage errors of 3.08%, 3.82%, and 5.43% for one-, ten-, and fifteen-step-ahead forecasts, outperforming benchmarks while providing interpretable insights.
temporal convolutional networkcausal graphmulti-resolution decompositiongraph-neural embeddingretail electricity forecasting
Pruning Binarized Neural Networks: A Dedicated Framework and Globally Weighted Algorithms
The authors introduce a PyTorch-based framework for pruning binarized neural networks, enabling rapid evaluation and prototyping of pruning strategies tailored to binary representations. They propose a novel pruning method that globally weights parameters across abstraction levels, optimizing the trade-off between accuracy and pruning rate. The method achieves a 70% pruning rate on VGG11 with maintained accuracy, surpassing state-of-the-art results of 41% in binarized settings. This framework facilitates efficient deployment on edge hardware like FPGAs and microcontrollers by combining extreme compression techniques.
binarized neural networkspruningfpgaedge hardwarepytorch
Group-Shared Low-Rank Approximation for Mobile-Efficient Pointwise Convolutions in Large-Kernel CNNs
Proposes Channel Group-Shared (CGS) low-rank approximation, an SVD-based parameter-sharing strategy to compress pointwise convolutions in large-kernel CNNs, which dominate parameter volume (>87% in RepLKNet-31B). CGS employs shared down/up-projection matrices across channel groups and group-specific diagonal matrices, reducing storage costs while maintaining performance. Experiments on RepLKNet, ConvNeXt, and SLaK show competitive accuracy with reduced memory bandwidth and loading latency, enabling edge deployment.
large-kernel cnnspointwise convolutionslow-rank approximationedge deploymentsingular value decomposition
Robust CurveMoE: Multi-Norm Adversarial Defense for Mixture-of-Experts Models via Mode Connectivity
Robust CurveMoE introduces a multi-norm adversarial defense framework for mixture-of-experts models via mode connectivity, addressing competing robustness objectives efficiently. The method connects norm-specialized experts through a low-loss path, selectively expertizes influential layers, and employs contribution-guided partial updating to reduce curve-construction costs. Theoretical bounds ensure minimal objective gap between partial and full curve optimization. Evaluations on CIFAR-100 and ImageNet-100 with WideResNet and Vision Transformer architectures demonstrate consistent improvements in clean, norm-specific, and Union accuracy, outperforming MSD and ERMC by 2.37 and 2.13 percentage points, respectively. Ablations validate the efficacy of partial updating, selective expertization, and robustness-constrained expert selection.
mode connectivitymixture-of-expertsmulti-norm adversarial defensecontribution-guided partial updatingrobustness-constrained expert selection
CardioFusion-AI: Robust ECG--PPG Fusion for Multimodal Physiological Monitoring Under Signal Degradation
CardioFusion-AI introduces a robust ECG-PPG fusion framework addressing signal degradation in multimodal physiological monitoring. The method combines R-peak/systolic-peak detection, a signal-quality index, and pulse transit time estimation, validated on 53 ICU recordings (1.61 bpm ECG/2.78 bpm PPG error) and fetal ECG data (R-peak F1 0.89-0.98). Synthetic degradation tests compared eight fusion strategies: attention fusion achieved lowest error (1.66±0.43 bpm), while signal-quality conditioning improved missing-PPG performance (1.56±0.59 bpm vs. 1.48 bpm unimodal). Results show modality availability and quality require distinct adaptive fusion approaches.
ecg-ppg fusionsignal degradationadaptive fusionpulse transit timesignal-quality index
Spectral Allocation: Why Muon Outperforms Adam, and How to Improve Muon
The paper introduces Spectral-Aware Muon (SAMuon), an improved optimizer that outperforms Muon and Adam by better utilizing the anisotropic spectral profile of Transformer loss landscapes. Through spectral probing, the authors show that loss landscapes exhibit a stable anisotropic structure with a volatile head requiring small steps and a tolerant bulk permitting larger steps. SAMuon amplifies bulk directions using a static spectral prior, achieving 13.3-24.0% faster convergence than Muon on models from 124M to 1B parameters, with negligible computational overhead.
spectral allocationoptimizertransformerloss landscapeanisotropy
Lost but not erased: Finding traces of a forgotten language in neural speech models
The study demonstrates that neural speech models retain functional traces of a forgotten first language (L1) after abrupt switching to a second language (L2), challenging biological critical-period explanations. Using automatic speech recognition models trained sequentially on L1 then L2, researchers found persistent L1 traces primarily in pre-phonemic layers. Models with prior L1 exposure re-learned L1 14% faster than naive models, an advantage eliminated by substituting earliest layers from non-adopted models. Results suggest critical-period effects stem from representational entrenchment rather than maturational plasticity loss.
neural speech modelscritical periodautomatic speech recognitionrepresentational entrenchmentpre-phonemic layers
When Pruning Meets Interpretability: Preserving Sparse Autoencoder Robustness in LLMs
The study analyzes how pruning impacts sparse autoencoder (SAE) interpretability in large language models (LLMs), demonstrating that perturbation energy—a covariance-weighted norm—governs SAE robustness. Theoretical analysis reveals magnitude pruning distorts activation geometry, while activation-aware methods (Wanda, SparseGPT) better preserve SAE functionality by implicitly controlling perturbation energy. Experiments across four architectures show middle layers are most pruning-sensitive, leading to a layer-wise sparsity allocation strategy that improves perplexity at equal average sparsity.
sparse autoencoderperturbation energymagnitude pruningactivation geometrylayer-wise sparsity
Continually learning neural-operator surrogate for three-dimensional airborne electromagnetic Bayesian inversion
The authors present a continually learning neural-operator surrogate for 3D airborne electromagnetic (AEM) Bayesian inversion, addressing computational bottlenecks in forward solves. The method leverages Maxwell's law invariance and employs continual learning across geological priors, using ensemble disagreement for validity checks. Results show the surrogate-driven Markov chain Monte Carlo sampler achieves credible intervals within 2.6 percentage points of full-solver accuracy, enabling rapid inversion of 2M+ soundings in the 2013 Capricorn TEMPEST survey.
neural-operatorbayesian inversioncontinual learningmarkov chain monte carloelectromagnetic
Controlling for Omitted Variable Bias in Deep Neural Networks
The paper introduces a control variable approach for deep neural networks to address omitted variable bias, specifically shortcut learning where models encode image-inferable covariates correlated with outcomes. The method employs generalised additive modelling with cross-fitting and ridge penalisation to orthogonalise covariate effects and marginalise predictions over covariate distributions. Evaluated on simulated images and real neuroimaging data, it achieves unbiased, interpretable predictions and recovers performance near unconfounded training levels, outperforming existing methods.
omitted variable biasshortcut learninggeneralised additive modellingcross-fittingridge penalisation
Classical and Hybrid Quantum Machine Learning for Trigger-Like Event Selection on CMS Open Data: An Eight-Qubit, PCA-Constrained Benchmark
The study benchmarks classical and hybrid quantum machine learning models for trigger-like event selection in high-energy physics using CMS open data. Four classical models (SVM, ANN, CNN, LSTM) and four quantum counterparts (QNN, QCNN, quantum-kernel, recurrent quantum) were evaluated on a binary classification task with invariant-mass window labels and physics-motivated input features. Under an eight-qubit constraint and PCA compression to 16 features, the classical ANN achieved 93.53% accuracy (0.9819 ROC-AUC), while the best quantum model (QCNN) reached 90.89% accuracy (0.9731 ROC-AUC), demonstrating competitive but not superior quantum performance.
quantum machine learningevent triggeringprincipal component analysisinvariant-mass windowroc-auc
SAMpLE: A SystemC-AMS Machine LEarning-based Framework for Virtual Prototyping
The paper introduces SAMpLE, an open-source SystemC-AMS framework for integrating ML models as Timed Dataflow components via standardized interfaces. The framework supports two execution backends: native C++ for online training of lightweight models and an offline backend for externally developed models using ONNX as an exchange format. SAMpLE enables evaluation of ML solutions within unified testbenches, promoting reproducibility and extensibility without modifying SystemC-AMS structure.
systemc-amstimed dataflowonnxvirtual prototypingmachine learning
Quantum-Inspired Modeling of Driving Behavior
The paper introduces a quantum-inspired density matrix representation for modeling heterogeneous, context-dependent driver behavior that evolves over time. The method encodes drivers as probabilistic density matrices that capture behavioral uncertainty, temporal dynamics, and variable interactions learned unsupervised from the I-24 MOTION dataset. Results show three interpretable driving profiles (free flow, transition, congestion) that align with macroscopic phenomena like the fundamental diagram and hysteresis loops, while enabling practical applications like context-aware car-following models and short-term motion forecasting for autonomous vehicles.
density matrixbehavioral uncertaintycontext-dependentfundamental diagramhysteresis loops
MetaSieve: Faster Relational Deep Learning through SQL-Based Metapath Selection
MetaSieve introduces a SQL-based metapath selection layer for Relational Deep Learning (RDL) to reduce subgraph sampling costs in GNN training. The method prunes uninformative metapaths by evaluating candidate extensions via database statistics and a novel scoring function, operating independently of GNN parameters. Evaluated on RelBench with multiple GNN backbones, MetaSieve reduces per-epoch training time significantly while maintaining or improving accuracy.
relational deep learningmetapath selectiongraph neural networksql joinsubgraph sampling
Forecasting Multiple Observables with SCROLL: Score-Trained Uncertainty for Stochastic Dynamics
SCROLL introduces a novel approach for forecasting multiple observables in stochastic dynamical systems by composing likelihoods via per-task free-routed last-layer beliefs on a shared backbone. This method absorbs unit-dependent loss scaling into likelihood parameters learned in a single gradient pass, leveraging stochastic dynamics to provide computable ground truth for predictive variance. Evaluated on the Ornstein-Uhlenbeck process, stochastic Lorenz-63, and real air-quality data, SCROLL achieves best single-run NLL on state and regime tasks, maintains calibration, and outperforms tuned grids at reduced cost. The state margin remains consistent across five rolling origins in real-world series.
stochastic dynamicspredictive variancelikelihood parametersornstein-uhlenbeck processnll
A General-Purpose Molecular Foundation Model Transfers Across Diverse Olfactory Tasks
A molecular foundation model fine-tuned on the GS-LF benchmark for odor descriptor prediction demonstrates cross-task transferability in machine olfaction. Using Uni-Mol2, the model achieves state-of-the-art performance on GS-LF and generalizes to four downstream tasks—cross-dataset prediction, odorous/odorless classification, enantiomer evaluation, and mixture discriminability—without additional training. The model's 3D representations outperform 2D graph models in distinguishing enantiomers, though perceptual stereochemistry prediction remains challenging. Results indicate that chemically pretrained representations enable a train-once, transfer-many paradigm for olfactory tasks.
molecular foundation modelmachine olfactionodor descriptor predictionenantiomer evaluationtransfer learning
Scalable Multi-GPU Simulation of 3D Multicellular Growth with RNN-Based Workload Balancing
A scalable multi-GPU framework for 3D multicellular growth simulation is introduced, combining GPU acceleration, spatial binning, domain decomposition, and an RNN-based load-balancing controller. The controller, trained offline in a differentiable surrogate environment, learns residual corrections to a reactive boundary-adjustment rule by observing per-rank execution times and partition states. Evaluations demonstrate GPU acceleration with spatial binning improves interaction computation by three orders of magnitude over serial CPU baselines. RNN-guided load balancing reduces mean global imbalance from 11.3% to 3.5%, lowers end-to-end runtime by 9.0%, and decreases slice migration by 7.7x compared to reactive baselines, enhancing workload balance while minimizing repartitioning.
multicellular growth simulationgpu accelerationspatial binningdomain decompositionrnn-based load-balancing
Efficient Estimation of High Information Projections using Nearest Neighbours
The authors propose an efficient method for dimensionality reduction by enhancing nearest neighbor relationships in multivariate data. The technique constructs a local covariance matrix from nearest neighbor pairs, whose spectral decomposition estimates the Density Information Matrix (DIM)—a non-parametric analog of the Fisher Information Matrix. Under standard regularity conditions, the estimator is consistent and computationally efficient compared to existing DIM estimators, which rely on surrogate densities. The method demonstrates practical utility in cluster analysis and outlier detection, linking to Independent Components Analysis and Sufficient Dimension Reduction in supervised settings.
dimensionality reductionnearest neighborsdensity information matrixspectral decompositionlocal covariance
How Edge of Stability Hinders SCAFFOLD in Federated Optimization
This work identifies Edge of Stability (EoS) dynamics as a key factor hindering SCAFFOLD's performance in federated optimization. Through empirical analysis across various architectures and hyperparameters, the authors demonstrate that EoS-like dynamics occur in both FedAvg and SCAFFOLD, with equilibrium sharpness inversely proportional to learning rate and influenced by data heterogeneity. Crucially, they show that SCAFFOLD's global gradient estimation capability degrades significantly at EoS, as evidenced by the correlation between sharpness and gradient estimation error. These findings explain SCAFFOLD's underperformance compared to FedAvg in deep learning contexts.
edge of stabilityfederated optimizationscaffoldgradient estimationsharpness
CEDAR: Controlled and Event-Driven Demand Forecasting via Residual Decomposition
The paper introduces CEDAR, a two-stage framework for decision-conditioned demand forecasting in e-commerce. Stage I employs an Action-Interleaved Transformer to model controllable state transitions under planned interventions, while Stage II uses a Residual Correction Module with event signals and LLM-assisted text representations to correct event-driven deviations. Evaluated on Alibaba 1688's dataset of 32M product trajectories, CEDAR outperforms traditional time-series forecasting methods in simulation accuracy and practical budget planning.
decision-conditioned forecastingaction-interleaved transformerresidual correction moduleevent-driven deviationsllm-assisted text representations
Precipitation Downscaling Using Foundation Model-Conditioned Diffusion
This work evaluates three conditioning strategies for denoising diffusion probabilistic models in precipitation downscaling: channel concatenation, cross-attention with a learned convolutional encoder, and cross-attention with the frozen Prithvi WxC weather foundation model. Experiments on the Colorado River Basin dataset show that concatenation achieves the lowest CRPS and MSE but produces over-smoothed fields, while cross-attention improves distributional realism and spectral fidelity, particularly for extreme events (>100mm/day). The Prithvi WxC-conditioned model retains over 50% of extreme events and matches learned encoder performance with only five years of training data, demonstrating the utility of foundation model representations in data-limited settings.
denoising diffusioncross-attentionprecipitation downscalingfoundation modelextreme events
Key Point Analysis Needs Structure Recovery: Task Definition, Dataset Diagnosis, and a Structure-Aware Benchmark
The paper redefines Key Point Analysis (KPA) as a structured prediction problem requiring semantic grouping recovery, key point generation, coverage assurance, and prevalence estimation. Through human-in-the-loop re-annotation, the authors construct a structure-aware benchmark addressing limitations in existing datasets (grouping quality, redundancy, coverage, and argument-key point mappings). Evaluations demonstrate improved coherence (human/LLM-judged), key point quality, coverage, and prevalence estimation reliability, accompanied by released annotation resources and a research agenda for true KPA.
key point analysisstructured predictionsemantic groupingprevalence estimationhuman-in-the-loop
FlowMoDL: Model-Based Deep Learning with Conjugate-Gradient Data Consistency for Highly Accelerated 4D Flow MRI Reconstruction
FlowMoDL introduces an unrolled neural network for highly accelerated 4D flow MRI reconstruction, optimizing both anatomical magnitude and phase-derived velocity accuracy. The method alternates a (3+1)D spatiotemporal denoiser with conjugate-gradient data-consistency updates, using a dual-pathway conditioning scheme to handle varying acceleration factors (10× to 50×). Evaluated on the CMRx4DFlow dataset, FlowMoDL outperforms CG-SENSE, MoDL, FlowVN, and FlowMRI-Net in magnitude SSIM, nRMSE, relative velocity error, and angular error, demonstrating robust convergence and superior gradient step efficiency.
unrolled neural network4d flow mriconjugate-gradientspatiotemporal denoiserdata-consistency
Learning Continuous Regional Temperature Fields with Lead-Time and Resolution Queries
The paper introduces Continuous Spatiotemporal Temperature Forecaster (CSTF), a neural field model for query-conditioned regional temperature forecasting that overcomes fixed-output constraints of existing methods. CSTF encodes multivariable ERA5 histories into latent states and decodes 2-m temperature (T2M) as a continuous field using spatial location, forecast lead time, and output resolution as explicit queries, supported by spatial-gradient, temporal-difference, and scale-consistency regularization. On the Southeast China 0-6 h ERA5-Land benchmark, CSTF reduces Bias by 17.0% and enables flexible lead-time and resolution-controllable inference.
neural fieldt2m forecastingera5-landspatiotemporal modelingquery-conditioned
Canalization Before Generalization: Grokking as a Dynamical Probe
The study investigates grokking in overparameterized neural networks as a dynamical probe for solution selection during training. By applying fixed-duration weight-decay (WD) pulses across the grokking plateau, the authors measure how these perturbations shift generalization timing. Results across three tasks reveal that WD effects transition from unordered to dose-ordered sensitivity before visible generalization, while test-loss barriers collapse. This phenomenon, termed 'canalization of function selection', demonstrates constrained solution selection precedes generalization.
grokkingoverparameterizationweight-decaygeneralizationcanalization
Geometry-Constrained Kolmogorov-Arnold Networks: Learning Edge Geometry via Banach Duality
The paper introduces geometry-constrained Kolmogorov-Arnold Networks (KANs), which learn edge activation functions via Banach duality maps parameterized by a scalar exponent $p > 1$ per edge, enabling adaptive function-space geometry. This approach outperforms fixed-basis KAN variants (splines, polynomials, Fourier features) on 50 symbolic-regression tasks, achieving median NRMSE of 0.030 and superior noise robustness (3.7× degradation vs. 21.6× for unregularized splines). Learned exponents provide interpretable geometric signatures, with sub-Euclidean values ($p < 2$) yielding threshold-like behavior and larger values producing flatter responses.
kolmogorov-arnold networksbanach dualitysymbolic regressionedge geometryfunction-space learning
TacForcing: Streaming Action Generation with Execution-Time Tactile Feedback
The paper introduces TacForcing, a streaming action-generation framework for contact-rich manipulation that incorporates real-time tactile feedback during execution. The method replaces standard action experts with a streaming variant and proposes Execution-Aware Tactile Attention (EATA) to align tactile conditioning with imminent actions, reducing temporal mismatch. Evaluated on six simulated UniVTAC tasks and three real-world manipulation tasks, TacForcing achieves average success rates of 65% and 69%, respectively, outperforming baselines.
tactile feedbackstreaming action-generationcontact-rich manipulationexecution-aware attentionreactive control
Cooperative Multi-Agent Reinforcement Learning for Adaptive Aggregation in Semi-Supervised Federated Learning with non-IID Data
The paper introduces pFedMARL, a Multi-Agent Reinforcement Learning (MARL) framework for adaptive aggregation in Federated Learning (FL) with non-IID data. The method employs server-side and client-side agents using Twin Delayed Deep Deterministic Policy Gradient (TD3) to dynamically adjust aggregation strategies and balance global-local updates. Evaluated on a semi-supervised audio spectrogram transformer, pFedMARL outperforms FedAvg, Ditto, and local training in non-IID scenarios and adversarial settings, demonstrating improved accuracy, robustness, and fairness.
federated learningmulti-agent reinforcement learningnon-iid dataadaptive aggregationtd3
EXAONE Tabular 1.0 : Technical Report
EXAONE Tabular 1.0 introduces a compact tabular foundation model family for classification and regression via in-context learning, eliminating dataset-specific gradient updates. The architecture redesign interleaves feature-axis attention with support-conditioned item-axis attention at each Transformer layer, mediated by item-summary and feature-summary tokens. Pretrained on synthetic SCM data, the 20.81M-parameter model achieves state-of-the-art performance: ranking first overall on TabArena classification, matching TabFM regression at 1/11 inference cost, and leading in ScoringBench metrics ($R^2$, RMSE, CRPS).
tabular foundation modelin-context learningstructural-causal-modeltransformer architectureprobabilistic regression
Drift-Aware Multimodal User Representation Learning via Multi-Scale Temporal Modeling and Sparse Mixture-of-Experts
The paper proposes DUMoE, a drift-aware multimodal user representation learning framework addressing temporal preference shifts in social media behavior. The model combines (i) a temporal dynamics-aware backbone integrating static profiles, short-term signals, and long-term dependencies with (ii) a sparse mixture-of-experts (MoE) interest adapter that disentangles multiple latent interests via specialized experts and adaptive routing. A three-stage training strategy decouples backbone learning, expert specialization, and gating optimization. Experiments on real-world datasets demonstrate superior performance on user interest and interaction prediction tasks compared to state-of-the-art methods.
user representation learningtemporal dynamicssparse mixture-of-expertsinterest driftadaptive routing
Large Language Model Few-Shot Prompting with Dilemma Training Outperforms Human Surrogates in Predicting Patient Preferences
P4-DT (Dilemma Training), a context-aware AI agent for predicting patient preferences in serious illness, outperforms human surrogates by employing bi-directional training with medical dilemmas to construct individualized decision policies. The method integrates contextual scenario decisions and open-ended text, improving prediction accuracy over static value ratings. In a study with 12 patient-surrogate dyads, P4-DT achieved 81.7% accuracy (OR = 5.61, p < .001), significantly surpassing unassisted surrogates (55.0%) and surrogates assisted by P4-DT (61.7%). Contextual decisions and text improved accuracy by 15.0 percentage points.
patient preference predictiondilemma trainingcontext-aware aibi-directional trainingmedical decision-making
Learning from waste: Machine Learning for health risk prediction and computer vision-based sorting in Ghana
The study presents two machine learning approaches for waste-related health risks in Ghana: (1) a Random Forest classifier predicting illness categories from waste disposal practices (macro F1=0.63, N=69), identifying disposal method as the top predictor, and (2) a MobileNetV2-based computer vision system for automated waste sorting (88.2% accuracy, macro F1=0.87, N=415). The vision system demonstrates feasibility for low-resource settings using simple camera inputs. Results quantitatively validate community-reported health-waste linkages while highlighting the necessity of institutional support alongside technical solutions for public health impact.
random forestmobilenetv2waste sortinghealth risk predictionlow-resource ml
LM-X: Explainable Action Modeling with Progress, Event, and Uncertainty Prediction for Generalist Robot Manipulation
LM-X introduces an explainable vision-language-action (VLA) policy for generalist robot manipulation, addressing bottlenecks in task progress, intermediate intent, and local reliability prediction. The model explicitly supervises three signals during execution: return-to-go (RTG) for task progress, event-to-go (ETG) for semantic transitions, and heteroscedastic action flow for local reliability estimation. Pretrained on 64 NVIDIA B200 GPUs and over 20,000 hours of real-robot trajectories, LM-X achieves 74.1% success on 50 RoboTwin2.0 tasks, outperforming GR00T N1.7 by 18.7 points. RTG and variance metrics provide interpretable internal estimates, demonstrating that multi-timescale predictive states enhance control transparency and performance.
vision-language-action policyreturn-to-goevent-to-goheteroscedastic action flowrobottwin2.0
Comparing Corrupted Constrained Learning Problems
The article demonstrates that the classical data processing inequality fails for constrained learning problems, where model classes exclude all measurable functions. It introduces a generalized data processing inequality requiring the constrained Bayes risk of a joint distribution to lower bound that of a stochastically modified distribution, independent of the distribution choice. The authors prove this inequality is equivalent to a set containment condition on the superprediction set induced by the loss and model class, and derive sufficient conditions for this containment.
data processing inequalityconstrained learningbayes risksuperprediction setstatistical uncertainty
A Constitutive Markov Physics-Informed Neural Operator (MPNO) for Autoregressive Stability in Transient Dynamics
The paper introduces a constitutive Markov physics-informed neural operator (MPNO) to address autoregressive instability in transient-dynamics PDEs with discontinuities. MPNO enforces stability by design through a Markov propagation operator with spectral radius ≤1, derived from a physics-coupled graph Laplacian. Evaluated on Burgers' equation and concrete-penetration scenarios, MPNO achieves stable rollouts with 0.7304±0.0008 relative L2 error, outperforming wavelet neural operators and matching Fourier neural operators at 25% parameter count, while offering 10^5x speedup over LS-DYNA.
neural operatorautoregressive stabilityspectral radiusgraph laplaciantransient dynamics
Why Does Graph Learning Fail to Fully Benefit from a Text Teacher?
The study investigates why multimodal graph-text models fail to achieve expected performance gains, identifying six key factors. The proposed method combines self-supervised GNN pretraining (transferable across datasets with varying node-feature dimensions) and alternating optimization (E-step: language model update, M-step: GNN update) instead of joint training. Experimental results reveal trade-offs in anchor strength, misaligned representation spaces, propagation effects, suboptimal geometric alignment, and conflicting optimization forces between self-supervised and teacher-guided objectives.
graph neural networksmultimodal learningself-supervised pretrainingalternating optimizationrepresentation alignment
Are LLM-Enhanced GNNs Privacy-Safe?
This work systematically evaluates privacy risks in LLM-enhanced GNNs through a unified framework comprising dataset preparation, victim model training, privacy attacks, risk assessment, and defense analysis. Experiments on six text-attributed graph datasets with 42 model configurations reveal that LLM-enhanced GNNs exhibit increased vulnerability to link, label, and membership inference attacks compared to shallow baselines, due to semantic enrichment amplifying sensitive signals. Differential privacy partially mitigates risks but incurs significant utility degradation, highlighting a privacy-utility trade-off.
llm-enhanced gnnsprivacy attacksmembership inferencedifferential privacysemantic enrichment
Multi-output Gaussian process prediction of physical fields under linear equality constraints
The authors propose a multi-output Gaussian process (GP) framework for predicting high-dimensional physical fields under linear equality constraints, addressing limitations of deductive approaches that introduce arbitrary output dependencies. Their method combines row-wise principal component analysis (PCA) to preserve constraints in latent space with a constrained multi-output GP using specialized kernel parametrization. Experiments on population dynamics and computational fluid dynamics (CFD) applications demonstrate improved predictive accuracy and uncertainty quantification compared to standard approaches.
gaussian process regressionmulti-field predictionlinear constraintsrow-wise pcauncertainty quantification
Fairness-Aware Test-Time Prompt Tuning
The paper introduces FairTPT, a fairness-aware test-time prompt tuning method for vision-language models that addresses bias under distributional shift without retraining. The approach jointly minimizes target marginal entropy while maximizing spurious marginal entropy via soft-prompt tuning during episodic test-time adaptation (TTA). Experiments on CLIP classification show standard TTA exacerbates demographic disparities, while FairTPT prevents catastrophic forgetting through linear regime monitoring, outperforming existing debiasing methods in fairness improvements while preserving accuracy.
test-time adaptationsoft-prompt tuningdistributional shiftdemographic biasvision-language models
Fast rates in Bayesian online learning with approximate posteriors
The paper establishes conditions under which approximate Bayesian online learning methods preserve the fast regret guarantees of exact Bayes prediction. By analyzing the interaction between Gibbs posterior contraction radius and Wasserstein approximation error, the authors prove that sufficiently accurate approximate posteriors inherit the original regret bounds. Three applications demonstrate this: (1) projected Langevin algorithms for linear models achieve logarithmic regret, (2) truncated exponential family models attain minimax rates with sublinear memory, and (3) sparse variational GPs match exact GP regret at lower cost.
bayesian online learningpredictive regretwasserstein distancegibbs posteriorvariational approximation
Tropospheric temperature and humidity profile retrieval from Meteosat Flexible Combined Imager based on deep learning
The authors propose a spatially aware deep learning framework for retrieving tropospheric temperature and humidity profiles from Meteosat Third Generation's Flexible Combined Imager (FCI) without relying on numerical weather prediction inputs. They train a Residual U-Net on 16-channel FCI observations and CERRA reanalysis targets, demonstrating that spatial context improves retrievals, particularly under clouds. Validation against radiosondes shows temperature biases below 0.4 K (1.5-1.9 K std) and relative humidity std deviations of 12-20%, with visible/near-infrared channels contributing despite their atypical use in physics-based inversions.
tropospheric profilingresidual u-netgeostationary imagerall-sky retrievalradiative transfer
Modeling spatio-temporal locality in multi-step forecasting of geo-referenced time series
The paper introduces SPALT, a method for multi-step forecasting of geo-referenced time series that captures spatio-temporal locality by modeling spatial autocorrelation among time series with similar trends. SPALT employs linear model trees to group time series exhibiting similar trends into nodes, selectively injecting spatial features, and proposes a novel pruning strategy based on Reduced Error Pruning that incorporates spatio-temporal locality. Evaluated on three real-world datasets of renewable energy production, SPALT outperforms both tree-based models and state-of-the-art neural networks incorporating temporal and spatial dimensions.
spatio-temporal localitylinear model treesspatial autocorrelationmulti-step forecastingreduced error pruning
TRACE: Retrospective Streaming Generation of Physical Fields under Sparse Structured Sensing
TRACE introduces a retrospective streaming generative reconstruction framework for physical fields under structured sensing, addressing limitations of batch-conditioned methods. The framework performs approximate Bayesian inference in a learned continuous-coordinate latent space, fusing sparse off-grid measurements with a state-space temporal prior via Kalman-style filtering and refining past frames through retrospective smoothing. Evaluations on active matter, ocean sound-speed fields, and supernova simulations demonstrate that TRACE outperforms frame-wise generative reconstructors, offline spatiotemporal methods, and streaming data-assimilation baselines in reconstruction quality under temporally sparse and spatially localized sensing conditions.
generative reconstructionbayesian inferencekalman filteringstate-space priorstructured sensing
📰 Industry Media (4)
Cohere Releases Parse 5 (parse-v5.0): A 2.3B Vision Language Model That Turns Enterprise Documents Into Markdown
Cohere introduces Parse 5 (parse-v5.0), a 2.3B-parameter vision-language model for enterprise document parsing, converting PDFs, PPTs, and JPEGs into structured Markdown with HTML tables and bounding box coordinates. Built on the North-Micro-Vision-Instruct architecture, it operates without separate OCR, supports 9 languages, and achieves a vendor-reported ParseBench score of 79.2 (averaging 3/5 dimensions). Priced at $1.50 per 1,000 pages via API, it targets document-heavy industries with applications in RAG ingestion and intelligent document processing.
vision-language modeldocument parsingmarkdown conversionnorth-micro-vision-instructparsebench
Best Agent Sandboxes in 2026: Cold Start, Per-Second Pricing, and Network Policy Across E2B, Daytona, Modal, Cloudflare, and Vercel
This systematic comparison evaluates agent sandbox platforms (E2B, Daytona, Modal, Cloudflare, Vercel) across four architectural dimensions: cold start latency under concurrency, filesystem persistence, egress policy granularity, and idle billing models. Using ComputeSDK's open-source sandbox leaderboard, it benchmarks Time to Interactive (TTI) across providers, revealing median TTI ranges from 0.27s (Daytona) to 5.06s (Cloudflare) under concurrent bursts. Pricing analysis normalizes per-second costs across CPU/memory billing models, showing Northflank as most economical ($1.67/1000 executions) for short bursts. Filesystem persistence mechanisms vary significantly, with E2B requiring explicit pause() to preserve state and Cloudflare maintaining ephemeral disks across sleeps.
cold starttime to interactivefilesystem persistenceegress policyidle billing
From In-Silico to Wet-Lab: Evaluating AI Protein Design Performance
This study evaluates AI-designed miniprotein binders using Anthropic’s claude-protein-binder-design dataset, which includes 1,440 designs tested against 16 targets. Computational predictions and wet-lab results from two independent labs were analyzed to assess structure predictors, ensemble scoring, and experimental hit rates. A target-aware classifier was trained to predict experimental success, revealing that target choice significantly influences hit rates. Results show AUC scores above chance but below optimal thresholds, with ensemble methods outperforming single predictors. The analysis highlights the practical utility of in-silico triage for wet-lab testing budgets.
miniprotein bindersensemble scoringhit ratestarget-aware classifierin-silico triage
Google Research Introduces GlucoFM: A 0.72M-Parameter Dual-Stream Foundation Model for Continuous Glucose Monitoring
Google Research and UNSW Sydney introduce GlucoFM, a 0.72M-parameter dual-stream foundation model for continuous glucose monitoring (CGM). The model decomposes CGM signals into a slow physiological state stream and a transient event stream, using a causal Gaussian filter and JEPA-style latent objectives (masked contextual prediction and next-patch dynamics). Pretrained on 109,066 hours of unlabeled CGM data from 477 subjects, GlucoFM achieves 58.8 task-averaged PR-AUC across 14 cohort–task evaluations, outperforming baselines by +4.1 points (7.5% relative). It excels in diabetes-risk and beta-cell-dysfunction tasks, with 21.88 mg/dL MAE for postprandial forecasting. The model is research-only, with no regulatory clearance or public checkpoint yet.
continuous glucose monitoringdual-stream architecturejepa-style objectivesmasked contextual predictionpr-auc
Generated automatically at 2026-08-28 02:59 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.