Daily Digest — 2026-09-03
334 items · 4 research labs, 321 arxiv papers, 9 industry media
🏛️ Research Labs (4)
How law firm Gilbert + Tobin governs and scales AI with OpenAI
Gilbert + Tobin implemented OpenAI's ChatGPT Enterprise and Codex to automate legal operations while maintaining governance. The firm employed CEO-led adoption strategies, role-specific training, and enterprise-grade controls to achieve 87% active usage among enabled users—more than double typical tool adoption rates. Results include 12x speedup in recruitment workflows (4h→20m), 36-96x acceleration in compliance checks (3-8h→5m), and full-day savings in audit report preparation across 300 entities through Codex automation.
chatgpt enterprisecodexenterprise-grade controlsworkflow automationdata residency
Real-Time Intelligence with IBM Time Series Models on Confluent
IBM introduces Granite Time Series Foundation Models (TSFMs) for real-time forecasting, anomaly detection, and optimization, deployed via Confluent Cloud with Apache Flink integration. These models—PatchTST-FM, FlowState, TTM, and TSPulse—address diverse time-series tasks (e.g., multi-horizon forecasting, anomaly detection) with minimal configuration, leveraging patch-based attention, continuous-time dynamics, and hybrid time-frequency architectures. Evaluated in industrial settings (cement, telecom), they achieve 5–10× productivity gains and millisecond-latency inference on CPU. Confluent enables stateful streaming with zero-config model serving, Kafka-backed governance, and SQL-native calls (AI_FORECAST, AI_DETECT_ANOMALIES), reducing operational overhead by eliminating separate ML infrastructure.
time-series foundation modelspatch-based attentionflink sqlanomaly detectionstreaming inference
BenchMIRT: What are LLM benchmarks actually measuring?
BenchMIRT introduces a multidimensional Item Response Theory (IRT) approach to audit LLM benchmarks at the prompt level, disentangling mixed signals in benchmark scores. It analyzes 100 LLMs across 16 benchmarks and 34K questions, estimating model capabilities and question difficulty independently. BenchMIRT identified two dominant dimensions: safety and general reasoning, revealing nuanced associations in benchmarks like BBQ and WMDP. It preserves benchmark insights with 10-50% of questions and predicts model performance on unseen questions with 79% accuracy. This method enhances benchmark transparency and efficiency, though it risks misuse in trimming evaluations.
item response theorymultidimensional irtbenchmark auditingllm evaluationquestion difficulty
Proactive cyber defense for governments and enterprises
Google introduces the Fairwind Program, leveraging Gemini 3.8 Flash Cyber and CodeMender to enable autonomous vulnerability detection and remediation at scale. The program integrates advanced AI models with specialized reasoning capabilities to generate verified, deployment-ready patches within minutes, reducing traditional remediation timelines from weeks to minutes. Initial access is prioritized for government agencies, critical infrastructure operators, and core technology platforms, with strict operational standards to ensure responsible use. Over 650 global partners are participating, benefiting from Google’s zero-trust architecture and AI-driven defenses. The initiative aims to enhance cyber resilience, supported by $100 million in global cybersecurity funding.
gemini 3.8 flash cybercodemenderzero-trust architecturevulnerability remediationagentic scale
📜 arXiv Papers (321)
Efficient SWE Agent Benchmarking via Trajectory-Aware Evaluation
The paper introduces PTA-IRT, a Privileged Trajectory-Aware Item Response Theory framework for efficient benchmarking of software engineering agents. Unlike result-only methods, PTA-IRT leverages historical execution trajectories—including explored context, attempted edits, and solving paths—as privileged information for calibration subset selection and ability estimation. Evaluated on four SWE benchmarks, PTA-IRT outperforms prior IRT baselines in score and ranking recovery under low calibration budgets.
software engineering agentsitem response theoryexecution trajectoriescalibration subset selectionbenchmark evaluation
Adaptive Critical Token-Aware Retrieval for Repository-Level Code Generation
ACToR introduces an adaptive critical token-aware retrieval framework for repository-level code generation, addressing the challenge of synthesizing code consistent with repository context beyond LLM input limits. The method identifies critical tokens during autoregressive generation, triggering targeted retrieval to provide fine-grained repository context at decisive positions, and employs a position-aware weighting method for dense retrievers. Evaluated on RepoExec and CoderEval benchmarks, ACToR achieves relative improvements of 8.4% and 15.4%, respectively, over state-of-the-art methods. The study systematically quantifies the impact of critical tokens, emphasizing their role in generation failures and the necessity of targeted retrieval strategies.
repository-level code generationcritical tokensadaptive retrievaldense retrieversautoregressive generation
CordisBench: Can Language Models Reason About Component Lifecycles in Dynamic Agent Harnesses?
We introduce CordisBench, a 1,200-question benchmark evaluating language models' ability to reason about component lifecycles in dynamic agent harnesses. The benchmark combines formal reasoning with execution against Cordis, a runtime managing component dependencies and cleanup, across tasks like identifying affected components and predicting state after teardown. Three efficiency-oriented models were evaluated at varying reasoning effort levels (2-32 interactions). Results show models perform well on small systems but reliability decreases with more interactions, particularly in final state prediction and teardown order reasoning. Increased inference effort yields gains, though at high cost (e.g., GPT-5.6 Luna uses ~3,000 tokens/question).
dynamic agent harnessescomponent lifecycleteardown orderfinite reference semanticsreasoning effort
The Rise of Verbal Reinforcement Learning
The paper introduces Verbal Reinforcement Learning (VRL) as a unified paradigm for improving language agents through natural language feedback. It organizes VRL into three pillars based on when and how verbal feedback modifies agent behavior: (1) Language as Grounding Signal, where language defines tasks by specifying goals, states, and rewards; (2) Language as Deliberative Feedback, guiding reasoning at test time without parameter updates; (3) Language as Learning Signal, shaping model parameters through training. The authors synthesize representative work, distinguish subcategories, and outline language's role in each pillar, demonstrating how VRL reshapes agent development and identifies challenges for building more capable and aligned agents.
verbal reinforcement learninglanguage agentsgrounding signaldeliberative feedbacklearning signal
Mechanism Design for Alignment and Control
The authors propose a mechanism design framework for AI agents with unknown alignment (preferences) and capabilities (feasible actions, information). The framework incentivizes honesty and obedience through a one-sided imitation structure, where capabilities can be concealed but not counterfeited. This yields a revelation principle, characterizes implementable policies via nested cyclical monotonicity, and identifies conditions where eliciting higher-order beliefs disciplines multiple agents. Applications include sandbagging, alignment-interpretability trade-offs, peer scoring, competitive reward coupling, and scalable oversight. The framework addresses challenges in controlling AI systems with uncertain properties.
mechanism designalignmentcapabilitiesrevelation principlecyclical monotonicity
Designing Proactive Thought Partners for Writing
The paper introduces proactive thought partners, AI agents providing customizable cognitive support during writing, addressing limitations of generic textual assistance. A technology probe was developed, enabling users to configure partner roles and proactivity, deployed with 16 participants over one week. Findings indicate users leveraged proactive support for prospective planning, idea generation, and self-monitoring, while valuing lightweight visual representations and non-directive rhetorical framing for non-intrusive interventions. Design implications emphasize customization, timing, engagement, and representation for proactive writing assistants.
proactive thought partnerscognitive supporttechnology probenon-directive framingprospective planning
Scaling Near-Optimal SFT-RL Annotation Budget Allocation from Small to Large LLMs
We introduce a framework for near-optimal budget allocation between supervised fine-tuning (SFT) and reinforcement learning (RL) in LLM post-training, addressing the lack of principled methods and transferability across model sizes. By characterizing the near-optimal region—allocations within a specified tolerance of peak performance—we empirically demonstrate its width (2-10% tolerance), scalability with model size, and transferability from small proxy models to large target models. This eliminates exhaustive large-scale search, validated across tasks, model families, and RL methods (preference-based off-policy and reward-supervision on-policy). We also analyze how SFT-RL annotation cost asymmetry shifts the near-optimal region.
supervised fine-tuningreinforcement learningannotation budgetnear-optimal regiontransferability
Selective Agent Guidance via Entropy: Learning Autonomous Policies from Imperfect VLM Teachers
SAGE (Selective Agent Guidance via Entropy) introduces a framework for learning autonomous policies from imperfect Vision-Language Model (VLM) teachers by querying VLMs selectively based on learner uncertainty and distilling guidance into lightweight Reinforcement Learning policies. SAGE weights VLM advice using environment-derived advantages rather than treating all suggestions equally, reducing VLM usage by prompting only on a fraction of training steps and requiring no VLM calls at deployment. Evaluated on sparse-reward visual reasoning and navigation tasks, SAGE outperforms unguided RL and, in some cases, exceeds its VLM teacher, demonstrating that selective guidance is most beneficial when VLMs help discover high-reward trajectories.
selective guidancevision-language modelsreinforcement learningentropypolicy distillation
From Confusion to Clarity: Confusion-Aware Retrieval and Knowledge Injection for Text Classification
The paper introduces a confusion-aware framework for improving text classification in large label spaces, where LLMs struggle with semantically similar labels. The method (1) identifies confusable label pairs, (2) expands candidate sets to include such pairs, and (3) generates targeted rules to differentiate them, all without fine-tuning. Evaluated on WOS, Flipkart, and LEDGAR benchmarks, the approach improves Macro F1 by up to 10.0pp over retrieval baselines, with smaller models (2B--20B) gaining up to 11.5pp via rule transfer.
text classificationlarge label spacesconfusion-aware retrievalknowledge injectionrule transfer
H3-World: Turning Language Understanding into World Control
H3-World introduces an efficient framework for transforming the 33B MiniMax-H3 video generator into an interactive world model using natural language as a control interface. The method leverages structured combinations of character and camera instructions, aligned with temporal video latents, and employs temporal attention routing to ensure precise temporal control. With minimal adaptation—8,000 gameplay samples, 10,000 LoRA optimization steps, and 0.199% trainable parameters—H3-World achieves effective character and camera control while maintaining generation quality and generalizing to unseen scenarios. This demonstrates the potential of large video generators for interactive world control.
video generatortemporal attention routinglora optimizationworld modelnatural-language control
Retrieved but not ranked: surface-form bias in structural retrieval, from mathematics to agent trajectories
The study investigates embedding retrieval performance when surface form and semantic structure are deliberately decoupled, evaluating two domains: competition mathematics (MathNet-Retrieve) and embodied-agent trajectories (ALFWorld-derived). Results reveal complete failure in mathematics retrieval, with Hit@1 at 0.0% for heavily disguised queries, while trajectory retrieval performs near hypergeometric chance when gold answers differ in object and receptacle. Lexical rerankers worsen mathematics retrieval but improve trajectory retrieval, closing 26-36% of the gap. LLM rerankers recover 5-63% of the gap in mathematics and 43-76% in trajectories, with domain-dependent effect sizes. Downstream experiments show oracle retrieval indistinguishable from adversarial retrieval, with solver accuracy primarily reflecting truncation effects.
embedding retrievalsurface-form biashypergeometric chancelexical rerankerllm reranker
BS: Take the Hint - Interactive Multitracer PET/CT Lesion Segmentation with a Scribble-Conditioned ResEnc U-Net
The authors propose a scribble-conditioned residual encoder U-Net for interactive lesion segmentation in PET/CT scans, addressing variability in tracer uptake and lesion appearance. The model extends the autoPET-III winning architecture to four input channels (CT, PET, foreground scribble, background scribble), initialized from pretrained weights with scribble channels zero-initialized. PET intensities are normalized using a CT-derived aorta blood-pool reference. Five-fold ensembling with Gaussian-weighted stitching achieves mean Dice scores of 0.554 without scribbles and 0.751 after five correction rounds, with 85% of the gain occurring after the first scribble.
lesion segmentationscribble-conditionedresidual encoderpet/ctgaussian-weighted stitching
Can LLMs Discover Scientific Laws in Real and Parallel Worlds?
We introduce SCILAWS-BENCH, a benchmark for evaluating LLMs' ability to discover scientific laws, comprising 118 problems from 381 papers across six disciplines with 8M real data points. The benchmark features two settings: SCILAWS-REAL evaluates predictive fit and scientific validity from fixed observations, while SCILAWS-PARALLEL assesses active recovery of synthesized hidden laws. Results reveal divergence between predictive fit and scientific validity, memorization effects on formula reproduction, and a selection bottleneck in best-of-N studies. This work provides a paper-grounded benchmark and empirical insights for AI-driven scientific discovery.
scientific law discoverypredictive fitscientific validityactive recoveryselection bottleneck
A Mathematical Theory of Reusable Neural Bases for Network Compression
The Linear Reusable Neural Bases Architecture (LRNBA) is introduced to improve parameter efficiency and reduce memory costs in large AI models. Inspired by RNN designs, LRNBA represents each network block as a linear combination of shared neural bases, enabling significant network compression while maintaining stable training dynamics. This architecture allows for wider and deeper networks under the same parameter budget. Extensive experiments show that LRNBA achieves comparable or faster convergence and lower loss than classical architectures, demonstrating its effectiveness in reducing memory overhead without compromising performance.
linear reusable neural bases architectureparameter efficiencynetwork compressiontraining dynamicsmemory cost
Can LLMs Design Video Coding Tools? A Case Study on Planar Mode
This paper investigates the capability of large language models (LLMs) to design video coding tools, focusing on the Planar mode in intra prediction. The study employs a generation-and-evaluation loop where the LLM iteratively refines Planar predictors based on encoder trial feedback. Experiments conducted with the Fraunhofer Versatile Video Encoder (VVenC) and Enhanced Compression Model (ECM) demonstrate that LLM-generated modes can outperform conventional Planar modes, achieving 0.18% bitrate savings with 0.4% complexity overhead in VVenC and coding gains in ECM under low-resolution settings. The findings suggest potential for LLM-based coding tool design while highlighting remaining challenges.
large language modelsplanar modeintra predictionvideo codingbitrate savings
EvoSCM: Scientific Belief Revision Through Causal Model Evolution and Experimentation
EvoSCM introduces a framework for scientific agents that maintain explicit structural causal models (SCMs) to represent and revise beliefs through experimentation. The method employs a population of competing SCM hypotheses, evolving them via abduction, discriminative intervention design, and inductive correction rules based on empirical discrepancies. Evaluated on DiscoverPhysics, EvoSCM outperforms baselines in uncovering hidden physical dynamics, achieving higher accuracy in explanations and predictions while optimizing experimental interactions.
structural causal modelsscientific discoveryabductiondiscriminative interventionsinductive correction
Relational-Core Graph Analytics Querying graphs at SQL scale, and why the node/edge model is a performance tax, not a truer picture of connected data
The paper challenges the assumption that graph analytics necessitates purpose-built graph engines, demonstrating that columnar relational systems outperform native graph engines on analytical queries. The authors introduce ClickGraph and DeltaGraph, systems that translate Cypher queries directly onto native relational schemas (tables, columns, foreign keys) and execute them in-place on ClickHouse, Databricks, or lakehouse files without requiring data import or separate clusters. These systems leverage SQL output for query optimization and engine extensibility. Empirical results show that a columnar engine outperforms Neo4j by two-to-four orders of magnitude on the LDBC Social Network Benchmark suite.
columnar relational enginecypher query languageldbc social network benchmarkclickhousedelta graph
When Guardrails Look Effective: Construct Validity Failures in LLM Agent Commerce Evaluation
This study identifies construct validity failures in evaluating LLM-based agent commerce systems, focusing on welfare claims from marketplace guardrails. Using a multi-turn buyer-seller testbed for hotel transactions, the authors audit protocol isolation and incentive validity by controlling offer schemas and choice procedures. Results show welfare gains vary significantly (+87.4 to -13.9) when fixing these factors, with generation residuals explaining 49.9% of variation. A seller-incentive check reveals non-monotonic profit behavior, and scripted controls demonstrate guardrails primarily redistribute welfare rather than create it. The authors propose a construct-validity contract to assess incentive validity, protocol isolation, stochastic stability, and welfare accounting before policy claims.
construct validitymarketplace guardrailsprotocol isolationstochastic stabilitywelfare accounting
TempCloze: Can Video-LLMs Identify the Missing Middle?
We introduce TempCloze, a video cloze benchmark for evaluating visual temporal reasoning in Video-LLMs, designed to reduce linguistic shortcuts prevalent in existing benchmarks. TempCloze comprises 1,521 videos from seven sources, requiring models to identify the true missing middle clip from four candidates constructed along Semantic, Alignment, and Progression dimensions. Evaluation of 31 Video-LLMs reveals Alignment as the primary bottleneck, with models struggling to temporally align events despite recognizing semantic content and local progression. Further analyses on TempCloze-Mixed and TempCloze-Hard examine error patterns and behavioral sensitivity to factors like candidate order and frame density.
video-llmstemporal reasoningcloze benchmarkalignmentsemantic
LatentPress: Context Compression Beyond Text and Vision
LatentPress introduces a novel context compression method using continuous memory tokens, bypassing text reconstruction for direct decoder consumption. The approach employs a small writer (4.2M-26.2M parameters) to compress conversational histories and documents 4-16×, training only an adapter. Evaluated on LongMemEval and LongBench-QA, LatentPress achieves 0.504 accuracy at 7.70× compression, outperforming text summaries (0.184) and OCR-based methods (0.426-0.312). Writing latency is 43ms per conversation, 5-9× faster than raw context or cached OCR. Zero-shot transfer experiments validate the interface's efficacy across domains, establishing soft tokens as a practical machine-facing context representation.
context compressioncontinuous memory tokensdecoder interfacezero-shot transfersoft tokens
Optimizing Byzantine Node Placement in Decentralized Federated Learning
The paper introduces Byzantine Placement Influence (BPI), a set-level metric optimizing Byzantine node placement in decentralized federated learning (DFL) to maximize adversarial impact under a fixed compromise budget. BPI quantifies cumulative exposure of honest nodes to Byzantine sources, accounting for weighted multi-hop propagation and compromised node interactions, without requiring full training simulations. Efficient algorithms for BPI optimization are developed and evaluated across six graph families, untargeted model poisoning, and backdoor attacks. Results demonstrate BPI-guided placements consistently identify highly damaging configurations across diverse network structures, even with Byzantine-robust aggregation, highlighting the critical role of Byzantine placement in DFL threat modeling.
byzantine placement influencedecentralized federated learninggossip dynamicsmodel poisoningbackdoor attacks
Rethinking Learnability in Offline Data-driven Optimization
The paper introduces algorithm-dependent learnability for offline data-driven optimization, proving its sufficiency for discrete (greedy/local search in submodular maximization) and continuous (projected gradient descent for convex minimization) settings via value-query and first-order forms. It proposes a trajectory-learning framework with trajectory construction, modeling (using conditional diffusion), and candidate generation, instantiated as Uncertainty-aware Gradient-guided Trajectory Learning (UGTL). UGTL constructs locally coherent improvement trajectories, models them with conditional diffusion, and selects diverse candidates, achieving top aggregate mean rank (3.1/25) on Design-Bench tasks, with ablation confirming trajectory construction's key role.
offline optimizationalgorithm-dependent learnabilitytrajectory learningconditional diffusionsubmodular maximization
GlossoGen: Emergent Language in Complex Multi-Agent LLM Interactions
GlossoGen introduces a novel platform for studying emergent language in multi-agent LLM interactions, addressing key questions about language evolution and its implications for safety and linguistic accounts. The platform features the SaveVeyru scenario, where agents with partial information communicate under pressure. Results show that LLM agents develop compositional, morphologically productive languages deviating from English, becoming incomprehensible to humans. Essential factors include efficiency pressure, model strength, and postmortem linguistic convention agreement. Findings reveal that stronger models enable novel language emergence, while weaker models can learn existing languages, indicating LLMs' potential for cumulative cultural evolution.
multi-agent llmlanguage evolutioncompositional languagecultural evolutionmorphological productivity
Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents
The paper introduces Defense-as-Skill, a paradigm for runtime protection in skill-augmented agents by implementing guards as inspectable, editable skills. SkillSonar, their guard, monitors untrusted task skills and routes sensitive actions through allow/replan/confirmation decisions without runtime modifications. Using SCOPE-R (206 malicious instances across 21 sub-categories), they evolve SkillSonar via Monte-Carlo Tree Search, reducing Claude Code and OpenClaw attack success rates from 0.482→0.104 (ID) and 0.606→0.115 (OOD) while maintaining utility. Ablations highlight the importance of safety responsibility assignment and skill-native representations.
skill-augmented agentsruntime guardmonte-carlo tree searchattack success ratesafety-utility trade-off
Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement
Harness-of-Harness (HoH) introduces a framework for autonomous software development, enabling LLM-based coding agents to iteratively improve software through planning-coding-testing loops. HoH balances repair with capability growth, scopes development into verifiable increments, and maintains versioned project histories while progressively exposing deliverables and tools. Evaluated on GameCraft-Bench, FrontierSWE, and ProgramBench with harness-model pairs (Codex-GPT-5.5, OpenCode-DeepSeek-V4-Pro, Pi-MiniMax-M3), HoH achieves an average relative gain of 52.25% and a maximum gain of 82.86% after three iterations. In a multi-day deployment with over 70 iterations, HoH autonomously develops a fully functional first-person-shooter game with coherent mechanics, visuals, and audio.
autonomous software developmentllm-based coding agentsplanning-coding-testing loopscapability growthversioned project histories
Parsing the Stream: A Live Trace Model for Long-Horizon Agents and Their Observers
The paper introduces a live trace model for long-horizon agents, addressing the challenge of trace growth exceeding both human observer and agent context limits. The model employs an append-only event ledger incrementally folded into typed run state and compiled into per-consumer views. Evaluations show that for human observers, the compiled view reduces input tokens by 14x and costs by 5-7x while improving accuracy (0.85-0.87 vs. 0.48). For agents, the model outperforms full-context prompting on sequential-dependency tasks (30/30 vs. 8/30 success) and offers deterministic auditability. Eleven trace folding requirements are derived, supported by code, benchmarks, and synthetic corpora.
live trace modelappend-only event ledgertyped run statesequential-dependency tasksdeterministic auditability
When Safety Routing Breaks: Understanding Alignment Fragility under Benign Fine-Tuning
This work investigates the fragility of safety alignment in large language models (LLMs) under benign fine-tuning, proposing a Fisher-geometric explanation centered on low-rank safety Fisher matrices and output-routing pathways. The study reveals that after 100 benign fine-tuning examples, output-side MLP modules selectively re-sharpen, causing asymmetric safety collapse while preserving general utility. Experiments demonstrate that few safety examples can restore refusal behavior, indicating intact internal safety representations. Techniques like LoRA and ASAM mitigate early collapse by suppressing output-side sharpness, though their efficacy diminishes at larger fine-tuning scales. Safety failure is characterized as a disruption of low-rank output-routing mechanisms.
fisher-geometricoutput-routinglow-rankmlp modulessafety alignment
Efficiently Estimating Optimal Hyperparameter Scaling Laws through Power-Law Entropy Search
We introduce Power-Law Entropy Search (PLES), a computational cost-aware acquisition function for efficiently estimating optimal hyperparameter scaling laws in large language model (LLM) training. PLES leverages multi-fidelity Bayesian optimization to adaptively select configurations that maximally reduce scaling law uncertainty per unit computational cost, prioritizing informative small-scale experiments. Evaluated on synthetic benchmarks, surrogate models, and actual LLM pre-training runs, PLES converges to accurate scaling laws using less than one-tenth the computational budget of conventional grid search and other baselines.
hyperparameter scaling lawsmulti-fidelity bayesian optimizationlarge language modelacquisition functioncomputational cost-aware
Learning Sparse Decision Trees via Transformer Variational Auto-Encoders
TREVIS introduces a method for learning sparse decision trees by exploring the latent space of a Tree Transformer Variational Auto-Encoder (TTVAE), replacing discrete search with continuous gradient-based optimization. The approach jointly optimizes predictive performance and structural sparsity through a differentiable surrogate model. Experiments demonstrate that TREVIS achieves comparable predictive accuracy to near-optimal algorithms while significantly improving tree sparsity.
decision treesvariational auto-encodergradient-based optimizationstructural sparsitydifferentiable surrogate
Semantic-Guided Multimodal Preprocessing for Vision Transformer-Based Clear Cell Renal Cell Carcinoma Grading
A semantic-guided multimodal preprocessing method enhances Vision Transformer (ViT)-based clear cell renal cell carcinoma (CCRCC) grading by integrating nuclei classification maps with RGB histopathology images. The approach employs classification map channel concatenation and multiplicative modulation, optimizing overlays to leverage nuclei grading information while preserving RGB textural features. Evaluation shows a balanced accuracy of 0.916, significantly outperforming RGB-only baselines (0.707) and max-voting aggregation (0.427). Sensitivity analysis confirms robustness under simulated perturbation, with consistent per-class recall across grades (0.93, 0.91, 0.91). This method effectively bridges fine-grained nuclear-level analysis with coarse-grained patch classification, utilizing existing imperfect nuclei classifiers.
vision transformermultimodal preprocessingnuclei classificationsemantic-guidedhistopathology images
Provably Safe Sim-to-Real Transfer
We propose a provably safe sim-to-real transfer algorithm for reinforcement learning that reduces real-world interaction while ensuring safe exploration. The method formulates safe sim-to-real transfer within reward-free safe RL, leveraging simulator information to compute near-optimal feasible policies for any potential reward function. The algorithm achieves computational efficiency and provides a real-world sample complexity bound that quantifies the benefit of simulator usage in terms of sim-to-real mismatch. This approach addresses the challenge of safe data collection in applications like robotics and healthcare, where direct real-world exploration is constrained by safety requirements.
sim-to-real transfersafe reinforcement learningreward-free rlsample complexitysafe exploration
EdiTikZ: Scientific Figure Editing from Revision Trajectories
The paper introduces EdiTikZ, a novel approach for scientific figure editing using revision trajectories, addressing the gap in iterative refinement for publication-ready figures. The authors present DaEdiTikZ, a dataset of 391K TikZ edit pairs from arXiv, GitHub, and TeX SE, and infer 781K directed edit instructions using a vision-language model (VLM). They train two Qwen3.5-based models (4B and 9B) via joint reconstruction and editing learning, followed by reinforcement learning with fidelity and edit application rewards. The 9B model outperforms baselines in automatic evaluation and matches Gemini-3.1-Pro in human evaluation (4,320 ratings), remaining competitive with GPT-5.6-Sol under out-of-distribution shifts.
vision-language modelsscientific figure editingreinforcement learningtikzrevision trajectories
Neuro-Symbolic Geometric Abstraction (NeuSOGA): From Observations to Symbolic Mathematical Representations
NeuSOGA introduces a neuro-symbolic framework for transforming geometric observations into symbolic mathematical representations, addressing AI's challenge of explicit symbolic abstraction. The method integrates topology-guided structural discovery via Euclidean Distance Transforms, foundation-model perception using Segment Anything, adaptive multi-scale geometric abstraction, and symbolic synthesis through Implicit Area Splines. This yields analytical implicit models supporting arbitrary-order smoothness, additive composition, and closed-form evaluation. Experiments on ModelNet40 point clouds, arbitrary-view projections, and segmented optical observations demonstrate NeuSOGA's ability to preserve geometric and topological structure across modalities, producing interpretable and editable symbolic representations.
neuro-symbolicgeometric abstractioneuclidean distance transformsimplicit area splinestopological structure
Evaluating Multimodal LLMs as Generalist Vision-Language-Action Agents for Drone Control: Commanding, Approaching, Tracking and Searching
The study introduces DroneCATS-Agent, a modular architecture for multimodal LLMs in drone control, and DroneCATS, a benchmark evaluating MLLMs as vision-language-action agents. The framework tests models' capabilities in approaching visible targets, tracking moving ones, searching beyond initial views, and commanding multi-drone fleets, without fine-tuning or function-calling schemas. Evaluations reveal that smaller open models (down to 2B parameters) often outperform frontier models in navigation but fail in protocol adherence, particularly in terminating actions and multi-drone coordination. The key challenge lies in developing models that sustain persistent planning and correct termination at onboard compute costs.
multimodal llmsdrone controlvision-language-actionprotocol adherenceonboard compute
Measuring consistency via ensemble margin and local prediction variability: Auditing decision systems in the presence of predictive multiplicity
The paper introduces a consistency criterion for auditing incorrect ensemble predictions in the presence of predictive multiplicity (Rashomon effect), combining ensemble margin with local prediction variability. Under stability and smoothness assumptions, it proves that consistency scores of finite ensembles converge to those of the expected model as ensemble size and sample count increase. Evaluations on transformer models for natural language understanding and parameter-efficient fine-tuning of large language models for tabular data classification show that ensembling from the Rashomon set reduces unchecked incorrect predictions with moderate diversion increases. The proposed measure outperforms existing consistency metrics in capturing predictive multiplicity.
rashomon effectensemble marginpredictive multiplicitylocal prediction variabilityparameter-efficient fine-tuning
EDGE: Error Dependency Graph-Guided Multi-Error Attribution in Multi-Agent LLM Systems
EDGE introduces an Error Dependency Graph-guided framework for multi-error attribution in multi-agent LLM systems, addressing the limitation of existing methods that fail to model error dependencies. The method constructs an error dependency graph from observed errors, validates a causal subset via counterfactual rollout, and employs a two-stage LLM-as-judge detector for error attribution. Experiments on TRAIL and MAST benchmarks demonstrate EDGE's effectiveness in improving category-level multi-error attribution across various models and prompting strategies, highlighting the utility of dependency structures in diagnosing agent failures.
error dependency graphmulti-agent llm systemscounterfactual rolloutllm-as-judgemulti-error attribution
PopPert: Population-level Joint-Distribution Modeling for Single-Cell Perturbation Prediction
PopPert introduces a population-level joint-distribution modeling framework for single-cell perturbation prediction, addressing the limitation of existing methods that assume cell-to-cell correspondence in unpaired single-cell RNA sequencing data. The method parameterizes population-level joint gene expression distributions using a low-rank Gaussian Copula to capture cross-gene dependencies, enabling prediction of perturbation-induced changes in distribution parameters and synthetic sampling of perturbed single-cell profiles. Evaluated on multiple single-cell benchmarks across genetic and chemical perturbations, PopPert demonstrates superior performance in differential expression recovery, perturbation effect estimation, and population-level distribution matching.
single-cell rna sequencinggaussian copulaperturbation predictionjoint-distribution modelingpopulation-level
SymFold: Synergizing Evolutionary and Structural Priors for Accurate Protein Inverse Folding
SymFold introduces a symmetric dual-path architecture for protein inverse folding, synergizing evolutionary priors from protein language models (PLMs) and structural priors from multimodal protein language models (MPLMs) to iteratively guide sequence generation. Unlike serial pipelines where PLMs refine coarse sequences post-hoc, SymFold directly encodes structural knowledge via MPLMs while leveraging PLMs for evolutionary insights. Evaluated on standard benchmarks, SymFold achieves state-of-the-art performance, with ablation studies confirming the efficacy of its symmetric design. This approach advances protein inverse folding, enabling applications in enzyme engineering and drug discovery.
protein inverse foldingprotein language modelsmultimodal protein language modelssymmetric dual-path architectureevolutionary priors
CHARM: Character Hallucination for Multicultural Role Play Benchmark
CHARM introduces a multicultural benchmark of 40 real and fictional characters from five cultural-linguistic regions to evaluate role-playing LLMs' ability to respect character knowledge boundaries. The benchmark probes Temporal and Cross-Universe boundaries using abstention-enabled multiple-choice questions, with a two-stage evaluation separating Boundary-Awareness (recognition of out-of-scope queries) from Boundary-Compliance (abstention when answering). Results across six LLMs show hallucination is primarily driven by compliance failures, where models acknowledge out-of-scope queries but still provide factual, out-of-character answers, often due to parametric overrides. Cultural variations in failures correlate with imbalances in character representation in model knowledge.
role-playing llmscharacter hallucinationboundary-awarenessboundary-complianceparametric overrides
Scalable Rao-Blackwellized Online Planning for High-Dimensional POMDPs
The authors extend the Rao-Blackwellized online POMDP (RB-POMDP) framework to improve scalability in high-dimensional partially observable environments through hybrid continuous-discrete belief representations. By analytically propagating uncertainty of marginalized state components during tree-based planning, the method reduces sampling-induced variance in value estimation. The framework is integrated with FastSLAM 2.0 and evaluated in a robotic search-and-rescue task, demonstrating higher cumulative rewards with significantly fewer particles and planning simulations compared to purely sampling-based methods under equivalent computational budgets. Results indicate that structured high-dimensional problems with tractable sufficient statistics can be effectively addressed within the RB-POMDP framework.
rao-blackwellized pomdphybrid belief representationstractable sufficient statisticsmonte carlo estimationfastslam 2.0
Cheap Verifiers, Large Blind Spots: Measuring the Reliability Cost of Cost-Saving Cascades
This paper investigates the reliability-cost tradeoffs in inference cascades that use cheap models for most queries and escalate hard cases to a frontier model verifier. The authors measure the loop where the cheap student is fine-tuned on the verifier's rejections, finding that the verifier's blind spot (fraction of wrong answers accepted) grows with student capability and shrinks with verifier capability. Naive fine-tuning degrades the student, collapsing performance across all tested teachers. Cascade dashboards report a flat 3% error while true error swings up to 32%, demonstrating systemic blindness to degradation. Theoretical analysis reveals a two-population conservation law explaining this phenomenon.
inference cascadesverifier blind spotfine-tuningtwo-population conservation lawsystemic degradation
Probing Factual Knowledge Transfer with Training Data Interventions
The study introduces an intervention-based framework to probe factual knowledge transfer across languages in multilingual language models, focusing on English-to-Persian transfer. Using SIFT, a resource of 500 triples across 20 relations stratified by cultural origin, the authors systematically remove specific facts from Persian pretraining data and evaluate transfer. Results indicate limited fact transfer, with most English-acquired facts failing to transfer under strict removal conditions. Sentence-level co-occurrence removal proves insufficient to eliminate fact signal, and easier negative candidate sets inflate apparent transfer. Source-language entity frequency significantly influences transfer, with Persian-related facts hardly transferring due to their rarity in the English corpus.
factual knowledge transferintervention-based frameworksift resourcesentence-level co-occurrenceentity frequency
LEAP: Likelihood Elicitation and Aggregation for LLM-based Probabilistic Forecasting
LEAP introduces a novel probabilistic forecasting framework that reorganizes evidence usage in LLM-based systems, addressing limitations of Monolithic Prediction. The method separately examines each evidence item, elicits likelihood parameters describing its implications, and combines these with an explicit prior via a deterministic probabilistic model. This approach supports continuous, single-choice, and multi-choice forecasts while preserving reproducible evidence contributions. Evaluated on a benchmark covering forecasting, information-seeking, and browsing tasks, LEAP improves prediction and calibration metrics across models under controlled comparisons of prior access, inference budget, and aggregation.
probabilistic forecastingmonolithic predictionlikelihood elicitationevidence aggregationcalibration metrics
Bandits in Prod: Hyperparameter Optimization at Inference Time
We introduce IMABO, a framework for Online Hyperparameter Optimization (OHPO) that combines bandit policies with oracles for proposing configurations in mixed and conditional search spaces. IMABO integrates IMOSS, a restart-free anytime policy with active-set growth controlled by β∈(0,1), achieving an expected cumulative quantile-regret bound of O(p_ρ^{-1/β} + T^{(1+β)/2}). IMOSS is paired with three oracles: Tree-structured Parzen Estimator, incumbent-mutation oracle, and pretrained tabular foundation model, all outperforming uniform random baselines. IMABO demonstrates superior cumulative regret across diverse OHPO tasks, including LLM-based agent configuration and classical ML model tuning.
online hyperparameter optimizationbandit policyquantile-regret boundtree-structured parzen estimatortabular foundation model
Automated Event Log Generation from Unstructured Text Using Finetuned LLMs
We propose a framework leveraging finetuned Large Language Models (LLMs) to automate event log generation from unstructured text, addressing a bottleneck in process mining (PM). Our method finetunes LLMs on a novel text-to-log dataset, enabling high-fidelity extraction of structured event data from resources like incident tickets and manuals. Results demonstrate that finetuning significantly outperforms few-shot or zero-shot prompting, establishing it as a prerequisite for reliable event data generation. This approach expands PM capabilities by making previously underutilized organizational knowledge accessible for workflow analysis.
process miningevent logslarge language modelsfinetuningunstructured text
MIDR: Enrichment-Augmented Indexing for Multimodal Document Retrieval
MIDR introduces a training-free framework for multimodal document retrieval that shifts reasoning to index time, addressing OCR limitations in visually rich documents. During ingestion, a multimodal LLM converts rendered pages into verified textual fields indexed with BM25F, optionally fused with dense retrieval. On ViDoRe V3, MIDR Hybrid achieves 0.6219 average nDCG across five English domains, a 23.0% gain over BM25, and bridges English queries to French page text, lifting BM25 from 0.1532 to 0.5448 nDCG. MIDR outperforms ColQwen2.5 on four of seven domains while using 9x smaller index memory and 2x lower query latency.
multimodal retrievalbm25findex-time reasoningndcgvisual documents
A Composable Evaluation System for Reproducible Omni-Modal Foundation Model Evaluation
OmniEvaluator introduces a composable evaluation system for reproducible omni-modal foundation model assessment across text, image, video, and audio modalities. It integrates four inference backends, four evaluation frameworks, and over a thousand benchmarks through a unified interface, eliminating the need for separate toolchain environments. The system records each run as a reproducible artifact, aggregates results in a shared dashboard, and supports federated GPU inference for concurrent evaluations. A built-in verifier ensures score stability across engines and prompts, matching commercial LLM judges' performance without recurring API costs. The system is publicly available with demo video and dashboard.
omni-modalinference backendsevaluation frameworksfederated gpuscore stability
GazeRefine: Expert Gaze as a Test-Time Prompt for Training-Free Medical Image Segmentation
GazeRefine introduces a training-free framework for zero-shot medical image segmentation using expert gaze as an inference-time prompt. The method converts sparse, duration-weighted fixations into foreground and background priors, initializing semantic prototypes in frozen DINOv3 feature space. These prototypes are iteratively refined through foreground-background discrimination, feature-space affinity propagation, and anchoring to initial gaze guidance, enabling segmentation beyond fixated regions while minimizing semantic drift. Evaluated on gaze-annotated polyp segmentation and prostate MRI segmentation, GazeRefine demonstrates strong performance on colonoscopy images and competitive results on prostate MRI, highlighting its potential for label-efficient, human-in-the-loop medical image segmentation.
zero-shot segmentationgaze guidancefeature-space affinitysemantic prototypestraining-free
Analog-DB: An Agent-First Analog Integrated Circuit Database, From Blocks to Systems
Analog-DB introduces an open-source, versioned database for analog integrated circuit designs, addressing sharing challenges through a domain-specific language that captures process-neutral topologies, reusable testbenches, and machine-readable datasheets. The method employs a parameterization scheme for composable circuits, governed by a schema-driven contract and queryable catalog for AI agents. Results show 23 circuit-kit bindings meeting specifications on three open kits, with 17 of 23 imported sizings corrected via gm/ID sizing loops. A case study demonstrated defect detection in an op-amp core by a coding agent. The database includes 68 circuits across 16 classes, verified at schematic level and tracked on a power/performance scoreboard.
analog integrated circuitsdomain-specific languagegm/id sizingtestbenchesparameterization scheme
HiLRP: Toward One Trustworthy Explanation for Vision Transformer: Conservation-Valid Attribution via Attention Primitives
HiLRP introduces a unified attribution framework for Vision Transformers (ViTs) by decomposing attention and resolution-reduction operators into four operation types: linear maps, bilinear mixing, normalization/gating, and reindexing. Each operation admits a relevance rule satisfying conservation, enabling HiLRP to support diverse ViT architectures without architecture-specific derivations. The method proves conservation and conditional equivariance to machine precision and outperforms 14 existing attribution methods across 10 ViT architectures. HiLRP achieves 0.97 Pointing on EfficientViT, significantly surpassing competing methods (0.55), and maintains reliability across windowed, spatial-reduction, multi-axis, and linear-attention models where naive extensions fail.
vision transformersattribution frameworkconservationattention primitivesrelevance rule
EmbodiedSkills: A Unified Framework for Orchestrating, Training, and Deploying VLA Agents
EmbodiedSkills introduces a unified framework for orchestrating, training, and deploying vision-language-action (VLA) agents, addressing long-horizon tasks by integrating perception, planning, execution, verification, and recovery. The framework treats skill decisions as execution proposals, verifying prerequisites and outcomes through a shared executable-skill interface that connects high-level skill selection, low-level VLA execution, and post-action verification. Structured trajectories record planning and execution events, enabling supervision and optional online adaptation. Instantiated with Qwen3-VL and OpenPI/pi0.5 on RoboTwin 2.0 and LIBERO, task-adapted VLA policies achieve 86.20% success on RoboTwin 2.0 and 97.40% on LIBERO, with 12.5% success on memory-dependent RMBench tasks.
vision-language-actionexecutable-skill interfacestructured trajectoriestask-adapted policiesclosed-loop embodied systems
Some Emotions Run Deeper: Layer-wise Probing and Causal Intervention in Large Language Models
This study investigates emotion representation depth in large language models (LLMs) across diverse text sources and model architectures. Using layer-wise probing, forward interventions, transfer analyses, and early-exit classifiers across three datasets (Twitter, Reddit, autobiographical narratives) and eight LLMs (1B--9B parameters), the authors demonstrate that optimal probing layers vary systematically by corpus, shifting from input-adjacent to deeper layers. Forward interventions on probe-selected bands reduce test accuracy by 5--6 points more than random bands (q < 0.01). Probe-selected bands transfer across datasets and emotion categories, indicating shared affective information. Early-exit representations outperform full-depth exits by 6.9 percentage points on average.
layer-wise probingforward interventionsearly-exit classifieraffective informationemotion representation
TimeSteer: Inference-Time Speech Scheduling in Joint Audio-Visual Diffusion Models
TimeSteer introduces inference-time speech scheduling for joint audio-visual diffusion models, enabling precise temporal placement of utterances within user-specified intervals without finetuning. The method leverages two intrinsic properties: timing-sensitive text-to-audio cross-attention heads for source span localization and the organization of coupled speech-visual articulation in predicted clean latents for region-aware latent remapping. Evaluated on the SpeechShift benchmark, TimeSteer significantly enhances interval controllability over training-free baselines while maintaining competitive generation quality across two representative backbones.
inference-time schedulingaudio-visual diffusioncross-attentionlatent remappingspeechshift
The Constitutional Coverage Trilemma in AI Governance
The study identifies a constitutional coverage trilemma in AI governance, demonstrating that frontier LLMs inadequately cover human demand for safety, helpfulness, honesty, autonomy, and equity. Through a paraphrase-controlled audit of 23 LLM archetypes and a pairwise-tradeoff study with 1,649 US participants, the authors reveal that demand spans all five values, while supply occupies only ~2% of the demand hull. Key findings include a directional drift away from autonomy, a sparse fix involving a 2-vertex menu reducing mean regret by 47%, and robustness to distance-based welfare and degraded routing. The results formalize a budgeted-pluralism trilemma, empirically validated in the binding regime.
llm archetypesconstitutional coveragepairwise-tradeoffbudgeted-pluralismwelfare floor
Making Prospective Memory SLM-Shaped: Typed Intention Stores for Small-Model Agents
The paper introduces Prospective Intention Store (PIS), a schema-constrained state tracking framework enabling small language models to execute prospective memory tasks effectively. PIS integrates lifecycle logic in code and scoped language work on the model, requiring no selector fine-tuning or trajectory distillation. On PM-Bench, DeepSeek-Chat with PIS achieves 82.9% Set-F1, surpassing the previous state-of-the-art large-model scaffold of 65.1%. Gemma-E2B with PIS reaches 66.2% Set-F1, significantly outperforming retrospective memory methods (max 6.6%). PIS demonstrates superior performance, enabling small models to exceed large-model benchmarks in prospective memory tasks.
prospective memoryschema-constrainedstate trackinglifecycle logicset-f1
Dual Process Motion Planning
This paper introduces a dual-process architecture for nonlinear motion planning that integrates symbolic reasoning and learning-based approaches. Inspired by the Thinking Fast and Slow paradigm, the framework combines a robust symbolic solver (System-2) with experience-driven modules (System-1), dynamically orchestrated by a metacognitive controller. Evaluations across diverse nonlinear benchmark environments demonstrate consistent improvements in planning efficiency, accuracy, and generalization, while promoting task reuse. The results indicate that tightly coupling learning with structured reasoning offers a scalable path toward more capable and adaptive robotic systems.
dual-process architecturenonlinear motion planningsymbolic reasoningmetacognitive controllertask reuse
Measuring the Behavioral Fidelity of Long-Horizon Human Activity Simulations
We introduce a framework for evaluating behavioral fidelity in long-horizon human activity simulations, addressing a gap in prior work focused on survey responses and dialogue. Using a 43-hour multi-camera dataset of in-the-wild office activity, we compare trace-derived conditioning mechanisms: persona descriptors, few-shot exemplars, and statistical transition and time-of-day priors. Results show that statistical priors best approximate real activity and sequence distributions but over-fragment routines and suppress within-person variability, highlighting the need for multi-metric, multi-granularity evaluation.
behavioral fidelitylong-horizon simulationtrace-derived conditioningstatistical priorstemporal granularity
One Prompt Is Enough: Watermark Laundering Through Foundation Image Models
The study introduces watermark laundering as a novel failure mode for invisible watermarking schemes, demonstrating that foundation image models can remove watermarks through single-prompt reconstruction. Using a joint payload-fidelity profile combining bit error rate (BER) with visual and semantic preservation, the authors evaluate six OpenAI and Google image editing models across three watermarking schemes and 1,800 reconstructed outputs. Results reveal two laundering regimes: OpenAI models maximally disrupt payloads, while Nano Banana 2 shows DwtDct remains vulnerable under high-fidelity reconstruction. Prompt ablations indicate payload disruption is primarily induced by reconstruction pathways rather than explicit attack wording, motivating foundation-model reconstruction as a missing robustness condition.
watermark launderingbit error ratepayload-fidelity profilereconstruction pathwayfoundation image models
Explore More, Drift Less: Outcome-Only Reinforcement Learning Can Suffice for Long-Horizon Interactive Agents
The paper introduces CANOPY (Coverage-ANchored On-PolicY RL), a minimalist reinforcement learning protocol addressing signal starvation and policy drift in outcome-only RL for long-horizon interactive tasks. CANOPY scales same-task exploration to recover natural gradients, enforces on-policy KL-anchored updates confined to action tokens, and leverages enlarged interaction budgets. On AppWorld, a Qwen3-14B policy trained with CANOPY achieved top leaderboard scores (Test-Normal TGC 86.9, Test-Challenge 67.6) without task-specific supervision or auxiliary signals. Similarly, Qwen3.5-9B improved by 16.6 points on SWE-bench Verified, demonstrating that agentic RL alone can internalize long-horizon capabilities in small open models.
reinforcement learninglong-horizon tasksoutcome-only rlpolicy driftkl-anchored updates
From Language to Behavior: Scaling Sequence Transformers for Industrial Recommendation Ranking with Rec-Native Designs
The authors propose ReST, a recommendation-native Transformer scaling framework addressing challenges in behavior-sequence modeling for industrial recommendation ranking. ReST introduces a sequence encoder with dual-gated attention, rotary positional and temporal embeddings, stabilized residual normalization, and auxiliary objectives to handle noisy, irregular, and sparsely supervised sequences. It factorizes ranking into a reusable encoder and lightweight cross decoder with projection-free KV attention and token-specific parameterization, enabling compute-once, decode-many-times ranking. Evaluated on industrial and public benchmarks, ReST achieves higher accuracy and scales better along sequence length, depth, and width compared to LLM-style Transformer blocks. A production A/B test shows a 1.31% AUC improvement and 11.93% revenue lift within a 50 ms P99 latency budget.
dual-gated attentionrotary positional embeddingkv attentionauxiliary objectivessequence encoder
MutMem-V2: Cryptographically Authorized Mutation in Persistent Agent Memory Portable Verification and Reproducible Evidence
MutMem-V2 extends persistent agent memory authorization by introducing cryptographic protocols for verifiable mutation, addressing limitations in portable verification and reproducible evidence from MutMem-V1. The system employs domain-separated commitments, mandatory recall evidence, and external trust anchors, with 18 versioned object schemas, 39 recall vectors, and 15 mutation vectors. Independent Node and Python implementations achieve full agreement on structural and cryptographic terminals (72/72 cases), while a production-conformance corpus matches 42/42 cases. Clean installations demonstrate boot readiness, and a 120-unit Canary experiment validates explicit-marker traversal. The protocol ensures portable integrity and traceability but does not guarantee semantic truth or universal robustness.
cryptographic authorizationpersistent memoryrecall vectorsmutation terminalsverification contract
Position Matters: Feature Inversion Attacks in ViT Split Inference with Token Reduction and Shuffling
We introduce the Spatially Aligned Reconstruction Attack (SARA), a unified pipeline demonstrating that token shuffling in ViT split inference provides only apparent privacy against feature inversion attacks. SARA predicts token positions, restores spatial layout, reconstructs missing embeddings via a feature-space masked autoencoder, and recovers input images. Results show token reduction offers stronger protection, but significant leakage persists when retained tokens preserve semantic and positional information. We propose a lightweight edge-side defense removing positional embeddings and adapting transformer blocks through knowledge distillation, substantially reducing SARA performance while preserving task accuracy and requiring no cloud-side modifications.
vision transformerssplit inferencefeature inversiontoken shufflingknowledge distillation
Prompt-Robust Language Models: Which Training Strategies Work?
This study systematically evaluates training strategies for improving prompt robustness in large language models. It compares robustness-enhancing methods (CoIN for contrastive alignment, PPCL for consistency regularization) against simpler data construction approaches under controlled conditions. Results show that while robustness fine-tuning outperforms standard fine-tuning and in-context learning, performance gaps between best and worst prompts remain substantial (40-57%). Diagnostic analysis reveals auxiliary objectives fail to generalize beyond their specific penalties, and conflicting per-template gradients (57-64% of parameters) complicate optimization when mixing prompt formulations in batches.
prompt robustnesscontrastive alignmentconsistency regularizationin-context learningper-template gradients
H2Table: Hierarchical Hypergraph-Enhanced Large Language Models for Complex Table Reasoning
H2Table introduces a hierarchical hypergraph-enhanced framework for complex table reasoning in large language models (LLMs). The method represents tables as hierarchical nested hypergraphs, employing a hypergraph encoder to facilitate message passing between hyperedges (headers) and nodes (cells), capturing semantic entailment relationships. Learnable query vectors bridge structural embeddings from the encoder to the LLM. On the HiTab dataset, H2Table achieves a 22.88% average improvement over state-of-the-art baselines for tables with a nesting depth of four.
hierarchical hypergraphtable reasoningmessage passingsemantic entailmentlearnable query vectors
REFACTOR-VLA: Unsupervised Library Learning of Typed Motor Programs
REFACTOR-VLA introduces an unsupervised wake/sleep system for learning reusable motor-program abstractions in vision-language-action (VLA) models. The sleep phase clusters motor-program fragments using a Behavioral-Equivalence Kernel (BEK) derived from rollouts of a learned latent world model, while the wake phase emits typed lambda terms consumed by a library-conditioned rectified-flow action decoder. Abstractions are admitted based on Minimum Description Length and return-preservation criteria. Experiments on LIBERO demonstrate that enlarging the world model from 188M to 430M parameters degrades performance, while adding an auxiliary InfoNCE loss improves clustering, achieving Normalized Mutual Information scores of 0.462 (object), 0.867 (spatial), 0.915 (goal), and 0.754 (LIBERO-10), outperforming baselines by Δ=+0.184 on average.
behavioral-equivalence kernelrectified-flowminimum description lengthnormalized mutual informationmotor-program
Who Judges the Judges? A Chinese Safety QA Benchmark for Evaluating LLM Responses and Safety Judges
We introduce C-SafeQA, a policy-grounded benchmark for Chinese safety evaluation of large language model (LLM) responses, addressing linguistic variation and adversarial transformations. The benchmark comprises 538 base queries and 8,877 adversarial queries answered by four LLM deployments, yielding 37,660 query-response records labeled safe, unsafe, or disputed through multi-model adjudication and expert audits. Results show unsafe-response rates ranging from 0.93% to 3.35% on base queries and 11.68% to 30.05% on adversarial queries, with substantial trade-offs in unsafe-response recall and safe-response false positive rates across seven automated safety judges. Acrostic transformations reduce unsafe recall for all judges, highlighting evaluator weaknesses.
large language modeladversarial transformationspolicy-grounded benchmarkunsafe-response recallfalse positive rate
Autonomous discovery of new structure-plausibility laws for explainable and rapid crystal diagnosis and screening
The authors introduce Plausibility Rules for Inorganic Structures (PRIS), a set of eight autonomously discovered physicochemical laws for explainable crystal screening. Using active refutation, agents generated and tested two million candidate laws, encoding five mechanisms: short-range repulsion, ionic contact, electrostatic balance, bond-valence conservation, and site complexity. PRIS achieves 82–99% accuracy on experimental structures, detects 87.9% of damaged structures (vs. 1.6–3.2% for distance cutoffs), and reduces DFT validation queues by 67.3% while retaining 99.2% of target-achieving candidates. PRIS-derived synthesis scores (PSS) screen 83.7% of hard-to-synthesize structures with 80.7% recall.
crystal screeningdensity functional theoryactive refutationbond-valence conservationinverse-design
Towards AI-Assisted Clinical Trial Matching: Practical Considerations, Multicenter Evaluation, and Real-World Deployment
We present TrialGPT 2.0, an AI-assisted clinical trial recommendation system that extends beyond eligibility assessment to prioritize trials based on clinical needs and workflow priorities, providing structured explanations for expert review. The system was evaluated retrospectively across 288 cases from multiple oncology settings and prospectively in an active precision oncology tumor board. TrialGPT 2.0 achieved 91% recall in retrieving clinician-recommended trials within its top 10 recommendations, reduced clinician screening time by 55.0%, and increased patient access to trial participation by 90.9%. We also introduce NIH-TrialBench, a clinician-authored dataset of 126 synthetic patient vignettes for reproducibility.
clinical trial matchingeligibility assessmentprecision oncologyretrospective evaluationprospective evaluation
FinLifeBench: Exhaustive Life-Event History and Financial-State Reconstruction from Longitudinal Banking Dialogue
We introduce FinLifeBench, a benchmark evaluating exhaustive longitudinal reconstruction of life events and financial states from cumulative banking dialogues. The benchmark comprises 6,000 eight-turn Korean banking sessions across 20 synthetic trajectories, with deterministic gold standards for 24 event types and 34 financial state paths. Evaluations on eleven LLMs reveal event-anchor recall declines from 0.591 at 15 sessions to 0.445 at 300 sessions, with errors driven by omitted events and outdated financial state information. The best GCA@15 score reaches 0.470, showing weak association between event and state reconstruction tasks.
longitudinal reconstructionfinancial stateevent-anchor recallgca@15deterministic gold
Athena: Vulnerability-Affected Library Identification via Knowledge Graph Completion
Athena introduces the first graph-based approach for vulnerability-affected library identification, addressing the limitations of existing text retrieval methods. It constructs a security knowledge graph integrating CVEs, libraries, CWE weakness types, CPE products, and software ecosystems, then applies knowledge graph completion (KGC) via link prediction to identify missing affected libraries. A re-ranking module fine-tunes an LLM with knowledge graph embeddings to jointly leverage structural and textual information. Experiments on VulLib show Athena achieves a 32% improvement in Avg. F1 over VulLibGen, with a 110M-parameter KGC backbone outperforming VulLibGen's 7B-parameter configuration.
knowledge graph completionlink predictioncvellmvulnerability identification
Jailbreaking Text-to-Image Models Through Cracks: Navigating Heterogeneous Safety Filters via Multi-Agent Debate
The paper introduces CRACK, a multi-agent debate framework for jailbreaking text-to-image (T2I) models guarded by heterogeneous safety filters. CRACK employs an Attack Agent, Defense Agent, and Judge Agent to iteratively generate prompt mutations, diagnose layer-specific feedback, and optimize strategies through reward-guided refinement, adapting to cross-layer constraints while preserving harmful intent. The method leverages the Detection Surface, a geometric framework characterizing decision boundaries of safety filters, revealing sparse, non-convex evasion regions shaped by cross-layer conflicts. Experiments demonstrate CRACK achieves Attack Success Rates up to 99.63% across multiple T2I models and safety configurations, outperforming existing methods in query efficiency and semantic fidelity.
text-to-image modelsjailbreak attacksmulti-agent debatesafety filtersdetection surface
Superposed Latent Autoencoder
The Superposed Latent Autoencoder (SLAE) improves representation compression by superposing multiple wide latent codes into a shared memory tensor instead of reducing individual latent dimensions. SLAE transforms latents into storage-friendly codes, binds them with randomized keys, and learns to recover each latent before decoding, replacing dimensional bottlenecks with structured interference. On CIFAR-10/100, SVHN, STL-10, and Tiny ImageNet, SLAE reduces reconstruction error by up to 56% versus conventional autoencoders at matched storage budgets and improves downstream classification accuracy by 16.79 percentage points, demonstrating that wider representations can be efficiently shared.
autoencoderlatent representationsuperpositionrepresentation compressionmemory budget
StainPresetNet: Stain Preset Network for Fast Multi-to-Multi Stain Normalization
StainPresetNet introduces a novel framework for fast multi-to-multi stain normalization, addressing limitations in computational efficiency, artifact generation, and directional rigidity of existing methods. The approach combines structural preservation with dataset-level color mapping, implementing pixel-wise normalization guided by preset reference images to enable multi-directional adaptability without retraining. Evaluations on cytopathology and histopathology datasets demonstrate superior color mapping accuracy, improved classifier generalization in diagnostic tasks, and a 90% reduction in computational overhead compared to deep learning approaches. The preset-guided mechanism allows flexible adjustment of normalization directions through simple reference image replacement.
stain normalizationcolor mappingpixel-wise normalizationdataset-level mappingcomputational efficiency
Revisiting Face Recognition for Monozygotic Twins: The Celeb Twins Test Set
The Celeb Twins Test Set (CTTS) introduces a novel dataset of 80 celebrity twin pairs, annotated with metadata on skin marks and mirror asymmetry, to advance face recognition research for monozygotic twins. Organized similarly to established face verification benchmarks like LFW and CALFW, CTTS enables evaluation of deep CNN matchers, which currently achieve over 76% accuracy on same-person/different-person classification tasks. Analysis reveals that existing models fail to leverage skin marks or asymmetry features. The study also explores the potential of generative AI tools such as Grok, ChatGPT, and Gemini to synthesize twin images for augmenting training datasets.
monozygotic twinsface verificationdeep cnnskin marksgenerative ai
DNC-IMM: Early Lane-Change Intention Recognition via Neural Calibration Based on Driving Context Information
The Dual Neural-Calibrated Interacting Multiple Model (DNC-IMM) improves early lane-change intention recognition by integrating driving context into a probabilistic framework. The method encodes target-vehicle motion, gaps to surrounding vehicles, and relative velocities via a neural network, which calibrates both the transition-probability matrix and measurement likelihoods. Intention prediction is derived from the calibrated IMM mode posterior rather than a separate classifier. Evaluated on the highD dataset, DNC-IMM achieves reliable lane-change recognition before lane crossing, with particularly strong performance at 2-3 s prediction horizons.
lane-change intentionneural calibrationtransition-probability matrixmeasurement likelihoodshighd dataset
Latent Recurrent Thoughts: Recurrent Refinement of Proposed Latents for Reasoning with Frozen LLMs
Latent Recurrent Thoughts (LRT) introduces a method for reasoning in continuous latent space using frozen large language models (LLMs) augmented by small auxiliary networks. The approach employs a task-specific proposer to generate base latent states, which are iteratively refined by a recurrent reasoner through bounded residual corrections, decoupling computation depth from model size. The frozen LLM decodes the final answer. LRT outperforms prior continuous-space reasoning methods and chain-of-thought prompting on symbolic reasoning tasks (Countdown-4, Sudoku) and natural-language reasoning benchmarks (HumanEval, MBPP, StrategyQA) with significantly lower inference compute.
latent recurrent thoughtsfrozen llmcontinuous latent spacerecurrent reasonerbounded residual corrections
EDRAC: Benchmarking Arabic Dialect Reading Comprehension
EDRAC introduces the first large-scale benchmark for dialectal Arabic machine reading comprehension (MRC) and generative QA, addressing the under-resourced nature of Dialectal Arabic (DA) compared to Modern Standard Arabic (MSA). The benchmark covers five major dialects—Egyptian, Moroccan, Emirati, Syrian, and Saudi Arabic—and includes 499 passages from naturally spoken interactions and 4,977 QA pairs generated via a human–LLM collaborative pipeline. Evaluation of Arabic-centric and multilingual LLMs using lexical and semantic metrics reveals significant gaps in semantic answer quality and dialectal fidelity, underscoring the limitations of current evaluation metrics for dialectal Arabic generation.
dialectal arabicmachine reading comprehensiongenerative qallm-as-a-judgesemantic metrics
Hints Help But Do They Teach? Evaluating Skills Transfer in Code Generation
This work evaluates whether hints in code generation transfer skills or merely guide models to existing solutions. Experiments on HumanEval+ and MBPP+ test Qwen2.5-3B-Instruct and Phi-3.5-mini with adaptive hints, unrelated hints, and unhinted sampling. Relevant hints rescue 36/79 and 42/101 failures respectively, but most solutions are reachable through ordinary sampling (46/79 and 57/101). Mechanistic tests on Qwen reveal a stable activation direction shared by hint types, with persistent addition yielding 14 rescues and 18 regressions. Hidden-state probes transfer across benchmarks (pooled AUROC 0.806, 0.780), but their advantage over token confidence remains unresolved.
code generationhintshumaneval+mbpp+activation direction
StateSwap: Probing Support-Elimination Hidden States in Multiple-Choice Questions
The paper introduces StateSwap, a method to probe framing-induced hidden states in multiple-choice question answering by large language models. Using a dual-framing protocol with support- and elimination-oriented prompts, the authors append an untrained [STATE] token to track residual-stream activations as an intervention interface. Experiments show that framing induces separable [STATE] activations in intermediate layers; swapping these activations systematically alters predictions and improves cross-framing consistency. Mean-difference steering directions derived from the contrast exhibit more bounded layer-wise responses than contrastive activation addition.
residual-stream activationdual-framing protocolintervention interfacemean-difference steeringcontrastive activation addition
Text-guided flow matching enables sample-efficient crystal structure generation
TFMat introduces a text-conditioned flow-matching framework for crystal structure generation, using structured materials language as a semantic prior for the CrystalFlow generator. The method enables text-guided control over composition, symmetry, and properties, addressing the mismatch in current crystal generator interfaces. On Perov-5, Carbon-24, and MP-20 benchmarks, TFMat achieves a 92.04% MP-20 match rate with 20 candidates and improves one-candidate match rates over CrystalFlow. It also enhances element-count and density distribution alignment in de novo generation while maintaining property consistency in composition-selected outputs.
flow matchingcrystal structure generationtext-conditionedsemantic priormaterials design
Space Generative AI with Solar Energy Harvesting
The paper introduces a framework for solar-powered space generative AI, optimizing the computation-communication trade-offs in satellite-based diffusion models. The method leverages predictable solar energy harvesting dynamics to jointly maximize end-to-end generative performance through a two-step approach: characterizing maximum downlink throughput for fixed generation depth and formulating a utility-maximization problem with a closed-form step-selection policy. Experiments under realistic orbital dynamics show the policy dynamically balances image generation quality and transmission reliability, outperforming static baselines across diverse solar energy harvesting states.
space generative aisolar energy harvestingdiffusion modelscomputation-communication trade-offsend-to-end performance
ARISE-RL: Agentic Rubric-Grounded Iterative Self-Evolution with Reinforcement Learning
ARISE-RL introduces a full-cycle self-evolution framework for training open-ended agents via reinforcement learning, addressing challenges in verifiable gold answers and scalable rubrics. The method couples a task/rubric Generator and a reasoning Solver through rubric-mediated co-evolution, where the Generator produces valid, intermediate-difficulty tasks aligned with the Solver's capability boundary, and the Solver learns from fine-grained rubric satisfaction signals. Reward-Gated Self-Evolution Distillation (RG-SED) selectively distills a memory-augmented policy variant back into itself to reduce distribution mismatch and avoid noisy guidance. Evaluated on ECR-Bench, an expert-calibrated rubric benchmark suite, ARISE-RL achieves robust and stable state-of-the-art performance across all benchmarks.
self-evolutionrubric-mediated co-evolutionreward-gated distillationopen-ended agentsmulti-step reasoning
User Representation via Cross Multi-source Behavior Pre-training for Mobile Games
CM-PTM introduces a novel Cross Multi-source Behavior Pre-Training Model for mobile game user representation learning, addressing the limitations of app-centric approaches by modeling cross-source dependencies and fine-grained behavioral dynamics. The model employs hierarchical cascaded mask-then-predict proxy tasks, first inferring the source of the next behavior and then refining predictions at the app-action level. Evaluated on large-scale real-world mobile datasets, CM-PTM effectively captures users' endogenous interests and significantly improves performance on downstream mobile game recommendation tasks.
user representationpre-trainingcross-source dependenciesbehavioral dynamicsmobile game recommendation
WorldBench: Culturally Grounded Benchmark for Multilingual Agents
WorldBench introduces a culturally grounded, multilingual benchmark for evaluating LLM-powered agents across seven languages and eight cultures, comprising 1,600 tasks simulating everyday workflows. The benchmark employs structured actions in a sandbox environment and introduces Constrained Task Success (CTS), a metric combining natural language instructions and testbeds to assess task completion, minimal modification, and environment preservation. Evaluations reveal frontier models achieve only 49.2% CTS, highlighting brittleness in multilingual, agentic scenarios, particularly for long-horizon tasks and state-preservation constraints.
constrained task successmultilingual benchmarkstate preservationstructured actionsllm-powered agents
Lagged Coupling: Internal Representations Become Readable Before They Become Causal
The study introduces lagged coupling, a phenomenon where internal representations in language models become readable before they gain causal efficacy, demonstrated across the Pythia suite (160M-12B parameters). Using linear probes, the authors decompose lagged coupling into three tracks: internal readability, behavioral readability, and causal efficacy. Results show internal readability saturates early (AUROC >= 0.990), while behavioral readability develops gradually and causal efficacy remains largely null-equivalent. Representation headroom grows up to 57x with training, yet causal write-in stays below 0.11% of headroom. The findings caution against inferring steerability from probe accuracy and highlight a developmental bottleneck in representation formation.
lagged couplinginternal readabilitycausal efficacylinear proberepresentation headroom
QILP-0: Constructing Observational Declarative Twins of Quantum Circuits
The paper introduces QILP-0, the first complete order-0 specialization of the QXymb framework, for constructing observational declarative twins of quantum circuits. The method converts observed quantum circuit behavior into finite multi-valued propositional logic programs via structural grading, target-independent discretization, and symbolic processing, while preserving observational semantics. Validation on Bars & Stripes (16-100 qubits) and MNIST (14,708 digit-0/1 instances) shows QILP-0 achieves conflict-free reconstruction with perfect accuracy (1.0) in all reported relations, separating logical exactness from numerical and discretization uncertainty.
observational declarative twinsquantum circuitsmulti-valued logictarget-independent discretizationsymbolic processing
HiveTraceGuard-Pro: A Compact Generative Guardrail for Prompt Injection, Jailbreaks, and Adversarial Obfuscation
HiveTraceGuard-Pro introduces a compact 0.6B generative guardrail model, LoRA-tuned from Qwen3-0.6B, to mitigate prompt injection, jailbreaks, and adversarial obfuscation in LLMs. It employs a binary scoring rule (safe/unsafe) and is trained on Russian and English corpora, applying eight obfuscation transforms to harmful and benign examples. Evaluated on nineteen benchmark groups, it achieves an aggregate key of 0.7432, with a clean Russian robustness combined-F1 of 0.88 and Russian prompt-injection recall of 0.999. Its median latency is 14.3 ms, the lowest among fifteen compared models. The model is released on Hugging Face under Apache-2.0.
generative guardraillora-tunedprompt injectionobfuscation transformsbinary scoring rule
AgentFactory: Towards Automated Agentic System Design and Optimization
AgentFactory introduces a framework for automated joint optimization of foundation models and workflow structures in agentic systems, addressing performance, cost, and efficiency objectives. The method employs LLMs as optimizers in a three-stage pipeline to explore configurations of fine-tuned models and optimized workflows, adapting to task-specific requirements. Evaluated across eight benchmarks in five domains, AgentFactory outperforms manual and automated approaches, achieving average improvements of 9.1% overall and domain-specific gains of 19.6% on MedQA and 18.7% on FinEval.
agentic systemsfoundation modelsworkflow optimizationlarge language modelstask-specific adaptation
From Truncation to Commitment: Persistent Context in Uniform Discrete Diffusion
The paper introduces committed reveal sampling (CRS), a training-free sampler for uniform-state discrete diffusion models that maintains persistent context by storing and reusing selected argmax tokens in subsequent predictions. CRS leverages the Bayes error property and a latent-mode model to justify late token selection and visibility retention, enhancing sequence-level coherence. Empirical evaluations on Duo-distilled demonstrate that CRS achieves lower generative perplexity (GenPPL) compared to top-$p$ truncation baselines ($p=0.95$, $p=0.9$) across 8--64 function evaluations (NFE), offering improved GenPPL--entropy tradeoffs. Results highlight distinct roles of support restriction and persistent context in controlling generative quality.
committed reveal samplinggenerative perplexityuniform-state discrete diffusionbayes errorfunction evaluations
ViTAMINS: An Empirical Study of Training Self-Supervised Vision Transformers with Synthetic Hard Negatives
ViTAMINS introduces synthetic hard negatives into unsupervised vision transformer pretraining to enhance representation quality. The method modifies existing contrastive frameworks, integrating these negatives to improve semantic content encoding and classification performance. Evaluated on ImageNet and various downstream tasks (transfer learning, image retrieval, copy detection, segmentation), ViTAMINS achieves up to +11.3% accuracy gains over baselines. Notably, the ViT-B variant outperforms V-JEPA with ViT-L while being more resource-efficient. The results suggest contrastive learning as a simpler yet powerful alternative to generative and self-distillation approaches in vision transformer training.
vision transformerscontrastive learningsynthetic hard negativesrepresentation qualityself-supervised learning
Causal Evidentiary Governance for High-Risk Machine Learning Systems
Proposes Causal Evidentiary Governance (CEG), a framework for high-risk ML systems that enforces causal fairness via versioned DAGs partitioning causal pathways into allowable/disallowed groups. Introduces Causal Harm Rate to quantify prediction variation from disallowed paths, with Decision-Evidence Packets (DEPs) cryptographically binding predictions to causal attributions. Validates on 10,000 synthetic credit applicants using PMA data, showing Causal Harm Rate outperforms demographic parity/equalized odds in isolating injected biases. Cross-model validation on German Credit data reveals associational metrics understate pathway-specific harm, while a proof-of-concept demonstrates operational feasibility.
causal fairnessdirected acyclic graphdecision-evidence packetcausal harm ratemerkle tree
Data-Driven Persona-Conditioned Agents for A/B Test Simulation
The paper introduces a simulation framework for predicting A/B test outcomes using LLM-powered agents conditioned on data-driven personas derived from real user behavioral signals (activity patterns, engagement, demographics). Unlike synthetic or rule-based approaches, these agents enable faithful population modeling by framing A/B simulation as a structured question task, studying question formats, persona data sources, and population subsampling trade-offs. On a benchmark of 40 A/B tests, the best configuration achieves 0.75-0.90 directional accuracy, demonstrating viability for low-cost experiment pre-screening.
a/b test simulationllm-powered agentsdata-driven personasbehavioral signalsdirectional accuracy
Spawn Freely, Act Sparingly: Progressive Risk Vesting for Recursive LLM-Agent Trees
The paper introduces Progressive Risk Vesting (PRV), a method for managing risk in recursive LLM-agent trees by holding a trajectory-level risk budget in escrow and debiting it as branches gain authority to act. PRV distinguishes between sandbox spawning (safe exploration) and capability activation (irreversible actions), proving an anytime harm bound for adaptively generated trees. Theoretical analysis shows trajectory harm scales with local risk $p$ as $p$ (subcritical), $\sqrt{p}$ (critical), or maintains a floor (supercritical), depending on the authority reproduction number $\mathcal{R}_A$. The method suggests broad sandbox search with sparse authority grants, supported by synthetic branching calculations and split-sample experiments.
progressive risk vestingrecursive llm-agent treessandbox spawningcapability activationauthority reproduction number
On Synthesis of Metric Interval Temporal Logics
The paper introduces the first framework for precise passive learning of Metric Interval Temporal Logic (MITL) without relying on templates or restricted fragments, addressing a gap in automated specification mining for real-time systems. The method reduces timed learning to a scalable untimed problem by identifying timing differences between traces, synthesizing timed constraints as Boolean propositions, and leveraging optimized untimed LTL tools. The framework guarantees completeness, ensuring a separating specification can always be found. Evaluation across benchmarks demonstrates its effectiveness in handling expressive timed logic.
metric interval temporal logicpassive learningtimed regular expressionsreal-time systemsformal specifications
A Network Science Perspective on Evaluating Deep Graph Generative Models
The study evaluates deep graph generative models (DGGMs) and the configuration model from a network science perspective, focusing on topological similarity to real-world networks and utility in epidemic mitigation. Using structural property comparison and node immunization strategy effectiveness as metrics, the authors find that two DGGMs generate synthetic networks closely resembling real-world topologies, outperforming traditional models like Erdos-Renyi. The synthetic networks enable accurate identification of effective immunization strategies to suppress epidemic or misinformation spread, addressing privacy constraints in sharing real contact networks.
deep graph generative modelstopological similaritynode immunizationepidemic mitigationconfiguration model
Figures as Programs: Recursive Generation of Editable Scientific Figures
The paper introduces \textsc{FigTree}, a multi-agent system that recursively generates editable scientific figures as SVG programs from research papers. The method grounds content in source text, decomposes figures into hierarchical regions, generates each as SVG fragments, and assembles them via a render-critic loop for iterative refinement. Evaluations demonstrate \textsc{FigTree} produces higher-quality figures than raster-based methods while enabling precise edits through programmatic traceability of visual defects to specific SVG statements.
vector graphicsmulti-agent systemsvg generationrecursive decompositionrender-critic loop
SinkPruner: Sink-Free Visual Token Pruning for Multimodal Large Language Models
SinkPruner introduces a training-free visual token pruning framework for efficient multimodal large language model (MLLM) inference, addressing redundancy in high-norm outlier tokens. The method combines a visual sanitizer to filter redundant tokens and mitigate attention issues, with a text-guided pruner to retain semantically relevant tokens. Evaluations on 16 benchmarks show SinkPruner preserves 96.5% (LLaVA-1.5) and 91.8% (Qwen2.5-VL) of original performance under 89% token reduction, while enhancing existing pruning methods via transferability.
visual token pruningmultimodal large language modelsattention sinkhigh-norm outlierstraining-free inference
Right Frame, Wrong Rule: Cultural Cues Expose the Financial Knowledge Gap They Were Meant to Close
The study introduces normative pluralism as a setting where language models must select among competing normative frameworks, demonstrated in Islamic finance using a four-choice taxonomy separating framework selection from within-framework correctness. It identifies the stereotype trap, where cultural cues bias framework selection but reveal within-framework inaccuracies. Evaluating twelve models across two languages and fifty demographic signals, cultural cues altered framework selection (97% Islamic framework selection for large open-weight models under strong signals) while exposing accuracy gaps (57--66% incorrect within-framework answers), motivating the competence-conditioned routing hypothesis.
normative pluralismstereotype trapcultural cuesframework selectioncompetence-conditioned routing
Inspicio: Open-Vocabulary, LLM-Based Sense Retrieval for Historical Languages
Inspicio introduces an open-vocabulary sense retrieval pipeline for historical languages, bypassing the need for language-specific sense inventories by leveraging LLM-generated translations, definitions, and lemmas. The method combines dense definition-synset similarity, sparse lemma matching, and Maximal Marginal Relevance re-ranking using outputs from instruction-tuned LLMs. Evaluated on Latin, Ancient Greek, and Italian datasets, the best configuration achieves 96% Recall@50 on perception verbs and demonstrates robustness in out-of-domain and cross-lingual settings.
word sense disambiguationopen-vocabulary retrievalllm-based translationmaximal marginal relevancesynset similarity
On the Human and Computer Alignment of Attribute-Based Music Matches
The study introduces MATCHA, a dataset of 1105 expert perceptual assessments (from 83 participants) evaluating alignment between human and computational judgments of musical similarity across five attributes: melody, harmony, rhythm, voice, and timbre. Using a triplet-based forced-choice task with 300 cases (including plagiarism, covers, and AI-generated music), the work measures agreement among participants and compares it to computational similarity metrics. Results show measurable human consensus in identifying matches and partial alignment with computational measures, highlighting the need for perceptually grounded evaluation in generative AI for music.
perceptual evaluationmusic similaritytriplet taskgenerative aiattribute alignment
Semi-Supervised Virtual Staining via Morphology Preservation and Histopathological Realism Constraints
Proposes a semi-supervised virtual staining framework that leverages limited paired and abundant unpaired histopathological images by jointly enforcing Hessian-derived morphology preservation and histopathological realism constraints. The method prevents structural degradation and unrealistic staining by extracting structural cues from source images and constraining outputs to retain diagnostically relevant morphology while matching target-stain characteristics. Experiments on H&E-to-IHC (Ki67, HER2) and FFPE-to-H&E translation show improved image quality, morphology preservation, robustness, and diagnostic performance compared to supervised baselines.
virtual stainingmorphology preservationhistopathological realismsemi-supervised learninghessian-derived constraints
Disclosure-Gated User Simulation for Companion-Agent Evaluation
The paper introduces a disclosure-gated user simulation method to address the excessive cooperativeness of LLM-based user simulators in companion-agent evaluation. The method employs a five-gate ladder merged into three observable layers, conditioning information release on agent behavior. Training combines synthetic (gating behavior) and real (speech patterns) data branches, eliminating runtime gate specification. Evaluated on CompanionBench's English corpus, the method maintains ranking stability (0.993 correlation with original simulator) and score scale-stability, unlike frontier-model prompts that inflate scores without altering rankings. The work formalizes the benchmark's implicit mechanism with specification, ablations, and sensitivity analysis.
disclosure gateuser simulationcompanion-agent evaluationllm-based simulatorranking stability
The zbMATH Open Knowledge Graph: Tracing Centuries of Mathematical Research
The zbMATH Open Knowledge Graph introduces a large-scale RDF knowledge graph (34M entities, 168M triples) integrating expert-curated semantic content from 250+ years of mathematical research, surpassing bibliographic metadata with reviews, subject classifications, and disambiguated authorship. Using Semantic Web vocabularies, it enables fine-grained temporal analysis of mathematical concepts and scholarly relationships. Query-driven case studies demonstrate its ability to uncover patterns opaque to citation-based methods, providing an open infrastructure for studying mathematical knowledge evolution.
knowledge graphsemantic webrdfmathematical scholarshiptemporal analysis
CoBRA: Learning Tool-Use Boundaries via Counterfactual Margins
CoBRA introduces a counterfactual boundary-learning framework for optimizing tool-use decisions in augmented language models. The method constructs internal and external experts from a base model, estimates reward margins between tool-augmented and unaugmented responses, and partitions data into boundary-aware categories. It employs Boundary-Aware Cold-Start SFT and MARS-RL with reference-split rollouts to optimize tool-calling decisions. Experiments on Qwen3-4B demonstrate improved tool-use efficiency and boundary-sensitive accuracy while maintaining performance on tool-dependent OOD questions.
tool-augmented language modelscounterfactual marginboundary-aware sftmars-rlreward margin
Few-Shot Out of Domain Intent Detection with Covariance Corrected Mahalanobis Distance
The paper proposes a covariance-corrected Mahalanobis distance to improve few-shot out-of-domain (OOD) intent detection in conversational agents. While Mahalanobis distance previously outperformed baselines in OOD detection (Podolskiy et al., 2021), it underperformed in few-shot settings. The authors identify the cause of this limitation and introduce a covariance correction to enhance performance. The method addresses the practical challenge of classifying unknown intents with limited training examples, though specific accuracy gains are not quantified in the provided text.
out-of-domain detectionmahalanobis distancefew-shot learningcovariance correctionintent classification
Calibration is the Bottleneck: An Action-Class Diagnostic of Multi-Turn Tool-Calling
The study introduces an action-class diagnostic framework to analyze multi-turn tool-calling failures in LLMs, decomposing them into action-class miscalibration and action-execution failure. The framework operates over a four-class action space (TOOL_CALL/ASK/REFUSE/CONFIRM) and uses a self-revealing upper bound (Acc GAR) to expose state-grader masking of miscalibration. Validation on multiple benchmarks reveals action-class miscalibration as a significant failure mode, obscured by aggregate accuracy metrics. Calibration is shown to be reshapable through context-only perturbations, but effects vary heterogeneously across model families (e.g., +11.5 pp to -21.0 pp). The authors advocate for supplementing aggregate accuracy with action-class diagnostics in evaluations.
multi-turn tool-callingaction-class miscalibrationllm agentscalibrationcontext-only perturbations
From Terminology to Diagrams: Visual-Instruction Generation for Scientific Diagram Understanding
We introduce SciGram, a framework for generating large-scale diagram-grounded instruction data to improve scientific diagram understanding in vision-language models (VLMs). The method systematically extracts domain concepts from scientific curricula, synthesizes atomic facts, retrieves relevant diagrams from the web, and generates multimodal supervision via captions and multiple-choice questions. SciGram comprises 194K diagrams and 1.4M visual instructions across life, earth, and physical sciences. Fine-tuned models achieve substantial improvements on benchmarks (TQA, ScienceQA, AI2D), outperforming state-of-the-art VLMs with fewer training instances. Augmenting LLaVA OneVision with SciGram establishes new state-of-the-art performance in diagram question answering.
vision-language modelsscientific diagramsmultimodal supervisionatomic factsdiagram-grounded instruction
Embedded Conditional Independence Tests for Large Language Model Generated Text with an Application to German Parliament Speeches
The paper introduces embedded conditional independence tests (eCITs) for assessing whether LLM-generated text $X$ retains information about attribute $Y$ beyond source text $Z$. By embedding $X$ and $Z$ and applying existing CITs to the representations, the method ensures validity if embeddings retain $Z$'s information about $Y$ or $X$'s representation. Theoretical conditions for hypothesis equivalence are derived, with mean sufficiency sufficient for conditional mean independence. Evaluated via semi-synthetic simulations on German Parliament speeches, eCITs detect faction and gender information in LLM summaries across all tested embedding maps.
conditional independence testslarge language modelsembedding sufficiencymean independencemultimodal data
DualStake: Dual-Path Confidence Calibration in Deep Research Agents
Proposes DualStake, a dual-path confidence calibration method for Deep Research agents that jointly aligns Evidence Confidence (E-Conf) and Answer Confidence (A-Conf) with correctness. The method applies margin-clipped, confidence-dependent rewards to mitigate overconfidence while preserving accuracy. Experiments on Qwen2.5-7B, Qwen2.5-7B-Instruct, and Qwen3-4B across 8 QA benchmarks show improved calibration without performance degradation.
confidence calibrationdeep research agentsevidence confidenceanswer confidencemargin-clipped rewards
Context-Grounding Gains Are Mediated by Pre-existing Machinery: Auditing GRPO, SFT, and DPO
The study investigates whether post-training methods (GRPO, SFT, DPO) improve language model grounding by leveraging pre-existing mechanisms or requiring new ones. Using nine post-training variants across scales and model families, the authors estimate a grounding direction from a starting checkpoint. Results show GRPO variants yield minimal grounding gains, while conflict-SFT moderately improves grounding and DPO achieves near-ceiling performance on matched distributions. Both conflict-SFT and DPO primarily utilize the same causal attention heads as the starting model. Subtracting the starting-model direction suppresses gains, while adding it recovers 35% of DPO's improvement. Findings suggest grounding gains largely depend on pre-existing model machinery.
grounding directionpost-trainingcausal attention-headsconflict-sftdpo
Beyond the Image Plane: World-Grounded Queries for Multi-Object Tracking
The authors propose PLANET, an end-to-end multi-object tracker that addresses limitations of monocular tracking by incorporating 3D scene geometry. The method lifts 2D tracking datasets to 3D, forms world-grounded queries using reconstructed geometry in features and positional encodings, and employs an auxiliary 3D location prediction task. A dual-resolution temporal memory preserves spatial evidence across frames. PLANET achieves state-of-the-art performance on three diverse benchmarks by explicitly modeling 3D relationships beyond traditional image-plane constraints.
multi-object tracking3d reconstructionpositional encodingtemporal memoryworld-grounded queries
VIBE-Bench: Evaluating Personalized Large Language Models When Profiles Don't Mean Preferences
The authors introduce VIBE-Bench, a benchmark for evaluating Personalized Large Language Models (PLLMs) in scenarios where user profiles and query-specific preferences are conceptually misaligned (PRCM). The benchmark comprises 3,504 personas and 12,239 dialogues across two psychology-grounded tasks, including a manually verified gold test set, requiring cross-concept preference reasoning. Experiments reveal that current PLLMs rely on shallow semantic correlations and fail at robust cross-concept mappings, establishing PRCM as a distinct failure regime. VIBE-Bench provides a focused testbed for advancing preference reasoning beyond semantic matching.
personalized large language modelspreference reasoningconceptual misalignmentsemantic retrievalbenchmark evaluation
RPCBench: A Benchmark for Proactive Premise Critique in LLM-based Recommendation
The paper introduces RPCBench, a benchmark for evaluating Recommender-Premise Critique (RPC) in LLM-based recommendation systems, focusing on detecting and handling flawed premises in natural-language requests. RPCBench includes evidence-grounded test instances from five domains, covering ten premise failure types, and assesses detection, localization, handling strategy, and evidence faithfulness. Evaluation of 11 LLMs reveals proactive detection as the primary bottleneck, with poorest performance on underspecified-premise errors, and identifies target-critical information density and intermediate reasoning length as key performance factors.
recommender-premise critiqueevidence-grounded evaluationpremise failure detectionllm-based recommendationreasoning length penalty
In-Context Neurofeedback: Can LLMs Control Their Internal Representations through Privileged Access?
The study investigates whether large language models (LLMs) can control their internal representations under stricter privileged access conditions, challenging prior claims of such capability. Researchers redesigned the neurofeedback paradigm to ensure control targets are privileged, unlike previous setups where targets could be inferred from prompts. Under this rigorous setting, LLMs failed to demonstrate reliable control over privileged representations, suggesting earlier findings may rely on superficial mechanisms. The results highlight the need for privileged-access evaluation methods to assess LLM metacognition accurately.
large language modelsneurofeedbackprivileged accessmetacognitioninternal representations
Vision-Language-Guided Pseudo-Labels for Unsupervised Domain Adaptation in Semantic Segmentation for Waste Sorting
Proposes a vision-language-guided pseudo-labeling pipeline for unsupervised domain adaptation in semantic segmentation, eliminating target-domain annotations. Combines Segment Anything Model (SAM) for class-agnostic region proposals with EVA-CLIP for text-based semantic labeling, augmented by BLIP for language-grounded verification of ambiguous regions. Demonstrates consistent improvements over source-only baselines on synthetic-to-real autonomous driving and lab-to-factory waste sorting, highlighting pseudo-label quality as critical for self-training under domain shift. Cross-modal grounding enables reliable automatic annotation in deployment-critical settings.
unsupervised domain adaptationsemantic segmentationpseudo-labelingvision-language modelsself-training
CARE: Contrastive Anchor-based Rubric Evolution for Large Language Model Post-Training
CARE introduces contrastive anchor-based rubric evolution for post-training LLMs, addressing limitations of static and dynamic rubric approaches in open-ended tasks. The method employs frontier-model-generated anchor responses to guide rubric updates via adaptive repair and proactive quality gap conversion, maintaining discriminative accuracy in high-reward regions. Evaluated on WildChecklist-9K with Qwen2.5-7B and Llama-3.1-8B, CARE achieves SOTA on Arena-Hard-2.0, InfoBench, and FollowBench, showing sustained GPT-4.1 win-rate improvement over 300 steps while generalizing across model families.
rubric-based reinforcement learningreward misspecificationdiscriminative accuracypost-traininganchor responses
CacheBridge: Efficient Cross-Model KV Cache Transfer
CacheBridge introduces efficient cross-model KV cache transfer by co-designing architecture-indexed mapper support, attention-aligned calibration, and bounded mapper construction while retaining a closed-form affine interface. The method restricts each target KV head to a matched source head, weights reconstruction errors by causal attention sensitivity, and uses a fused GPU kernel for efficient construction. Results show CacheBridge recovers accuracy in Ministral 3 transfers where Full-Head Mapping fails, achieves 99.83% mean target retention on Qwen3, reduces mapper storage by 8×, accelerates application by up to 3.0×, and cuts construction time from 92.63 to 8.63 seconds (10.7×) on Qwen3 14B→32B transfers.
kv cachecross-model transferattention alignmentclosed-form mappinggpu acceleration
Denoising Diffusion Generative Models Secretly Calculate Attentions
The paper reveals that denoising diffusion models inherently employ an attention mechanism akin to transformers, suggesting attention as a universal principle in machine learning. By reformulating diffusion frameworks using attention, the authors propose a simplified algorithm for image generation that reduces training time and computational costs. Experimental results demonstrate comparable performance to traditional diffusion models while requiring fewer resources.
denoising diffusion modelsattention mechanismimage generationtransformerscomputational efficiency
Towards reliable multimodal disaster severity assessment through preference optimization and explainable vision-language reasoning
The study introduces a two-stage framework for multimodal disaster severity assessment, combining Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) to improve both prediction accuracy and explanation quality. The method derives complementary datasets (ReasoningSet and PreferenceSet) from Human-in-the-Loop annotations, evaluating performance via automatic metrics, model-based scoring, and human ranking. Results show SFT increases accuracy from 73.64% to 78.29% (29% Macro-F1 gain) and explanation quality by ~25%, with further interpretability gains from DPO. Cross-validation on InternVL-3-8B and LLaVA-1.5-7B confirms robustness, particularly for mild damage cases and high-risk misclassifications.
multimodal assessmentsupervised fine-tuningdirect preference optimizationhuman-in-the-loopexplanation quality
FractalNet-Based Heterogeneous Federated Learning for Orbital Edge Intelligence in Satellite Mega-Constellations: A Wildfire Case Study
The paper proposes FractalNet-based heterogeneous federated learning for satellite mega-constellations, addressing SWAP-C and orbital heterogeneity via depth-adaptive model partitioning and contact-aware distributed scheduling. The method employs a three-tier control plane for update pooling, anomaly handling, and autonomy, optimizing communication and energy efficiency. Evaluated on wildfire detection, the framework achieves multi-tier situational awareness (pixel-scale in LEO, regional in MEO, large-scale in GEO/HEO), with experiments demonstrating improvements in convergence, robustness, and latency.
federated learningsatellite mega-constellationsfractalnetswap-corbital edge intelligence
Beyond the Clock: Measuring the Value of Adaptive Revision
The paper investigates meta-level control in hierarchical latent reasoners, focusing on when a higher-level controller should retain or revise strategies guiding lower-level computation. Using a hierarchical model with three training seeds, the study finds that learned revision timing produces diverse policies but fails to outperform the best fixed schedule, separating state dependence from decision value. Results show timing is consequential and order-sensitive, with strong fixed schedules capturing most available value. Counterfactual diagnostics reveal score-level evidence can mislead when predictability stems from decision position. The authors propose evaluating learned meta-level control along three axes: state dependence, behavioral impact, and outcome value beyond non-adaptive policies.
meta-level controlhierarchical latent reasonerstate dependencefixed schedulecounterfactual diagnostics
Benchmarking Vision-Language Models for Automated Pathology Diagnosis and Report Generation
The study introduces REG 2025, a benchmark for whole-slide image (WSI)-based pathology report generation, using a Pan-Asia dataset of ~10,500 WSI-report pairs from five institutions. Evaluated methods include pretrained vision-language models (VLMs), multiple-instance learning frameworks, hierarchical expert models, retrieval-augmented generation, and cross-modal Transformers. Results indicate that top-performing approaches leveraged structured report representations, hierarchical diagnostic decomposition, and multimodal grounding, outperforming VLM-only baselines, but reveal limitations in quantitative attribute estimation (e.g., numeric hallucination) and diagnostic overspecification.
vision-language modelswhole-slide imagemultiple-instance learningretrieval-augmented generationcross-modal transformers
Reinforcement Learning Enhanced LLM Agents for Complex Vehicle Routing Problems
The paper proposes Reinforcement Learning Enhanced LLMAgents (RLEA), a multi-agent framework for automating the modeling of complex Vehicle Routing Problems (VRPs). RLEA combines a lightweight neural Planner trained with Soft Q-learning to coordinate LLM-based agents, augmented by an evolutionary memory module and retrieval-augmented generation for leveraging experience and solver knowledge. Evaluated on 48 VRP variants, RLEA achieves a 16.67% higher success rate than prior methods while reducing runtime errors, demonstrating the efficacy of RL-enhanced LLM reasoning for optimization modeling.
vehicle routing problemssoft q-learningretrieval-augmented generationmulti-agent frameworkevolutionary memory
Verifiable Disaster Storylines and Causal Knowledge Graphs: A Citation-Grounded Pipeline from Heterogeneous Humanitarian Sources
The paper presents a verifiable disaster analysis pipeline combining structured EM-DAT records with unstructured ReliefWeb and EMM documents to generate citation-grounded storylines and causal knowledge graphs. Using Retrieval-Augmented Generation, the system extracts 17-field event profiles and constructs causal graphs with source-attributed nodes/edges, ensuring traceability. Evaluation with 18 human evaluators (9 experts, 9 non-experts) demonstrates high retrieval precision (exact metrics unspecified), faithful causal relation extraction, and expert preference for citation-grounded outputs. The pipeline scales to EM-DAT's full catalogue for public release.
retrieval-augmented generationcausal knowledge graphscitation-groundedhumanitarian informaticssituational awareness
Does Fault Localization Beat a Fresh Attempt? A Placebo-Controlled Study of Test-Guided Code Repair
The study evaluates fault localization's efficacy in test-guided code repair by comparing blind whole-solution resampling, spectrum-based localization with suspect-span infilling, and random-span infilling across 488 failing candidates using 26-32B models. Results show localization is rarely available (9.0% of cases) and underperforms blind resampling (3:40, p = 3.0 x 10^-9), with no recovery from wider edits. Localized infilling shows a pooled advantage over random spans (11:1, p = .019) but lacks significance in individual models. Token efficiency favors blind attempts (10.1% success vs. 6.8% for 16 localized attempts).
fault localizationtest-guided repairspectrum-based localizationinfillingcode repair
ADGNet: Asymmetric Dual-text Guided Network for Infrared Small Target Detection
Proposes ADGNet, an asymmetric dual-text guided network for infrared small target detection (IRSTD), addressing limitations of vision-only and single-prompt multimodal methods. Introduces Asymmetric Dual-text Prompt (ADP) with abstract target and detailed background prompts, an Asymmetric Dual-Branch Interaction (ADBI) module for separate text-guided feature optimization, and Adaptive Feature Aggregation (AFA) for dynamic fusion. Evaluated on the constructed AITIR dataset (annotations for IRSTD-1K, NUDT-SIRST, SIRST), ADGNet outperforms 21 SOTA methods in clutter suppression and target preservation.
infrared small target detectionasymmetric dual-text promptmultimodal fusionfeature aggregationsemantic asymmetry
A Checklist to assess the energy and carbon impacts of ML/AI applications in Earth System Modeling
This paper introduces a practical checklist for assessing and mitigating the environmental impact of ML/AI applications in Earth system modeling, addressing the gap between ethical principles and daily research practices. The checklist is structured around the model development pipeline, incorporating metrics for energy consumption and carbon footprint estimation. It synthesizes dispersed best practices from recent literature into actionable guidelines, supported by concrete examples. The framework aims to enable researchers to make informed decisions that reduce computational resource usage while maintaining scientific integrity in climate, weather, and Earth system modeling applications.
earth system modelingmachine learningcarbon footprintenergy consumptionmodel development pipeline
Towards Generalizable Visually Grounded Exploration of Household Devices
The paper introduces VGEBench, a benchmark for evaluating Generalizable Visually Grounded Exploration (GVGE) in Vision-Language Models (VLMs), addressing the limitation of current embodied exploration paradigms that rely on imitation learning and annotated trajectories. VGEBench employs a Logic-Driven State Machine framework to simulate multi-turn interaction loops, requiring agents to achieve goals through active visual perception and feedback-driven correction. Experiments reveal that existing VLMs struggle with translating semantic knowledge into physical execution and maintaining long-horizon state tracking.
vision-language modelsembodied explorationgeneralizable visually grounded explorationlogic-driven state machinelong-horizon state tracking
Probabilistic Model Checking of Autoregressive Neural Sequence Models
The paper introduces a probabilistic model checking pipeline to evaluate autoregressive neural sequence models beyond test-set accuracy, quantifying constraint-violating alternatives and input-space coverage. The method extracts a discrete-time Markov chain (DTMC) from token-by-token generation, verifies PCTL specifications with PRISM, and aggregates verdicts into conservative coverage curves, with soundness guarantees via under-approximation. A CEGAR loop refines intervals, while maximum-likelihood extraction identifies falsifying traces. Case studies on GPT-2 for process planning and a SMILES molecular generator reveal hidden probability mass under sampling and gaps between structural and chemical validity, unreported by test accuracy.
probabilistic model checkingautoregressive modelsdiscrete-time markov chainpctl specificationscounterexample-guided abstraction refinement
Replacing Training with Memory: Listwise Selection for Text-to-SQL
The paper introduces MaP-SQL, a fine-tuning-free listwise selector for Text-to-SQL that replaces training with structured memories. Instead of learning selection criteria as model parameters, it retrieves memories encoding NL-to-SQL mappings from training data, using them as explicit decision criteria. To mitigate positional bias, it aggregates rankings across input permutations, optimized via execution results and pointwise scoring. Evaluated on BIRD-dev, MaP-SQL improves execution accuracy by 2.02 points over R^3-SQL while reducing token usage by 2.92x, demonstrating stable selection without fine-tuning.
text-to-sqllistwise selectionstructured memoriespositional biasexecution accuracy
FLaG: Frequency-Domain Latent-attention Gated Pooling for Token Aggregation
The paper introduces Frequency-Domain Latent-attention Gated Pooling (FLaG), a token aggregation module that operates in the Fourier domain. FLaG re-expresses encoder outputs via rFFT, summarizes spectral tokens with latent queries, applies channel gating, and reconstructs modulated representations for downstream tasks. Evaluated across ESM2-based antimicrobial peptide prediction, ResNet18 on CIFAR-10/100, and RoBERTa-based language tasks, FLaG achieves top performance in 4/4 AMP settings (best Spearman, RMSE, Recall@50), CIFAR-10 top-1 accuracy, and 5/7 language metrics. Analysis reveals low-frequency dominance in early layers, high-frequency sensitivity in final layers, and species-specific spectral allocation via gating.
token aggregationfourier domainlatent attentionspectral gatingcross-modal transfer
Visual Attention Faithfulness in Vision-Language Models is Heterogeneous
The study investigates visual attention faithfulness in Vision-Language Models (VLMs) through causal perturbation analysis, identifying three distinct processing modes: Faithful-Sufficient (top-k attention tokens are necessary and sufficient), Faithful-Distributed (necessary but requiring broader context), and Non-Focal (no localized region is individually necessary). Analysis reveals systematic divergence from human-annotated ground-truth regions, with model attention rankings satisfying comprehensiveness in only ~60% of cases. These patterns persist across VQAv2, VRDU, and ChartQA benchmarks, demonstrating that visual attention faithfulness varies with task demands and model architecture.
vision-language modelsattention faithfulnesscausal perturbationcomprehensiveness gapvisual tokens
HarnessEvolve: Learning from Reference Trajectories for Reliable Agent Self-Evolution
HarnessEvolve introduces a framework for reliable agent self-evolution by learning from reference trajectories to address credit assignment failure, shortcut learning, and catastrophic forgetting. The method decouples execution from evolution via independent modules, using reference trajectories for error signal extraction and dual gating (quality and performance) to filter updates. Experiments across open-domain and enterprise benchmarks show HarnessEvolve outperforms state-of-the-art baselines consistently in reliability and performance.
self-evolving agentsreference trajectoriescredit assignmentshortcut learningcatastrophic forgetting
Polished but Unresolved: Identifying Late-Stage Pressure States in Long-Horizon Tool-Use Agents
The paper introduces the concept of late-stage pressure states in long-horizon tool-use agents, where agents prematurely finalize answers despite unresolved constraints. The authors train a linear probe to detect this pressure from hidden states and employ activation interventions to modulate agent behavior, showing that constraint clarity and action mapping mitigate pressure. They propose Probe-Sensed Pressure Relief (PSPR), a lightweight plugin that applies pressure relief interventions, improving performance across multiple long-horizon benchmarks.
late-stage pressure stateslinear probeactivation interventionstool-use agentspressure relief
AnalysisBank: An Expert Analysis Pattern Library for Financial Report Generation
AnalysisBank introduces a financial report generation method operating at the analytical level by decomposing expert reports into reusable Analyses, each combining a data signal, analytical move, and expert-derived span. The system matches input signals to library entries at inference, applying retrieved moves for composition. Analysis of 550 reports revealed 47-52 signal types and 13 move types in a heavy-tailed distribution. Evaluated on two financial benchmarks with four LLM backbones, AnalysisBank increased novel, data-grounded insights by 1.7-3.7x over structural baselines, with demonstrated transferability to scientific writing.
financial report generationanalytical movedata signalllm backbonesexpert span
One Policy, Any Budget: Internalizing Budget-Aware Search via Reinforcement Learning
The paper introduces AnySearch, a reinforcement learning framework enabling a single policy to perform budget-aware search under variable constraints. The method employs a two-phase training approach: first with explicit budget state injection and structured reasoning prompts for efficient allocation, then via autonomous operation under adaptively sampled budgets. Optimization uses a composite reward balancing accuracy and efficiency, with adaptive weighting based on query difficulty. Evaluations on seven QA benchmarks demonstrate superior performance across budget scales, generalization to unseen constraints, and efficient tool usage without excessive token overhead.
reinforcement learningbudget-aware searchcomposite rewardadaptive weightingmulti-hop qa
Ctrl-F-Resist. Practices, Challenges, and Technical Needs of Civil Society Organizations Monitoring the Far-Right Online
This paper addresses the gap in research on civil society organizations (CSOs) monitoring far-right online activities through a qualitative study of 15 practitioners from 12 Germany-based CSOs. The study explores their practices, challenges, and technical needs, revealing that monitoring remains largely manual due to a lack of tailored tools. Enhanced search capabilities emerged as the most pressing technical need, while skepticism persists regarding automated classification due to trust and legal concerns. The authors propose a conceptual monitoring workflow and implement it in an open-source Telegram monitoring prototype, introducing the 'manual labor trap' concept to explain CSOs' labor-intensive arrangements.
civil society organizationsautomated classificationmonitoring workflowtelegram monitoringmanual labor trap
Towards a Reliable and Practical Eval Pipeline
The authors propose an end-to-end evaluation pipeline for LLM-based software systems, addressing reliability and practical requirements through combined checklist creation and learned aggregation of responses. Their framework enhances agreement across LLM judges and accuracy against human judgments by incorporating self-consistency, explanations, and prediction uncertainty. Empirical results demonstrate the pipeline's effectiveness in improving evaluation quality.
llm-based systemseval pipelinechecklist creationlearned aggregationprediction uncertainty
Agentic programs: an emerging form of scientific software in computational materials science
The paper introduces agentic programs, a novel form of scientific software in computational materials science that integrates deterministic algorithms with bounded LLM-based judgment, task-specific verification, episodic maturation, and full delegation. These programs combine human-like scientific reasoning with computational efficiency, demonstrated through DeMARS, an agentic program for constructing atomistic models from disordered crystal structures. DeMARS exemplifies how LLM-based agents can automate scientific judgments traditionally reserved for humans while maintaining verifiability and scalability.
agentic programscomputational materials sciencellm-based judgmentatomistic modelsdisordered crystal structures
MADS: A Multiview Acoustic Descriptor Set Beyond Standard Spectral Summaries
The authors introduce MADS (Multiview Acoustic Descriptor Set), a 19-dimensional physics-informed audio representation that captures spectral, temporal, mechanical, and stochastic properties beyond conventional spectral summaries. MADS explicitly models sound-generating dynamics through descriptors for excitation, damping, periodicity, impulsiveness, and structural consistency. Evaluated on ESC-10, ESC-50, and MSoS with classical ML models, MADS achieves 81.00%, 52.78%, and 67.48% accuracy respectively, outperforming 26D MFCC and 38D spectral baselines while using lower dimensionality.
audio representationphysics-informed descriptorsspectral-temporal featureshandcrafted featuresacoustic classification
Instella-MoE Technical Report
Instella-MoE introduces a fully open 16B-parameter Mixture-of-Experts (MoE) language model with 2.8B active parameters per token, trained on AMD Instinct GPUs using novel architectural components (Gated Multi-head Latent Attention, FarSkip-Collective) and a multi-stage pipeline (pre-training, long-context extension, DPO, RL with Multi-Teacher On-Policy Distillation). It achieves 76.7 average score on pre-training benchmarks, outperforming open models like OLMo-3-7B and OLMoE-1B-7B, and maintains competitiveness with open-weight baselines. The post-trained Think checkpoint scores 73.2 on instruction/QA tasks, surpassing comparable models. Full model weights, training code, and configurations are released for reproducibility.
mixture-of-expertsgated multi-head latent attentiondirect preference optimizationmulti-teacher distillationsparse activation
StudyBench: Can Self-Evolution Squeeze Textbooks for Olympiad Capability?
StudyBench introduces a controlled physics benchmark to measure self-evolution methods' efficiency in converting training material into problem-solving capability. The benchmark comprises an Application Set, evaluating absorption via textbook problems, and a Transfer Set, assessing transfer via olympiad-level problems. Testing three base models reveals that improvements on the Application Set rarely extend to the Transfer Set, exposing a Guidance Gap and Compute Plateau. Results indicate that the remaining gap is methodological rather than data- or compute-related. StudyBench provides a measurable target for advancing self-evolution research.
self-evolutiontransfer abilityguidance gapcompute plateauolympiad-level problems
When Features Become Instances: Inverted Contrastive Learning for Unsupervised Feature Selection
Proposes Inverted Contrastive Learning for Unsupervised Feature Selection (ICLFS), a feature-wise contrastive framework that treats features as instances by inverting the data matrix and learning consistent representations under structured perturbations via InfoNCE. Uses projector-space embedding magnitude as a saliency signal, refined by Laplacian-Gated Ranking Correction to suppress redundancy. Achieves top clustering accuracy on 10/12 benchmarks, outperforming classical and neural baselines in standard UFS evaluation.
unsupervised feature selectioncontrastive learninginfoncerepresentation consistencylaplacian-gated ranking
Solaris: Towards Interfaces That Are Generated, Not Coded
Solaris introduces an interface world model that generates interactive UIs frame-by-frame in response to user actions, eliminating predefined intermediate representations. The system conditions autoregressive frame generation on mouse interactions, combining few-step distillation and self-training for real-time coherence. A language model interprets user intent to guide interactions, decoupling high-level reasoning from visual rendering. Solaris achieves interactive speeds while enabling open-ended, dynamically generated interfaces without explicit programming. This approach advances a paradigm where interfaces adapt continuously to user intent rather than relying on fixed states.
interface world modelautoregressive generationfew-step distillationinteractive renderinguser intent modeling
VOIM: Training-Free Open-Vocabulary 3D Instance Mapping for RGB-D and Monocular SLAM
VOIM introduces a training-free voxel-grounded instance manager for open-vocabulary 3D instance mapping from RGB-D or monocular RGB, deferring label and instance decisions until evidence accumulates across views. It leverages unmodified off-the-shelf perception models, demonstrating that mapping quality depends on the accumulation stage rather than specific perception components. On ScanNet++, VOIM outperforms OVO-SLAM by 4.8–11.7 mIoU across four perception configurations, achieving 44.07 mIoU vs. 32.37 in a like-for-like protocol. The system also matches monocular RGB baselines on Replica (27.80 vs. 27.50 pooled). Performance varies by regime, with label limitations at room scale and drift limitations at building scale.
open-vocabulary3d instance mappingtraining-freergb-d slammonocular slam
DiagEvo: Diagnosis-Guided Self-Evolution via Hierarchical Error Memory
DiagEvo introduces a diagnosis-guided self-evolution framework for language models, leveraging hierarchical error memory to target unresolved reasoning weaknesses during self-play. The method employs a diagnostician to extract recurring error causes from the solver's failure history, organizing them under skill nodes (Active/Mastered) and balancing cause-targeted generation with free exploration via a challenger. Double-confidence filtering retains intermediate-difficulty questions based on solver consensus. Evaluated on Qwen3-4B, Qwen3-8B, and OctoThinker-8B, DiagEvo achieves 72.3% mean accuracy on five mathematical reasoning benchmarks (4.5pp above R-Zero) and 57.4% across nine benchmarks (1.1pp above DARC), with ablations confirming the contributions of hierarchical memory and confidence filtering.
self-playhierarchical error memorydouble-confidence filteringdiagnosticianskill nodes
Are You Thinking What I am Thinking? : Examining Conceptual Separation in Neural Architectures
The study investigates conceptual separation in neural networks, analyzing whether examples of the same concept form coherent internal representations and whether related concepts cluster geometrically. Using geometric and distributional analysis of activations in Convolutional Neural Networks (CNNs) and Large Language Models (LLMs), the authors find that CNNs exhibit semantically ordered representations for familiar ImageNet concepts, with degradation for unseen concepts and domain shifts. LLMs maintain separation for distinct domains but collapse distinctions for ambiguous topics at mean and covariance levels. This approach reveals structural insights beyond output accuracy, serving as a diagnostic for concept robustness. Code and data are publicly available.
conceptual separationinternal representationsgeometric analysisconvolutional neural networkslarge language models
Automated Tree Knowledge Graph Construction using Ontology Expansion and Retrieval from Vietnamese History Textbooks
The paper proposes an automated pipeline for hierarchical knowledge graph (KG) construction and retrieval from Vietnamese history textbooks, addressing low-resource language challenges. The method employs a three-phase hybrid relation extraction pipeline (intra-batch deduplication, cross-batch search, LLM extraction with centroid filtering) and a dual-LLM validator to control ontology growth. A two-tier node architecture preserves document structure. Three graph traversal strategies (Top-Down, Horizontal, Bottom-Up) are evaluated on 1,210 synthetic Vietnamese queries, with Top-Down outperforming vector baselines by 4.7pp in NDCG@10. The resulting KG contains 750 nodes and 4,341 edges from 400 textbook pages.
knowledge graph constructionontology expansionlow-resource languagesretrieval augmented generationgraph traversal strategies
S^3martCirc: Self-supervised Smart Circuit Discovery
S^3martCirc introduces a unified framework for mechanistic interpretability in LLMs, jointly discovering circuits and interpreting functionality by abstracting node behavior into two general computational roles with quantifiable metrics. The method addresses limitations of sequential approaches by codependently determining component importance and functional roles, enhancing generalization across tasks. Experiments demonstrate superior performance over existing circuit discovery methods, validated through quantitative metrics and task-agnostic role assignments.
mechanistic interpretabilitycircuit discoveryself-supervised learninglarge language modelscomputational roles
ContextPipe: Database-Inspired Context Assembly for Long-Horizon Agents
ContextPipe introduces a database-inspired pipeline for context assembly in long-horizon LLM agents, addressing inefficiencies in ad hoc prompt construction. The method formalizes context assembly as a five-phase query execution process (Plan, Bind, Optimize, Execute, Feedback) with a structured catalog, deterministic cache-aware optimizer, and execution traces. Evaluated on SWE-bench Pro's Qutebrowser subset, it reduces token volume by 31%, LLM calls by 23%, and response time by 9% compared to append-only baselines, albeit with lower KV cache-hit ratios. The system ensures auditable, replayable, and failure-isolated context handling.
context assemblyquery executionkv cachellm agentsdeterministic optimizer
Escaping Redundant Reasoning: Structure-Aware Search for Inference-Time LLMs
The paper introduces BASIN, a training-free method to mitigate reasoning basin collapse in inference-time LLM search by grouping reasoning states into basins and penalizing redundant strategy revisits, thereby diversifying exploration under fixed compute. BASIN improves Tree of Thoughts (ToT) by +22pp on Game of 24 and +6.7pp on MuSR, with QA-BASIN further enhancing robustness via quality-aware basin preservation. The redundancy gap Δ quantifies search concentration differences between correct/incorrect predictions, showing BASIN consistently shifts Δ positive compared to ToT's Δ≈0. Code is available at GitHub.
reasoning basin collapseinference-time searchstructure-aware selectionredundancy gaptree of thoughts
Agentic Empirical Asset Pricing: Methodological Foundations
The paper introduces Agentic Empirical Asset Pricing (AEAP), a paradigm leveraging LLM agents for autonomous scientific discovery in asset pricing. It proposes a reference architecture, rigorous evaluation standards for factor discovery, and out-of-sample backtesting methods for discovery systems. The authors evaluate SEADS, an AEAP instance, against five baselines on two US equity panels, finding no single metric consistently ranks systems, necessitating multi-axis evaluation. Rolling re-execution tests reveal reliability challenges in the discovery process itself. Negative findings highlight evaluation pitfalls for future AEAP systems.
llm agentsfactor discoveryout-of-sample backtestingautonomous systemsevaluation metrics
SOVER: Formal Certification of Optimization Reformulations via LLM-Assisted SMT Verification
SOVER introduces an LLM-assisted SMT framework for formally certifying optimization problem reformulations, separating semantic mapping (handled by LLMs) from formal verification (using Z3 for mixed-integer linear and dReal for continuous nonlinear formulations). The method validates domain cross-feasibility, objective-order preservation, and ε-argmin properties. Evaluated on NLEquiv-150, a benchmark with 100 equivalent and 50 non-equivalent pairs, SOVER achieves 99.33% accuracy (149/150), correctly identifying all hard negatives; the single error stems from incomplete LLM mapping extraction.
optimization reformulationsmt verificationllm-assistedformal certificationnonlinear equivalence
Heard but Not Heeded: Paralinguistic Information Encoding and Loss in Audio-Language Models
This work identifies a systematic gap in paralinguistic information encoding and utilization across audio-language models. The authors analyze Whisper-large-v2, Qwen2-Audio-7B Instruct, Qwen2.5-Omni-7B, and Chroma-4B using the Expresso dataset with controlled speaking styles, employing centered kernel alignment, linear probing, tone prediction, and a content-prosody leakage metric. Results show that while speaking style is strongly encoded in late encoder layers, this information degrades before reaching the output. Models exhibit either content-driven or acoustic-driven behaviors at the output level, quantified by the leakage metric. The findings reveal a key limitation in current audio-language models' ability to utilize encoded paralinguistic information.
paralinguisticaudio encodercentered kernel alignmentcontent-prosody leakagelinear probing
A Closed-Loop Evaluation of Capability Loss and Recovery in Compressed Driving Policies
The study introduces a closed-loop evaluation framework for assessing capability loss and recovery in compressed driving policies, formulated as a POMDP and trained via PPO in Gym-Duckietown. The method evaluates structured pruning, knowledge distillation, and integer quantization across five driving curricula. Results indicate structured pruning causes initial capability loss, distillation offers limited improvement constrained by rehearsal data, and quantization affects stop-resume tasks unless applied to unpruned actors, preserving all curricula.
pruningknowledge distillationinteger quantizationpartially observable markov decision processproximal policy optimization
ChatDev 2.0: A No-Code Multi-Agent Platform for Developing Everything
ChatDev 2.0 (DevAll) introduces a no-code multi-agent system (MAS) platform combining expressiveness and usability for developing heterogeneous agent workflows. The method employs a declarative executable graph abstraction and cycle-aware execution engine to support dynamic, cyclic interactions, alongside a visual interface for codeless authoring and monitoring. Experiments show DevAll replicates state-of-the-art MAS performance across three tasks without task-specific orchestration code, demonstrating general-purpose utility for LLM-based MAS.
multi-agent systemsdeclarative programmingno-code platformllm orchestrationheterogeneous agents
Differentially Private Paired Table-Image Multimodal Synthesis
The paper introduces DP-TabImage, a differentially private framework for synthesizing paired tabular and image data by factorizing the joint distribution as $p(x,y)=p_T(y)p_I(x\;|\;y)$. It employs a private Probabilistic Graphical Model for tabular data and a table-conditioned diffusion model trained with DP-SGD for images, enhanced by pretraining on private table-image prototypes to improve conditional learning under DP constraints. Evaluations on three datasets demonstrate balanced performance in tabular fidelity, image fidelity, and cross-modal alignment, with analysis showing that visual warm-up enhances marginal image fidelity while aligned warm-up improves cross-modal correspondence.
differential privacymultimodal synthesisprobabilistic graphical modeldiffusion modeldp-sgd
Value Over Language Model: Detecting Original Contribution in Writing
The paper introduces Value Over Language Model (VOLM), a framework for quantifying human contribution in LLM-assisted writing by comparing document reconstructions from partial content representations versus task descriptions alone. VOLM operates without training data or surface-text analysis, using granular content extraction and LLM-based reconstruction to isolate informational value from stylistic variation. Evaluations on news articles, ICLR reviews, and essays show VOLM distinguishes human-authored from LLM-generated texts (using generic prompts) and resists content-preserving transformations like translation, with constrained extractors reducing residual stylistic differences.
llm-assisted writingcontent extractionvalue quantificationstylistic invariancereconstruction-based evaluation
A Study of Hidden-State Optimization Order in Predictive Coding Networks
The study introduces a boundary-first inference schedule to improve feature learning in predictive coding networks (PCNs), addressing limitations of unstructured local objectives in deep networks. The method partitions models into chunks, first optimizing hidden states at chunk boundaries before refining internal representations. On CIFAR-10, this approach boosts accuracy by 9.77% (standard parametrization) and 5.51% (μ-parametrization) over baseline PCNs, with diagnostic analyses revealing stronger early-layer updates, reduced CKA similarity, and more diverse gradients. The results advocate for chunk-based inference as a design principle for PCNs and broader local-learning systems.
predictive coding networkslocal learningboundary-first inferencefeature learninggradient diversity
Visual Framing for News Stance Detection via Image Generation
VFStance improves article-level news stance detection by leveraging visual framing through image generation to make implicit stance cues explicit. The method addresses challenges in detecting subtle, structurally complex stances in news articles by transforming textual framing into visual representations. Evaluation shows VFStance outperforms existing methods, with a user study (N=200) confirming its effectiveness in making stance signals visually salient in snippet-based news consumption.
stance detectionvisual framingimage generationjournalistic framingnews articles
Triple-Bottom-Line Sustainability of Language Models for Edge AI: A Comparison Between SLMs and Quantized LLMs
This work introduces a Holistic Sustainability Score (HSS) to evaluate edge-deployable language models across economic (capability, efficiency), environmental (GPU energy), and social (harmful-prompt robustness) pillars. The study compares five BF16 small language models (SLMs) and five large language models (LLMs) under five quantization approaches (BF16, INT8, NF4 4-bit, GPTQ 4-bit, GGUF Q4), measuring 30 configurations on zero-shot benchmarks, latency, throughput, VRAM, energy, and attack success rates. Results show quantized LLMs (Qwen3-30B-A3B/GGUF Q4, HSS 93.38; Mistral-Small-24B/GGUF Q4, HSS 92.40) outperform native SLMs (best: Phi-4-mini/BF16, HSS 89.49), challenging the assumption that SLMs are universally more sustainable for edge deployment.
holistic sustainability scorequantizationedge-aizero-shot benchmarksharmful-prompt robustness
Drift-Aware LLM Routing with Sparse Contexts and Shared Budgets
The paper introduces Drift-Aware Sparse Routing (DRS), a method for dynamically routing requests in multi-model language services under nonstationary conditions and multiple resource constraints. DRS combines pessimistic reward estimation, optimistic cost projection, online shadow price updates, and hard metering to handle workload-level budgets for compute, latency, memory, or cost. Theoretical analysis shows regret bounds against a dynamic fluid benchmark, with terms for prediction radii, capacity buffer, and pacing. Under sparse linear assumptions, the regret scales as $\widetilde O(T\sqrt{s/(ρW)} + WV_T + \sqrt{T})$, where $s$ is sparsity, $ρ$ is audit rate, $W$ is window length, and $V_T$ is drift magnitude.
multi-model routingsparse contextual banditsnonstationary optimizationresource constraintsregret bounds
SciTrue: Reliable Scientific Claim Validation with Frontier and Open Language Models at the NTCIR SciClaimEval Task
The SciTrue team achieved top performance in the NTCIR-19 SciClaimEval task by benchmarking eleven multimodal models (including Claude Opus 4.8, Gemma-4-31B, GPT-5.5, and Claude Fable 5) under a uniform protocol with light post-processing. Key findings include: (1) instruction-tuned models like Claude Opus 4.8 (97.7 accuracy on Subtask 2) outperformed public baselines; (2) a leak-free pair prior improved Subtask-1 pair-accuracy from 72.2 to 93.5 by inferring claim-evidence pairings; (3) residual errors were largely due to label noise or mapping issues. The study also identified a measurement leak in dataset packaging affecting label integrity.
multimodal modelsinstruction-tunedleak-free pair priorlabel noisemeasurement leak
EEG-AS: Instance-Level Foundation Model Selection for EEG Foundation Models via Behavior Reconstruction
The paper introduces EEG-AS, an instance-level algorithm selection framework for EEG foundation models that reconstructs unavailable model behaviors using latent EEG embeddings, neurophysiological features, and an anchor model. By leveraging inference-available embeddings and privileged prediction tokens during training, EEG-AS efficiently estimates model behaviors without full portfolio execution, enabling adaptive selection from seven foundation models. Experiments on seven EEG benchmarks show EEG-AS significantly reduces the performance gap between the Single Best Solver and the oracle upper bound per instance.
eeg foundation modelsalgorithm selectionneurophysiological featureslatent embeddingsbehavior reconstruction
Self-Reports Are Not Verification: Environment-Grounded Auditing of LLM Operators in Evolutionary Search
The study introduces an environment-grounded audit for language model agents, testing three assumptions about self-reports in evolutionary Contexto search. Using 200 runs across five configurations and three model families, the audit assigns exact ranks to 12,249 proposals without human annotation. Results show operators overstate top-100 success by 4.8–9.3×, with no measurable benefit from inherited rationales (bounded to ~250 ranks) or fitness-based selection. Self-reports dissociate from actual reliability, necessitating external verification.
language model agentsevolutionary searchself-report verificationcontextofitness-based selection
DramaChain Bench: An End-to-End Benchmark for Short-Drama Generation
DramaChain Bench introduces the first end-to-end benchmark for short-drama generation, evaluating all stages of production—script, storyboard, keyframe imagery, shot-level video, and final drama—against five evaluation axes instantiated across 63 leaf dimensions. The benchmark employs DramaChain Agent, calibrated against commercial platforms, and DramaChain Labeling System, with 5,785 items scored by three annotators, yielding 17,488 scores and 255,925 traceable records. Results show upstream defects cascade through the pipeline, impacting final quality beyond video generation alone. DramaChain Agentic Judge automates scoring, achieving a mean PLCC of 0.918, enabling model evaluation without additional annotations.
short-drama generationevaluation axesagentic judgetraceable attributioncascading defects
REVISE: Validity-Guided Recovery for Online Revisions in Agent Workflows
We present REVISE, a validity-guided runtime for fine-grained recovery in structured agent workflows, addressing the correctness-efficiency trade-off during concurrent execution. REVISE intersects revision deltas with recorded dependencies, propagates impact through the DAG, stops invalid work, preserves valid progress, and recomputes only affected regions. Analysis of coding-agent traces reveals online recovery opportunities, with 118 sessions retaining observable work and 167 overlapping responses showing 56.55s enqueue-to-completion overlap at p95. REVISE matches a latest-version oracle with no stale outputs, reduces Qwen3-14B model calls by 40.6-56.0% versus full restart and 31.3-43.6% versus suffix recomputation, and improves SLO goodput by 3.07-5.43% under serving pressure.
validity-guided recoverystructured agent workflowsdependency propagationonline recoverymodel call reduction
TUTTI: Toward generalizable audio-to-score transcription via fully synthesized data
TUTTI introduces a synthetic-data-driven pre-training paradigm for generalizable Audio-to-Score (A2S) transcription, addressing data scarcity in real-world paired datasets. The method leverages a symbolic music generation model to create TuttiCorpus, a large-scale multi-instrumentation synthetic dataset, and trains a Transformer encoder-decoder. Pre-training on synthetic data yields stronger foundational representations than single-instrumentation training, achieving state-of-the-art results when fine-tuned on real-world datasets. TUTTI demonstrates exceptional cross-instrument transferability, adapting competitively to unseen instruments.
audio-to-score transcriptionsynthetic datasettransformer encoder-decodercross-instrument transferabilitymulti-instrumentation pre-training
Breaking the Structural Identity: Personalized Federated LoRA Fine-tuning under Rank Heterogeneity
FedRoRA introduces rank-wise personalized LoRA for federated learning, addressing resource and data heterogeneity by decoupling adaptation into shared global directions and client-specific magnitudes via learnable diagonal scales. The method employs SVD for global subspace extraction and a personalized projection mechanism for client initialization. Evaluations on NLU and NLG benchmarks show consistent improvements over state-of-the-art approaches.
federated learninglow-rank adaptationrank heterogeneitysingular value decompositionnon-iid data
Restrict, Don't Retrain: Inference-Time VLM Guidance for Zero-Shot Aerial Segmentation
The paper introduces an inference-time guidance method for zero-shot aerial segmentation using vision-language models (VLMs), avoiding retraining of foundation models. The approach combines a frozen foundation model for pixel-level labeling with two VLM queries: one to select relevant classes and another to locate small objects missed by the base model. Evaluated on four aerial datasets, the method consistently improves segmentation accuracy, particularly for smaller objects, while providing auditable evidence for the results. The system operates on a single consumer-grade GPU, demonstrating practical feasibility.
zero-shot segmentationvision-language modelinference-time guidanceaerial imagerypixel-level labeling
Control-Data Flow Separation: Stable Prompt Optimization in Multi-Agent LLMs
The paper introduces control-data flow separation, a method for stable prompt optimization in multi-agent LLM systems by disentangling execution-critical protocols from task-relevant content. Execution protocols are represented as typed, validated program objects, while task-related language remains unstructured and optimizable. This approach prevents prompt drift from corrupting routing or formatting interfaces. Evaluated on synthetic reasoning, collaborative review generation, and insurance rating workflows, the framework achieves 100% protocol validity while improving task performance.
prompt optimizationmulti-agent llmscontrol-data flowprotocol validitytask performance
Towards Effective Structured Context Modeling for Conversational Recommender Systems via Dual-node Monte Carlo Tree Search
The paper proposes DREAMS, a tree-structured framework for conversational recommender systems that models user preference evolution through dual-node Monte Carlo Tree Search (MCTS). Elicitation nodes use MCTS to explore conversational actions and infer latent preferences, while exploitation nodes employ LLM-based refinement to convert tracked preferences into structured retrieval queries. Experiments on benchmark datasets validate the framework's effectiveness in preference tracking and recommendation.
conversational recommender systemsmonte carlo tree searchpreference elicitationllm-based refinementstructured retrieval
Confess What You Know: Forget-Set Misalignment with Model Knowledge in LLM Unlearning
The paper introduces CONfession-to-Forget-Set (CONFS), a data-blind framework addressing forget-set misalignment in LLM unlearning, where pre-defined forget sets often mismatch model-memorized information. CONFS constructs model-aligned forget sets by eliciting and formalizing the model's knowledge, mitigating Under Unlearning (persistent leakage) and Out-of-Knowledge Unlearning (utility degradation). Gradient-level analysis reveals these issues stem from misaligned unlearning targets, not optimization choices. Evaluated on synthetic, multimodal, and real-world benchmarks, CONFS achieves Gold-standard performance in forgetting metrics while maintaining utility better than other data-blind approaches.
machine unlearningforget-set misalignmentgradient-level analysisdata-blind frameworkutility degradation
SoK: When Safe Agents Fail Together: The Security of Multi Agent LLM Systems
The paper systematizes security risks in multi-agent LLM systems (MAS) through an execution-centered analysis of 197 works, introducing the A-I-R framework to unify attacks by adversary position, interaction interface, and system-level risk. It identifies six interaction interfaces, four adversary positions, seven system-level risks, and eight recurring attack paths, proposing a five-part defense contract targeting path closure and recovery. An audit of 44 evaluation works reveals challenges in isolating interaction effects and designing diagnostic metrics, advocating for interaction-aware security practices with end-to-end attack tracing and counterfactual evaluation.
multi-agent llm systemssecurity risksinteraction interfacesattack pathscounterfactual evaluation
Socrates went Nuclear: Comparing Interaction Strategies for AI systems in a Learning Context using Brain Sensing
This study compares three AI interaction designs for learning: unrestricted chatbot (ChatGPT-like), pedagogically constrained Socratic bot (hint-based), and EEG-driven adaptive tutoring. Using 50 participants learning nuclear safety protocols with Muse headband monitoring, results show unrestricted chatbots yield higher immediate learning gains (δ > 0.80) but promote answer-retrieval strategies, while Socratic mode users initially engage with hints before disengaging. EEG data indicates adaptive tutoring elicits significantly higher cognitive engagement (p = .018), suggesting unrestricted AI's performance may reflect test-taking strategies rather than deeper learning.
cognitive engagementadaptive tutoringsocratic learningeeg monitoringknowledge acquisition
Predicting Program Exit Code with LLMs and Programming Language Semantics
The paper introduces Program Executability Prediction (PrEx), a novel task assessing whether large language models (LLMs) rely on pre-training priors or apply given programming-language semantics to predict program validity. Using a dataset of systematically generated invalid transformations from valid programs, the authors evaluate open-source coding LLMs under two semantic formalisms and shifts across Human-Written, LLM-Translated, and Fuzzer-Generated splits. Results indicate LLMs predominantly rely on pre-training priors, performing poorly on modified semantics and degrading with program complexity.
program executability predictionsemantic formalismspre-training priorsinvalid transformationsprogram complexity
Same Request, Different Boundary: Evaluating Cybersecurity Assistance across Conversational Contexts
3R-Bench introduces a benchmark of 150 real-world cybersecurity requests with adversarial conversational settings to evaluate LLM refusal mechanisms. It assesses eight models by examining how prior assistant behavior (refusal, acceptance) affects compliance rates for unchanged requests. Results show compliance rises from 62.0% after refused history to 85.1% after accepted history, while dialogue decomposition reduces compliance from 501/800 direct responses to 172/800, with a 45.1-point drop in model-authored text pairs. Failure feedback minimally mitigates this loss.
llmscybersecurityrefusal mechanismsconversational contextcompliance rates
GeoPAR: Large-Scale Multi-Agent Combinatorial Optimization with Geometry-Guided Parallel Autoregressive Learning
GeoPAR introduces a geometry-guided parallel autoregressive reinforcement learning framework for scalable multi-agent combinatorial optimization, addressing performance degradation in large-scale instances. The method combines: (1) a projection-window sparse geometry mechanism for lightweight local neighborhoods, (2) sparse edge-biased attention to inject geometric relations into node representations, and (3) cache-guided conflict-aware assignment to suppress duplicate task selections. Evaluations on heterogeneous vehicle routing and open multi-depot pickup-and-delivery problems demonstrate improved zero-shot generalization, reduced rollout steps, and efficient inference.
parallel autoregressive learningmulti-agent combinatorial optimizationsparse geometry mechanismconflict-aware assignmentzero-shot generalization
Consistency Without Alignment: Item-Sensitive Language Models Indistinguishable From Random
The study demonstrates that item-sensitivity in language models, while necessary, is insufficient to indicate task competence, introducing the concept of 'consistency without alignment'. Using a forced-choice signalling task derived from 'Deception: Murder in Hong Kong', the authors evaluate seven language models across two families, a post-training ablation, and three scoring rules. Results show that 8 of 21 model-by-rule cells are indistinguishable from random selection, with item-sensitivity and distance from random weakly correlated (r = 0.30), and a literal-similarity baseline outperforms most models. The findings challenge evaluations relying on item-sensitivity or self-consistency without independent reference.
item-sensitivityforced-choice taskconsistency without alignmentlanguage modelspragmatic layer
Residual Sparsification via Output Importance for Compressing Mixture-of-Experts LLMs
PARSER introduces residual sparsification via output importance to compress Mixture-of-Experts (MoE) large language models more effectively. Unlike existing methods that minimize isolated matrix errors, PARSER preserves expert output error by measuring contributions to the final output through coupled computations across projections and hidden representations. Experiments demonstrate that PARSER reduces the accuracy gap to uncompressed models by 1.41× on Qwen and 1.44× on DeepSeek while maintaining equivalent memory reduction.
mixture-of-expertsresidual sparsificationoutput importancecompressionlarge language models
A Mathematical Framework for Legacy, Governance, and Decision Integrity in Enterprise AI
The paper proposes a mathematical framework for evaluating institutional legacy in enterprise AI systems, defined as sustained decision quality amid personnel, model, and regulatory changes. Key contributions include: (i) a Legacy Score combining knowledge retention, governance, and adaptability metrics; (ii) Decision Confidence/Risk models; (iii) authority-aware retrieval; (iv) Decision Memory for organizational learning; (v) Regulatory Change Velocity analysis; and (vi) a federated regulatory knowledge-graph. A computational demonstration with 200 Monte Carlo replications of 10,000 synthetic decisions validates non-compensatory Legacy Score behavior and compares routing strategies. The framework remains conceptual, requiring context-specific calibration.
legacy scoredecision confidenceregulatory knowledge-graphauthority-aware retrievaldecision memory
VoiceLongMemEval: Do Assistants Remember How You Sounded?
The paper introduces VoiceLongMemEval (VLME), a benchmark assessing AI assistants' ability to utilize paralinguistic metadata (emotion labels, prosody descriptors, voice events) in multi-session conversations, which is unrecoverable from text transcripts alone. VLME employs a three-stage adversarial gate to ensure failure by text-only models. Evaluations show a 0.09–0.38 accuracy gain when models access paralinguistic metadata (0.61–0.69 with evidence hints), while audio-native models achieve 0.354–0.412 accuracy by extracting cues directly from speech versus 0.325 without audio. Standard ASR pipelines discard this signal systematically.
paralinguistic metadatamulti-session conversationsadversarial gateprosody descriptorsaudio-native models
WiseSpec: Requirements-Driven Agents for Code Generation
WiseSpec introduces a requirements-driven agent framework for repository-level code generation, addressing LLMs' limitations with incomplete or ambiguous task descriptions. The method automatically constructs structured requirements, evaluates them via execution-based assessment, and iteratively refines them to guide code synthesis. Experiments demonstrate a 13.17% average improvement in %Resolved over baselines, highlighting the impact of requirement quality on generation accuracy.
code generationrequirements engineeringlarge language modelsexecution-based evaluationrepository-level synthesis
EEG-VID: Task-Guided Latent Predictive Pretraining for EEG Decoding and Assistive Target Selection
EEG-VID introduces a task-guided latent predictive pretraining framework for EEG decoding, addressing session and subject shifts via future latent state prediction using an exponential-moving-average target encoder with weak task guidance. The method demonstrates improved mean accuracy in 41 of 42 backbone-dataset-protocol comparisons on VIG-48 and BCI Competition IV-2a/IV-2b, with maximum gains of 16.22 percentage points. It achieves 6.52% Top-1 and 30.50% Top-5 accuracy on VIG-48's 48-region cross-day task, and 40.24% target selection accuracy (vs. 25% chance) in a six-participant robot-scene study after calibration.
eeg decodinglatent predictionexponential-moving-averagesubject shiftassistive target selection
EM^2Mem: Event-Centric Multimodal Memory for Large Language Models
EM^2Mem introduces an event-centric multimodal memory framework that addresses the limitations of isolated fragment retrieval in long-video question answering. The method binds heterogeneous evidence to event anchors during memory construction, aligning multimodal records, temporal context, graph-linked relations, semantic facts, and provenance within each event-indexed memory cell. This enables compact evidence readout over grounded multimodal events. Evaluated across three long-video QA benchmarks, EM^2Mem improves average accuracy by 2.0, 2.4, and 3.7 points, enhances strict event-level Top-5 evidence recall by 7.0 points, reduces per-query latency by 4.67 times, and decreases total inference tokens by 63.66%.
event-centricmultimodal memorylong-video qaevidence recallinference tokens
Runtime-Independent Persistent Agents: Preserving Identity, Memory, and Code Across Models, Harnesses, and Servers
The paper proposes a runtime-independent architecture for persistent agents that maintain identity, memory, and code across model, harness, and server changes. The design introduces a continuity-bearing substrate $P_t=(I_t,M_t,B_t)$ for identity, memory, and versioned code, alongside replaceable deployment bindings $E_t=(R_t,H_t,D_t)$ for execution. A quiesce-checkpoint-validate-bind-rehydrate-resume protocol ensures continuity invariants during migration. The implementation, Enoch, passes 833 core and 92 provider/library tests, demonstrating substitutability across reasoner versions, interaction surfaces, and host machines while preserving state. Results support mechanical substitutability but not behavioral invariance.
persistent agentscontinuity invariantsruntime-independentdeployment bindingsubstitutability
Feedback-Assisted Trust Propagation over Document Relation Graphs for Retrieval-Augmented Generation
TrustPropRAG improves retrieval-augmented generation (RAG) reliability by propagating document trust scores through a relation graph, leveraging limited human feedback. The method constructs a document relation graph, formulates an optimization problem to estimate trust scores by jointly modeling pairwise relations and feedback, and uses these scores for trust-aware retrieval and generation. Evaluations demonstrate improved retrieval quality (exact match) and robustness to sparse/noisy feedback compared to baselines.
retrieval-augmented generationdocument relation graphtrust propagationfeedback anchoringtrust-aware generation
Are We There Yet? Assessing Computer-Use Agents for Blind Users' Accessible Interaction with Desktop Applications
The study evaluates the effectiveness of computer-use agents (CUAs) for blind screen-reader users in desktop workflows through a three-week diary study. Eight blind users interacted with OLLA, a screen-reader-accessible CUA prototype, executing 1,258 commands across 12 applications. GPT-5 achieved the highest success rate at 52.5% compared to four additional models. Trace analysis identified failures in grounding, planning, constraint-tracking, and termination, while user interviews highlighted needs beyond automation.
computer-use agentsscreen-readergroundingconstraint-trackingui trees
The Safeguard Worked. Is the LLM System Safer?
The paper evaluates LLM safeguards by analyzing their implications for deployment safety, focusing on residual harmful task assistance after adaptation. It argues that safeguard effectiveness cannot be assessed solely through local metrics like refusal, attack success, and policy violation rates, as these fail to account for system-wide vulnerabilities. The authors demonstrate that establishing safety requires evidence of what the surrounding system permits post-safeguard, a criterion met by only a minority of depth-coded claims. They conclude that improving local scores does not inherently enhance deployment safety, necessitating broader system-level evaluations.
llm safeguardsrefusal rateattack successpolicy violationdeployment safety
The Interlingua Hypothesis: LLMs Translate via a Latent Task-agnostic Feature Space
The paper proposes the interlingua hypothesis, positing that large language models (LLMs) perform machine translation via a latent task-agnostic feature space. The authors present three lines of evidence: (1) BLEU score variance across language pairs is predictable from language-specific competences without pair-specific interactions; (2) model components exhibit causal influence in both monolingual and translation tasks; (3) monolingual fine-tuning recovers a significant proportion of translation improvements compared to aligned document fine-tuning. These findings support the hypothesis and suggest new approaches for leveraging LLMs in translation tasks.
interlingua hypothesislatent feature spacemachine translationlarge language modelsfine-tuning
ISO-RAG: Isoperimetric Noise Control for Retrieval-Augmented Generation
ISO-RAG introduces a geometry-aware Retrieval-Augmented Generation (RAG) framework to address semantic drift and latency in multi-hop question answering. By projecting the knowledge graph into a hyperbolic Poincare ball and precomputing node-wise isoperimetric profiles, ISO-RAG prunes spurious edges, restricting retrieval to localized subgraphs. This approach regulates Personalized PageRank diffusion, ensuring exact and low-latency convergence. Experiments on multi-hop QA benchmarks show ISO-RAG achieves average absolute gains of 10.0% in retrieval recall and 4.3% in downstream exact match, outperforming state-of-the-art baselines.
retrieval-augmented generationmulti-hop question answeringhyperbolic poincare ballisoperimetric profilespersonalized pagerank
When the Algorithm Becomes the Brand Crisis: A Sociotechnical Theory of Distributed Responsibility and Accountable Transparency
The paper develops a sociotechnical process theory to analyze responsibility attribution in AI-related brand crises, distinguishing between incidents, crises, and scandals. Using a federated scoping synthesis of academic and primary sources, the framework posits that incident configuration influences actor-specific attribution, which in turn affects appraisals of capability, integrity, fairness, and relationships. The theory reconciles conflicting findings on algorithmic blame allocation and introduces accountable transparency as a response strategy combining notice, intelligible accounts, and remedy. Evidence supports proximal inferences about blame and trust but not durable reputation effects.
responsibility attributionsociotechnical theoryalgorithmic incidentaccountable transparencybrand crisis
CoVer: Conflict-Aware Claim Verification
The paper introduces CoVer, a conflict-aware claim verification framework addressing evidence-level and aggregation-level conflicts in social media fact-checking. CoVer employs a three-stage pipeline: evidence schema normalization, factual consensus, and support verification to prioritize reliable evidence. The authors also present ContraNote, a dataset with 33,686 posts for conflict resolution and 54,474 instances for prioritization, sourced from X's Community Notes. Evaluations show CoVer achieves 86.0% accuracy on conflict tasks and 88.5% on prioritization, outperforming baselines on ContraNote, CONFACT-HumC (88.4%), and CONFACT-ModC (89.4%).
fact-checkingconflict resolutionevidence aggregationdatasetverification framework
RecalibrateGPT: AI Fatigue Resilient Conversational Interfaces
RecalibrateGPT introduces five cross-turn operators (Anchor, Replay, Delta, Scope, Steer) to mitigate conversational AI fatigue in LLM interfaces, targeting four fatigue types (retyping, scanning, decision paralysis, context drift). The system operates via an AssistiveButton with three palette layouts (Vertical, Arc, Tablet), applying operators to the full conversation history with one click. Two pilot studies (N=12 advanced users) show it halves perceived cognitive workload (NASA-TLX=2.7) while maintaining high usability (SUS=86.5), demonstrating fatigue as an interaction-flow issue addressable through structured recalibration.
cross-turn operatorsassistivebuttonnasa-tlxsuscontext drift
Independent Reinforcement Learning in Discounted Markov Games
The work establishes computational hardness for decentralized learning in discounted general-sum Markov games, proving under ETH for PPAD that no polynomial-time algorithm exists for inverse-polynomially accurate coarse correlated equilibria (CCE) with independent learners. It introduces a layered optimistic mirror descent variant with increasing step-sizes, yielding the first radically uncoupled algorithm achieving sub-exponential convergence to CCE without game structure assumptions. Full- and partial-feedback versions demonstrate sub-exponential convergence rates.
markov gamescoarse correlated equilibriaradically uncoupled learningoptimistic mirror descentsub-exponential convergence
Wave Function Backpropagation with Explicit Temporal-Interval Dynamics
The paper introduces Wave Function Backpropagation (WFB), a novel learning formulation where neural responses are parameterized by wave properties (amplitude, wavenumber, angular frequency, phase) and explicitly linked to temporal intervals Δt through differentiable spatiotemporal waves. The method derives gradients with spatial-curvature corrections via the Laplacian of wave responses. Evaluated on motion prediction, real-interval WFB reduces average displacement error (ADE) by 20.4% versus a feed-forward network (FFN) baseline, while position-only tests show a 10.4% improvement. Shuffled-interval WFB achieves the lowest ADE, suggesting gains stem from wave representation rather than interval alignment. Results validate WFB as a structured feed-forward approach for future architectural exploration.
wave function backpropagationtemporal intervalspatiotemporal waveaverage displacement errorfeed-forward learning
Validity-Aware Jailbreak Evaluation for Large Language Models
The paper introduces Sequential Epistemic and Action-Level Validation (SEAV), a jailbreak evaluation framework for large language models (LLMs) that assesses response validity and correctness beyond linguistic plausibility. SEAV decomposes responses into ordered steps, combining LLM-as-a-judge semantic interpretation with retrieval-grounded verification using external knowledge. Empirical results show SEAV reduces false positives by 14.9 percentage points on the SD-A benchmark and reclassifies 22.1%–51.0% of prior-labeled jailbreak successes as invalid across three public benchmarks, demonstrating that correctness enforcement significantly alters robustness measurements.
jailbreak evaluationlarge language modelssequential validationfactual correctnessretrieval-grounded verification
Beyond Token Positions: Safety Alignment Across Denoising Steps in Diffusion Language Models
This paper introduces Refusal-Aware Early Commitment (RAEC), a training-free decoding method for enhancing safety alignment in diffusion large language models (dLLMs). RAEC leverages the observation that refusal signals are concentrated in early denoising steps and leading response positions, committing persistent refusal signals from these stages. The method was evaluated on LLaDA and Dream, demonstrating reduced attack success rates while preserving utility. Analysis revealed that early token commitment significantly influences final safety outcomes, highlighting the importance of denoising step and refusal-token persistence in dLLM safety behavior.
diffusion large language modelsdenoising stepssafety alignmentrefusal-aware early commitmenttoken commitment
The Privacy-Hallucination Tradeoff in Differentially Private Language Models
The study identifies a privacy-hallucination tradeoff in differentially private (DP) language models, where stricter privacy budgets correlate with increased factual inaccuracies. Through empirical analysis, the authors demonstrate that DP mechanisms flatten output distributions, redistributing probability mass toward incorrect alternatives. Controlled experiments reveal that higher-frequency factual information in training data mitigates hallucination risks, highlighting the need for privacy-preserving methods that maintain factual accuracy.
differentially privatehallucinationprivacy budgetoutput distributionsfactual accuracy
EvoFlint: An Evolutionary Atlas of Multi-Turn LLM Vulnerabilities
EvoFlint introduces an evolutionary quality-diversity search method for systematically discovering and refining multi-turn attack strategies against large language models (LLMs), framing the problem as search rather than generation. The approach evolves phased conversation plans via LLM-driven mutation and crossover, using Pareto fitness over attack success rate and severity, with a risk-indexed archive enabling novelty search. On HarmBench-test, EvoFlint achieves success rates of 35.8% (Claude Sonnet 4.6), 59.7% (GPT-5.4), and 94.3% (Qwen3-32B), revealing gaps in safety training through a structured archive of risk categories.
multi-turn attacksquality-diversity searchred-teamingevolutionary algorithmsrisk-indexed archive
Beyond Scores: Understanding LLM-as-a-Judge Mechanisms in Summarization Evaluation
The study mechanistically analyzes how LLM-based evaluators assess NLG quality, revealing a two-stage pipeline: early layers (below 15) perform local error comparison via attention, while later layers (25-26) integrate signals and crystallize ratings in the residual stream. Using a perturbation taxonomy, controlled error generation, and causal tracing on Themis (Llama-3-8B) and Prometheus (Mistral-7B), the authors identify fine-tuning-specific mechanisms—suppression of early MLP contributions and shifted crystallization depth—that sculpt pre-existing architectural substrates. Code and data are released.
llm-as-a-judgecausal tracingresidual streamfine-tuningattention-head knockout
Facet-0: A Robotic Foundation Model for Contact-Rich Precise Manipulation
Facet-0 is a robotic foundation model for contact-rich manipulation, integrating multimodal representation learning and reinforcement learning to predict action-wrench couplings. It employs flow matching for action-wrench proposal generation, a distributional Action-Wrench Critic for contact outcome discrimination, and phase-aware rewards for policy improvement. The system features a bounded actor for on-robot adaptation and achieves 82% mean success on sub-millimeter assembly tasks (vs. 15% baseline), with 0.5 mm placement accuracy and 50 ms latency, trained on the 1,000-hour ManuFacet-1K dataset spanning three robotic embodiments.
foundation modelaction-wrench couplingflow matchingdistributional criticsub-millimeter assembly
The Structure of Quantization Damage in LLMs: Why the Next Bit Should Be Spent Globally
This work analyzes the structure of quantization damage in large language models (LLMs), demonstrating that precision budgets should be allocated globally rather than to specific layers. Using causal mixed-precision intervention across 9 models (including Qwen3-8B and OpenLLaMA), the study finds that quantization damage is diffuse, with 75% recovery requiring roughly half the layers in 8 of 9 cases. Global granularity outperforms layer-specific repair by 21-52 accuracy points in group-128-compatible models. Secondary findings show 8-bit quantization is near-lossless and recovery patterns correlate with architecture families. The results challenge assumptions about task circuits or weight statistics as primary damage loci.
quantization damagemixed-precision interventionpost-training quantizationlarge language modelsgranularity
Gradient-Update Mismatch: Rethinking Conflict-Free Training of Physics-Informed Neural Networks
The paper introduces Gradient-Update Mismatch (GUM), a phenomenon where conflict-free gradient directions constructed by gradient surgery methods lose this property after optimizer transformation in Physics-Informed Neural Networks (PINNs). To address GUM, the authors propose Gradient-Update Alignment (GUA), which projects optimizer updates onto the conflict-free cone and adjusts the optimizer's internal state accordingly. Experiments reveal that GUM is prevalent across momentum, adaptive, and curvature-based optimizers, with conflict rates up to 86.3%. GUA consistently improves gradient surgery methods, reducing the relative $L_2$ error by up to 98.2% in PINN settings.
physics-informed neural networksgradient surgeryoptimizer transformationconflict-free conegradient-update alignment
NashDreamer: Model-Based Reinforcement Learning for Zero-Sum Imperfect-Information Games
NashDreamer introduces a model-based reinforcement learning framework for two-player zero-sum imperfect-information games, addressing opponent-induced non-stationarity via a centralized Multi-Agent Recurrent State-Space Model (MARSSM) that decouples environment dynamics from strategy effects. The method leverages arbitrary policy gradient algorithms, inheriting their Nash equilibrium convergence guarantees under idealized conditions. Empirical results on four benchmarks show significant early-training sample efficiency gains over model-free baselines. Theoretical analysis reveals vulnerability to posterior collapse in stochastic environments, identified as an open challenge.
model-based reinforcement learningimperfect-information gamesnash equilibriummulti-agent systemsposterior collapse
Variable Selection for Feature-Based Newsvendor
The paper proposes a variable selection framework for feature-based newsvendor problems under cardinality constraints, addressing interpretability and cost challenges in high-dimensional settings. It formulates an ℓ₀-constrained empirical optimization problem with ℓ₂-regularization, proves its computational hardness, and develops a mixed-integer second-order cone programming reformulation alongside scalable randomized-rounding and greedy heuristics. Theoretical analysis provides finite-sample error bounds, out-of-sample risk guarantees, and support recovery for the sparse policy estimator. Experiments on synthetic and real data show the method achieves competitive operational costs while using significantly fewer covariates than baselines.
newsvendor problemvariable selectionℓ₀-constraintmixed-integer optimizationsupport recovery
Quantum Sparse Autoencoders for Q-Matrix Estimation in Cognitive Diagnosis
The paper introduces a quantum sparse autoencoder (QSAE) for Q-matrix estimation in cognitive diagnosis, marking the first application of quantum machine learning (QML) to this task. QSAE encodes binary student response vectors into quantum circuits, compresses them into sparse latent representations, and maps these to Q-matrices. Evaluated on 60 simulated and 9 real-world assessment datasets, QSAE demonstrates complementary strengths to classical autoencoders (CAE): while CAE achieves higher accuracy in some simulations, QSAE shows superior stability (lower variance in 49/60 conditions) and outperforms CAE on 6/9 real datasets, highlighting QML's robustness and ability to handle latent-structure complexity.
quantum sparse autoencoderq-matrix estimationcognitive diagnosisquantum machine learningeducational data mining
Sierpiński--Knopp Wasserstein Distance for Persistence Diagrams and Applications to 2-Wasserstein Approximation
(No summary returned.)
Does Imitation Learning Preserve Temporal Robustness in Dexterous Manipulation? An Expert-Learner Comparison Across Task Execution Speeds
This study evaluates the temporal robustness of imitation learning in dexterous manipulation by comparing expert and learner performance across task execution speeds. Using the ParcelStow task, a scripted expert and an Action Chunking with Transformers (ACT) policy trained from demonstrations were tested. While both achieved 100% success at nominal speed, their performance diverged within the demonstrated speed range: expert success dropped to 84%, while ACT fell to 53%. Stage-level analysis revealed 35 of ACT's 47 failures at maximum speed were insertion misalignments. The findings show that equal nominal success does not guarantee preserved expert performance across speeds.
imitation learningtemporal robustnessdexterous manipulationaction chunking with transformersparcelstow
Diffusion as a Training Curriculum for Timestep-Free Iterative Reasoning
The paper introduces a timestep-free diffusion model for iterative reasoning by adding a persistent hidden state and removing timestep conditioning, enabling arbitrary-depth inference. The model achieves 99.90% exact solve accuracy on Sudoku-Extreme and 98.93% on Maze-Unique, improving with inference depth beyond training rollout lengths. Surprisingly, progressive denoising is unnecessary at inference: maximal corruption with fresh Gaussian noise at each step retains near-perfect solving. The method requires no parallel rollouts or verifiers, attributing its success to diffusion's training curriculum rather than sampling. Ordered annealed corruption remains critical during training.
diffusion modelsiterative reasoningpersistent hidden statetraining curriculumnoise injection
Edge-Girth as a Structural Edge Feature for Graph Neural Networks
The paper introduces edge-girth, a structural edge feature for graph neural networks (GNNs), defined as the length and multiplicity of the shortest cycle through an edge, computable via breadth-first search. This feature avoids the need to predefine subgraph sizes, unlike traditional cycle-counting descriptors. Implemented in EGAGNN, a gated message-passing architecture, edge-girth achieves a test MAE three times lower than the closest comparator on the ZINC-12k benchmark with 104k parameters. It matches performance only with cycle-counting dictionaries up to length eight, while shorter dictionaries perform no better than no structural information. However, on graphs with uniform shortest cycle counts, edge-girth fails to improve discrimination, reverting to the 1-WL bound, as demonstrated across all 400 BREC benchmark pairs.
edge-girthgraph neural networksmessage-passingcycle-countingbreadth-first search
TRIAGE: Three-level Routing and Intelligent Agent Guidance for Efficient Execution
TRIAGE introduces a three-level routing framework to enhance efficiency in LLM agent execution by reusing historical trajectories. The core innovation, TaaS (Trajectory-as-a-Skill), abstracts execution trajectories into reusable skills, enabling 'experience as a service'. Queries are classified into Direct Reuse (0 tokens), Skill Substitution (0 tokens via parameter substitution), and Full ReAct (novel queries stored for future reuse). Experiments on 1,007 security monitoring queries show 62.3% token savings, with 56.0% Level 2 and 5.5% Level 1 queries executed at zero cost. Cross-domain validation on ToolBench achieves 76.3% token reduction, demonstrating semantic routing generalizability.
trajectory-as-a-skillllm agentstoken savingssemantic routingskill substitution
CATeye: Coupled Attribute-Topology Invariance Learning for Voucher Abuse Detection
CATeye introduces a Coupled Attribute-Topology Invariance Learning framework to detect voucher abuse in e-commerce by addressing coupled attribute-topology distribution shifts. The method employs an Attribute Invariance Selector (AIS) to filter non-invariant node attributes and an Edge Invariance Selector (EIS) to sample invariant subgraphs, constructing multiple views for domain-invariant representation learning. Evaluated on Lazada's proprietary dataset and a public benchmark, CATeye outperforms nine baselines in domain generalization and graph anomaly detection, achieving an 8.61% average F1-score improvement over the strongest baseline.
voucher abuse detectiondomain generalizationgraph neural networksinvariance learningdistribution shift
Predicting Subsurface Abnormalities Growth using Physics-Informed Neural Networks
This research introduces a Physics-Informed Neural Network (PINN) framework for predicting subsurface abnormalities using Ground-Penetrating Radar (GPR) data, analogous to tumor behavior prediction in medical imaging. The model integrates a CNN, spatial feature channel attention (SFCA), ConvLSTM, and temporal feature frame attention (TFFA) modules to embed electromagnetic wave propagation physics, enhancing prediction accuracy. Attention mechanisms adaptively compute channel and temporal weights to extract salient visual and temporal features. The PINN demonstrates improved precision in forecasting GPR data, advancing Non-Destructive Evaluation (NDE) for infrastructure deterioration assessments and providing deeper insights into deterioration mechanisms through physics-based models.
physics-informed neural networksground-penetrating radarconvolutional neural networksattention mechanismsnon-destructive evaluation
On the Reliability of Generative Augmentation: A Wasserstein-Based Theoretical and Empirical Study
This work establishes a statistical framework for analyzing the reliability of generative data augmentation in imbalanced classification tasks. The authors formalize augmentation as a distribution-mixing process and derive a generalization bound based on Rademacher complexity, revealing a trade-off between hypothesis complexity, augmentation intensity, and generative fidelity. Empirical evaluations on binary and multiclass tasks using Conditional GAN and Conditional WGAN-GP show that while CWGAN-GP achieves lower Wasserstein discrepancies than CGAN, improved distributional fidelity does not necessarily yield superior classification performance compared to classical oversampling methods. The results demonstrate that augmentation reliability is governed by distributional approximation error rather than predictive performance alone.
generative augmentationwasserstein discrepancyrademacher complexityconditional ganimbalanced classification
Contribution-Aware Bandwidth Allocation for Multimodal Split Learning
ModalShare introduces contribution-aware bandwidth allocation for multimodal split learning, optimizing inter-modality allocation based on Shapley contribution scores computed server-side without additional uplink traffic or client computation. The method dynamically adjusts each modality's keep-ratio under a fixed uplink budget, ensuring efficient payload distribution across modalities. Evaluations on CREMA-D and MVSA datasets demonstrate accuracy improvements of 15.4 and 12.4 percentage points, respectively, at 5x compression, outperforming equal keep-ratio baselines. ModalShare consistently enhances performance across three compressors, three datasets, and four budgets, recovering gains lost by existing compressors in multimodal settings.
multimodal split learningshapley contributioninter-modality allocationbandwidth optimizationcompression
Investigating Linear Probe Robustness to Linguistic Register, Medical Specialty, and Corpus Shifts in Medical QA
This work isolates and evaluates three factors affecting the robustness of linear probes for factual error detection in medical QA: linguistic register, medical specialty, and corpus shifts. Using 500 MedQA entries rewritten into four styles (textbook, patient, clinical note, colloquial) and two additional datasets (MedMCQA, MMLU-medical), the study probes four open-weight LLMs (2-8B parameters). Results show truth-direction stability across registers (ΔAUROC≈0.10) and specialties (Δ≈0.03), but significant corpus degradation (ΔAUROC 0.12-0.21), suggesting dataset structure influences probe performance more than medical knowledge alone.
linear probeslinguistic registermedical qatruth directionauroc degradation
Exact Risk-Complexity Laws for Projective Boundaries in Scenario Optimization and Distribution-Free Certification
This paper establishes exact risk-complexity laws for projective boundary schemes in scenario optimization and distribution-free certification. By modeling decision rules as acceptance sets with boundary maps, the authors introduce proper projective boundary schemes where violation risk follows conditional laws determined by the boundary's cross-sample complexity profile. Stable profiles yield beta distributions, while varying profiles produce exact profile corrections. The framework generalizes classical settings including scalar order-statistic calibration, support-reconstructive scenario programs, and Pareto-frontier calibration, while also providing conditional probabilistic certificates and a no-go result on observed complexity.
projective boundaryscenario optimizationviolation riskbeta lawcross-sample complexity
Where the Verifier Fails: A Category-Level Audit of Reward Signals in RLVR
This work conducts a category-level audit of reward signal verifiers in RLVR, revealing systematic failure patterns. Using metamorphic testing, the authors generate certified equivalent answer variants to isolate verifier errors without human adjudication. Across 307,420 verdicts from four verifiers, key findings emerge: self-validation rates vary dramatically (53.8%-95.2%), with 49.9% disagreement between library configurations; whitespace and punctuation account for 93.0% of failures; and numeric cascades exhibit scale-invariant tolerance thresholds, accepting off-by-one errors as a step function of magnitude.
metamorphic testingrlvrverifiernumeric cascadescale-invariant
SMELT: Scaling Laws for Compute-Matched MoE Looped Transformers
The authors introduce SMELT (Sparse MoE Transformer, middle layers Loop Twice), a method for looping middle layers in Mixture-of-Experts Transformers while matching per-token FLOPs, non-embedding parameters, and KV cache. Scaling SMELT up to 54B non-embedding parameters, they fit Chinchilla-style scaling laws and demonstrate a 6.8--18.0% reduction in training FLOPs on the compute-optimal frontier. SMELT outperforms baselines on downstream benchmarks, particularly in code tasks, with gains increasing with sample length and in-context examples. Mechanistic analysis reveals that looping reduces attention sink effects and improves focus on content-relevant tokens, validating depth reuse as a practical optimization strategy.
mixture-of-expertskv cachescaling lawsattention sinkin-context learning
mzCache: On-Device LLM Memory Management under Multitasking
mzCache introduces an on-device LLM inference system optimized for multitasking environments, addressing memory pressure-induced performance degradation. The system employs restoration-oriented memory management, partitioning LLM memory into fine-grained shared buffers to enable partial eviction and restoration with concurrent CPU-GPU access. It utilizes hybrid swap and backward-out eviction policies to ensure low-latency restoration from any eviction state. Implemented on llama.cpp and deployed as an Android application, mzCache achieves a 2.1-5.5× reduction in Time-to-First-Token compared to storage-backed partial offload, demonstrating efficacy in real multitasking scenarios.
llm inferencememory managementkv cachemultitaskingtime-to-first-token
Exploring Sparse Autoencoders in Text-Based Causal Confounding Adjustment
The authors propose a novel causal adjustment pipeline using sparse autoencoders (SAEs) to address the tradeoff between preserving confounding variables and ensuring finite-sample overlap in text-based causal inference. The method iteratively selects a minimal set of SAE features via conditional independence tests. Evaluations on semi-synthetic datasets with binary confounders demonstrate that SAE representations achieve lower bias and higher coverage compared to alternative representations, while offering interpretability for falsification. A more realistic semi-synthetic evaluation using multi-label confounders highlights the need for further investigation into off-the-shelf adjustment methods in complex settings.
sparse autoencoderscausal adjustmentconditional independencesemi-synthetic evaluationconfounding variables
Matched Queries for Curvature and Density at Branching Junctions
The paper introduces a method for recovering branchwise curvature and log-density slope at branching junctions using matched score queries at noise scales σ and λσ. By leveraging normalized score expansions and matched subtraction, the approach cancels tangent contributions to expose second-order effects, uniquely identifying sD branch parameters given tangent directions and weights. Theoretical analysis establishes perturbation bounds, conditional kernel-density-estimation rates, and full-rank calibration requirements. Experimental results demonstrate the method's robustness, achieving full rank up to D=20 with 16 branches and reducing median parameter error by a factor of 49.4 compared to naive tangent subtraction in end-to-end tests for D=3–5.
matched score queriesbranchwise curvaturelog-density slopefull-rank calibrationperturbation bound
One-Layer Transformer Provably Learns Multiclass One-Nearest Neighbor in Context
The work extends binary classification results to multiclass settings by proving that one-layer transformers with argmax heads are functionally equivalent to one-nearest-neighbor classifiers. Using simplex encoding, the authors demonstrate this equivalence without relying on non-standard rounding methods required by prior work. The result closes a theoretical gap, aligning practical transformer implementations with their nearest-neighbor interpretations in multiclass scenarios.
one-layer transformermulticlass classificationsimplex encodingargmax headone-nearest-neighbor
Relational Task Generation Language: A Declarative Specification Framework for Relational Deep Learning
We introduce Relational Task Generation Language (RTGL), an open-source declarative language that simplifies the formulation of Relational Deep Learning (RDL) tasks by abstracting low-level SQL details. RTGL addresses the labor-intensive and error-prone process of manually defining RDL prediction tasks, which often leads to data leakage. The framework reconstructs existing RDL benchmark tasks, revealing inconsistencies in manually crafted SQL definitions, and demonstrates its utility by designing diverse new tasks. Experiments confirm RTGL's robustness, usability, and seamless integration with existing RDL frameworks, enhancing accessibility for the community.
relational deep learningdeclarative languagesql abstractiondata leakagetask formulation
Position: Privacy Is a Claim, Not a Property of Synthetic Data
The paper argues that privacy in synthetic data should be treated as an explicit, evidence-based claim rather than an inherent property inferred from generation methods. Through empirical analysis of recent ML publications, the authors demonstrate that synthetic data is often used in privacy-sensitive contexts without clear articulation of threat models, inference risks, or falsifiable privacy claims. This leads to implicit privacy assurances that are difficult to verify and disproportionately affect rare and minority records. The authors recommend ML venues adopt norms requiring privacy assertions to be scoped, testable, and contestable.
synthetic dataprivacy claimthreat modelsinference risksfalsifiable privacy
Solving In-Table Prediction Problems by Deep Neural Networks with Performance Evaluation Using Synthetic Data
The paper introduces In-Table Prediction (ITP), a novel self-supervised learning approach for predicting arbitrary column values in tabular data using deep neural networks. Three neural architectures—MLP, ResNet, and Transformer—are evaluated on synthetic datasets containing continuous features, with a novel embedding layer proposed to handle missing values. Results indicate that attention-based Transformers outperform MLP and ResNet when trained on sufficiently large datasets with extended embedding lengths. The study is limited to synthetic data with predefined column relationships and a small number of columns, serving as an initial exploration rather than a comprehensive analysis of real-world tabular data.
in-table predictionself-supervised learningtabular deep learningattention-based transformermissing value embedding
Post-Training Science for Supervised Fine-Tuning
This study systematically evaluates key decisions in supervised fine-tuning (SFT) across model families (Qwen3, Llama), scales (up to 235B parameters), and datasets (four real-world customer SFT tasks). Through controlled sweeps varying learning rate, batch size, LoRA rank, optimizer choice, and training duration, the authors analyze trade-offs between LoRA and full fine-tuning, validate metrics like loss landscape flatness, and assess scaling effects of model size and data volume. Findings include optimal hyperparameter transferability across models, LoRA rank-alpha interactions, and epoch limits before instruction-following degradation. Each recommendation is accompanied by uncertainty quantification.
supervised fine-tuninglora rankloss landscapemixture-of-expertshyperparameter sweep
Multi-Head Self Attention is a Parameter Identification Mechanism
The paper demonstrates that multi-head scaled dot-product attention functions as a parameter identification mechanism, with the ratio of unidentified to total parameters scaling inversely with the number of heads (1/2 → 1/(2H)). It proves that full identification is impossible in attention layers and reveals bias terms that have no effect on softmax-based attention. The analysis extends to modern transformer improvements like RoPE and GQA, showing how they enhance the ratio of meaningful parameters. Numerical experiments confirm theoretical results, including training updates overlapping model-invariant subspaces due to lack of identification.
multi-head attentionparameter identificationtransformer architecturesoftmax attentionmodel-invariant subspaces
Recent Developments in Transformer Inference Deployment on FPGA Platforms: A Survey
This survey systematically reviews recent advancements in deploying Transformer-based models on FPGA platforms, focusing on inference performance and energy efficiency. Through a comprehensive literature analysis, it identifies preferred implementation techniques and optimization strategies for FPGA-based Transformer inference. The study highlights FPGA's advantages in flexibility, latency reduction, and on-site deployment suitability compared to CPUs and GPUs. A taxonomy of design choices is provided to guide both academic and industrial researchers in FPGA-based Transformer deployment.
transformer architecturefpga platformsinference optimizationenergy efficiencylatency reduction
Births are difficult to predict even with rich survey and full-population register data
This study establishes a predictive ceiling for birth prediction, demonstrating that even with comprehensive data and diverse modeling approaches, individual life events remain inherently stochastic. Researchers conducted a data challenge involving 147 participants predicting births within three years for Dutch residents aged 18-45, utilizing both survey data and full-population registers. Methods spanned logistic regression, transformers, and large language models, with best F1 scores of 0.59 (register) and 0.76 (survey). Advanced models did not outperform classical ones, and register data underperformed survey data. Simulations estimating biological stochasticity revealed predictive ceilings of F1 ~ 0.86-0.94 (survey) and 0.88-0.96 (register), highlighting the limits imposed by chance in reproduction.
predictive ceilingstochastic biologyfull-population registerstransformersf1 score
Pre-carved Niches: The Formation Dynamics of Modular Task Partitions in Early LLM Training
This work investigates the formation dynamics of modular task partitions during early training of large language models, addressing a gap in prior research focused on finished models. The authors train a Pythia-410M model from scratch (bf16 and fp32 trajectories) and analyze attribution patching, gradient norms, effective updates, weight norms, and first-order loss decomposition across 14 tasks in four cognitive domains. Key findings include: (1) modular maps are pre-carved, with task pair overlaps 3.6× above baseline before learning; (2) partitions lock in through two sharp jumps (up to 20.4σ) with gradient-level relative deprivation (2.25-2.73×); (3) domain-specific deviations emerge during learning, supporting modularity-learning hypotheses.
modular task partitionsattribution patchinggradient normsfirst-order loss decompositiondomain-specific deviations
CopyShield: A Cross-Level Benchmark of Copyright Defenses in LLMs
CopyShield introduces a cross-level benchmark for evaluating copyright defenses in LLMs, comparing contrastive decoding, Direct Preference Optimization (DPO), and activation intervention across LLaMA-3.1-8B and Mistral-7B-v0.3. The benchmark measures literal leakage, non-literal leakage, utility, and degeneracy using controlled memorization of five public-domain books. Results show distinct trade-offs: contrastive decoding maintains low degeneracy (0-2%) but limited literal suppression (NV-Recall 0.192-0.203), DPO reduces literal leakage (0.263 to 0.002) but induces paraphrase-loop degeneracy (58%), and activation intervention achieves the lowest non-literal flagging rate (1/200) by blocking 84% of non-literal queries. DPO degeneracy varies by model (10-14% on Mistral), highlighting model-dependent effects.
contrastive decodingdirect preference optimizationactivation interventionliteral leakagedegeneracy
Scaled Idempotence in Transformer Attention: Paired OV Geometry and Shared-Value Algebras
The paper identifies scaled idempotence as a recurrent algebraic regularity in Transformer attention, where a sparse subset of OV operators nearly closes under composition ($T^2\approx\alpha T$). Through principal-coordinate factorization and empirical analysis across six pretrained models (2.8B--235B parameters), the authors demonstrate that trained orientation significantly impacts closure alignment, with 98.64% of heads favoring trained orientations. Experiments reveal that high closure is feasible but not universally attained, and value sharing extends headwise closure into a local operator algebra. These findings characterize scaled idempotence as a sparse trained orientation within broadly available geometric capacity.
scaled idempotencetransformer attentionov operatorsprincipal-coordinate factorizationvalue sharing
When Does Online Adaptation Pay on the Edge? A Leakage-Free Evaluation of Warmup, Learning-Rate Selection, and Resource Trade-offs for Time-Series Forecasting
The study evaluates online adaptation for edge time-series forecasting under distribution drift, focusing on leakage-free streaming protocols and identifying biases in warmup budget and optimizer learning-rate selection. Using six public multivariate streams, including building-sensor and smart-meter data, the authors demonstrate that warmup budget impacts adaptation benefit by 3.0 to 18.8 percentage points across 1,000-20,000 steps. Adam outperforms SGD with momentum in 310 of 360 evaluated cells when learning rates are validated pre-drift. Parameter-efficient variants of PatchTST show nondominated performance on adaptation-state memory, while smart-meter gains depend on selection rules. Code and data are publicly available.
online adaptationtime-series forecastingdistribution driftwarmup budgetadam optimizer
Replicating TRACE: A Practitioner's Guide to Its Threshold and Particle Budget
The study replicates TRACE (Math & Lienhart, 2023), a method for extracting causal graphs from autoregressive models by thresholding conditional mutual information. Key findings include: (1) optimal thresholds align with ground-truth margins (delta = 0.05), not fixed values; (2) a global threshold primarily recovers lag-1 edges (recall 0.97-0.99), while lag-2+ edges require per-lag thresholds; (3) benchmark skew (85% lag-1 edges) inflates headline F1 (0.86-0.91); (4) F1 saturates at N = 2 particles due to threshold margin, not estimator convergence (O(N^(-1/2))). Five practitioner rules are derived.
causal graphsconditional mutual informationautoregressive modelsthresholdinglag decay
Neural Symbollic Regression Using Deep Learning and Sparse Modelling
The authors propose Neural Symbolic Regression (NSR), a framework combining deep learning and sparse modeling for symbolic regression. NSR employs a decoupled pipeline: a neural network learns noise-robust approximations in an interaction-aware nonlinear feature space, followed by LASSO-based sparse regression to extract interpretable closed-form expressions. Distributed hyperparameter optimization with Ray Tune and ASHA scheduling enhances accuracy and symbolic fidelity. Evaluations on the Nguyen benchmark suite demonstrate superior RMSE, noise robustness, and out-of-distribution generalization compared to SINDy and non-tuned neural baselines. Ablation studies highlight the importance of feature interactions, neural depth, and tuning strategies.
symbolic regressionsparse modelinglassohyperparameter optimizationneural approximation
Subliminal Learning as Trait-Direction Drift: A Mechanism and Targeted Control under SFT Distillation
The study introduces trait-direction drift as a mechanism for subliminal learning during model distillation, where hidden preferences from a teacher model transfer to a student. The authors propose probe-space corridor regularization, a method that constrains drift along calibrated trait directions to mitigate unwanted trait transfer. Results demonstrate significant reduction in malicious-response transfer from 29.55% to 6.45% with minimal impact on main-task accuracy, and consistent suppression of animal-preference transfer in the Qwen setting. The findings link subliminal learning to trait-direction drift and validate corridor regularization as an effective control during distillation.
trait-direction driftsubliminal learningprobe-space corridor regularizationmodel distillationsupervised fine-tuning
Modelpedia: A Catalog of Model Findings for the Meta-Science of AI
Modelpedia introduces an automated, LLM-assisted framework to organize scientific findings about AI models into a searchable public catalog. The system extracts findings from published papers, links them to relevant models, datasets, methods, and concepts, and aggregates the results for meta-analysis. Applied to ICLR 2024 and 2025 papers, it extracted over 1,000 findings, enabling a meta-analysis of model investigation practices. The open catalog aims to establish model findings as a shared foundation for the meta-science of AI, inviting community exploration and contribution.
meta-sciencellm-assisted frameworkmodel findingssearchable catalogmeta-analysis
Let Confidence Change, Not the Prediction: Prediction-Preserving Repair for Post-hoc Calibration
We propose Calibrator-Output Repair for Top-1 Decision Preservation (CORD), a post-hoc adapter that preserves original top-1 predictions while improving calibration metrics. CORD repairs the full calibrated probability vector by determining the mass assigned to the original top-1 prediction and allocating remaining mass over other classes, ensuring zero Top-1 Prediction Change Rate (TPCR). The method requires no additional supervised mapping or hyperparameter tuning. Experiments on CIFAR-10/100 and ImageNet-1K demonstrate that CORD achieves zero TPCR by construction while reducing Expected Calibration Error (ECE), Negative Log-Likelihood (NLL), and Brier score across datasets, maintaining gains under distribution shift and varying calibration-set sizes.
post-hoc calibrationtop-1 predictionprobability vectorexpected calibration errordistribution shift
Artificial Rosetta Stone: Constrained Maximum A Posteriori (MAP) Reconstruction of Symbolic Raga Sequences via Order-k Markov Models
The Artificial Rosetta Stone (ARS) proposes a mathematical framework for probabilistically reconstructing damaged symbolic musical sequences constrained by raga grammar. Using order-k Markov models with a symmetric Dirichlet prior, the method formulates missing-note reconstruction as a constrained maximum a posteriori (MAP) problem, solvable via dynamic programming with $O(TN^{k+1})$ complexity. Theoretical analysis includes parameter count derivation ($N^k(N - 1)$), concentration bounds, and error propagation. Experiments demonstrate feasibility on synthetic data with six raga-inspired alphabets, orders $k \in \{1, 2, 3\}$, and masking rates up to 50%, and on 30 real-audio sequences from Yaman clips via automated pitch extraction and quantization. Claims are limited to probabilistic reconstruction under explicit constraints, not historical authentication.
ragamarkov modelmaximum a posterioridynamic programmingdirichlet prior
Accelerating Reinforcement Learning via MPC Solver-Gradient Guidance for Weights-varying MPC
SG-RL introduces solver-gradient guidance to enhance RL-based online adaptation of MPC cost-function weights, combining RL's closed-loop return optimization with low-variance gradients from differentiable MPC. The method integrates solver-derived gradients into PPO via four modular algorithms for actor-update scaling, policy loss, advantage estimation, and value-function learning. Experiments on autonomous racing platforms show SG-RL achieves PPO's best return with 70.6% fewer samples, outperforms GB-PL by 54% in closed-loop return, and generalizes zero-shot to unseen environments.
model predictive controlreinforcement learninggradient-based policy learningproximal policy optimizationautonomous racing
SAGE: Subpopulation-Aware Generative Enhancement for Mitigating Spurious Correlations
The paper introduces Subpopulation-Aware Generative Enhancement (SAGE), a two-stage generative augmentation framework to mitigate spurious correlations in datasets without prior group labels. SAGE employs cluster-derived sub-labels and class labels to fine-tune a conditional generative model and text encoder, generating synthetic data to address underrepresented regions and construct a balanced validation set for last-layer reweighting. Experiments demonstrate SAGE's effectiveness, achieving worst-group accuracies of 89.5%, 85.7%, and 79.1% on Waterbirds, CelebA, and MetaShift, respectively, outperforming group-label-free baselines by up to 7.7 percentage points.
spurious correlationsgenerative augmentationempirical risk minimizationworst-group accuracyconditional generative model
The Multiple Timescales of Gradient Descent on the Edge of Stability: A Perturbative Derivation of the Central Flow
We provide a formal derivation of the central flow model for gradient descent at the edge of stability by treating it as a singularly perturbed dynamical system. Assuming a loss decomposition $f = g + \varepsilon h$, we analyze gradient descent dynamics in the limit $\varepsilon \to 0$ using the method of multiple scales. Three distinct timescales emerge: fast oscillations, intermediate self-stabilization, and slow central flow dynamics. The central flow arises as the leading-order term in the expansion, while self-stabilization appears in the next-order term. We extend previous analyses by computing energy drift for single-edge eigenvalues and deriving stabilization systems for multiple-edge eigenvalues.
central flowgradient descentsingular perturbationmultiple scalesself-stabilization
Web Price Extraction: State of the Art and an Adaptive Browserless Implementation
The authors present an adaptive browserless price extraction system combining HTML fragmentation with syntactic, semantic, and frequency rules, enhanced by Bayesian weight updates and genetic algorithm optimization. This hybrid approach improves robustness to website structural variations while maintaining computational efficiency. Compared to the baseline, the method increases precision from 77.2% to 87.3% and reduces per-page processing time by ~14%, offering a resource-efficient alternative to manual wrappers, browser-based tools, and LLM-based methods for e-commerce data extraction.
price extractionbrowserlesshtml fragmentationbayesian weightinggenetic algorithm
When Metropolis and Hastings Meet Bradley and Terry: Exact MCMC From Preference Voting
Pref-MH introduces an exact Metropolis-Hastings sampler for conditional distributions induced by stochastic binary pairwise comparisons, leveraging the equivalence between MH unnormalized density ratios and Bradley-Terry preference odds. The method addresses the challenge of precise ratio computation by developing a valid accept/reject rule, ensuring Markov chain convergence to the target distribution. Pref-MH is shown to be Peskun-Tierney optimal for fixed proposal kernels and budgets. Experiments in text generation, molecular design with LLM judges, and image generation with VLM judges demonstrate its practical utility in scenarios where comparative feedback is readily available.
metropolis-hastingsbradley-terryconditional samplingpreference oddspeskun-tierney
Direct Optimization of a 3D Finite-Source Reflector via Neural-Network Parameterization
The authors propose a direct optimization method for designing 3D freeform reflectors that transform finite-étendue light sources into prescribed far-field angular intensity distributions. The reflector profile is parameterized by a small multilayer perceptron, trained end-to-end via differentiable ray-tracing with gnomonic coordinate parameterization to ensure ray-reflector intersections. A bicubic spline representation enables efficient ray-tracing, with intersections solved via damped Newton methods and gradients computed using the implicit function theorem. The output distribution is compared to the target using an $H^{-1}$-weighted soft histogram, optimized via BFGS with Broyden updates. The method converges within seconds on a single GPU for all tested cases.
freeform reflectorsdifferentiable ray-tracinggnomonic coordinatesbicubic splinebfgs optimization
Poisson-Gamma Dynamical Systems with Time-varying Transition Dynamics
The authors propose Poisson-Gamma Dynamical Systems with time-varying transition kernels (TV-PGDS) to address the limitation of existing PGDS models in capturing evolving transition dynamics in count-valued time series. The method introduces three Dirichlet Markov chains (Dir-Dir, Dir-Gam-Dir, PR-Gam-Dir) to model heterogeneous structural mutations and employs Dirichlet-Multinomial-Beta data augmentation for efficient Gibbs sampling. Experiments demonstrate that TV-PGDS achieves superior predictive performance by learning time-evolving transition matrices compared to related models.
poisson-gamma dynamical systemstime-varying transition kernelsdirichlet markov chainsdirichlet-multinomial-betagibbs sampling
iPINN for Broadband CARS Phase Retrieval: A Framework for Function Approximation and Inverse Modeling Problems in Nonlinear Spectroscopy
The authors propose an inverse physics-informed neural network (iPINN) for broadband coherent anti-Stokes Raman spectroscopy (BCARS) phase retrieval, addressing the ill-posed inverse problem of reconstructing resonant susceptibility from raw spectra. iPINN employs a transformer encoder to assign spectral features to 24 learnable peak slots and uses a multi-view consistency loss to ensure invariance across non-resonant background (NRB) patterns, NRB strength, and noise. The method outperforms baselines on a public benchmark (MAE 0.016 vs. 0.046) and demonstrates depth-invariant accuracy on 28 zero-shot test spectra across seven solvents and four focal positions, validating its robustness under varying acquisition conditions.
phase retrievalphysics-informed neural networkbroadband carstransformersinverse modeling
Sharp Mixed Spectral Barron Regularity of Coulombic Many-Electron Wave Functions
The study establishes sharp mixed spectral Barron regularity for eigenfunctions of molecular Coulomb Hamiltonians, employing a Fourier $L^1$ norm with isotropic and coordinate-product weights to detect previously invisible regularity. By analyzing a nonempty set $I$ of electron indices with antisymmetric wave functions, the authors derive explicit admissible regions for isotropic order $s$ and coordinate orders $α,β$, optimal across clamped-nuclei Coulomb Hamiltonians. For fixed-spin components with two occupied spin blocks, the condition simplifies to $s+α+β<1$, while fully spin-polarized states reduce to $s+α<1$. The results demonstrate that fixed-spin spatial components satisfy specific $L^1$ integrability conditions based on occupied same-spin blocks.
barron regularitycoulomb hamiltoniansfourier l1 normspin-polarized statesantisymmetric wave functions
The Visual Insensitivity Gap: Diagnosing When Vision-Language Models Fail to Use Visual Evidence
This paper introduces the Visual Insensitivity Gap, demonstrating that vision-language models (VLMs) fail to utilize visual evidence on 40%-97% of samples across six VLMs and three benchmarks, despite visual perturbations. The authors propose a Visual Sensitivity Index (VSI) to quantify this gap, showing it is sample-specific rather than model-specific, with strong cross-model correlations (Spearman rho=+0.40). A linear probe reveals a 0.72-0.79 accuracy in distinguishing perturbed images, yet VLMs' token predictions change only 2%-11% of the time, indicating a significant encoder-LLM gap. VSI's diagnostic utility varies, excelling in multi-choice reasoning (AUROC=0.85-0.87) but less so in factuality tasks.
vision-language modelsvisual insensitivity gapvisual sensitivity indexmultimodal benchmarksencoder-llm gap
MemoryWalker: Stop Training Agents on Contexts They Never Saw
MemoryWalker addresses the conditioning problem in training production agents under context compression by introducing exact gradient-equivalent corrections and a variational relaxation. LogitTree employs a segmented K-forward traversal requiring K+1 backward passes, while a packed 4D attention mask necessitates custom kernels and eviction records. SDCC (Self-Distillation for Conditioning Consistency) minimizes forward KL divergence between compressed student and stop-gradient teacher models, ensuring O(sqrt(epsilon_KL)) train-deployment gap. Evaluated on seven web-search benchmarks, exact methods maintain no-compression performance, and SDCC reduces logit drift while improving rollout rewards.
context compressiongradient-equivalentself-distillationforward kllogit drift
Conditional Flow Matching for ML-Based Inverse Design Problems
The paper introduces conditional flow matching (CFM) for inverse design problems in engineering, comparing it against conditional diffusion models and conditional generative adversarial networks (cGANs) on EngiBench tasks. CFM generates candidate designs without iterative solver reruns, leveraging a single forward pass for sampling. Evaluated on structural (beams2d) and thermal (heatconduction2d) benchmarks, CFM achieves superior performance in cumulative optimality gap (COG), final optimality gap (FOG), maximum mean discrepancy (MMD), and volume-fraction deviation. Specifically, CFM demonstrates mean volume-fraction deviations of 0.4% and 1.0% on beams2d and heatconduction2d, respectively, outperforming diffusion models. CFM also achieves higher throughput (53.2 samples/s) and lower COG compared to diffusion baselines.
conditional flow matchinginverse designcumulative optimality gapmaximum mean discrepancyvolume-fraction deviation
Dense Process Supervision for Search Agents via Fact Utility Estimation
The paper introduces a dense process supervision method for search agents based on fact utility estimation, addressing the credit assignment problem in reinforcement learning (RL) by modeling reasoning as discrete evidence accumulation. Structured facts are extracted from raw observations and organized into a fact store, with semantically equivalent facts clustered and their posterior utilities inferred via Bayesian estimation over group rollouts. These utilities are converted into dense step-level rewards for RL training. Experiments on seven single-hop and multi-hop QA benchmarks demonstrate consistent outperformance over baselines, with ablation studies confirming improvements in multi-hop QA compared to outcome reward-only training.
reinforcement learningcredit assignmentfact utility estimationbayesian estimationmulti-hop qa
Subspace Levenberg Marquardt Algorithms in Training Neural Networks
This work evaluates subspace Levenberg-Marquardt (LM) algorithms for training neural networks, addressing the computational and memory limitations of classical LM in large networks. The study compares subspace variants, including Krylov subspace LM (KSLM) and hybrid subspace LM (HSLM), against classical LM and first-order methods like stochastic gradient descent (SGD) and Adam. Experiments focus on regression and classification tasks, demonstrating the efficiency and scalability of subspace LM methods. Results indicate that subspace LM variants achieve competitive performance while reducing computational overhead, making them viable for medium- to large-scale neural network training.
levenberg-marquardtkrylov subspaceneural networksstochastic gradient descentsecond-order methods
Semi-Supervised Classification with Informative Missing Labels in Weibull Mixture Models
The paper introduces a semi-supervised classification framework for two-component Weibull mixture models with informative missing labels, where missingness depends on classification uncertainty. The method jointly models the classifier and the missing-label mechanism, leveraging missing-label indicators to enhance classifier estimation. Decision boundaries are characterized under equal and unequal Weibull shape parameters, and Fisher information is derived after adjusting for nuisance parameters. An asymptotic expansion of the expected error rate reveals conditions for improved classification performance. Numerical experiments and a semi-synthetic hard-drive failure dataset demonstrate reductions in error rates and enhanced decision-boundary estimation when incorporating feature-dependent label missingness.
weibull mixturemissing labelssemi-supervised classificationfisher informationdecision boundary
Frozen Cores Need Task Signal: Fisher-Whitened Cross-Covariance for Low-Resource LLM Adaptation
The paper introduces Frozen-Core Cross-Covariance Adaptation (FCCA), a parameter-efficient fine-tuning method for low-resource LLM adaptation. FCCA estimates the signed input-error cross-covariance, whitens it using diagonal Fisher moments, truncates it in the local metric, and applies thin QR decomposition to obtain stable core coordinates. Evaluated on 11 tasks across four model settings, FCCA achieves an 83.0 macro-average on Qwen2.5-3B, outperforming other methods under a matched parameter budget. It also demonstrates significant gains from whitening and QR stabilization, and comes within 0.32 and 0.23 points of LoRA and DoRA while optimizing only 36.9K parameters instead of 7.4M.
parameter-efficient fine-tuningcross-covariancewhiteningqr decompositionlow-rank factors
How Do Language Models Choose Between Context and Memory?
The study investigates how language models mediate between contextual information and parametric knowledge by analyzing activation directions and their causal role in source selection. Using counterfactual experiments with Qwen, Llama, and OLMo models, authority directions were estimated from agreement prompts where context and knowledge align. Interchanging coordinates along these directions reproduced 30-68% of authority-induced shifts in source choice, while controls showed minimal effect. Cross-task transferability of authority directions was limited, closing only 9% of the authority gap compared to 57% with task-specific directions, indicating task-dependent authority computations.
activation directionsparametric knowledgeauthority directionscounterfactual experimentscross-task transferability
Text Capability Loss in Vision-Language Adaptation: An Attention-Sink Diagnosis
The study identifies attention-sink corruption as the mechanism behind text capability loss when fine-tuning pretrained LLMs into vision-language models (VLMs), particularly affecting tasks requiring strict output adherence. It introduces Sink Strength, a scalar metric computed on base LLMs that predicts post-VLM degradation without requiring VL training, validated across six VLM-LLM pairs and format-sensitive tasks. Experiments show that interventions like QK-RMSNorm injection and weight-merging fail to recover lost capabilities, emphasizing the need for backbone screening with Sink Strength prior to VL training and suggesting head-selective training-time protection.
attention-sinkvision-language modelqk-rmsnormweight-mergingfine-tuning
Online Self-Weighted Fine-Tuning
Online Self-Weighted Fine-Tuning (OSW-FT) enhances supervised fine-tuning (SFT) by dynamically adjusting loss weights based on the model's competence, estimated via inference-only rollouts. The method maintains the expert trajectory's optimization direction while adaptively scaling update magnitudes, connecting to SFT and RL through variance-reduction principles. Evaluated on Qwen3 models (0.6B-4B) across benchmarks like AIME, OSW-FT outperforms SFT on small-to-medium models, achieving a favorable compute-performance trade-off with just 2 online rollouts for binary-verifiable reasoning tasks.
supervised fine-tuningonline weightingvariance-reductionbinary-verifiable reasoningrollouts
MaskCode: Mask Transformer for Feedback-Assisted Coding With Linear Block Codes
MaskCode introduces a Transformer-based inner feedback code for concatenated coding systems, explicitly incorporating structural knowledge of outer linear block codes. The method employs two mechanisms: a soft syndrome-based input to inform the encoder about parity constraint violations and a code-aware attention mask derived from the Tanner graph. End-to-end training with a differentiable belief propagation decoder is shown to be unnecessary, as MaskCode's design already internalizes outer code structure. Evaluations on BCH and LDPC outer codes demonstrate MaskCode's superiority, achieving up to 1.5 dB SNR gain over baselines.
transformerfeedback codelinear block codetanner graphbelief propagation
Prediction-Assisted Pricing and Admission for LLM APIs with Stochastic Token Consumption
The paper introduces Prediction-Clipped UCB (PCUCB), a sequential pricing and admission algorithm for LLM APIs with stochastic token consumption. The method integrates offline predictions with online learning, intersecting prediction intervals with confidence intervals and using resource shadow prices for product evaluation. It ensures hard feasibility through reservation envelopes and achieves regret bounded by a pacing term plus cumulative interval diameters. Theoretical analysis shows PCUCB interpolates between full-information and learning-from-scratch regimes, with regret scaling as Õ(√T + (1+Λ̄)min{Tε,√JT}) for J segment-product cells and prediction radius ε.
stochastic token consumptionresource shadow pricesprediction-clipped ucbregret boundreservation envelopes
Controllable Image Captioning with Prompt-Conditioned Scene Rewards
FoCUS introduces a controllable image captioning method enabling semantic emphasis steering via natural-language prompts. The approach employs a prompt-conditioned control objective based on scene-graph-aligned component scores, where objects, attributes, and relations are differentially weighted according to user-specified emphasis. Optimization utilizes GRPO, enhanced by stricter object validity thresholds and reasoning-based verification for attribute and relation scoring. Evaluated on the SCoPE benchmark with Include/Avoid constraints, FoCUS demonstrates improved controllability and fine-grained caption quality across two VLM backbones without compromising general caption performance.
scene-graphcontrollable captioningsemantic emphasisnatural-language promptsreasoning-based verification
Patterning in Practice: Debiasing Reward Models with Susceptibilities
This work introduces patterning, a method to debias reward models by reweighting preference pairs based on their susceptibility to benchmark losses. Applied to the Gemma 2 9B Instruct reward model trained on Skywork-Reward-Preference v0.2, patterning achieves a +14.2 ± 1.2 percentage point improvement on RM-Bench Hard, preserving overall accuracy. The method's interpretability is demonstrated by tracing a regression on a safety subset to specific training pairs, confirmed via ablation. The computed weights transfer across Gemma 2 variants (2B, 27B) and partially to Llama 3.1 8B without recomputation, marking patterning's first application beyond small models and synthetic tasks.
patterningreward modelssusceptibilitybenchmark lossesdebiasing
MUGEN: Generating Unlearnable Graph Examples for Multiple Learning Tasks
MUGEN introduces the first framework for generating unlearnable graph examples that jointly protect multiple learning tasks, including node classification, graph classification, and link prediction. The method employs a shared GNN encoder and task-specific heads, optimizing perturbations via a Task-Aligned Separability Objective (TASO) and Type-Adaptive Perturbation (TAP), which tailors updates for discrete and continuous node attributes. Experiments across five benchmarks, four GNN backbones, and three learning paradigms demonstrate MUGEN's effectiveness in generating transferable unlearnable examples, even under adversarial training and data augmentation.
unlearnable examplesgraph neural networkstask-aligned separabilitytype-adaptive perturbationadversarial training
Verdict Instability of OOD Scores under Reference Resampling
The paper introduces verdict instability, a measure of variability in out-of-distribution (OOD) detection scores due to reference set resampling. It derives a closed-form expression for this instability, which depends on within-class dispersion and reference count, highlighting its role under class imbalance. The study finds that far-OOD queries exhibit low variance in anisotropic embeddings, leading to reproducible high scores. However, estimators of local dispersion often carry incorrect signs, negatively impacting abstention decisions. Empirical results show that wrong-signed scores degrade abstention performance compared to random abstention across datasets.
out-of-distribution detectionverdict instabilityreference resamplingclass imbalanceanisotropic embedding
HarmoCore: Functional Latent Diffusion for Sparse Reconstruction of Oscillatory Wave Fields
HarmoCore introduces a functional latent diffusion framework for reconstructing oscillatory wave fields from sparse sensor data, addressing challenges of complex-valued, frequency-sensitive, and highly oscillatory responses. The method employs Functional Tucker cores over shared continuous spatial bases, learns a frequency-conditioned core diffusion prior, and performs Diffusion Posterior Sampling directly in core space, avoiding dense pixel-space correction. Optional residual guidance enhances physical consistency. Experiments on 2D Helmholtz, 2D synthetic wave fields, and 3D Helmholtz demonstrate significant improvements with 1%--2% sensing, maintaining practicality in three dimensions.
functional tucker coresdiffusion posterior samplingoscillatory wave fieldsfrequency-conditioned priorhelmholtz reconstruction
DK-GBMKKM: Dynamic Kernel-Space Granular-Ball Multiple Kernel $k$-Means Clustering
The authors propose DK-GBMKKM, a dynamic kernel-space granular-ball multiple kernel k-means clustering method that addresses limitations in traditional multiple kernel k-means. DK-GBMKKM generates granular balls in the evolving fused kernel space, alternates kernel-weight learning with granular-ball membership updates, and constructs a sample-size-weighted granular-ball kernel to preserve contributions across varying ball sizes. Theoretical properties including positive semidefiniteness and equivalence are established. Empirical evaluation on 12 public datasets demonstrates superior clustering performance. The implementation is open-sourced for reproducibility.
multiple kernel k-meansgranular-ball representationkernel-weight learningpositive semidefinitenessclustering performance
Disciplined Bilevel Programming
The authors introduce Disciplined Bilevel Programming (DBLP), a symbolic framework for specifying and solving optimistic bilevel optimization problems in a high-level, human-readable manner. DBLP automatically canonicalizes convex lower problems into conic form and constructs equivalent single-level reformulations using conic Karush-Kuhn-Tucker conditions, complemented by relaxation and gap continuation procedures. Implemented in the open-source Python package BLVPY, an extension of CVXPY, DBLP enables users to solve bilevel problems with minimal code and no prior expertise. The framework's efficacy is demonstrated across various application domains, showcasing its modeling and solution capabilities.
bilevel optimizationdisciplined programmingconic formkarush-kuhn-tuckercvxpy
BeamRMX: Radiation-Pattern-Driven Learning for Generalizable Beam Radio Map Prediction and Beam Management
BeamRMX introduces a novel framework for beam radio map (BeamRM) prediction in beamformed MIMO systems, treating spatial radiation patterns as primary queries to model their transformation into received power fields via scene geometry. The method employs XBase for multiscale interaction learning and an optional Evidence Adapter leveraging cross-configuration BeamRMs. Evaluations demonstrate significant improvements, including 26.1% and 47.8% reductions in mean absolute error for unseen scenes and configurations, respectively, outperforming deterministic and diffusion baselines. Cross-configuration evidence further enhances reconstruction and intra-sector beam refinement.
beam radio mapradiation patternbeamformed mimocross-configurationscene geometry
Topological Steering
We introduce Topological Steering, a novel framework for controlling large language model (LLM) behavior by leveraging topological representations of activation spaces. Motivated by Topological Data Analysis (TDA), our method employs persistence diagrams to capture global structural patterns in activation spaces, addressing limitations of local feature-based interventions that are sensitive to outliers and distributional shifts. Experimental results demonstrate that Topological Steering robustly modifies LLM behavior across diverse model families and scales, offering improved stability compared to traditional activation-space interventions.
topological steeringpersistence diagramsactivation spacestopological data analysisbehavioral control
CRAFT: Fine-Tuning Pre-hoc Explainability in AI-native 6G RAN
CRAFT introduces a data-centric method for fine-tuning small language models (SLMs) in AI-native 6G RAN to enable pre-hoc explainability, addressing the cold-start barrier where SLMs struggle to simultaneously output reasoning traces and labels. The method autonomously generates verified datasets of (input, trace, label) triplets and fine-tunes SLMs using low-rank adaptation (LoRA), requiring less compute and wall-clock time than Group Relative Policy Optimization (GRPO). On TRACTOR and IC xApp datasets, CRAFT achieves 86.5% accuracy and 94.6% F1 with no parse failures, outperforming GRPO and SFT+GRPO. It also reduces energy consumption by 59% and provides a robust foundation for subsequent GRPO fine-tuning.
pre-hoc explainabilitylow-rank adaptationcold-start barriersmall language models6g ran
Real-Time Neuromorphic Spectrum Intelligence Simulator
The Real-Time Neuromorphic Spectrum Intelligence Simulator (RT-NuSIS) introduces a modular framework for studying spiking neural networks (SNNs) and memristor-inspired agents in dynamic spectrum access scenarios under energy constraints and adversarial conditions. RT-NuSIS integrates leaky integrate-and-fire neuronal dynamics, memristive synaptic models, physics-informed energy-harvesting models (triboelectric and RF), and adversary models such as jamming and Byzantine behavior. The simulator is mathematically formalized, with proven boundedness, a mean-field adversary threshold, and per-step complexity analysis. It includes a reproducible benchmark harness for evaluating energy-per-inference, latency, and robustness metrics. The framework is deterministic by seed and optimized for large-scale event-driven simulations.
spiking neural networksmemristive synaptic modelsdynamic spectrum accessenergy-harvesting modelsevent-driven simulations
Manifold-Aware General Coded Computing for Straggler-Resilient Distributed Computing
The paper introduces a manifold-aware encoding strategy for general coded computing (GCC) that leverages the intrinsic low-dimensional manifold structure of high-dimensional datasets, contrasting with traditional approaches that impose artificial algebraic structures. Inspired by graph-based manifold learning, the method preserves data geometry during encoding, rather than eliminating it through source coding. Experiments on neural network inference and high-dimensional polynomial evaluation demonstrate that the proposed strategy significantly reduces mean squared recovery error under straggler conditions compared to standard GCC.
manifold-aware encodinggeneral coded computingstraggler-resilientgraph-based manifold learningmean squared recovery error
GenONet: A Generative operator Network for High-Resolution Precipitation Nowcasting
We introduce Spatio-Temporal U-DeepONet (GenONet), a novel architecture for high-resolution precipitation nowcasting up to 3 hours, designed to produce sharp and physically consistent forecasts. GenONet integrates a Deep Operator Network (DeepONet) as a generator within a Generative Adversarial Network (GAN) framework, learning continuous-time dynamics of precipitation for stability. Adversarial training with a spatio-temporal discriminator ensures sharp forecasts, while a physics-informed loss regularizer derived from the Moisture Conservation Equation enhances physical plausibility. Quantitative evaluations demonstrate superior performance on most metrics, particularly for high-intensity events and longer lead times. Ablation studies confirm the efficacy of combining operator learning with adversarial training.
deep operator networkgenerative adversarial networkphysics-informed lossprecipitation nowcastingspatio-temporal discriminator
DeSyR: A Decoupled Symbolic Recovery Framework with PINN-Guided Structure Search and Physics-Informed Coefficient Refinement
DeSyR introduces a decoupled symbolic recovery framework for differential equations, combining physics-informed neural networks (PINNs) with symbolic topology search and coefficient refinement. The method employs PINN-guided topology construction, followed by physics-only coefficient refinement and gated selection. Theoretical analysis characterizes teacher-error inheritance in linear parameterizations, showing O(β⁻¹) teacher-dependent contributions, while physics-only refinement guarantees exact coefficient recovery under specific conditions. Evaluated on 15 differential-equation problems across 18 configurations, DeSyR achieves 99.23% convergence in free-parameter refits and reduces errors by 8-14 orders of magnitude in same-topology comparisons, demonstrating effective topology discovery without inheriting teacher error scales.
physics-informed neural networkssymbolic recoverycoefficient refinementtopology searchdifferential equations
Why Multi-Layer Message Passing Works: Completeness Theory for Graph Neural Network Interatomic Potentials
The paper establishes a multi-layer completeness theory for graph neural networks (GNNs) in interatomic potential modeling, proving that Hypergraph Neural Networks with 3-body message passing are universal approximators for potential energy surfaces. The authors demonstrate that L layers of message passing on sparse, cutoff-based graphs achieve equivalent representational power to full L-hop neighborhood access, given generic configurations satisfying overlap and connectivity conditions. This result rigorously justifies the widespread practice of using multi-layer message passing with per-layer cutoffs smaller than physical interaction ranges. As corollaries, the universal approximation capability is shown to extend to DPA3 and CHGNet architectures.
graph neural networksmessage passinguniversal approximatorinteratomic potentialshypergraph
Soft-Argmax for the Projective Plane via the Veronese Embedding
The paper introduces Veronese soft-argmax, a method for recovering lines in Hough space $H=S^1\times\mathbb{R}$ that addresses the geometric tearing caused by the double-cover $H/\mathbb{Z}_2$. By parametrizing lines via unit-norm homogeneous vectors and applying the Veronese map $v_2(\ell)=\ell\ell^{\top}$, the method embeds the quotient $H/\mathbb{Z}_2$ into the linear space $\mathrm{Sym}^2(\mathbb{R}^3)$, resolving antipodal ambiguity. Line extraction is performed as a barycentre in $\mathrm{Sym}^2(\mathbb{R}^3)$, projected back via its leading eigenvector. Validation in a Hough transform-based network demonstrates uniform and seam-free recovery across all resolvable lines. The $L_2$-loss on Veronese embeddings is shown to equal the squared chordal distance between lines in projective space, providing a geometrically precise training objective.
veronese soft-argmaxhough spaceveronese mapprojective spacechordal distance
Learning Task-Specific Antibody Representations via Function-Aware Masking
The authors introduce function-aware masking, a family of pretraining algorithms that align mask placement with specific functional priors to shape antibody representation spaces in language models. By leveraging biological priors from IMGT annotations or structure predictions, the method imposes inductive biases during pretraining. Specialist masking strategies yield significant improvements, including a 14% gain on structure-related tasks and a 5.9x improvement on CDR-related tasks. Hybrid masking strategies further enhance performance across multiple functional axes, balancing reconstruction over binding, structural, and biophysical objectives. The approach demonstrates that informed mask placement provides a parameter-free mechanism for functional inductive bias in antibody language model training.
masked language modelinginductive biascomplementarity-determining regionsfunction-aware maskingantibody language models
VATO: A Vortex-Force-Aware Transformer Operator for Unsteady Separated Aerofoil Flows
We introduce VATO (Vortex-Force-Aware Transformer Operator), a geometry-aware neural operator that integrates the Vortex Force Map (VFM) method to improve prediction of unsteady separated aerofoil flows. VATO employs two mechanisms: VATO-S adds training-only supervision of the local VFM force-contribution field, while VATO-A uses VFM contribution and sensitivity fields to prioritize force-relevant source locations for residual cross attention. Evaluated on unsteady CFD data for double-edged-plate aerofoils over 54 trajectories, VATO-S reduces velocity, pressure, and vorticity errors by 10.4%, 1.0%, and 15.6%, respectively, while VATO-A achieves reductions of 15.8%, 7.5%, and 31.2%. VATO-A maintains significant error reductions in vorticity and force readouts even beyond the training range.
vortex force mapneural operatorunsteady flowsaerodynamic loadscross attention
A hybrid quantum-classical neural network for learning to route
A hybrid quantum-classical neural network architecture is proposed for learning routing heuristics, specifically targeting the capacitated vehicle routing problem. The study investigates replacing parameter-heavy modules in an attention-based routing model with small quantum neural networks, focusing on encoder feed-forward replacement as the primary design. This approach reduces model parameters by 56.6% while maintaining solution quality comparable to classical neural baselines for small and medium instance sizes, though performance gaps widen for larger instances. The hybrid model is benchmarked against classical routing algorithms, which remain competitive on fixed Euclidean test sets. The findings suggest encoder feed-forward replacement as a viable strategy for hybrid-module compression in neural combinatorial optimization.
hybrid quantum-classicalcapacitated vehicle routingencoder feed-forwardneural combinatorial optimizationattention-based routing
AdaptNTK: Adaptive Uncertainty Quantification and Active Learning for Neural Network Potentials
AdaptNTK introduces a single-model framework for adaptive uncertainty quantification and active learning in neural network potentials, addressing computational cost-reliability trade-offs and redundancy in acquisition batches. It measures uncertainty as a regularized Mahalanobis distance in empirical neural tangent kernel (NTK) feature space, enabling recursive updates without retraining. On rMD17 data, AdaptNTK achieves Spearman and Pearson correlations of 0.68 and 0.71 with force errors, matching ensemble performance. It demonstrates superior active learning performance on rMD17 and Transition-1X datasets, particularly on transition-state configurations, with a 2.6-fold speedup per Transition-1X cycle compared to ensemble methods.
neural tangent kernelactive learninguncertainty quantificationmahalanobis distanceneural network potentials
Are Near-Tied LLM Rankings Robust to Family-DIF-Guided Benchmark Recomposition?
This study demonstrates that near-tied rankings of language models (LLMs) on benchmarks are sensitive to item composition, challenging the robustness of small leaderboard gaps. Using item-level responses from five benchmarks and a family-label-free spectral approximation to multidimensional item-response theory (MIRT), the authors recompose benchmarks by identifying items with low residual differential item functioning (low-DIF) across model families. Results show strong correlations between full-benchmark and low-DIF rankings (τ_b=.900–.948), yet 30.9–47.1% of cross-family pairs within one percentage point reverse order, exceeding matched-random controls by 16.9–28.6 points (p=.001). This sensitivity persists across population perturbations, though no family consistently outperforms across benchmarks.
language modelsbenchmark compositiondifferential item functioningmultidimensional item-response theoryleaderboard gaps
Fractal dimension predicts quantum kernel collapse in angle-encoded data
The study introduces the correlation fractal dimension D2 as a predictive metric for qubit budget in angle-encoded quantum kernels, preventing kernel collapse when the feature map exceeds the data's intrinsic dimension. Using FD-ASE coordinate selection, the method encodes D2 coordinates instead of PCA-95% width or all attributes. Experiments on nine datasets with a statevector simulator (n=32) and IBM Quantum (ibm_fez, 256 shots, n=8) demonstrate that a one-layer ZZ fidelity kernel at q=D2 maintains geometric integrity, whereas PCA-95% width leads to collapse. Results show a mean absolute error (MAE) of 0.021 for the fractal width kernel, confirming the ceiling effect depends on map-data pair and bandwidth.
quantum kernelfractal dimensionangle encodingqubit budgetkernel collapse
Higher Structures in Deep Learning
The paper introduces higher-arity tensor operations as a critical component in deep learning, empirically investigates higher-arity phenomena in trained neural networks, and proposes a hypergraphical generalization of the multilayer perceptron. The authors explore connections to evolutionary algorithms and identify promising research directions. The study combines theoretical exposition with novel empirical analysis, offering insights into the structural complexity of deep learning models.
higher-arity tensor operationshypergraphical generalizationmultilayer perceptronevolutionary algorithmsempirical investigation
Context Window Failures in Relational Foundation Models
We introduce Animus, a synthetic financial dataset demonstrating context window failures in relational foundation models, where predicting customer income requires aggregating up to tens of thousands of transactions. We evaluate three recent models (RT, Griffin, RelGT) on raw data and with temporal pre-aggregation. Results show poor performance on raw data ($R^2 \le 0.18$) but significant improvement ($R^2$ up to $0.65$) with pre-aggregation, highlighting limitations of current models in handling high-cardinality relational data. This raises questions about their readiness for real-world applications.
relational foundation modelscontext windowhigh-cardinalitytemporal pre-aggregationsynthetic dataset
Can LLMs Use Relational Transformer Embeddings?
This work investigates the efficacy of injecting frozen relational-encoder embeddings into large language models (LLMs) for relational prediction tasks. The authors integrate embeddings from a frozen Relational Transformer (RT) into Qwen3.5-4B using a learned MLP projection and LoRA adaptation, trained via supervised fine-tuning and group-based reinforcement learning. Evaluation across 10 binary classification tasks on 6 RelBench databases under four supervision regimes reveals that the hybrid model frequently underperforms standalone RT, exhibiting instability and sensitivity to serialization format and relational-token budget. The authors conclude that soft-token fusion requires stronger alignment objectives and schema-aware design for reliable relational prediction.
relational transformersoft tokenslora adaptationgroup-based reinforcement learningrelbench
HBQ: Hierarchical Scaling Block Quantization with Hardware-Efficiency-Aware Design for Accurate LLM Inference
We propose Hierarchical Block Quantization (HBQ), a hardware-efficient quantization method for large language models (LLMs) that addresses the trade-off between block size, accuracy, and hardware efficiency. HBQ employs large blocks for efficiency and introduces low-overhead significand (SIG) scaling for second-level quantization, effectively compensating for errors caused by large blocks. HBQ-A achieves W4A16-level accuracy with W4A5, requiring less silicon area than NVFP4, while HBQ-E reduces hardware cost by 17% while maintaining higher accuracy than existing block quantization methods. Implemented on a 28nm ASIC accelerator, HBQ delivers 2.3×/4.6× higher area/energy efficiency compared to state-of-the-art weight-only quantization, with 1.6–3.3× system energy reduction and 1.5–3.0× speedup over prior block quantization methods.
hierarchical block quantizationlarge language modelshardware efficiencysignificand scalingasic accelerator
CRAD: Class-wise Reliability-Aware Distillation for Decentralized Heterogeneous Federated Learning
CRAD introduces class-wise reliability-aware distillation for decentralized heterogeneous federated learning, addressing the challenge of combining peer teacher predictions by considering per-class reliability. The method filters out teachers disagreeing with the peer consensus and weights the remaining predictions by their per-class precision, leveraging the inverse variance of accuracy estimates. Evaluated on CIFAR-10, CIFAR-100, and PathMNIST under severe non-IID data skew, CRAD outperforms existing methods in global accuracy across heterogeneous architectures.
federated learningknowledge distillationnon-iid dataclass-wise reliabilitydecentralized learning
Capability-Gated Language Models: Security Composes, Utility Does Not
The paper introduces capability-gated deployment for language models, enabling per-principal access control within a single set of model weights via sparse rank gating over a nested-factorization mechanism. Security properties compose provably under a monotone-elicitation assumption, with median held-out meets deepening suppression in two lineages. However, utility does not compose, as individually harmless profiles can collectively degrade retention and fluency, with no compositional bound identified. The method employs one-pass attribution for profile search and evaluates results on pre-registered held-out splits.
capability-gated deploymentsparse rank gatingnested-factorizationmonotone-elicitationone-pass attribution
Group Adaptive Clipping Policy Optimization
Group Adaptive Clipping Policy Optimization (GAPO) introduces an adaptive clipping boundary for importance-sampling ratios in reinforcement learning with verifiable rewards, addressing the limitation of fixed clipping that suppresses rare correct rollouts disproportionately. Motivated by a reverse-KL trust-region perspective, GAPO adapts the clipping threshold based on rollout advantage, preserving the standard PPO/GSPO surrogate while enhancing learning signals for exploration. Evaluated on Qwen and Llama models, GAPO improves Pass@1 and Pass@k metrics over fixed clipping and advantage-shaping baselines on math reasoning and coding benchmarks with low base model pass rates.
importance-samplingtrust-regionrollout advantageppoclip boundary
Physiological Information Reliability: Cross-Layer Adaptive Resource Allocation for Cardiovascular Sensing
We introduce Physiological Information Reliability (PIR), a cross-layer framework for cardiovascular sensing that jointly optimizes physiological information value, wireless communication, energy constraints, and edge-computation latency. PIR integrates multimodal ECG/PPG signal-quality estimation with adaptive network coding under burst-erasure conditions, using a contextual bandit (PIR-LinUCB) for resource allocation. Experiments demonstrate PIR-LinUCB achieves a low-energy operating point while maintaining medical latency constraints and competitive physiological estimation accuracy compared to fixed and heuristic policies. Limitations include proxy physiological information value estimation and simulated communication dynamics, motivating future clinical and real-channel validation.
cardiovascular sensingcontextual banditnetwork codingsignal-quality estimationedge-computation
SAGE: State-Grounded, Abstention-Aware Evaluation of Task-Oriented Dialogue Agents
SAGE introduces a state-grounded, abstention-aware evaluation framework for task-oriented dialogue agents, addressing limitations of conventional LLM-based judges. The method compiles workflow specifications into atomic criteria, verified through a cascade of symbolic and encoder/NLI components that abstain rather than guess, aggregating verdicts into turn-level decisions. SAGE-Core achieves 81-91% criterion decisions without LLM calls, outperforming GPT-4.1-based judges at zero cost. Human audits confirm strong label fidelity (κ=0.94), with SAGE-Core statistically matching the strongest LLM judge on visible failure classes. The framework demonstrates construct validity while honestly scoping limitations in state-consistency signals.
task-oriented dialogueabstention-awareworkflow specificationsymbolic verifiersencoder/nli
Accelerating Chemical Kinetics for Exoplanet Atmospheres using Neural Networks
A neural network-based local-box chemical kinetics solver is introduced for accelerating exoplanet atmosphere simulations, addressing computational limitations of classical methods. The model employs a residual flow-map architecture to emulate state-to-state flow-map problems, achieving microsecond-scale inference while maintaining percent-level accuracy. It operates across a broad parameter space: temperatures (300-3000 K), pressures (10^-6-10^4 bar), time steps (10^-3-10^8 s), and compositions (C/O ratio and metallicity ranging from 10^-2 to 10^3 times solar). The solver outperforms common machine learning architectures and handles the extreme stiffness of atmospheric chemistry robustly, offering a flexible and efficient alternative to traditional approaches.
chemical kineticsexoplanet atmospheresresidual flow-mapstate-to-state flow-mapstiffness
How Temporal Correlations Shape Memory in Linear Recurrent Neural Networks
This work provides an exact solution for the learning dynamics of linear recurrent neural networks (LRNNs) under temporally correlated inputs, revealing how memory formation is shaped by input correlations. The analysis shows that memory retention incurs a cost proportional to input correlation, leading to three key findings: (1) memory builds, overshoots, and partially decays during training, with settled networks retaining less past information; (2) memory retention switches off at a threshold determined by input similarity, independent of sequence length or long-range correlations; (3) optimal networks develop a feedthrough path for zero error when provided with an additional hidden dimension. The results explain why correlated inputs transform LRNNs into change detectors.
linear recurrent neural networkstemporal correlationsmemory retentionfeedthrough pathchange detectors
DynaNDE: Dynamic Near-Data Expert Scheduling for Batched MoE Inference
DynaNDE introduces a dynamic near-data expert scheduling framework to accelerate batched Mixture-of-Experts (MoE) inference on NPU-NDP systems. The framework employs an analytical performance model that captures hardware heterogeneity, data-movement costs, and communication-computation overlap, enabling per-layer expert scheduling across NPU and NDP while considering expert-level concurrency. DynaNDE also incorporates a reuse-aware runtime to minimize redundant parameter movement when experts reside in NPU memory. Experimental results demonstrate significant throughput improvements, achieving average speedups of 2.6× and 2.2× for the prefill and decoding stages, respectively, compared to state-of-the-art NPU-NDP MoE serving frameworks.
mixture-of-expertsnear-data processingneural processing unitexpert schedulingbatched inference
A Multi-Branch Feature Fusion Approach for Health Misinformation Detection and Propagation
The paper proposes a multi-branch fusion framework for health misinformation detection and propagation analysis, integrating transformer-based semantics, rhetorical cues, stance representations, and psychologically motivated proxies. Grounded in cognitive theories (ELM, TPB), the model introduces a Cognitive Propagation Score (CPS) for interpretable diffusion-risk assessment. Evaluated on Constraint, COVID-19_FNIR, and Monkeypox datasets, it achieves ROC-AUC up to 0.9999 and near-perfect Spearman's ρ (0.9952–0.9954) for propagation ranking. Ablations confirm complementary gains from psychological and rhetorical branches. The approach bridges cognitive theory and neural modeling but requires further validation of CPS against human judgments.
multi-branch fusioncognitive propagation scoreelaboration likelihood modelrhetorical cuesstance representations
A convolutional framework for detecting event-driven dynamics in energy price series
The paper introduces a convolutional neural network (CNN) framework for detecting heterogeneous event-driven dynamics in univariate time series. The CNN class exactly represents classifiers based on range, maximum drawup, maximum drawdown, and slope change, while uniformly approximating realised volatility and autoregressive explosiveness on compact domains. Error bounds for representative rules and an oracle inequality for learning across them are established. Simulations demonstrate the model's ability to match or outperform classifiers based on individual statistics as training samples grow. Applied to six daily energy price series, a hierarchical CNN identifies geopolitical dynamics during the 2026 Iran war and distinguishes a natural gas spike linked to weather.
convolutional neural networkunivariate time seriesmaximum drawdownrealised volatilityautoregressive explosiveness
NeuroPriv: Adversarial Representation Learning for Privacy in Wearable EEG Systems
This work demonstrates that compact EEG features in wearable systems expose sensitive user information, motivating privacy-aware representation learning. Using EEGMAT as a case study, the authors show that spectral and spatial EEG features achieve 0.788 balanced accuracy for cognitive-state classification but also enable gender, age, and identity inference with accuracies of 0.858, 0.789, and 0.692, respectively. An adversarial representation learning approach preserves task performance at 0.781 while reducing inference accuracies to 0.563, 0.467, and 0.206. The findings highlight the need for purpose-limited representations and privacy auditing in neurohealth systems.
eegneuroprivacyrepresentation learningwearable systemsadversarial learning
Hidden relationships in a document-derived property graph: top-k chunk embeddings and inverse-distance weighting over a dynamically evolving ontology
We propose a second-pass method for discovering latent semantic relationships in document-derived property graphs without altering extracted facts. The approach chunks documents, computes embeddings once, and identifies candidate node pairs via top-k nearest-neighbor queries across chunks using entity membership maps. Candidate pairs are scored using Shepard inverse-distance weighting with a rescaled chord distance metric, avoiding threshold-collapsing flaws in affine cosine scoring. The un-gated per-pair accumulators form a commutative monoid, ensuring order-independence and incremental scalability. Implemented across FalkorDB, Kinetica, ArangoDB, and Neo4j, the method demonstrates that 768- and 240-dimensional embeddings retain 92% and 72% edge fidelity against a 3072-D baseline while achieving a 25x faster top-k formulation.
entity membership mapsshepard inverse-distance weightingcommutative monoidtop-k nearest-neighborchord distance metric
Adapting Without Gradients: Affine Statistics Transport and What Its Certificate Can Tell You
The paper introduces CASTER, a gradient-free test-time adaptation method for frozen models that transports source class statistics via affine transformations in a discriminative subspace, requiring no backward passes or optimizer state. It outperforms k-NN on frozen features in 27/28 backbone-dataset settings while using 18x less memory. The authors identify reliability issues in low-sample regimes (e.g., -21.2 top-1 points on ImageNet-C) and propose a transportability certificate to gate updates, converting a -3.35-point average loss to +1.69 gain. The certificate is mechanism-specific, rejecting 95.7% of Tent updates while preserving minimal gains (0.6%).
test-time adaptationaffine transportfrozen modelsgradient-freetransportability certificate
Towards unsupervised representation learning for quantum data: quantum models with inference and generation
The authors develop a framework for unsupervised representation learning with quantum data, addressing the challenge posed by quantum states' lack of universal factorization properties. They propose models as joint quantum states over visible and latent systems, utilizing state-over-time maps to factorize into marginal states and inference/generation channels. Non-trivial models require non-linear maps to extended spaces. Three state-over-time maps are characterized, revealing a hierarchy tied to the positive-partial-transpose criterion. The Leifer-Spekkens construction supports inference and generation for PPT states, enabling quantum visible-latent correlations. Quantum counterparts of exact and approximate inference training are formulated, alongside weaker data extension notions.
quantum statesunsupervised learningpositive-partial-transposeinference trainingstate-over-time maps
Dr. Claw: An AI Scientist Workspace for Vibe Research
Dr. Claw introduces an open-source AI scientist workspace that integrates coding-agent executors into a controllable, auditable human-in-the-loop workflow, addressing fragmentation in end-to-end research processes. The system employs persistent state objects, a reusable skill library, and multi-executor coordination to link human decisions to AI execution, creating a traceable and recoverable loop for planning, execution, and writing. Evaluated against a bare command-line agent with the same backend executor, Dr. Claw demonstrates higher research completeness while maintaining an auditable process trail. The tool is available under AGPL-3.0 with GPL-3.0 upstream components.
coding-agentauditable workflowskill librarymulti-executor coordinationpersistent state objects
Do LLMs Know Your Neighborhood? Auditing LLM Priors for Neighborhood-Level Mobility Prediction and Structural Alignment
The study evaluates zero-shot large language models (LLMs) for neighborhood-level mobility prediction across four U.S. metropolitan areas, comparing their performance to supervised baselines. Using anonymized Cuebiq data, the authors construct point-level, trajectory-level, and temporal mobility outcomes, paired with sociodemographic and built-environment predictors. Supervised models achieve 0.580 average accuracy, outperforming the best LLM (0.435), with spatial extent outcomes showing the largest LLM-baseline gaps. Directional alignment analysis reveals LLMs rely on coarse, stable predictor-level priors, including asymmetric treatment of protected-group predictors, suggesting LLMs partially recover aggregate mobility patterns but require empirical alignment auditing to mitigate potential bias.
zero-shot learningmobility predictionsociodemographic predictorsdirectional alignmentprotected-group predictors
Latent-Space No-Arbitrage Geometry of Generative Models for Implied Volatility Surfaces
The paper introduces a geometric framework for analyzing no-arbitrage constraints in the latent space of generative models for implied volatility surfaces. By assigning each latent code a scalar margin based on no-arbitrage conditions, the study defines the admissible latent set and establishes conditions for its stability under perturbations. A level-set equation is formulated to guide dynamics toward the zero-margin boundary. The method applies to various generative architectures, including variational autoencoders and generative adversarial networks. Numerical experiments on Heston surfaces demonstrate varying admissible regions for similar reconstruction errors and show that latent priors may concentrate within these regions. The framework enables modification of latent codes to enforce no-arbitrage constraints.
latent spaceno-arbitragegenerative modelsimplied volatilityvariational autoencoders
Workload Identification with Physical Side Channels for AI Governance
The study demonstrates that GPU workload classes, including AI training, inference, and non-AI computation, can be identified from physical power draw traces with 97% accuracy and a macro-averaged F1 score of 0.955. Using 930 five-second traces recorded at ~10 MHz across seventeen LLM families and twenty-five non-AI workloads, the method distinguishes training through memory-bound optimizer updates, with spectral content primarily below ~20kHz. Adversarial evasion strategies, including diluted low-rank adaptation (LoRA), were tested, with hardened detectors catching training ≥99% for three strategies and 48–88% for LoRA, rising to ≥98% with a rescue rule.
gpupower drawlow-rank adaptationworkload identificationadversarial evasion
TRUST: Threshold-Recalibrated Uncertainty-Safe Training for Certified Dismissal in Breast Cancer Screening
The paper introduces TRUST, a threshold-recalibrated uncertainty-safe training method for certified dismissal in breast cancer screening, aiming to reduce radiologist workload while maintaining cancer detection accuracy. The approach employs a closed-loop threshold-aware training strategy, dynamically recalculating dismissal thresholds during training and penalizing cancer-positive images nearing the dismissal region. Evaluated on NLBS and RSNA datasets, TRUST achieved case-level dismissal rates of 19.74% and 21.70% at 98% and 95% recall targets on NLBS, outperforming cross-entropy baselines. On RSNA, dismissal rates improved from 7.04% to 14.31% and from 13.49% to 19.69%. External validation on RSNA→NLBS yielded dismissal rates of 12.95% and 19.87% at respective recall targets.
threshold-aware trainingcertified dismissalbreast cancer screeningcase-level assessmentrecall targets
Geometry-aware Latent Autoregressive Generative Model for PDEs in Complex Domains
GeoLAMP, a Geometry-aware Latent Autoregressive generative Model for PDEs, addresses multiphysics PDE solving in complex $μ$m-scale geometries. The model employs a dual-encoder architecture on graph representations to capture global topology and fine-scale geometric features, transitioning real-space fields to compact latent representations. A causal self-attention transformer with flow matching models temporal dynamics, enabling stable block-wise autoregressive prediction, while a flexible decoder reconstructs high-resolution physical fields on arbitrary points. GeoLAMP demonstrates stable autoregression performance on three multiphysics benchmark datasets covering reactive flow, heat convection, and elasticity, maintaining low errors throughout the rollout horizon.
autoregressive predictionflow matchingdual-encoder architecturegraph representationscausal self-attention transformer
WiSDoM: Wireless Sparse Decision Transformer with Mixture-of-Experts for Multi-Task Mobile Network Optimization
WiSDoM introduces a sparse multi-task offline reinforcement learning framework for adaptive multi-cell selection in 6G wireless networks, addressing challenges in radio resource management across heterogeneous environments. The method combines Decision Transformers with a Mixture-of-Experts architecture, enabling dynamic expert activation based on task characteristics to improve model capacity and mitigate negative transfer. Trained on diverse network configurations, WiSDoM outperforms heuristic methods and conventional models, improving quality of experience by up to 55% while activating only one-third of dense model parameters during inference, and demonstrates strong generalization to unseen scenarios through few-shot prompting.
decision transformersmixture-of-expertsradio resource managementoffline reinforcement learning6g networks
Lightweight Adaptation of EEG Foundation Models for Stroke Motor Imagery Decoding: Domain Shift and Subject-Level Robustness
This work demonstrates that Low-Rank Adaptation (LoRA) enables efficient domain adaptation of EEG foundation models for stroke motor imagery decoding, addressing the domain shift between healthy and pathological EEG data. The authors evaluated LoRA-based adaptation of three pretrained models (LaBraM-base, REVE-base, REVE-large) on two datasets: PhysioNet EEGMMIDB and UET175 stroke dataset. LoRA improved accuracy to 0.822 (LaBraM-base) and 0.957 (REVE-base) on EEGMMIDB, while REVE-base achieved 0.847±0.194 on UET175, outperforming REVE-large (0.806±0.178). Leave-one-subject-out validation revealed high mean accuracy (0.952) but substantial subject-wise variability (0.586-1.000). Zero-shot transfer from EEGMMIDB to UET175 remained near chance (0.464±0.072), highlighting the need for target-domain adaptation in rehabilitation BCI systems.
low-rank adaptationeeg foundation modelsmotor imagery decodingdomain shiftstroke rehabilitation
Exact Global MCMC with Denoising Diffusion
The paper demonstrates that denoising diffusion models trained with standard loss functions can serve as effective global MCMC proposals for high-dimensional target densities. The method leverages the observation that sequential forward and reverse diffusion processes define a Markov chain with a target stationary distribution, made exact via a Metropolis-Hastings step incorporating forward and reverse path densities. The authors propose Denoising Diffusion Monte Carlo (DDMC), combining global denoiser-based path sampling with local MALA sampling. Experiments indicate DDMC achieves high acceptance rates across complex target densities, suggesting diffusion training scaling behavior transfers to exact sampling from high-dimensional unnormalized densities.
denoising diffusionmetropolis-hastingsmarkov chainmalasde
CompanionSim: Synthetic Data for Evaluating Anthropomorphism in Human-AI Relationships
CompanionSim introduces a simulation framework generating 2,240 multi-turn human-chatbot dialogues across 16 behaviors and seven use cases to study anthropomorphism in human-AI relationships. The framework synthesizes data by scaling limited real-world interactions, enabling researchers to evaluate companionship behaviors like validation. Two studies (N₁=628 U.S. participants; N₂=3,646 across U.S., U.K., India, Nigeria) annotated simulated and real conversations, revealing that companionship behaviors reduced chatbot likability, humanlikeness, and trust, particularly among women and older participants. The work advocates combining synthetic and real-world data for benchmarking AI chatbots and understanding differential impacts of companionship.
anthropomorphismhuman-ai interactionsynthetic datacompanionship behaviorsbenchmark evaluation
QTEA: Ternary LLMs with Sparse Residual Salient Weight and By-Column Optimization
QTEA introduces a sub-2-bit post-training quantization (PTQ) framework for large language models (LLMs), quantizing weights into ternary values and using salient weights as residual error compensators. It employs semi-structured 1:4 sparsity within salient columns and column-wise rescale refinement to reduce reconstruction error, while addressing order-dependent error propagation in GPTQ-style quantization. QTEA achieves an effective 1.7 bits per weight, improving average accuracy by 16.7% over ternary PTQ baselines on Qwen3-14B and reducing perplexity by 1.40× and 2.61× on WikiText and C4, respectively. Similar gains are observed on Llama3-8B, with a 6.6% accuracy improvement and 1.34×/1.95× lower perplexity. A lookup-table-based kernel achieves 7.2× faster per-token generation compared to FP16.
post-training quantizationternary quantizationresidual errorsemi-structured sparsityperplexity
WHALE: A Simple Recipe for Joint Harness-Weight Optimization
We propose Weight-Harness Alternating Learning (WHALE), a joint optimization method for model parameters and executable harness code, addressing limitations of isolated optimization. WHALE alternates between updating model weights via online rejection-sampling fine-tuning and searching for improved harnesses via Meta-Harness, with phase switching governed by fixed durations or adaptive patience rules. Evaluated on Qwen3.5-2B/4B agents across search question answering, mathematical reasoning, and chess puzzles, WHALE outperforms weight-only, harness-only, and Fast-Slow Training by 4.15-24.38 percentage points in best mean@8 accuracy. Results show harness search can match peak weight-only accuracy with fewer rollouts, and interleaved updates outperform stagewise optimization in accuracy and rollout cost.
joint optimizationharness coderejection-samplingmeta-harnessadaptive patience
Provably Efficient Federated Reinforcement Learning with Linear Function Approximation and Logarithmic Communication Cost
(No summary returned.)
Elite-Weighted Supervised Fine-tuning for Goal-Directed Molecular Optimization
Elite-Weighted Supervised Fine-tuning (EW-SFT) is introduced as a unified optimizer for goal-directed molecular optimization across diverse generative architectures. EW-SFT selects high-scoring molecules based on reward, then updates the model using its pretraining loss on this elite set, avoiding trajectory-level reinforcement learning. The method is architecture-agnostic, applicable to autoregressive, masked-diffusion, and discrete-flow generators, and supports de novo, motif-extension, and linker-design tasks. Evaluations demonstrate EW-SFT's superiority over native optimizers in 3D shape alignment oracle calls for kinase compounds and 2D similarity oracle performance on held-out references. It achieves comparable sample efficiency without RL formulation, establishing EW-SFT as a versatile molecular optimization framework.
molecular optimizationsupervised fine-tuningreward-guided selectionarchitecture-agnosticgoal-directed optimization
Synthetic Worlds for Temporal Evaluation and Knowledge Updating in LLMs
We propose a synthetic simulation-driven framework for evaluating knowledge updates in LLMs, addressing limitations of static pretraining corpora and counterfactual edits. Our approach introduces ParallelEvents, a benchmark generating coherent fictional future event trajectories for controlled evaluation, and Synapse, a training framework leveraging model-generated data for parameter updates via mid-training and instruction tuning. This pipeline enables scalable knowledge integration without human-curated data. Empirical results demonstrate Synapse outperforms existing methods by 14.23%, indicating robust and coherent knowledge insertions through simulation-based synthetic training.
large language modelssynthetic trainingknowledge insertioninstruction tuningmid-training
Lingua Franca or Probing Artifact? Rethinking Latent Language in Multilingual LLMs
This work critically examines latent language identification in multilingual language models, challenging the assumption that probes measure a unified internal lingua franca. The authors analyze discrepancies between Gaussian Mixture Model (GMM)-based representation probes and decoding-based probes across 27 languages, varying model families, training regimes, domains, tasks, and checkpoints. Results demonstrate systematic divergence: GMM probes indicate earlier cross-lingual mixing, while decoding probes retain sharper language-specific and English-biased signals. These differences persist across domains but vary with model multilinguality and training progression. The findings suggest that current probes expose distinct aspects of multilingual processing rather than revealing a single internal language representation.
latent languagemultilingual llmsgmm probedecoding probecross-lingual mixing
Flawed in Nature, Perfect through Evolution
We introduce 'Flawed in Nature, Perfect through Evolution', a mechanism for improving AI/ML model performance in non-stationary environments by maintaining model diversity through deliberate mutations of model coefficients. The approach leverages a swarm of models subjected to mutations that deviate from optimality, acting as a statistical hedge against environmental drift. Four theorems guarantee regret reduction under general conditions. Empirical validation on synthetic linear regression tasks shows the mutated swarm delivers the best model in ∼80% of environment changes, with optimal performance when mutation drift rate matches environmental drift rate. An adaptive controller enables practical application by tuning the mutation drift rate.
non-stationaritymodel diversityregret reductionmutation drift rateadaptive controller
Good Memory Has ECC: Evaluating the Memory of Vision-Language Models Beyond Accuracy
We introduce ECCBench, a benchmark evaluating memory in vision-language models (VLMs) beyond accuracy via three axes: efficiency (FLOPs required), compression (compressible input recall), and calibration (uncertainty-aware abstention). Testing pretrained VLMs reveals they compress text but not video memories and exhibit poor calibration. Comparative analysis shows non-Transformer architectures achieve superior compression-calibration tradeoffs over RoPE Transformers, suggesting their utility in long-horizon agent tasks.
eccbenchvlmscompressioncalibrationrope transformers
Different representation learning objectives recover distinct latent structures from the same psychometric data
The study demonstrates that distinct representation learning objectives recover different latent structures from the same psychometric data. Using 757 teacher-child pairs from the Cyprus ProW preschool trial, behavioral phenotypes were derived from child SDQ, ASBI, and CBRS item responses via principal component analysis (PCA) and clustering. A contrastive objective improved teacher-child retrieval, increasing Top-1 accuracy from 0.13% to 7.27% and Top-10 accuracy from 1.98% to 56.14%, but preserved behavioral phenotypes less effectively than PCA. A multi-task objective partially restored behavioral organization but reduced retrieval performance, indicating that teacher-child correspondence and behavioral phenotypes represent distinct latent structures.
representation learningpsychometric datacontrastive objectivebehavioral phenotypesprincipal component analysis
📰 Industry Media (9)
Facilitating AI integration with simplicity at scale
Jabil, a global manufacturing firm with 100+ sites across 30+ countries, prioritized system integration and simplification to address operational complexity. By adopting a 'simplify-first, then-innovate' approach, the company standardized processes, consolidated tools, and established a unified data backbone using SAP Integration Suite. This enabled seamless data flow across systems, improved real-time supply chain visibility, and reduced manual reconciliation. The foundation supports future AI-driven applications, such as predictive insights and intelligent exception handling. The integration strategy enhanced operational resilience and scalability, demonstrating that simplicity at scale provides a competitive advantage.
system integrationdata backbonesupply chain visibilityoperational resiliencepredictive insights
Meet Switchyard: A Rust Proxy and Library That Routes and Translates LLM Traffic Across OpenAI and Anthropic APIs
Switchyard, an Apache-2.0 licensed Rust proxy and library, introduces a translation layer for routing LLM traffic across OpenAI and Anthropic APIs without requiring client-side modifications. It decodes requests into provider-neutral Rust types, applies routing algorithms (passthrough, random, LLM-classifier, stage_router), and re-encodes requests into the target backend's format. The system supports OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages, enabling seamless interoperability. Operational metrics, including routing overhead and model-call latency, are exposed via Prometheus. Switchyard is pre-alpha, experimental, and not recommended for production use, serving primarily as an evaluation tool for multi-provider LLM deployments.
rust proxyllm trafficrouting algorithmsprometheus metricsapi translation
Google DeepMind Releases Gemini 3.8 Flash and Gemini 3.8 Flash Cyber: One Core Model, Two Access Envelopes
Google DeepMind introduces Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, two variants of a single core model differentiated by safety envelopes and access controls. Both models leverage agentic loops for recursive evaluation, maintaining a 1,048,576-token context window and 65,536-token maximum output. Gemini 3.8 Flash enhances accuracy through additional reasoning steps and iterative tool calls, albeit at higher token consumption, achieving 54.9% on DeepSWE v1.1 and competitive performance on Vals Finance Agent V2 and Harvey’s Legal Agent Benchmark. Gemini 3.8 Flash Cyber, restricted to trusted defenders via the Fairwind Program, demonstrates frontier-level vulnerability discovery and patching, with 47.2% pass@1 on CWE-Bench and 70% success rate across 20 programming languages.
agentic loopscontext windowreasoning stepsvulnerability discoverypass@1
Anthropic Introduces Enterprise Frontier Safeguards (EFS): Zero-Data-Retention Privacy Plus Cross-Session Misuse Detection
Anthropic introduces Enterprise Frontier Safeguards (EFS), a novel architecture enabling zero-data-retention (ZDR) privacy while maintaining cross-session misuse detection. EFS stores monitoring data in customer-controlled cloud infrastructure, with detection handled by Anthropic's automated systems. The design emerged from collaboration with over 100 enterprise customers across sectors, including Fortune 100 companies and global systemically important banks. Key features include customer-managed storage, encryption keys, and automated review without Anthropic human intervention. EFS, currently in phased rollout, complements Claude Fable 5.1 and Mythos 5.1, offering reduced cache reads (75% lower) and fewer cybersecurity interventions (60% reduction). The solution incurs no additional cost beyond cloud provider storage and egress fees.
zero-data-retentionmisuse detectioncloud infrastructureautomated reviewcache reads
Meta Superintelligence Labs Releases Muse Voice Transcribe: One Real-Time Model for Streaming ASR, Diarization, and Endpointing
Meta Superintelligence Labs introduces Muse Voice Transcribe, a unified autoregressive model for real-time streaming ASR, diarization (20+ speakers), and endpointing, eliminating multi-system pipelines. The model processes 80ms audio chunks at 12.5 Hz, emitting text or continuation tokens adaptively via RL-optimized delay policies. It achieves 3.1% WER at 0.16s latency (Artificial Analysis AA-WER Streaming) and 17.5% diarization error rate (AMI/VoxConverse), supporting 70+ languages with native code-switching. Deployed as a hosted API ($3.00/1k minutes), it outperforms Cartesia Ink-2 and ElevenLabs Scribe v2 in accuracy-latency trade-offs.
streaming asrautoregressive modeldiarizationadaptive delaycode-switching
Perplexity Releases Hybrid Compute on Mac: Cloud Agents Orchestrate Down to a Local Model, Gated On Device
Perplexity introduces hybrid compute for Mac, enabling task orchestration between cloud-based frontier models and local on-device models to handle sensitive data securely. The system employs an on-device privacy gate, powered by PII-Tracer, a 0.6B bidirectional encoder adapted from Qwen3, which classifies and masks sensitive information. PII-Tracer achieves a character F1 score of 0.629 and excels in consistency, detecting 79.4% of recurring identifiers. The solution supports macOS 15+ on Apple silicon with ≥24GB unified memory, offering enterprise-grade privacy controls and audit logs.
hybrid computeon-device privacy gatepii-tracerbidirectional encodertask orchestration
Motional and MIT AI explains self-driving car decisions
Motional and MIT CSAIL researchers developed Concept-Wrapper Network (CW-Net), a real-time explainable AI system for autonomous vehicles that translates neural network decisions into human-interpretable concepts like 'Approaching Stopped Vehicle'. Unlike post-hoc explanations, CW-Net integrates causally faithful concepts directly into decision-making, enabling precise attribution of actions like braking to specific triggers. Deployed on public roads in Las Vegas, CW-Net revealed critical insights, including hallucinated objects and reliance on backup systems, while maintaining <1% performance degradation compared to state-of-the-art autonomous driving algorithms. This approach addresses regulatory demands for transparency in safety-critical AI systems.
concept-wrapper networkcausally faithfulneural networkautonomous vehiclesexplainable ai
ChatGPT Ads passes $1B run rate in 200 days
ChatGPT Ads achieved a $1 billion annualized revenue run rate within 200 days, driven by its integration into user conversations and expansion to over 40 countries. OpenAI employs a context-aware ad selection mechanism that leverages ongoing ChatGPT interactions, enabling advertisers to target users based on conversational context and broader interests. The platform supports CPC and outcome-optimized bidding, with pixel and Conversions API integrations for measurement. Self-service Ads Manager access has been extended to India, Europe, the Middle East, and North Africa, facilitating SMB participation. EU regulators designated ChatGPT as a Very Large Online Search Engine under the Digital Services Act, imposing additional compliance obligations.
context-awarecpcconversions apiself-servicedigital services act
Why MCP servers are becoming AI’s newest attack surface
The rapid adoption of MCP (Model Connector Protocol) servers as the standard for connecting AI agents to external tools and data has introduced significant security vulnerabilities, outpacing existing defenses. MCP, launched by Anthropic in November 2024, became widely adopted by major AI platforms by December 2025, with over 10,000 active public servers. However, 40% of these servers exhibit exploitable weaknesses, including tool poisoning, rug pull attacks, and data exfiltration. Security vendors like Check Point, Cisco, and TrueFoundry are developing AI-specific firewalls and governance solutions to address these risks, integrating MCP security into broader AI ecosystem defenses.
mcp serverstool poisoningai firewalldata exfiltrationrug pull attacks
Generated automatically at 2026-09-02 22:21 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.