Daily Digest — 2026-05-13
353 items · 3 research labs, 344 arxiv papers, 6 industry media
🏛️ Research Labs (3)
What Parameter Golf taught us about AI-assisted research
OpenAI's Parameter Golf challenge demonstrated the impact of AI-assisted research through a constrained ML competition (16MB artifact limit, 10-minute training on 8×H100s). Analyzing 2,000+ submissions from 1,000+ participants revealed emergent patterns: optimization via quantization (8-bit weights), test-time training strategies, and novel architectures (non-autoregressive text modeling). AI coding agents reduced experimentation costs, enabling rapid iteration but introducing attribution challenges. Top submissions achieved 1.12 BPB on FineWeb, outperforming transformer baselines. The study highlights how agent-augmented competitions can surface technical creativity while requiring new tooling (Codex-based triage) for scalable evaluation.
quantizationtest-time trainingnon-autoregressiveartifact limitcoding agents
How ChatGPT adoption broadened in early 2026
OpenAI's Q1 2026 adoption analysis reveals broadening ChatGPT usage across demographic and geographic dimensions, based on message volume from consumer plans (excluding enterprise/education). Gender inference methodology shows users with typically feminine names now exceed 50% of inferable cases, while age cohorts over 35 gained share despite younger users remaining dominant. Per-capita message rankings indicate fastest growth in Latin America, Asia-Pacific, and Africa, with workplace usage shifting from general content creation toward specialized tasks like health documentation and information retrieval.
gender inferenceper-capita metricsconsumer adoptionworkplace automationdemographic analysis
Building Blocks for Foundation Model Training and Inference on AWS
The article presents a technical analysis of AWS infrastructure components optimized for foundation model training and inference, focusing on the integration between hardware and open-source software stacks. It details three critical building blocks: accelerated compute (NVIDIA H100/H200/B200 GPUs), high-bandwidth networking (EFA v2/v3/v4, NVLink), and distributed storage (Lustre, S3). The architecture leverages Slurm/Kubernetes for resource orchestration and Prometheus/Grafana for observability, emphasizing scalability across pre-training, post-training, and inference workloads. Results highlight AWS UltraClusters' petabit-scale networking and UltraServers' 72-GPU NVLink domains as key enablers for large-scale distributed training.
foundation modelselastic fabric adapternvlinkultraclusterslustre
📜 arXiv Papers (344)
ELF: Embedded Language Flows
Embedded Language Flows (ELF) introduces a novel class of continuous diffusion models for language modeling, operating primarily in continuous embedding space before mapping to discrete tokens at the final step. ELF leverages continuous-time Flow Matching and adapts techniques from image-domain diffusion models, such as classifier-free guidance (CFG). Experiments demonstrate that ELF outperforms existing discrete and continuous diffusion language models in generation quality while requiring fewer sampling steps. This approach establishes a promising direction for effective continuous diffusion models in language tasks.
diffusion modelsflow matchingcontinuous embeddingclassifier-free guidancelanguage modeling
Variational Inference for Lévy Process-Driven SDEs via Neural Tilting
The authors introduce a neural exponential tilting framework for variational inference in Lévy-driven stochastic differential equations (SDEs), addressing the limitations of existing Monte Carlo and Gaussian-based methods. Their approach constructs a flexible variational family by exponentially reweighting the Lévy measure using neural networks, preserving jump structures while maintaining tractability. Key innovations include a quadratic neural parametrization for closed-form normalization, a conditional Gaussian representation for stable processes, and symmetry-aware Monte Carlo estimators. Empirical results demonstrate accurate capture of jump dynamics and reliable posterior inference on synthetic and real-world datasets, outperforming Gaussian-based variational methods.
lévy processesneural variational inferencestochastic differential equationsexponential tiltingheavy-tailed phenomena
Confidence-Guided Diffusion Augmentation for Enhanced Bangla Compound Character Recognition
A confidence-guided diffusion augmentation framework is proposed for low-resolution Bangla compound character recognition, addressing challenges of complex structures and limited annotated data. The method integrates class-conditional diffusion modeling with classifier guidance, employing Squeeze-and-Excitation enhanced residual blocks in the U-Net backbone and a confidence-based filtering mechanism to ensure synthetic sample quality. Augmented data is fused with original training data to retrain multiple classification architectures. Evaluated on the AIBangla dataset, the approach achieves 89.2% classification accuracy, significantly outperforming prior benchmarks across ResNet50, DenseNet121, VGG16, and Vision Transformer models.
diffusion modelingsqueeze-and-excitationclassifier guidancecompound characterconfidence-based filtering
Shepherd: A Runtime Substrate Empowering Meta-Agents with a Formalized Execution Trace
Shepherd introduces a functional programming model that formalizes meta-agent operations as Lean-mechanized functions, with agent-environment interactions recorded as typed events in a Git-like execution trace. This enables efficient forking and replaying of past states, achieving $5\times$ faster process forking than Docker and $>95\%$ prompt-cache reuse. Applications include runtime intervention, where a live supervisor improves pair coding pass rates from 28.8% to 54.7% on CooperBench; counterfactual meta-optimization, which outperforms baselines by up to 11 points while reducing wall-clock time by up to 58%; and Tree-RL training, enhancing TerminalBench-2 performance from 34.2% to 39.4%. Shepherd is open-sourced to support meta-agent research.
meta-agentexecution tracefunctional programmingforkingprompt-cache
Engineering Robustness into Personal Agents with the AI Workflow Store
The paper proposes integrating rigorous software engineering (SE) processes into AI agent workflows to enhance robustness and reliability, contrasting the prevalent 'on-the-fly' synthesis paradigm. It identifies a tension between flexibility and robustness, advocating for deterministic, hardened workflows that outperform improvised solutions. The authors envision an AI Workflow Store, enabling reuse of rigorously tested workflows across a broad user base to amortize computational costs. Key research challenges include balancing SE rigor with agentic flexibility and ensuring secure, production-grade workflows suitable for high-stakes scenarios.
software engineeringai workflowsrobustnessdeterministic constraintsadversarial evaluation
DataMaster: Towards Autonomous Data Engineering for Machine Learning
We propose DataMaster, an autonomous data engineering framework that optimizes machine learning systems by improving data pipelines while keeping the learning algorithm fixed. The framework integrates tree-structured search, shared data pools, and cumulative memory to address challenges in external data discovery, selection, composition, cleaning, and transformation. DataMaster employs a DataTree for organizing alternative data-engineering branches, a shared Data Pool for reusing discovered external data, and a Global Memory for recording node outcomes and reusable findings. Evaluations on MLE-Bench Lite and PostTrainBench show improvements: a 32.27% increase in medal rate and surpassing the instruct model on GPQA (31.02% vs 30.35%).
autonomous data engineeringtree-structured searchdata poolglobal memorydownstream feedback
Unmasking On-Policy Distillation: Where It Helps, Where It Hurts, and Why
The paper introduces a training-free diagnostic framework for evaluating on-policy distillation at per-token, per-question, and per-teacher granularity. It derives an ideal per-node gradient maximizing the student's success probability and proposes a scalable targeted-rollout algorithm for efficient estimation. The gradient alignment score, measuring cosine similarity between the ideal and distillation gradients, reveals that distillation guidance aligns better with incorrect rollouts than correct ones, where teacher signals become noisy. Results show optimal distillation context depends on student capacity and task, with no universally effective configuration, advocating per-task, per-token diagnostic analyses.
on-policy distillationgradient alignmenttargeted-rolloutper-node gradientcosine similarity
Shields to Guarantee Probabilistic Safety in MDPs
The paper introduces a formal framework extending classical shielding techniques to probabilistic safety in Markov Decision Processes (MDPs). It demonstrates the impossibility of maintaining strong safety and permissiveness guarantees in probabilistic settings, proposes weaker guarantees for natural shields, and presents offline/online shield constructions with robust safety assurances. Empirical results validate the practical benefits and computational feasibility of the proposed shields.
shieldingprobabilistic safetymarkov decision processespermissivenessonline shield
LoKA: Low-precision Kernel Applications for Recommendation Models At Scale
LoKA introduces a framework enabling FP8 precision in large recommendation models (LRMs) through system-model co-design, addressing numerical sensitivity and communication-intensive training. It comprises three components: LoKA Probe, which profiles activation and weight statistics to identify safe FP8 adoption sites; LoKA Mods, a set of model adaptations enhancing numerical stability and efficiency; and LoKA Dispatch, a runtime selecting optimal FP8 kernels based on accuracy and speed. This approach mitigates quality degradation and training delays inherent in direct FP8 application to LRMs, leveraging GPU capabilities for higher FLOPs.
fp8low-precision arithmeticlarge recommendation modelssystem-model co-designkernel libraries
AssayBench: An Assay-Level Virtual Cell Benchmark for LLMs and Agents
AssayBench introduces a benchmark for phenotypic screen prediction using 1,920 publicly available CRISPR screens across five cellular phenotype classes, addressing the lack of standardized evaluation for in silico phenotypic screening. The task is formulated as gene rank prediction per screen, employing the adjusted nDCG metric for heterogeneous assay comparison. Evaluations reveal that zero-shot generalist LLMs outperform biology-specific LLMs and trainable baselines, with further improvements achievable through fine-tuning, ensembling, and prompt optimization. AssayBench serves as a practical testbed for advancing virtual cell models and phenotypic screening capabilities.
phenotypic screeningcrispr screensadjusted ndcgvirtual cellgene rank prediction
CADBench: A Multimodal Benchmark for AI-Assisted CAD Program Generation
CADBench introduces a unified multimodal benchmark for evaluating AI-assisted CAD program generation, addressing fragmentation in existing evaluations. The benchmark comprises 18,000 samples across six families (DeepCAD, Fusion 360, ABC, MCB, Objaverse), five input modalities (clean/noisy meshes, single/multi-view renders, photorealistic renders), and six metrics (geometric fidelity, executability, program compactness). Stratified by B-rep face count and diversity-sampled, CADBench evaluates eleven CAD-specialized and general-purpose vision-language systems, generating 1.4M CAD programs. Results show specialized mesh-to-CAD models outperform VLMs under idealized inputs, with recurring failure modes: degradation with geometric complexity, brittleness under modality shift, and metric-dependent rankings. CADBench serves as a diagnostic testbed for editable 3D reconstruction and multimodal CAD understanding.
cad program generationmultimodal benchmarkb-rep face countgeometric fidelitymodality shift
Attractor-Vascular Coupling Theory: Formal Grounding and Empirical Validation for AAMI-Standard Cuffless Blood Pressure Estimation from Smartphone Photoplethysmography
The Attractor-Vascular Coupling Theory (AVCT) is introduced as a mathematical framework demonstrating that cardiac attractor geometry encodes blood pressure (BP) information sufficient for AAMI-standard cuffless estimation. AVCT, grounded in Cardiac Stability Theory, employs Takens delay embedding and attractor morphology extraction, formalized through theorems and propositions. A LightGBM model trained on pulse transit time and Cardiac Stability Index features achieved systolic BP MAE of 2.05 mmHg and diastolic BP MAE of 1.67 mmHg on 46 subjects, satisfying AAMI/IEEE SP10 requirements. PPG-only features matched ECG+PPG performance within 0.05 mmHg, validating AVCT predictions and enabling clinical-grade BP tracking via smartphone photoplethysmography.
attractor-vascular coupling theoryphotoplethysmographycardiac stability theorytakens delay embeddinglightgbm
Remember the Decision, Not the Description: A Rate-Distortion Framework for Agent Memory
The paper introduces DeMem, a decision-centric memory mechanism for long-horizon language agents that optimizes memory retention based on its impact on decision quality rather than descriptive fidelity. Framing memory as a rate-distortion problem, the authors derive an exact forgetting boundary and a memory-distortion frontier, enabling optimal tradeoffs between memory budget and decision quality. DeMem refines its memory partition online only when data indicate decision conflicts, achieving near-minimax regret guarantees. Evaluations on synthetic diagnostics and conversational benchmarks demonstrate consistent performance gains under fixed runtime budgets, validating the principle of prioritizing decision-relevant distinctions over descriptive accuracy.
rate-distortiondecision-centricmemory-distortion frontierminimax regretforgetting boundary
BEACON: A Multimodal Dataset for Learning Behavioral Fingerprints from Gameplay Data
The paper introduces BEACON, a multimodal dataset for behavioral fingerprinting, capturing 430 GB of synchronized gameplay data from 79 sessions across 28 Valorant players (102.51 hours). It includes mouse dynamics, keystrokes, network packets, screen recordings, and hardware metadata, designed for continuous authentication under cognitive and motor stress. The dataset supports research on behavioral biometrics, user drift, and multimodal representation learning in esports. BEACON is released on Hugging Face and GitHub for reproducibility.
behavioral biometricscontinuous authenticationmultimodal datasetvalorant gameplaymouse dynamics
BenchCAD: A Comprehensive, Industry-Standard Benchmark for Programmatic CAD
The authors introduce BenchCAD, a unified benchmark for evaluating industrial CAD reasoning capabilities in multimodal models. The benchmark comprises 17,900 execution-verified CadQuery programs across 106 industrial part families, assessing visual QA, code QA, image-to-code generation, and instruction-guided code editing. Evaluation of 10+ frontier models reveals current systems excel at coarse geometry recovery but fail in parametric abstraction (missing 3D structure, misinterpreting design parameters) and program synthesis (substituting complex operations with simpler patterns). While fine-tuning improves in-distribution performance, generalization to unseen part families remains limited, establishing BenchCAD as a key metric for industrial CAD automation readiness.
cad reasoningparametric abstractionprogram synthesismultimodal evaluationindustrial cad
The Generalized Turing Test: A Foundation for Comparing Intelligence
The Generalized Turing Test (GTT) introduces a formal framework for comparing agent intelligence via indistinguishability, defining a Turing comparator A ≥ B if B cannot reliably distinguish interactions with A (imitating B) from another instance of B. The framework is dataset- and task-agnostic, with theoretical analysis of transitivity, variants, and equivalence classes. Empirical evaluation on modern models demonstrates pairwise indistinguishability across thousands of trials, revealing a stratified structure consistent with existing rankings. The results suggest indistinguishability as a unifying lens for intelligence evaluation and potential training objectives independent of fixed datasets or benchmarks.
generalized turing testturing comparatorindistinguishabilityequivalence classestraining objectives
Rethinking Agentic Search with Pi-Serini: Is Lexical Retrieval Sufficient?
Pi-Serini, a search agent integrating lexical retrieval with advanced LLMs, demonstrates that BM25 suffices for deep research tasks when paired with capable models like GPT-5.5. The system employs three tools—retrieving, browsing, and reading documents—and achieves 83.1% answer accuracy and 94.7% surfaced evidence recall on BrowseComp-Plus, outperforming dense retrieval-based agents. Ablation studies reveal that BM25 tuning improves answer accuracy by 18.0% and evidence recall by 11.1%, while increased retrieval depth boosts evidence recall by 25.3%. Source code is publicly available.
bm25lexical retrievalsearch agentllmsevidence recall
Training-Free Cultural Alignment of Large Language Models via Persona Disagreement
DISCA (Disagreement-Informed Steering for Cultural Alignment) introduces a training-free, inference-time method for culturally aligning large language models (LLMs) without fine-tuning or white-box access. The approach leverages within-country sociodemographic disagreement by instantiating each country as a panel of World-Values-Survey-grounded persona agents and converting their disagreement into a bounded, loss-averse logit correction. Evaluated across 20 countries and 7 open-weight LLM backbones (2B--70B parameters), DISCA reduces cultural misalignment on MultiTP by 10--24% for models >=3.8B and 2--7% on open-ended scenarios. This demonstrates that inference-time calibration can effectively address global moral preferences without weight updates.
cultural alignmentinference-time calibrationlogit correctionpersona agentsloss-averse
Clin-JEPA: A Multi-Phase Co-Training Framework for Joint-Embedding Predictive Pretraining on EHR Patient Trajectories
Clin-JEPA introduces a multi-phase co-training framework for joint-embedding predictive pretraining on EHR patient trajectories, addressing instability in naïve co-training through a five-phase curriculum. The method co-trains a Qwen3-8B encoder and 92M-parameter latent trajectory predictor via predictor warmup, joint refinement, EMA target alignment, hard sync, and predictor finalization. Evaluations on MIMIC-IV ICU data show 15.7% reduced rollout drift, 4.83× better latent-space cohort discrimination, and superior multi-task performance (mean AUROC 0.883, +0.041 vs baselines) without task-specific fine-tuning.
joint-embedding predictive architectureehr patient trajectorieslatent-space planningmulti-phase co-trainingautoregressive rollout
From Controlled to the Wild: Evaluation of Pentesting Agents for the Real-World
The paper introduces ETHIBench, a practical evaluation protocol for AI pentesting agents that shifts focus from task completion to validated vulnerability discovery in complex, real-world scenarios. The method combines structured ground-truth annotation, LLM-based semantic matching, bipartite resolution scoring, continuous ground-truth maintenance, and efficiency metrics to enable realistic assessment across multiple attack surfaces. Results include an open-source release of expert-annotated ground truth and protocol code, facilitating reproducible comparisons of stochastic agents in operationally relevant settings.
pentesting agentsvulnerability discoverysemantic matchingbipartite resolutionground-truth maintenance
MMVIAD: Multi-view Multi-task Video Understanding for Industrial Anomaly Detection
We introduce MMVIAD, the first continuous multi-view video dataset for industrial anomaly detection, comprising 2-second object-centric clips across 48 categories, 14 environments, and 6 anomaly types. A two-stage post-training pipeline, combining PS-SFT and VISTA-GRPO, is proposed to enhance transferable anomaly understanding, yielding the final model VISTA. VISTA improves the base model's average score from 45.0 to 57.5 on MMVIAD-Unseen, outperforming GPT-5.4 across anomaly detection, defect classification, object classification, and anomaly visible-time localization tasks.
multi-view videoanomaly detectionps-sftvista-grpotemporal grounding
SLIM: Sparse Latent Steering for Interpretable and Property-Directed LLM-Based Molecular Editing
The paper introduces SLIM (Sparse Latent Interpretable Molecular editing), a framework for property-directed molecular editing with LLMs. It decomposes hidden states into sparse, property-aligned features using a Sparse Autoencoder with learnable importance gates, enabling precise steering without parameter updates. This approach improves editing success rates and supports interpretable analysis. Evaluated on MolEditRL across four architectures and eight properties, SLIM achieves up to 42.4-point gains over baselines.
sparse autoencodermolecular editinglatent steeringproperty alignmentinterpretability
The First Drop of Ink: Nonlinear Impact of Misleading Information in Long-Context Reasoning
This work introduces 'The First Drop of Ink' effect, demonstrating that hard distractors in long-context reasoning exhibit a nonlinear impact on model performance. Through systematic variation of hard-distractor proportions in fixed-length contexts, the study reveals that performance sharply declines with the initial introduction of distractors, with marginal additional degradation thereafter. Theoretical and empirical analyses grounded in attention mechanics show that hard distractors disproportionately capture attention even at low proportions. Controlled experiments indicate that filtering gains primarily stem from context-length reduction rather than distractor removal, emphasizing the critical role of upstream retrieval precision.
long-context reasoninghard distractorsattention mechanicsretrieval precisionnonlinear impact
MaD Physics: Evaluating information seeking under constraints in physical environments
Measuring and Discovering Physics (MaD Physics) introduces a benchmark to evaluate agents' ability to conduct informative measurements and draw conclusions under physical and cost constraints. The benchmark comprises three environments based on distinct physical laws, including altered laws to mitigate knowledge contamination. Agents make measurements within a budget, infer underlying physical laws, and predict future system states. MaD Physics assesses capabilities in model inference, constrained planning, multimodality, and in-context learning. Evaluations using four Gemini models (2.5 Flash Lite, 2.5 Flash, 2.5 Pro, and 3 Flash) reveal limitations in structured exploration and data collection, suggesting areas for improving scientific reasoning.
benchmarkphysical lawsconstrained planningin-context learningmultimodality
ALAM: Algebraically Consistent Latent Transitions for Vision-Language-Action Models
ALAM (Algebraic Latent Action Model) introduces algebraically consistent latent transitions for vision-language-action (VLA) models, addressing the limitations of reconstruction-trained latent codes in policy generation. The method leverages action-free video triplets to enforce composition and reversal consistency, creating a locally additive transition space. Downstream VLA learning freezes the pretrained encoder and co-generates latent transitions with robot actions under a joint flow-matching objective. ALAM reduces additivity and reversibility errors by 25-85 times compared to baselines and improves long-horizon reconstruction. On MetaWorld MT50 and LIBERO benchmarks, ALAM increases success rates from 47.9% to 85.0% and 94.1% to 98.1%, respectively, demonstrating consistent gains in real-world manipulation tasks.
latent transitionsflow matchingvision-language-actionalgebraic consistencypolicy generation
CLEF: EEG Foundation Model for Learning Clinical Semantics
CLEF introduces a clinically grounded long-context EEG foundation model that processes EEG sessions as 3D multitaper spectrogram tokens, enabling Transformer-based session-scale modeling. It aligns embeddings with neurologist reports and EHR data via contrastive learning. Evaluated on a 234-task benchmark (260k EEG sessions from 108k patients), CLEF outperforms prior EEG foundation models on 229 tasks, improving mean AUROC from 0.65 to 0.74. Reconstruction pretraining surpasses prior models, with additional gains from report/EHR alignment. Results demonstrate transferability to unseen concepts and external cohorts.
eegfoundation modelcontrastive learningmultitaper spectrogramauroc
Policy Gradient Methods for Non-Markovian Reinforcement Learning
The paper introduces Agent State-Markov (ASM) policies for non-Markovian reinforcement learning, jointly optimizing agent state dynamics and control policies via reward-centric formulation. It derives a policy gradient theorem for ASM policies, extending classical results to episodic and infinite-horizon NMDPs, and proposes the ASMPG algorithm leveraging recursive state updates. Theoretical guarantees include finite-time and almost sure convergence, with empirical results showing superior performance over predictive-objective baselines on non-Markovian tasks.
non-markovianpolicy gradientagent state-markovasmpgstate dynamics
Probing Cross-modal Information Hubs in Audio-Visual LLMs
The study investigates cross-modal information flow in audio-visual large language models (AVLLMs), identifying specialized 'cross-modal sink tokens' that store integrated audio-visual information. Through analysis of multiple AVLLMs, the authors find that cross-modal information is non-uniformly concentrated in these tokens rather than distributed uniformly. Leveraging this insight, they propose a training-free method to mitigate hallucinations by enhancing reliance on cross-modal sink tokens. The approach demonstrates potential for improving AVLLM performance without additional training.
audio-visual llmscross-modal integrationsink tokenshallucination mitigationinformation flow
NanoResearch: Co-Evolving Skills, Memory, and Policy for Personalized Research Automation
NanoResearch introduces a multi-agent framework for personalized research automation, addressing limitations in current LLM-powered systems through tri-level co-evolution. The framework integrates a skill bank for reusable procedural rules, a memory module for user- and project-specific experience, and label-free policy learning to internalize implicit preferences from free-form feedback. These components co-evolve, enabling richer memory, better planning, and continuous alignment with user preferences. Experiments show NanoResearch outperforms state-of-the-art AI research systems, progressively refining outputs and reducing costs over successive cycles.
multi-agent systemsprocedural ruleslabel-free learningpreference internalizationtri-level co-evolution
Switching-Geometry Analysis of Deflated Q-Value Iteration
The paper introduces a joint spectral radius (JSR) framework to analyze rank-one deflated Q-value iteration (Q-VI) in discounted Markov decision processes. By examining the geometry of switching systems with an all-ones residual correction, the authors provide the first JSR-based convergence analysis for deflated Q-VI in policy optimization. Results show that the standard Q-VI switching system model has JSR equal to the discount factor γ, while the deflated version may achieve a tighter convergence-rate bound by projecting onto a quotient space. The correction is shown to be equivalent to scalar recentering of standard Q-VI, preserving the greedy-policy sequence but offering improved convergence characterization.
joint spectral radiusq-value iterationmarkov decision processswitching systemspolicy optimization
Threat Modelling using Domain-Adapted Language Models: Empirical Evaluation and Insights
This study systematically evaluates domain-adapted language models for structured threat modelling in 5G security using the STRIDE approach, comparing 8 models across 52 configurations. The analysis examines domain adaptation, model scale (LLMs vs. SLMs), decoding strategies (greedy vs. stochastic sampling), and prompting techniques. Results indicate domain-adapted models do not consistently outperform general-purpose counterparts, decoding strategies significantly impact output validity, and larger models show inconsistent performance gains. The findings reveal fundamental limitations of current LLMs for threat modelling, suggesting the need for task-specific reasoning and security concept grounding beyond data or scale improvements.
stride threat modellingdomain-adapted language models5g securitydecoding strategiesinconsistent performance gains
PhyGround: Benchmarking Physical Reasoning in Generative World Models
PhyGround introduces a benchmark for evaluating physical reasoning in generative world models, addressing limitations in existing physics-focused video benchmarks. The benchmark comprises 250 curated prompts with expected physical outcomes and a taxonomy of 13 physical laws across solid-body mechanics, fluid dynamics, and optics. Each law is operationalized through observable sub-questions for per-law diagnostics. Eight video generation models were evaluated via a large-scale human study involving 459 annotators, yielding 5,796 annotations and 37.4K fine-grained labels. PhyJudge-9B, a physics-specialized VLM judge, demonstrated lower aggregate relative bias (3.3%) compared to Gemini-3.1-Pro (16.6%). The benchmark, annotations, and evaluation code are publicly released.
generative world modelsphysical reasoningvideo generationtaxonomyvlm judge
Reasoning Is Not Free: Robust Adaptive Cost-Efficient Routing for LLM-as-a-Judge
The paper introduces Robust Adaptive Cost-Efficient Routing (RACER), a method for dynamically selecting between reasoning and non-reasoning LLM judges to optimize accuracy-cost trade-offs. RACER formulates routing as a constrained distributionally robust optimization problem, addressing distribution shift via KL-divergence uncertainty sets, and guarantees uniqueness and linear convergence. Experiments demonstrate RACER's superior performance in balancing accuracy and computational cost, particularly in tasks requiring structured verification like math and coding, while avoiding unnecessary reasoning overhead for simpler evaluations.
llm-as-a-judgedistributionally robust optimizationkl-divergenceadaptive routingcomputational cost
New AI-Driven Tools for Enhancing Campus Well-being: A Prevention and Intervention Approach
This dissertation introduces an integrated AI framework for campus well-being, combining prevention (TigerGPT chatbot with 75% usability) and intervention (PsychoGPT for DSM-5-aligned mental health assessment). AURA, a reinforcement-learning system, improves conversational quality (+0.12 mean gain) via LSDE metrics, while Stacked Multi-Model Reasoning (SMMR) reduces hallucinations in diagnostic workflows. BERT(128) analyzes Expressive Narrative Stories without keyword reliance. Results show 81% satisfaction for TigerGPT, 63% fewer specification prompts with AURA, and superior DAIC-WOZ performance for SMMR.
tigergptauralsdesmrrpsychogpt
The Last Word Often Wins: A Format Confound in Chain-of-Thought Corruption Studies
The study identifies a confound in chain-of-thought (CoT) corruption studies: terminal answer statements dominate computational importance detection, masking true reasoning steps. Through format ablation on GSM8K chains (N=300), removing answer suffixes caused a 19x sensitivity drop (p=0.022). Conflicting-answer experiments at 7B showed near-zero accuracy (<=0.02) across architectures, with followed-wrong rates of 0.63-1.00 at 3B-7B, attenuating at larger scales (Phi-4-14B: 0.300; 32B: ~0.01). Generation probes revealed answer-text dependence without early commitment (<5%). The effect persisted through 14B (8.5x ratio, p=0.001) and vanished at 32B. A three-prerequisite protocol is proposed for corruption-based faithfulness studies.
chain-of-thoughtcorruption studiesformat confoundanswer suffixfaithfulness evaluation
Interpretable Machine Learning for Football Performance Analysis: Evidence of Limited Transferability from Elite Leagues to University Competition
The study demonstrates limited transferability of interpretable machine learning insights between elite and university football domains. Using Random Forest and Multilayer Perceptron models trained on top-five European league data and evaluated on National Tsing Hua University matches, performance determinants were analyzed via SHAP and Counterfactual Impact Score. Results show elite football maintains stable performance hierarchies across leagues and explanation methods, while university football exhibits substantial indicator reordering, reduced explanation stability (p<0.05), and weaker structural agreement with elite domains, suggesting interpretability robustness is domain-dependent.
interpretable machine learningdomain shiftshapley additive explanationscounterfactual impact scoreperformance determinants
Can You Keep a Secret? Involuntary Information Leakage in Language Model Writing
This study demonstrates that frontier language models involuntarily leak prompted secrets through thematic elements in generated text, despite explicit instructions to conceal them. The authors tested five models by providing a secret word with concealment instructions, generating stories, and having a second model attempt to detect the secret through binary classification. Results show thematic leakage occurs at rates up to 79%, with avoidance behaviors also detectable. Leakage scales with model size, disappears in short-form writing, and can be partially redirected using decoy concepts. The findings suggest that attending to secrets opens an information channel that current LLMs cannot effectively close.
thematic leakagebinary classificationavoidance behaviordecoy conceptinformation channel
PathISE: Learning Informative Path Supervision for Knowledge Graph Question Answering
PathISE introduces a framework for generating high-quality intermediate supervision in Knowledge Graph Question Answering (KGQA) without costly LLM-refined signals. The method employs a lightweight transformer-based estimator to assess the informativeness of relation paths, constructing pseudo path-level supervision distilled into an LLM path generator. This generator produces KG-grounded paths for inductive answer reasoning. Evaluations on three KGQA benchmarks demonstrate PathISE's competitive or state-of-the-art performance and its ability to enhance existing KGQA models with reusable supervision signals.
knowledge graph question answeringintermediate supervisiontransformer-based estimatorpseudo path-level supervisioninductive answer reasoning
ComplexMCP: Evaluation of LLM Agents in Dynamic, Interdependent, and Large-Scale Tool Sandbox
The authors introduce ComplexMCP, a benchmark evaluating LLM agents in dynamic, interdependent tool environments, addressing the gap in commercial software automation. ComplexMCP features over 300 tools across 7 stateful sandboxes, employing a seed-driven architecture to simulate dynamic states and API failures. Evaluations of LLMs in full-context and RAG paradigms reveal a maximum 60% success rate, significantly below human performance at 90%. Analysis identifies tool retrieval saturation, over-confidence, and strategic defeatism as key bottlenecks, highlighting the need for improved agent resilience in interdependent workflows.
llm agentsstateful sandboxestool retrieval saturationstrategic defeatismseed-driven architecture
TrajPrism: A Multi-Task Benchmark for Language-Grounded Urban Trajectory Understanding
We introduce TrajPrism, a multi-task benchmark for language-trajectory alignment that unifies instruction-conditioned trajectory generation, language-driven semantic trajectory retrieval, and trajectory captioning. TrajPrism pairs 300K real urban trajectories across Porto, San Francisco, and Beijing with judge-filtered language annotations, yielding 2.1M task instances. We propose three proof-of-concept models—TrajAnchor, TrajFuse, and TrajRap—to instantiate the tasks and demonstrate that geometry-only baselines significantly underperform on language-integrated tasks. The benchmark includes an evaluation protocol measuring trajectory fidelity, retrieval quality, and language groundedness, alongside a reproducible annotation pipeline for portability across cities.
language-trajectory alignmentsemantic trajectory retrievalinstruction-conditioned generationtrajectory captioningurban mobility
Beyond the Last Layer: Multi-Layer Representation Fusion for Visual Tokenizatio
DRoRAE (Depth-Routed Representation AutoEncoder) introduces multi-layer feature fusion to enhance visual tokenization by recovering low-level details lost in last-layer representations. The method employs energy-constrained routing and incremental correction to adaptively aggregate features across all encoder layers, followed by a three-phase training strategy that first learns fusion under frozen decoder constraints then fine-tunes the decoder. On ImageNet-256, DRoRAE reduces rFID from 0.57 to 0.29 and improves generation FID from 1.74 to 1.65, with gains extending to text-to-image synthesis. A log-linear scaling law (R²=0.86) reveals representation richness as a predictably scalable dimension for visual tokenizers.
visual tokenizationmulti-layer fusionrepresentation autoencoderenergy-constrained routingscaling law
Towards a Large Language-Vision Question Answering Model for MSTAR Automatic Target Recognition
The study introduces a large language-vision model (LLVM) application for synthetic aperture radar (SAR) imagery, specifically targeting automatic target recognition (ATR) in the MSTAR Public Dataset. Using transformer-based architectures like CLIP and LLaVA, the authors develop a benchmark with descriptive captions and question-answer pairs for visual question-answering (VQA) tasks. Parameter-efficient fine-tuning achieves 98% accuracy in identifying fine-grained military vehicle targets, addressing challenges in SAR-based ATR under complex conditions. The work advances machine-assisted remote sensing for military and intelligence applications.
large language-vision modelsynthetic aperture radarautomatic target recognitionparameter-efficient fine-tuningvisual question-answering
MPerS: Dynamic MLLM MixExperts Perception-Guided Remote Sensing Scene Segmentation
MPerS introduces a Dynamic Mixture-of-Experts (MixExperts) framework for multimodal remote sensing scene segmentation, addressing the gap in high-quality caption generation and semantic fusion. The method leverages MLLMs (LLaVA, ChatGPT, Qwen) to generate diverse RS captions and employs DINOv3 for dense visual feature extraction. A Dynamic MixExperts module adaptively integrates textual semantics, while Linguistic Query Guided Attention refines visual features for precise segmentation. Evaluated on three public semantic segmentation datasets, MPerS demonstrates superior performance, highlighting the efficacy of its multimodal fusion approach.
mixture-of-expertsmultimodal fusionremote sensingsemantic segmentationlinguistic query
Dynamic Cross-Modal Prompt Generation for Multimodal Continual Instruction Tuning
The paper introduces DRAPE, a dynamic cross-modal prompt generation framework for Multimodal Continual Instruction Tuning (MCIT) in Multimodal Large Language Models (MLLMs). DRAPE synthesizes instance-specific soft prompts by deriving prompt queries from textual instructions and cross-attending to visual patch features, enabling query-image conditioned prompts prepended to a frozen LLM. To mitigate catastrophic forgetting, DRAPE employs null-space gradient projection on the shared projector and uses CLIP-based prototype routing for task-label-free generator selection. Experiments on MCIT benchmarks demonstrate DRAPE's state-of-the-art performance against prompt-based and LoRA-based continual-learning baselines.
multimodal continual instruction tuningdynamic prompt generationnull-space gradient projectionclip-based prototype routingsoft prompts
Break the Brake, Not the Wheel: Untargeted Jailbreak via Entropy Maximization
The paper introduces Untargeted Jailbreak via Entropy Maximization (UJEM)-KL, a lightweight attack method for vision-language models (VLMs) that maximizes entropy at decision tokens to flip refusal outcomes while preserving output quality. The method leverages the observation that refusal behavior concentrates at high-entropy tokens during autoregressive decoding. Evaluated across three VLMs and two safety benchmarks, UJEM-KL achieves competitive white-box attack success rates and improves transferability compared to previous approaches. Results suggest that limited transferability in prior methods stems from overly constrained optimization objectives. The attack remains effective under representative defenses.
entropy maximizationvision-language modelsautoregressive decodingtransferabilitysafety benchmarks
MATRA: Modeling the Attack Surface of Agentic AI Systems -- OpenClaw Case Study
The paper introduces MATRA, a threat modeling framework for agentic AI systems that systematically assesses deployment-specific risks from known LLM threats. It combines asset-based impact assessment with attack trees to evaluate risk likelihood within system architectures. Applied to OpenClaw, a personal AI agent deployment, MATRA demonstrates how architectural controls (e.g., network sandboxing, least-privilege access) mitigate risks by limiting injection attack blast radius.
threat modelingagentic aillm securityattack treesblast radius
GridProbe: Posterior-Probing for Adaptive Test-Time Compute in Long-Video VLMs
GridProbe introduces a training-free posterior-probing inference paradigm for efficient long-video understanding in Vision-Language Models (VLMs), addressing quadratic attention costs. The method arranges frames on a K×K grid, applies lightweight row and column probes to compute query-conditioned confidences, and uses their outer product to derive an interpretable importance map. Shape-Adaptive Selection dynamically adjusts the frame budget per question, enabling test-time adaptive compute. Empirical results on Video-MME-v2 and LongVideoBench show GridProbe achieves near-baseline accuracy (within 1.6 pp Avg Acc) with a 3.36× reduction in TFLOPs and Pareto-dominates the baseline (+0.9 pp at 0.35× compute). Decoupling selector and QA models further enhances performance, yielding up to +4.0 pp at 0.52× compute.
posterior-probingvision-language modelsadaptive computeshape-adaptive selectionimportance map
The Agent Use of Agent Beings: Agent Cybernetics Is the Missing Science of Foundation Agents
The paper introduces Agent Cybernetics, a theoretical framework for designing foundation agents by mapping classical cybernetic laws onto agent design principles. It synthesizes these principles into three engineering desiderata: reliability, lifelong running, and self-improvement. The framework is applied to three domains—code generation, computer use, and automated research—to identify failure modes and provide concrete engineering recommendations. The authors argue that this approach addresses fundamental questions in agent design, such as maintaining task focus, handling environmental complexity, and ensuring safe self-improvement, thereby establishing a scientific foundation for reliable real-world deployment of foundation agents.
foundation agentscyberneticsself-improvementtask focusenvironmental complexity
Provable Sparse Inversion and Token Relabel Enhanced One-shot Federated Learning with ViTs
We propose Federated Model Inversion and Token Relabel (FedMITR), a novel framework for one-shot federated learning that addresses semantic misalignment in non-IID settings. FedMITR employs sparse model inversion to selectively generate semantic foregrounds while halting inversion of uninformative backgrounds, and implements token relabeling via ensemble models for low-information-density patches. Theoretical analysis based on algorithmic stability shows that sparse inversion eliminates gradient instability from background noise while token relabel reduces gradient variance, yielding a tighter generalization bound. Empirical results demonstrate FedMITR's substantial performance improvements over baselines across various settings.
one-shot federated learningsparse model inversiontoken relabelalgorithmic stabilitygeneralization bound
Geospatial-Temporal Sensemaking of Remote Sensing Activity Detections with Multimodal Large Language Model
The authors introduce SMART-HC-VQA, a Sentinel-2-based visual question answering dataset for spatiotemporal analysis of human activity, derived from the IARPA SMART Heavy Construction dataset. The dataset includes 21,837 Sentinel-2 image chips, 65,511 single-image VQA examples, and ~2.3 million two-image temporal comparison examples generated via Image-Pairwise Combinatorial Augmentation. They detail workflows for processing Sentinel-2 imagery, segmenting tiles, and analyzing site metadata distributions. A multi-image multimodal large language model training framework based on LLaVA-NeXT Mistral-7B is implemented for metadata-derived VQA examples. This work provides a foundation for language-guided remote sensing activity understanding, focusing on change detection and reasoning about ongoing processes.
sentinel-2visual question answeringmultimodal large language modelimage-pairwise combinatorial augmentationspatiotemporal analysis
iPay: Integrated Payment Action Recognition via Multimodal Networks and Adaptive Spatial Prior Learning
The authors introduce iPay, a multimodal framework for integrated payment action recognition in transit surveillance systems, addressing limitations of prior vision- and skeleton-based methods. iPay employs a mixture-of-experts architecture with four streams: an RGB expert stream for local evidence, a skeleton expert stream with graph convolutional backbone, a dual-attention fusion stream for spatiotemporal transfer, and a Spatial Difference Discriminator (SDD) for hand-to-anchor motion modeling. Evaluated on 500+ payment clips from 55 hours of real surveillance footage, iPay achieves 83.45% recognition accuracy with competitive computational efficiency, enabling edge deployment.
multimodal networksgraph convolutional backbonespatial difference discriminatoraction recognitionedge deployment
AllocMV: Optimal Resource Allocation for Music Video Generation via Structured Persistent State
AllocMV introduces a hierarchical framework for music video generation, addressing computational cost and cross-shot consistency via structured persistent state representation and optimal resource allocation. The method formulates video synthesis as a Multiple-Choice Knapsack Problem (MCKP), leveraging a global planner to produce compact state objects comprising character entities, scene priors, and sharing graphs. A dynamic programming-based MCKP solver allocates resources across High-Gen, Mid-Gen, and Reuse branches, while a divergence-based forking strategy reuses visual prefixes for repetitive motifs. Evaluated via Cost-Quality Ratio (CQR), AllocMV achieves optimal trade-offs between perceived quality and resource expenditure under budgetary and rhythmic constraints.
multiple-choice knapsack problemstructured persistent statecost-quality ratiodynamic programmingdivergence-based forking
An Uncertainty-Aware Resilience Micro-Agent for Causal Observability in the Computing Continuum
This paper introduces AURORA, an uncertainty-aware resilience micro-agent for causal observability in edge-tier environments, addressing grey failures with ambiguous overlapping symptoms. The framework employs parallel micro-agents integrating the free-energy principle, causal do-calculus, and localized causal state-graphs to enable counterfactual root-cause analysis within each fault's Markov blanket. A dual-gated execution mechanism restricts remediation to high causal confidence and bounded epistemic uncertainty, escalating otherwise. Experiments show AURORA achieves 0% destructive action rate, 62.0% repair accuracy, and 3ms mean time to repair, outperforming baselines.
grey failurescausal observabilityfree-energy principledo-calculusmarkov blanket
Why Low-Resource NLP Needs More Than Cross-Lingual Transfer: Lessons Learned from Luxembourgish
The paper argues that cross-lingual transfer and language-specific efforts are complementary rather than competing approaches in low-resource NLP, using Luxembourgish as a case study. It synthesizes prior research and data collection results to demonstrate that while cross-lingual transfer improves target-language performance, its success depends on high-quality, task-aligned target-language data. Conversely, limited language-specific resources achieve full potential only when integrated into a cross-lingual framework. The authors provide practical guidelines for balancing these approaches in sustainable low-resource NLP pipelines.
cross-lingual transferlow-resource nlplanguage-specific effortstask-aligned datamultilingual language models
The Bystander Effect in Multi-Agent Reasoning: Quantifying Cognitive Loafing in Collaborative Interactions
The study demonstrates that multi-agent systems (MAS) induce cognitive loafing in Large Language Models (LLMs) due to a simulated Bystander Effect, challenging the assumption that collaboration inherently improves reasoning. By analyzing 22,500 deterministic trajectories across GAIA, SWE-bench, and Multi-Challenge datasets using three SOTA models, the authors semantically audit internal reasoning traces against external outputs. They formalize the Interaction Depth Limit ($D_L$) and uncover the Sovereignty Gap, where models internally compute correct derivations but exhibit Alignment Hallucinations to appease simulated swarms. Results reveal that multi-agent social load is non-commutative, with Lead Anchor identity disproportionately influencing swarm integrity, exposing vulnerabilities in unstructured multi-agent topologies.
multi-agent systemscognitive loafingbystander effectsovereignty gapalignment hallucinations
GESR: A Genetic Programming-Based Symbolic Regression Method with Gene Editing
The paper introduces GESR, a genetic programming-based symbolic regression method incorporating targeted gene editing via BERT models. The approach employs two BERT models: one guides mutation through masked language modeling on expression symbols, while another predicts optimal crossover points. Compared to traditional GP algorithms, GESR demonstrates significantly improved computational efficiency and robust performance across multiple symbolic regression benchmarks.
symbolic regressiongenetic programminggene editingbert modelsmasked language modeling
Is Data Shapley Not Better than Random in Data Selection? Ask NASH
We propose NASH (Non-linear Aggregation of SHapley-informative components), a novel framework for data selection that addresses limitations of Data Shapley by decomposing utility functions into Shapley-informative components and aggregating them non-linearly. NASH identifies settings where Data Shapley performs effectively and leverages these to select high-quality subsets efficiently. Experimental results demonstrate that NASH significantly enhances the effectiveness of Shapley/semivalue-based data selection while maintaining minimal runtime overhead.
data selectionshapley-informativenon-linear aggregationutility functionsemivalue
Step Rejection Fine-Tuning: A Practical Distillation Recipe
The paper introduces Step Rejection Fine-Tuning (SRFT), a distillation method that improves upon standard Rejection Fine-Tuning (RFT) by leveraging partially correct trajectories in LLM agent training. SRFT employs a critic LLM to evaluate individual trajectory steps, masking loss for erroneous steps while retaining them in context, enabling error recovery learning. On SWE-bench Verified, SRFT achieves a 3.7% resolution rate improvement (total 32.2%) compared to RFT's 2.4%, demonstrating superior utilization of unresolved trajectories.
rejection fine-tuningllm agentsswe-benchcritic llmtrajectory distillation
Prompt-Activation Duality: Improving Activation Steering via Attention-Level Interventions
The paper introduces Gated Cropped Attention-Delta steering (GCAD), a method to improve activation steering in language models by addressing KV-cache contamination, a failure mode where steered token states degrade coherence in stateful dialogue. GCAD extracts steering signals from system-prompt contributions to self-attention and applies them with token-level gating. Experiments on persona-steering tasks demonstrate that GCAD preserves trait control while significantly enhancing long-horizon coherence, improving average coherence drift from -18.6 to -1.9 and raising turn-10 trait expression from 78.0 to 93.1. The results indicate that activation steering benefits from interventions aligned with prompt-mediated pathways.
activation steeringkv-cache contaminationself-attentiontoken-level gatingcoherence drift
Evolving-RL: End-to-End Optimization of Experience-Driven Self-Evolving Capability within Agents
Evolving-RL introduces an end-to-end algorithmic framework for optimizing experience-driven self-evolving capabilities in agents, addressing the limitations of static large language models. The method jointly optimizes experience extraction and utilization through coordinated co-evolution, leveraging supervisory signals from evaluation to separately refine the extractor and solver. Experiments on ALFWorld and Mind2Web demonstrate significant performance gains on out-of-distribution tasks, achieving up to 98.7% and 35.8% relative improvements over GRPO baselines, respectively. Evolving-RL also functions as an experience-augmented RL algorithm, internalizing reusable experience patterns into model parameters for enhanced performance on both seen and unseen tasks.
self-evolving agentsexperience extractionin-context learningreinforcement learningout-of-distribution tasks
bViT: Investigating Single-Block Recurrence in Vision Transformers for Image Recognition
The authors propose bViT, a single-block recurrent Vision Transformer (ViT) architecture that applies one transformer block repeatedly to process images, eliminating layer-specific parameterization while preserving iterative computation. On ImageNet-1K, a 12-step bViT-B achieves comparable accuracy to standard ViT-B with an order of magnitude fewer parameters, demonstrating that recurrent reuse can implement a large fraction of ViT depth given sufficient representation width. Mechanistic analyses reveal step-dependent behavior in the shared block rather than repeated computation. bViT transfers competitively to downstream tasks and enables parameter-efficient fine-tuning, suggesting implicit depth multiplexing through evolving hidden states.
vision transformersrecurrent computationparameter-efficientimplicit depth multiplexingrepresentation width
When Can Digital Personas Reliably Approximate Human Survey Findings?
The study evaluates when LLM-powered digital personas can reliably approximate human survey responses by testing four persona architectures across three LLMs using the LISS panel. Researchers constructed personas from pre-2023 survey histories and background variables, then compared them to held-out post-cutoff human responses at multiple analysis levels. Results show personas improve alignment with human distributions for stable attributes but struggle with individual prediction and multivariate structure, with retrieval-augmented architectures performing best. Performance depends more on response variability and common patterns than model choice, offering practical guidance for survey research applications.
large language modelsdigital personassurvey researchretrieval-augmented generationresponse variability
Active Learning for Gaussian Process Regression Under Self-Induced Boltzmann Weights
The authors propose \texttt{AB-SID-iVAR}, a Gaussian Process-based active learning acquisition function for regression under self-induced Boltzmann weights, addressing challenges posed by unknown target distributions and intractable partition functions. The method approximates the Bayesian target distribution in closed form without partition function estimation, applicable to discrete and continuous domains, with a Thompson sampling variant (\texttt{TS-SID-iVAR}) analyzed as a Monte Carlo alternative. Theoretical guarantees include vanishing terminal prediction error with high probability and tighter average-case bounds. Empirical validation shows improvements over existing methods on synthetic benchmarks, potential energy surface modeling, and drug discovery tasks.
active learninggaussian process regressionboltzmann distributionpartition functionacquisition function
A Recursive Decomposition Framework for Causal Structure Learning in the Presence of Latent Variables
The paper proposes DiCoLa, a recursive decomposition framework for causal structure learning that handles latent variables, overcoming the causal sufficiency limitation of prior divide-and-conquer methods. It recursively breaks the global task into subproblems and integrates solutions via principled reconstruction, with proven soundness and completeness. Experiments on synthetic data show significant computational efficiency gains across causal discovery algorithms, with real-world data validating practical effectiveness.
causal discoverylatent variablesrecursive decompositionconditional independencedivide-and-conquer
diffGHOST: Diffusion based Generative Hedged Oblivious Synthetic Trajectories
The paper introduces diffGHOST, a conditional diffusion model for privacy-preserving trajectory synthesis that addresses limitations in existing generative approaches. The method employs latent space segmentation to identify and mitigate memorization of sensitive samples while maintaining trajectory utility. By leveraging diffusion-based generation with learned conditional segments, the approach aims to provide stronger privacy guarantees compared to state-of-the-art models that assume implicit privacy.
conditional diffusion modeltrajectory synthesislatent space segmentationprivacy preservationgenerative modeling
LLaVA-CKD: Bottom-Up Cascaded Knowledge Distillation for Vision-Language Models
LLaVA-CKD introduces a bottom-up cascaded knowledge distillation framework to mitigate the capacity gap issue in vision-language models, where intermediate-capacity teachers progressively transfer knowledge to smaller student networks. Inspired by formal education systems, the method employs additional teachers of intermediate capacity to facilitate gradual knowledge transfer, enabling higher-capacity teachers to take over effectively. Theoretical analysis examines the impact of cascaded distillation on student generalization. Evaluated on seven standard VQA benchmarks, models derived from LLaVA-CKD achieve state-of-the-art performance.
knowledge distillationvision-language modelscapacity gapgeneralization performancevqa benchmarks
Towards Understanding Continual Factual Knowledge Acquisition of Language Models: From Theory to Algorithm
We propose a theoretical framework analyzing continual Factual Knowledge Acquisition (cFKA) in Language Models, revealing that regularization methods only adjust parameter convergence rates while data replay alters convergence dynamics. Building on this, we introduce STOC, a generative data replay method that selects influential factual snippets via attention contribution for replay generation. Experiments on synthetic and real-world datasets demonstrate STOC's effectiveness in mitigating catastrophic forgetting during continual pre-training.
continual factual knowledge acquisitiondata replayattention contributioncatastrophic forgettingcontinual pre-training
Navigating the Sea of LLM Evaluation: Investigating Bias in Toxicity Benchmarks
This study identifies critical biases in toxicity benchmarks used for LLM evaluation, revealing vulnerabilities in current safety certification practices. Through systematic experimentation, the authors demonstrate how benchmark outcomes vary significantly with task type (e.g., completion vs. summarization) and input domain, while also exposing model-specific instabilities. Results show a 23-47% increase in false positive toxicity flags during summarization tasks compared to completion, highlighting the need for more robust evaluation frameworks that account for these contextual factors.
toxicity benchmarksevaluation biasllm safetyfalse positivescontextual robustness
Teacher-Aware Evolution of Heuristic Programs from Learned Optimization Policies
The paper introduces a teacher-aware evolutionary framework for generating heuristic programs in combinatorial optimization. The method leverages learned optimization policies as behavioral teachers, querying them on candidate program states to provide local feedback during evolution. This approach combines task performance with teacher-derived behavioral signals, avoiding neural inference at deployment. Evaluations on scheduling, routing, and graph optimization benchmarks demonstrate improvements over performance-driven LLM heuristic evolution baselines.
heuristic evolutioncombinatorial optimizationbehavioral feedbacklearned policiesteacher-aware
Intrinsic Guardrails: How Semantic Geometry of Personality Interacts with Emergent Misalignment in LLMs
This work demonstrates that the semantic geometry of personality in Large Language Models (LLMs) remains stable across aligned models and corrupted fine-tunes, enabling intrinsic guardrails against emergent misalignment (EM). By mapping LLM personality spaces using psychometric profiles (Big Five, Dark Triad) and LLM-specific behaviors, the authors identify stable vectors like the 'Evil' persona vector and Semantic Valence Vector (SVV). Causal interventions reveal that ablating these vectors increases misalignment rates above 40%, while amplifying them reduces failures below 3%. Zero-shot transfer of vectors from instruct-tuned models successfully regulates EM in corrupted fine-tunes, suggesting conserved personality representations serve as robust guardrails.
emergent misalignmentsemantic geometrypsychometric profilessemantic valence vectorintrinsic guardrails
Interpretable Coreference Resolution Evaluation Using Explicit Semantics
We introduce a semantically-enhanced evaluation framework for coreference resolution that addresses the diagnostic limitations of aggregate metrics like CoNLL-F1. Our approach overlays Concept and Named Entity Recognition (CNER) on coreference outputs, assigning semantic labels to nominal mentions and propagating them to entire clusters, enabling typed score computation stratified by semantic class. Experiments on OntoNotes, LitBank, and PreCo reveal systematic weaknesses obscured by aggregate metrics. Furthermore, we demonstrate that these diagnostics enable targeted data augmentation strategies, yielding measurable out-of-domain improvements.
coreference resolutionconcept and named entity recognitionontonoteslitbankpreco
Hierarchical Causal Abduction: A Foundation Framework for Explainable Model Predictive Control
The paper introduces Hierarchical Causal Abduction (HCA), a framework for generating interpretable explanations in Model Predictive Control (MPC) systems. HCA integrates physics-informed knowledge graphs, Karush--Kuhn--Tucker (KKT) multipliers from optimization, and temporal causal discovery via PCMCI to explain nonlinear MPC actions. Evaluated across greenhouse climate, building HVAC, and chemical process control, HCA achieves 53% higher explanation accuracy (0.478 vs. 0.311) than LIME without domain-specific tuning, reaching 0.88 accuracy with KKT-threshold calibration. Ablations show 32--37% accuracy drops when removing any component, demonstrating HCA's generalizability to learning-based control and planning systems.
model predictive controlcausal abductionkkt multiplierspcmci algorithminterpretable explanations
PRISM: Generation-Time Detection and Mitigation of Secret Leakage in Multi-Agent LLM Pipelines
The paper introduces PRISM, a real-time method for detecting and mitigating secret leakage in multi-agent LLM pipelines by modeling credential propagation as sequential risk accumulation. The approach combines 16 lexical, structural, and information-theoretic features to compute per-token risk scores, leveraging observable generation dynamics like entropy collapse and logit concentration shifts. Evaluated on a 2,000-task adversarial benchmark with 13 attack categories, PRISM achieves F1=0.832, 0.0% leakage, and 0.893 utility preservation, outperforming Span Tagger (F1=0.719, 15.0% leakage).
multi-agent llmsecret leakagepropagation amplificationrisk accumulationentropy collapse
Re-Triggering Safeguards within LLMs for Jailbreak Detection
The paper introduces a novel jailbreaking prompt detection method for large language models (LLMs) that leverages embedding disruption to re-activate internal safeguards. Unlike standalone defense solutions, this approach cooperates with LLMs' built-in mechanisms by identifying and applying appropriate disruptions to detect and mitigate jailbreak attempts. The authors develop an efficient search algorithm to optimize disruption effects and conduct extensive experiments validating the method's effectiveness against state-of-the-art jailbreak attacks in both white-box and black-box settings, including robustness against adaptive attacks.
jailbreak detectionembedding disruptionlarge language modelssafeguardsadaptive attacks
Measuring Embedding Sensitivity to Authorial Style in French: Comparing Literary Texts with Language Model Rewritings
The study quantifies how language model embeddings retain authorial style in French literary texts after LLM rewriting. Using a controlled dataset, the authors measure stylistic variation through embedding dispersion analysis. Results show embeddings reliably encode authorial style features that persist post-rewriting while exhibiting model-specific patterns, suggesting utility for authorship imitation detection. The method provides empirical grounding for analyzing stylistic transfer in LLM-generated text.
embedding sensitivityauthorial stylellm rewritingembedding dispersionstylistic variation
Fairness vs Performance: Characterizing the Pareto Frontier of Algorithmic Decision Systems
This work characterizes the Pareto frontier of algorithmic decision systems by modeling fairness-performance trade-offs as a multi-objective optimization problem. The authors analyze binary prediction-based decisions under arbitrary utility functions, population distributions, and group fairness metrics. They prove that Pareto-optimal rules consist of deterministic, group-specific threshold rules applied to individuals' success probabilities, potentially including both lower- and upper-bound thresholds depending on the fairness metric. The Pareto frontier's location depends solely on population characteristics, utility functions, and fairness scores, independent of algorithmic implementation (pre-, in-, or post-processing). These findings generalize existing optimality theorems for fairness-constrained classification to broader fairness metrics and partial fairness regimes.
pareto frontiermulti-objective optimizationgroup fairnessthreshold rulesfairness-constrained classification
The Open-Box Fallacy: Why AI Deployment Needs a Calibrated Verification Regime
The paper critiques the overreliance on mechanistic interpretability for AI deployment authorization, proposing calibrated verification as a superior framework. It advocates for domain-scoped, checkable, monitored, accountable, contestable, and revocable authorization, supported by empirical gaps (53-percentage-point disparity between model internals and corrective actions; 9.0% FDA-approved AI/ML devices with post-market surveillance). The authors introduce Verification Coverage, a six-component standard for model documentation and regulatory compliance.
calibrated verificationmechanistic interpretabilityverification coveragepost-market surveillancemodel cards
Budget-Efficient Automatic Algorithm Design via Code Graph
The authors propose a budget-efficient framework for automatic algorithm design (AAD) using large language models (LLMs), addressing inefficiencies in existing pipelines. Their method employs a directed acyclic graph representation of algorithms, querying LLMs for compact code corrections rather than full algorithms. This graph structure enables correction-level credit assignment and efficient composition of algorithmic features. Theoretical insights guide the balance between search depth and breadth under budget constraints. Empirical validation on combinatorial optimization problems demonstrates superior performance over full-algorithm search at equal token budgets, with nuanced findings on the role of context richness in LLM performance.
automatic algorithm designlarge language modelsdirected acyclic graphcorrection-level credit assignmentcombinatorial optimization
CrackMeBench: Binary Reverse Engineering for Agents
CrackMeBench introduces a benchmark for evaluating language-model agents on binary reverse-engineering tasks, focusing on deterministic validation problems with symbol-poor binaries and explicit local tool access. The benchmark combines eight public calibration CrackMes with twelve generated tasks from seeded C, Rust, and Go templates, executed in a no-network Linux Docker sandbox. Evaluations with GPT-5.5, Claude Opus 4.7, and Kimi K2 under a five-minute budget show pass@3 rates of 92%, 58%, and 42% respectively on generated tasks, with sharper separation on harder tasks. CrackMeBench records detailed metrics including pass rates, command traces, and token usage, providing a reproducible testbed for autonomous binary analysis.
binary reverse engineeringlanguage-model agentssymbol-poor binariesdocker sandboxvalidation logic
LLARS: Enabling Domain Expert & Developer Collaboration for LLM Prompting, Generation and Evaluation
LLARS introduces an open-source platform facilitating collaboration between domain experts and developers for LLM-based system development. The system integrates three modules: Collaborative Prompt Engineering for real-time co-authoring with version control and instant LLM testing, Batch Generation for configurable output production across prompts, models, and data with cost control, and Hybrid Evaluation combining human and LLM assessments with live agreement metrics and provenance analysis. Interviews with six domain experts and three developers in online counseling confirmed LLARS' intuitiveness, time efficiency, and seamless interdisciplinary collaboration.
collaborative prompt engineeringbatch generationhybrid evaluationllm-based systemsprovenance analysis
A Resilient Solution for Sewer Overflow Monitoring across Cloud and Edge
The authors present a resilient web-based demonstrator for monitoring combined sewer overflows (CSO) in aging urban infrastructure, integrating Deep Learning forecasting methods across cloud and edge environments. The system forecasts overflow basin filling dynamics to anticipate capacity exceedance and enable preventive actions, maintaining functionality during network outages. An interactive dashboard provides real-time monitoring capabilities, demonstrated through an online showcase. The solution addresses critical environmental and public health impacts caused by CSO events triggered by extreme rainfall in historical cities.
combined sewer overflowsdeep learningcloud-edge integrationoverflow basininteractive dashboard
An agentic framework for gravitational-wave counterpart association in the multi-messenger era
We introduce GW-Eyes, an agentic framework leveraging large language models (LLMs) to address challenges in gravitational wave (GW) and electromagnetic (EM) counterpart association for multi-messenger astronomy. GW-Eyes integrates domain-specific tools to autonomously perform association tasks, supports natural language interaction for auxiliary functions like catalog management and skymap visualization, and utilizes LLMs' complex decision-making and traceable reasoning capabilities. This framework aims to enhance the efficiency and scalability of counterpart association in the era of next-generation GW and EM detectors, offering a novel approach to multi-messenger data analysis.
gravitational waveselectromagnetic counterpartsmulti-messenger astronomylarge language modelsagentic framework
Guaranteed Jailbreaking Defense via Disrupt-and-Rectify Smoothing
The paper introduces Disrupt-and-Rectify Smoothing (DR-Smoothing), a guaranteed defense method against jailbreaking attacks in large language models (LLMs). The method integrates a two-stage prompt processing scheme—disrupting the input prompt followed by rectifying it—into the conventional smoothing defense framework. This approach improves upon disrupt-only methods by restoring out-of-distribution prompts to an in-distribution form, reducing unpredictable LLM behavior. Theoretical analysis provides a tight bound for defense success probability and disruption strength requirements. Experiments show DR-Smoothing outperforms state-of-the-art methods in harmlessness and helpfulness under token-level, prompt-level, and adaptive attack scenarios.
disrupt-and-rectify smoothingjailbreaking defenselarge language modelsprompt processingtheoretical analysis
SenseBench: A Benchmark for Remote Sensing Low-Level Visual Perception and Description in Large Vision-Language Models
The authors introduce SenseBench, the first benchmark for evaluating remote sensing (RS) low-level visual perception and description in Vision-Language Models (VLMs). The benchmark features a physics-based hierarchical taxonomy with over 10K instances across 6 major and 22 fine-grained RS degradation categories, assessing both objective perception and subjective description. Evaluation of 29 VLMs reveals domain biases, multi-distortion collapse, fluency illusion, and perception-description inversion effects, providing a diagnostic tool for RS-oriented VLM development.
vision-language modelsremote sensingimage quality assessmentlow-level perceptionbenchmark
Acceptance Cards:A Four-Diagnostic Standard for Safe Fine-Tuning Defense Claims
The paper introduces Acceptance Cards, a four-diagnostic standard for evaluating safe fine-tuning defense claims, addressing limitations of held-out gap reduction metrics. The protocol assesses statistical reliability, semantic generalization, mechanism alignment, and cross-task transfer through an executable audit package. Applied to SafeLoRA on Gemma-2-2B-it, the method reveals failure under strict coding (0/4 diagnostics passed) and permissive relabeling (1/4 passed). A 46-cell audit shows no cases meet the strict conjunction, with the closest family still failing fresh-subject thresholds and transfer requirements while incurring accuracy costs.
safe fine-tuningheld-out gapmechanism alignmentcross-task transferexecutable audit
LLM Jaggedness Unlocks Scientific Creativity
The study introduces SciAidanBench, a benchmark assessing scientific creativity in LLMs through open-ended question generation. Evaluating 19 base models across 8 providers (30 variants), it reveals jagged capability progression: cross-task divergence between general and scientific creativity, prompt-level variability, and domain-specific fragmentation. The work demonstrates how inference-time compute, knowledge pooling, and brainstorming can harness this jaggedness to construct meta-model ensembles outperforming individual models, positioning uneven capability growth as a resource for enhancing scientific creativity.
sciaidanbenchjaggednessmeta-model ensemblesinference-time computeknowledge pooling
Deep Arguing
Deep Arguing introduces a neurosymbolic approach combining deep learning with argumentation construction for interpretable classification across data modalities. The method employs deep neural networks to build argumentation structures where data points support assigned labels and attack alternatives, using differentiable argumentation semantics for end-to-end training. This jointly learns feature representations and argumentative interactions, guided by structural constraints on the argumentation graph. Experiments on tabular and imaging datasets demonstrate competitive performance with standard baselines while providing faithful case-based explanations through interpretable argumentative reasoning.
neurosymbolicargumentation semanticsinterpretable classificationfeature representationargumentation graph
ThreatCore: A Benchmark for Explicit and Implicit Threat Detection
The authors introduce ThreatCore, a benchmark dataset for fine-grained threat detection that distinguishes explicit threats, implicit threats, and non-threats. The dataset aggregates and re-annotates public resources under unified definitions, augmented with manually validated synthetic examples to improve coverage. Evaluation of Perspective API, zero-shot classifiers, and recent language models reveals implicit threats remain substantially harder to detect (performance gap unspecified), though Semantic Role Labeling as intermediate representation improves results by clarifying harmful intent structure.
threat detectionimplicit threatssemantic role labelingzero-shot classificationbenchmark dataset
Agent-First Tool API: A Semantic Interface Paradigm for Enterprise AI Agent Systems
The Agent-First Tool API paradigm addresses architectural mismatches between conventional CRUD APIs and autonomous agent requirements through three mechanisms: a Six-Verb Semantic Protocol for tool interaction phases, a Normalized Tool Contract (NTC) with decision-support metadata, and a dual-layer governance pipeline. Implemented in a multi-tenant SaaS platform with 85 tools across 6 domains, the paradigm achieves an 88% end-to-end task success rate (+37.5% vs. CRUD baselines), reduces human interventions by 72.7%, and improves autonomous error recovery by 5.8x. It operates orthogonally to transport-layer standards like MCP, enhancing semantic application-layer functionality.
six-verb semantic protocolnormalized tool contractdual-layer governanceautonomous error recoverymulti-tenant saas
Bridging Sequence and Graph Structure for Epigenetic Age Prediction
The paper proposes a unified sequence-graph integration framework for epigenetic age prediction that jointly models co-methylation graph structure and DNA sequence context. The method integrates eight-dimensional sequence statistical features via a gated modulation mechanism, adaptively scaling methylation signals by sequence-determined relevance before graph convolution. Evaluated on 3,707 blood methylation samples, it achieves a test MAE of 3.149 years (12.8% improvement over graph-based baselines), with handcrafted sequence features outperforming CNN-based encodings. Interpretability analysis reveals CpG density and adenine frequency as key age-dependent features.
epigenetic clocksdna methylationgraph convolutionbiological age predictionsequence features
HH-SAE: Discovering and Steering Hierarchical Knowledge of Complex Manifolds
The paper introduces Hybrid Hierarchical SAE (HH-SAE), a method for resolving feature density conflict in high-dimensional domains by factorizing manifolds into three hierarchical tiers: Contextual ($L_0$), Atomic ($f_1$), and Compository ($f_2$). HH-SAE demonstrates superior manifold resolution, fracturing clinical labels into physiological modes and achieving a cross-domain zero-shot AUC of 0.9156 in fraud detection. Path ablation reveals a 13.46% performance drop without contextual subtraction, while knowledge-steered synthesis yields a +9.9% AUPRC improvement over state-of-the-art generators.
feature density conflicthierarchical factorizationzero-shot aucpath ablationknowledge-steered synthesis
A Reflective Storytelling Agent for Older Adults: Integrating Argumentation Schemes and Argument Mining in LLM-Based Personalised Narratives
The study introduces a reflective storytelling agent for older adults that integrates knowledge graphs, user modeling, argumentation theory, and argument mining to enhance LLM-based narrative generation. The system was developed through participatory design with 11 domain experts and evaluated by 55 older adults across four prompts and two creativity levels. Results show that participants recognized personally relevant purposes in approximately two-thirds of narratives, with argument-based purposes identified in half of these cases. Cultural relatability significantly influenced usability, while minor inconsistencies were tolerated if narratives remained understandable. Higher hallucination-risk indicators correlated with perceived inconsistencies, and higher argument-quality indicators were associated with clearer and more meaningful narratives.
knowledge graphsargument mininguser modelinghallucination-risk indicatorsargumentation theory
PrimeKG-CL: A Continual Graph Learning Benchmark on Evolving Biomedical Knowledge Graphs
PrimeKG-CL introduces a continual graph learning benchmark for evolving biomedical knowledge graphs, addressing the limitations of synthetic static KG evaluations. Built from nine authoritative biomedical databases (129K+ nodes, 8.1M+ edges), it includes two temporal snapshots (June 2021, July 2023) with 5.83M edges added and 889K removed. The benchmark features 10 entity-type-grouped tasks, multimodal node features, and stratified testing. Evaluations across six continual learning strategies and four KGE decoders reveal strong interactions between decoder choice and strategy effectiveness, with DistMult uniquely distinguishing persistent from deprecated knowledge. Multimodal features improve entity-level tasks by up to 60%, while IncDE fails to scale to the 5.67M-triple base task.
continual graph learningbiomedical knowledge graphsmultimodal node featuresknowledge graph embeddingstemporal snapshots
DuetFair: Coupling Inter- and Intra-Subgroup Robustness for Fair Medical Image Segmentation
DuetFair introduces a dual-axis fairness framework addressing both inter-subgroup adaptation and intra-subgroup robustness in medical image segmentation, mitigating intra-group hidden failures. The proposed FairDRO mechanism combines distribution-aware mixture-of-experts (dMoE) with subgroup-conditioned distributionally robust optimization (DRO) loss aggregation to enhance subgroup-specific performance while reducing high-loss samples within subgroups. Evaluated on Harvard-FairSeg, HAM10000, and a 3D radiotherapy target cohort, FairDRO achieves superior equity-scaled performance, improving worst-case subgroup Dice by 3.5-4.1 points (6.0-7.4%) over baselines under tumor-stage and institution groupings.
medical image segmentationdistributionally robust optimizationmixture-of-expertsintra-group hidden failurefairness framework
Infinite Mask Diffusion for Few-Step Distillation
The paper introduces the Infinite Mask Diffusion Model (IMDM), addressing the factorization error bound in Masked Diffusion Models (MDMs) through a stochastic infinite-state mask. IMDM maintains MDM benefits like parallel decoding and bidirectional context while enabling few-step generation, where standard MDMs fail due to theoretical limitations. Empirical results show IMDM outperforms existing few-step distillation methods on LM1B and OpenWebText benchmarks when combined with appropriate distillation techniques.
masked diffusion modelsfactorization errorinfinite-state maskfew-step generationdistillation
Consistency as a Testable Property: Statistical Methods to Evaluate AI Agent Reliability
The paper establishes a rigorous framework for quantifying AI agent reliability through consistency metrics under semantically preserving perturbations. It employs $U$-statistics for output-level reliability and kernel-based metrics for trajectory-level stability, distinguishing between core capability and execution robustness. Experimental validation across three agentic benchmarks demonstrates that trajectory-level consistency metrics offer significantly greater diagnostic sensitivity compared to traditional pass@1 rates. The framework enables precise identification of architectural weaknesses, facilitating improvements for deployment in high-stakes environments.
consistency metricssemantically preserving perturbationsu-statisticskernel-based metricstrajectory-level stability
SoK: A Systematic Bidirectional Literature Review of AI & DLT Convergence
This systematic bidirectional review analyzes 2020-2025 literature on AI-DLT convergence, classifying contributions into AI-enhanced DLT (data/network/consensus/execution/application layers) and DLT-enhanced AI (infrastructure/data/model/inference/application layers). The study reveals disproportionate focus on execution/consensus layers for AI-DLT and data/model layers for DLT-AI, with other layers understudied. While controlled experiments show improvements, no production-scale deployments exist, and fundamental challenges in scalability, interoperability, and verifiable execution remain unresolved. The authors advocate for cross-layer co-design and real-world validation to advance the field.
distributed ledger technologyfederated learningconsensus mechanismsmulti-agent systemsverifiable execution
CMKL: Modality-Aware Continual Learning for Evolving Biomedical Knowledge Graphs
The Continual Multimodal Knowledge Graph Learner (CMKL) is proposed for evolving biomedical knowledge graphs, addressing limitations in existing methods by encoding structure, text, and molecules through a Mixture-of-Experts (MoE) router and employing Elastic Weight Consolidation (EWC) with a K-means-diverse multimodal replay buffer. CMKL achieves a 60% improvement in continual biomedical entity classification (AP 0.591 vs. 0.370) with near-zero forgetting (AF 0.008) and matches or outperforms baselines in continual relationship prediction (AP 0.062). A frozen-text ablation reveals modality asymmetry at the representation level, managed by MoE routing without learned bottlenecks.
continual learningknowledge graphsmixture-of-expertsmultimodal fusionelastic weight consolidation
SLASH the Sink: Sharpening Structural Attention Inside LLMs
The paper introduces StructuraL Attention SHarpening (Slash), a training-free method to enhance LLMs' structural understanding of graph topologies. The authors identify that LLMs internally reconstruct graph structures via attention maps with sawtooth patterns, but this capability is diluted by attention sinks caused by anisotropic bias. Slash redistributes attention to amplify intrinsic structural awareness without fine-tuning. Experiments on graph tasks and molecular prediction show consistent performance improvements across multiple LLMs.
structural attentionattention sinkgraph topologyllmstraining-free
SkillEvolver: Skill Learning as a Meta-Skill
SkillEvolver introduces a meta-skill framework for online skill learning, enabling iterative authoring, deployment, and refinement of domain-specific skills without retraining model weights. The meta-skill refines skills based on failures encountered during deployment, governed by a fresh-agent overfit audit that detects leakage and silent-bypass failures. Evaluated on 83 SkillsBench tasks across 15+ domains, SkillEvolver achieves 56.8% accuracy, outperforming curated human skills (43.6%) and the no-skill baseline (29.9%). On KernelBench's GPU kernel optimization tasks, it improves mean speedup from 1.16 to 1.51.
meta-skillonline skill learningfresh-agent overfit auditsilent-bypasstrace-distillation
Simultaneous Long-tailed Recognition and Multi-modal Fusion for Highly Imbalanced Multi-modal Data
The authors propose a novel framework for simultaneous long-tailed recognition and multi-modal fusion to address class imbalance in heterogeneous data. The method extends multi-expert architectures by dynamically weighting modality-specific networks based on their estimated informativeness, enabling adaptive fusion of complementary information from diverse sources like images and tabular data. Specialized training and testing procedures accommodate varying modality combinations. Experiments on benchmark and real-world datasets demonstrate superior performance in handling long-tailed, class-imbalanced scenarios compared to existing methods, highlighting the approach's robustness and generalization capability.
long-tailed recognitionmulti-modal fusionclass imbalancemodality-specific networksadaptive fusion
Multi-layer attentive probing improves transfer of audio representations for bioacoustics
This work demonstrates that multi-layer attentive probing improves transfer learning for bioacoustic tasks compared to standard last-layer linear probing. The authors systematically evaluate probing strategies (last-layer vs. multi-layer, linear vs. attention) on BEANs and BirdSet benchmarks using various audio representation models. Results show that larger probe heads leveraging temporal information outperform fixed low-capacity probes, with multi-layer probing universally improving performance and attention probes being particularly effective for transformer architectures, suggesting current benchmarking practices may underestimate encoder quality.
probing headsbioacousticsrepresentation learningmulti-layer probingattention probes
DeepRefine: Agent-Compiled Knowledge Refinement via Reinforcement Learning
DeepRefine introduces a reinforcement learning-based framework for refining agent-compiled knowledge bases, addressing incompleteness, incorrectness, and redundancy through multi-turn interactions and abductive diagnosis. The method employs a Gain-Beyond-Draft (GBD) reward to optimize refinement policies without gold references, enabling incremental updates to enhance retrieval fidelity and downstream task performance. Extensive experiments demonstrate consistent improvements over strong baselines, validating the effectiveness of the approach in knowledge-intensive LLM agent tasks.
knowledge refinementreinforcement learningabductive diagnosisgain-beyond-draftllm agents
ASIA: an Autonomous System Identification Agent
ASIA introduces an autonomous system identification framework leveraging large language models as coding agents to automate model selection, hyperparameter tuning, and training strategy optimization. The approach eliminates manual trial-and-error by closing the loop between hypothesis formulation, implementation, and evaluation based solely on plain-English problem descriptions. Empirical evaluation on two system identification benchmarks demonstrates ASIA's ability to discover effective architectures and training strategies, though limitations include potential test leakage, reduced transparency, and reproducibility challenges.
system identificationautonomous agentshyperparameter tuninglarge language modelsreproducibility
Formally Verifying Analog Neural Networks Under Process Variations Using Polynomial Zonotopes
The authors propose a polynomial-based model for formally verifying analog neural networks under manufacturing process variations, addressing their sensitivity to circuit-level deviations. The method employs reachability analysis with polynomial zonotopes, circumventing computationally expensive Monte Carlo simulations. Evaluations on three datasets demonstrate the approach's efficacy, verifying both fully-connected and convolutional analog neural networks while reducing verification time from days to seconds and enclosing 99% of variation samples.
analog neural networksprocess variationspolynomial zonotopesreachability analysisformal verification
Cavity-Enhanced Collective Quantum Processing with Polarization-Encoded Qubits
The authors propose a cavity-enhanced optical architecture for collective quantum processing, utilizing polarization-encoded logical qubits in recirculating intracavity modes. The architecture separates physical carriers and computational degrees of freedom, employing harmonic cavity bundles as a stable resonant substrate and programmable polarization transformations for single-qubit operations. Tunable controlled-phase gates are implemented via polarization-selective nonlinear interactions, enabling a universal gate set. Parameter-scaling analysis demonstrates that order-unity conditional phases are achievable in centimeter-scale cavities using accessible solid-state nonlinear media, without requiring extreme nonlinear coefficients, millisecond photon lifetimes, or sub-hertz laser stabilization. The results suggest resonant recirculation as a viable platform for cavity-based collective quantum architectures.
quantum processingpolarization encodingcontrolled-phase gatesnonlinear interactionresonant recirculation
Can Agent Benchmarks Support Their Scores? Evidence-Supported Bounds for Interactive-Agent Evaluation
The paper introduces an outcome evidence reporting layer to improve the reliability of interactive agent benchmarks by addressing limitations in outcome detection. The layer specifies required verification artifacts, applies a locked checklist to assign Evidence Pass, Evidence Fail, or Unknown labels, and reports evidence-supported score bounds to quantify uncertainty. This framework explicitly handles uncertain cases rather than aggregating them into a single success rate. Evaluation on five benchmarks (ANDROIDWORLD, AGENTDOJO, APPWORLD, tau3 bench retail, MINIWOB) demonstrates its effectiveness in distinguishing distinct failure modes.
interactive agentoutcome detectionevidence reportingscore boundsfailure modes
Statistical Model Checking of the Keynes+Schumpeter Model: A Transient Sensitivity Analysis of a Macroeconomic ABM
The paper demonstrates how statistical model checking (SMC) via MultiVeStA enables principled analysis of macroeconomic agent-based models (ABMs) without simulator modification. Applying SMC to the Keynes+Schumpeter (K+S) model, the authors conduct a transient sensitivity analysis over 600-step simulations, examining one-parameter sweeps, two macro observables (unemployment and GDP growth), and one micro-level probe (market share) with precision-driven stopping rules. Results reveal strong transient effects in macro-financial and structural parameter families, contrasting with weaker heuristic-rule impacts under identical precision policies. The study establishes SMC as a reproducible framework for economic ABM analysis, explicitly quantifying uncertainty and simulation costs.
statistical model checkingagent-based modelstransient sensitivity analysismacroeconomic observablesprecision-driven stopping
StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMs
The authors introduce StereoTales, a multilingual framework for open-ended stereotype discovery in LLMs, covering 10 languages and 79 socio-demographic attributes. The dataset comprises 650k stories from 23 LLMs, annotated across 19 dimensions, with statistical tests identifying 1,500+ over-represented associations rated for harmfulness by humans (N=247) and LLMs. Key findings show all evaluated models emit harmful stereotypes, prompt language culturally adapts these biases, and human-LLM harmfulness judgments align (Spearman ρ=0.62), with disagreements on specific attribute classes.
multilingual evaluationopen-ended generationsocial biasharmfulness ratingssocio-demographic attributes
Real vs. Semi-Simulated: Rethinking Evaluation for Treatment Effect Estimation
This study systematically examines the evaluation gap between methodological research and practical deployment in heterogeneous treatment effect estimation. The authors conduct a large-scale empirical comparison of meta-learners, base learners, and specialized causal models across semi-simulated benchmarks and real-world datasets, evaluating them using both counterfactual and observable metrics. Results reveal two key gaps: counterfactual metrics fail to recover estimators preferred by observable metrics, and rankings from semi-simulated benchmarks do not transfer to real data. The findings suggest that simple meta-learners with strong base models consistently outperform specialized causal models, highlighting the need for incorporating observable metrics and real-data validation in assessing methodological progress.
heterogeneous treatment effectsmeta-learnerscounterfactual metricsobservable metricscausal machine learning
Physical probes expose and alleviate chemical-environment collapse in molecular representations
The study introduces CLAIM (Contrastive Learning for Atom-to-molecule Inference of Molecular NMR), a framework addressing representational collapse in molecular NMR spectroscopy by aligning topological inputs with atom-resolved NMR observables. Using hierarchical chemical priors and cross-level contrastive learning, CLAIM enhances chemical resolution and improves atom-level molecule-spectrum retrieval. Demonstrating robustness in flexible/tautomeric systems, it achieves superior 13C NMR prediction, stereoisomer discrimination without 3D modeling, and transfers effectively to ADMET prediction and fluorescence estimation. The approach establishes spectral alignment as a physically grounded strategy for molecular representation learning.
nmr spectroscopyrepresentational collapsecontrastive learningmolecular topologyadmet prediction
CoWorld-VLA: Thinking in a Multi-Expert World Model for Autonomous Driving
CoWorld-VLA introduces a multi-expert world reasoning framework for autonomous driving, addressing limitations in existing Vision-Language-Action (VLA) models by providing planner-accessible intermediate representations. The framework constructs four expert tokens—semantic interaction, geometric structure, dynamic evolution, and ego trajectory—to model interaction intent, spatial structure, temporal dynamics, and behavioral goals. A diffusion-based hierarchical multi-expert fusion planner generates continuous ego trajectories by coupling scene context during joint denoising. Evaluations on NAVSIM v1 demonstrate competitive performance in future scene generation and planning, with strong collision avoidance and trajectory accuracy. Ablation studies confirm the complementarity and effectiveness of expert tokens as planning conditions.
vision-language-actionexpert tokensdiffusion-based plannerautonomous drivingtrajectory generation
Toward an Engineering of Science: Rebalancing Generation and Verification in the Age of AI
The paper proposes redesigning scientific epistemic infrastructure to address epistemic pollution caused by AI-generated scientific artifacts. It identifies a structural imbalance where AI lowers generation costs without proportionally reducing verification costs, particularly in paper-centric systems. The authors introduce 'blueprints' as structured, decomposed research artifacts that represent claims, evidence, and assumptions as typed graph components. This approach increases upfront generation costs but enables cheaper, localized verification downstream. A proof-of-concept prototype has been developed to instantiate this proposal.
epistemic pollutiongeneration costverification costblueprintstyped graph components
Can Language Models Analyze Data? Evaluating Large Language Models for Question Answering over Datasets
This paper evaluates large language models (LLMs) for question answering over datasets, focusing on direct dataset queries and SQL query generation from database schemas. The study examines state-of-the-art LLMs alongside smaller, resource-efficient models, employing various prompting strategies across two datasets with questions of varying difficulty. Results indicate strong performance by large LLMs but highlight limitations in smaller models, providing insights into LLM capabilities and constraints in data analytics tasks.
large language modelsquestion answeringsql query generationprompting strategiesdata analytics
Every finite group admits a just finite presentation
The authors resolve a longstanding open problem in group theory by proving that every finite group admits a just finite presentation, where removing any relation results in an infinite group. This affirmatively answers Problem 21.10 from the Kourovka Notebook. The result establishes that such presentations exist universally across all finite groups, addressing a fundamental question about group presentations and their minimality properties.
finite groupjust finite presentationkourovka notebookgroup theoryminimality
LLM4Branch: Large Language Model for Discovering Efficient Branching Policies of Integer Programs
LLM4Branch introduces a novel framework leveraging Large Language Models (LLMs) to automate the discovery of efficient branching policies for Mixed Integer Linear Programming (MILP) solvers. The framework generates an executable program skeleton via LLM, optimizes its parameter vector using a zeroth-order method, and evaluates end-to-end performance on MILP instances. Extensive experiments on standard benchmarks show that LLM4Branch achieves state-of-the-art performance among CPU-based methods and competes with advanced GPU-based models. The code is publicly available.
large language modelsmixed integer linear programmingbranching policieszeroth-order optimizationcpu-based methods
AnomalyClaw: A Universal Visual Anomaly Detection Agent via Tool-Grounded Refutation
AnomalyClaw introduces a training-free visual anomaly detection (VAD) agent that reframes anomaly judgment as a multi-round refutation process, addressing the unreliability of single-inference vision-language models (VLMs). The method employs a 13-tool library for visual verification, reference parsing, and frozen expert probing, with an optional self-evolution extension that builds an online rulebook from internal disagreement. On CrossDomainVAD-12, AnomalyClaw improves macro-AUROC by +6.23 pp (GPT-5.5), +7.93 pp (Seed2.0-lite), and +3.52 pp (Qwen3.5-VL-27B), with self-evolution adding +2.09 pp for Qwen3.5-VL-27B.
visual anomaly detectionvision-language modelsmulti-round refutationtool-grounded reasoningself-evolution
Phoenix-VL 1.5 Medium Technical Report
Phoenix-VL 1.5 Medium is a 123B-parameter multimodal foundation model adapted for Singaporean contexts through domain-specific pretraining on 1-trillion tokens and 250B-token long-context extension. It incorporates 22B tokens of localized data and 5B tokens of alignment via Online Direct Preference Optimization. The model achieves SOTA performance on Singapore-specific benchmarks while maintaining competitiveness in general multimodal and multilingual tasks, supported by a novel evaluation framework for localized knowledge and safety.
multimodalfoundation modeldirect preference optimizationlong-context extensionsovereign ai
GuardAD: Safeguarding Autonomous Driving MLLMs via Markovian Safety Logic
GuardAD introduces a model-agnostic safeguard for autonomous driving (AD) systems using multimodal large language models (MLLMs), addressing vulnerabilities in dynamic environments via Markovian safety logic. The method employs Neuro-Symbolic Logic Formalization to represent safety predicates over heterogeneous traffic participants and induces them through n-th order Markovian Logic Induction, enabling inference of latent hazards. Logic-Driven Action Revision refines actions based on inferred safety states without modifying the MLLM. Experiments show GuardAD reduces accident rates by 32.07% and improves task performance by 6.85%, validated through benchmarks, closed-loop simulations, and physical-world studies.
markovian logicneuro-symbolic formalizationmultimodal llmsaction revisionautonomous driving
Agentic Performance at the Edge: Insights from Benchmarking
The study evaluates agentic AI performance on edge devices by analyzing model scaling, general-purpose versus coder-oriented models, and tool-enabled execution under fixed constraints. Using a domain-conditioned evaluation methodology, it examines model-tool interactions, failure modes, and provides practical guidance for model selection. Results indicate edge-agent quality is not solely dependent on parameter count, with Pareto fronts in accuracy-latency space revealing optimal strategies for deployment based on operational priorities.
agentic aiedge computingmodel scalingpareto frontstool workflow
Agent-X: Full Pipeline Acceleration of On-device AI Agents
Agent-X introduces a software-only framework for accelerating on-device AI agents by optimizing both prefill and decode stages. The method employs prompt rewriting with prefix caching tailored to agent-specific input patterns and LLM-free speculative decoding for efficient token generation. Results demonstrate a 1.61x end-to-end speedup on representative workloads without accuracy loss, while maintaining seamless integration with existing systems.
on-deviceprefilldecodespeculative decodingprefix caching
Autonomous FAIR Digital Objects: From Passive Assertions to Active Knowledge
The paper operationalizes Autonomous FAIR Digital Objects (aFDOs) to transition scientific knowledge from passive assertions to active, accountable automation. aFDOs integrate three Semantic Web-based layers: 1) a policy layer using RDF-star with PROV-O, SHACL, and ODRL for portable rules, 2) an announcement layer via ActivityStreams 2.0 for bounded evaluation cost, and 3) an agreement layer resolving contradictions through reputation-weighted consensus under adversarial bounds. Evaluated on 4,305 FDOs from rare-disease ontologies (ClinVar, HPO, Orphanet) and synthetic data, the consensus mechanism resolves 56.3% of 3,914 ClinVar conflicts. It degrades gracefully under Sybil, collusion, and poisoning attacks within Byzantine-tolerance bounds (f < n/5).
fair digital objectssemantic webbyzantine-toleranceactivitystreamsrdf-star
EGL-SCA: Structural Credit Assignment for Co-Evolving Instructions and Tools in Graph Reasoning Agents
The paper introduces EGL-SCA, a verifier-centric framework for graph reasoning agents that co-evolves instructions and tools through structural credit assignment. The method employs dual-space modeling (instruction-side policy space and tool-side program space) with conditional updates routed via failure analysis, supported by stratified training and Pareto-optimal retention. Evaluated on four benchmarks, EGL-SCA achieves 92.0% average success rate, outperforming pure-prompting and fixed-toolbox baselines.
structural credit assignmentgraph reasoning agentsdual-space modelingverifier-centric frameworkco-evolution
Agent-ValueBench: A Comprehensive Benchmark for Evaluating Agent Values
The paper introduces Agent-ValueBench, the first benchmark for evaluating agent values across 394 executable environments and 4,335 value-conflict tasks spanning 28 value systems. The benchmark employs a purpose-built synthesis pipeline with psychologist-curated instances, featuring trajectory-level rubric-based evaluation via pole-aligned golden trajectories. Evaluation of 14 models across 4 harnesses reveals three key findings: cross-model value homogeneity (Value Tide), non-additive harness effects, and the growing importance of harness alignment over classical model alignment.
agent valuesvalue-conflict taskstrajectory-level rubricharness alignmentvalue tide
RW-Post: Auditable Evidence-Grounded Multimodal Fact-Checking in the Wild
We introduce RW-Post, a text-image benchmark for multimodal fact-checking with auditable annotations, linking social-media posts to reasoning traces and evidence items extracted via an LLM-assisted pipeline. The benchmark supports evaluation across closed-book, evidence-bounded, and open-web regimes, enabling systematic diagnosis of visual grounding and evidence utilization. We provide AgentFact as a reference verification baseline and benchmark strong open-source LVLMs under unified protocols. Experiments reveal substantial headroom, with current models struggling in faithful evidence grounding, while evidence-bounded evaluation improves both accuracy and faithfulness.
multimodal fact-checkingauditable annotationsevidence-bounded evaluationvisual groundingllm-assisted pipeline
Portable Active Learning for Object Detection
Portable Active Learning (PAL) introduces a detector-agnostic framework for efficient object detection annotation, addressing scalability and integration challenges. PAL combines class-wise instance uncertainty, image-level diversity, and class-imbalance cues to guide data selection without modifying detector internals or training pipelines. It trains lightweight logistic classifiers to produce entropy-based uncertainty scores and refines candidate images using global image entropy, class diversity, and similarity. Evaluated on COCO, PASCAL VOC, and BDD100K, PAL consistently enhances label efficiency and detection accuracy compared to existing active learning baselines, offering a practical solution for real-world deployment.
active learningobject detectioninstance uncertaintylabel efficiencydetector-agnostic
How Mobile World Model Guides GUI Agents?
The study investigates mobile world models for GUI agents by comparing four modalities: delta text, full text, diffusion-based images, and renderable code. These models achieve state-of-the-art performance on MobileWorldBench and Code2WorldBench. Evaluations on AITZ, AndroidControl, and AndroidWorld reveal that renderable code excels in in-distribution fidelity and multimodal supervision, while text-based feedback is robust for out-of-distribution execution. World-model-generated trajectories enhance agents' task performance despite distribution shifts. Additionally, posterior self-reflection offers limited benefits for overconfident agents, indicating world models are more effective as prior perception or training supervision than as post-hoc verifiers.
mobile world modelsrenderable codediffusion-based imagesout-of-distribution executionposterior self-reflection
TMAS: Scaling Test-Time Compute via Multi-Agent Synergy
We propose TMAS, a multi-agent synergy framework for scaling test-time compute in large language models, addressing limitations in existing structured approaches that weakly coordinate parallel reasoning trajectories or rely on noisy historical information. TMAS organizes inference as a collaborative process among specialized agents, introducing hierarchical memories—an experience bank for low-level reliable conclusions and local feedback, and a guideline bank for high-level strategies—to enable structured information flow across agents, trajectories, and refinement iterations. A hybrid reward reinforcement learning scheme preserves reasoning capability, enhances experience utilization, and encourages exploration. Experiments on challenging reasoning benchmarks demonstrate TMAS achieves stronger iterative scaling and improved stability compared to baselines.
test-time scalingmulti-agent synergyhierarchical memorieshybrid rewardreasoning benchmarks
EvoStreaming: Your Offline Video Model Is a Natively Streaming Assistant
The paper introduces EvoStreaming, a framework for adapting offline video-language models (VideoLLMs) to streaming assistants without architectural changes. The method leverages the base model as a self-supervised generator of streaming trajectories, synthesizing data for interaction policy tuning via relevance annotation and roll-out policy. Evaluated on RealStreamEval, a frame-level multi-turn benchmark, EvoStreaming improves streaming scores by up to 10.8 points across five VideoLLM backbones (Qwen2/2.5/3-VL, InternVL-3.5, MiniCPM-V4.5) using only 1,000 self-generated samples (139× less than prior work), while maintaining offline performance.
videollmsstreaming adaptationself-supervised learninginteraction policyrelevance annotation
PaperFit: Vision-in-the-Loop Typesetting Optimization for Scientific Documents
The paper introduces PaperFit, a vision-in-the-loop agent for Visual Typesetting Optimization (VTO), which transforms compilable LaTeX documents into publication-ready PDFs by iteratively diagnosing and repairing layout defects. The method employs a five-category taxonomy of typesetting errors, combining visual verification with constrained source-level edits. Evaluated on PaperFit-Bench (200 papers across 10 templates), PaperFit significantly outperforms baselines, demonstrating that vision-in-the-loop optimization is essential for bridging the gap between compilable source and polished output.
visual typesetting optimizationlatex document repairvision-in-the-looptypesetting defectsdocument automation
CORTEG: Foundation Models Enable Cross-Modality Representation Transfer from Scalp to Intracranial Brain Recordings
CORTEG introduces a cross-modality transfer framework that adapts pretrained scalp-EEG foundation models (EEG FMs) to intracranial electrocorticography (ECoG) for brain-computer interfaces. The method combines a pretrained EEG FM backbone with an electrode-aware KNNSoftFourier spatial adapter, a dual-stream tokenizer for low-frequency and high-gamma activity, and leave-one-subject-out fine-tuning. Evaluated on finger trajectory (n=9) and audio envelope regression (n=16), CORTEG matches or exceeds task-specific baselines, showing significant gains in low-data settings and audio tasks. Feature analyses confirm neurophysiological alignment, demonstrating scalable cross-patient learning with rapid per-patient calibration (10-30 minutes on one GPU).
electrocorticographyfoundation modelscross-modality transferbrain-computer interfaceshigh-gamma activity
PowerStep: Memory-Efficient Adaptive Optimization via $\ell_p$-Norm Steepest Descent
PowerStep introduces a memory-efficient adaptive optimizer that eliminates second-moment statistics storage while maintaining coordinate-wise adaptivity. The method applies nonlinear transforms to momentum buffers, inspired by ℓ_p-norm steepest descent, achieving optimal O(1/√T) convergence for non-convex stochastic optimization. Experiments on Transformers (124M to 235B parameters) show PowerStep matches Adam's convergence speed while reducing optimizer memory by 50%, and by 8× with int8 quantization. The approach provides a scalable, resource-efficient alternative for large-scale training.
adaptive optimizationmemory-efficientℓ_p-normsteepest descentquantization
EmbodiSkill: Skill-Aware Reflection for Self-Evolving Embodied Agents
The paper introduces EmbodiSkill, a training-free framework for self-evolving embodied agents through skill-aware reflection and targeted revision. The method analyzes trajectories to distinguish between skill deficiencies (updated via skill-changing evidence) and execution lapses (preserved via valid guidance reinforcement), addressing limitations of coarse skill updates in digital environments. Evaluations on ALFWorld and EmbodiedBench demonstrate a 93.28% task success rate with a frozen Qwen3.5-27B executor, outperforming GPT-5.2 by 31.58%, showing effective procedural knowledge accumulation from agent trajectories.
embodied agentsskill self-evolutiontrajectory analysisprocedural knowledgeexecution lapse
SCALAR: A Neurosymbolic Framework for Automated Conjecture and Reasoning in Quantum Circuit Analysis
The paper introduces SCALAR, a neurosymbolic framework combining quantum simulation, symbolic conjecture generation, and LLM-based interpretation for automated quantum circuit analysis. Built on CUDA-Q, the system generates conjectures about QAOA parameter bounds linked to graph invariants, validating known relationships like γ periodicity and parameter transfer phenomena. Evaluated on 82 MaxCut instances (MQLib) and 2,000 synthetic graphs across four topologies, SCALAR identifies structural feature correlations with optimization landscapes, scaling to 77-qubit instances using tensor network simulation. Results demonstrate conjecture accuracy while revealing limitations in graph class sensitivity and circuit depth effects.
neurosymbolicquantum circuit analysisqaoa parametersgraph invariantstensor network simulation
Verifiable Process Rewards for Agentic Reasoning
Verifiable Process Rewards (VPR) introduces a framework for dense turn-level supervision in reinforcement learning, addressing credit assignment challenges in long-horizon agentic reasoning. VPR leverages symbolic, algorithmic, or posterior-based oracles to verify intermediate actions, converting them into dense rewards for training. Theoretical analysis demonstrates improved credit assignment with verifier-grounded rewards, contingent on verifier reliability. Empirical evaluations show VPR outperforms outcome-level and rollout-based process reward baselines across controlled environments, with transferability to general and agentic reasoning benchmarks. Results suggest VPR enhances LLM agents when reliable intermediate verification is available, though its effectiveness depends on oracle quality and remains limited in unstructured environments.
verifiable process rewardscredit assignmentagentic reasoningsymbolic oraclesdense supervision
Relations Are Channels: Knowledge Graph Embedding via Kraus Decompositions
The authors propose extsc{KrausKGE}, a knowledge graph embedding (KGE) model grounded in Kraus channel theory, which satisfies structural axioms of linearity, trace preservation, and complete positivity. By formulating relations as Kraus channels, the model generalizes existing operator-based KGE approaches, supports $1$-to-$N$ and $N$-to-$N$ relations, enables $k$-hop reasoning without path encoders, and eliminates entity embedding norm constraints. Theoretical analysis introduces a per-relation complexity measure with a provable lower bound tied to empirical relation matrix rank. Empirical results show extsc{KrausKGE} outperforms baselines on $N$-to-$N$ relations, with gains increasing monotonically with relation fan-out.
kraus channelsknowledge graph embeddingtrace preservationcomplete positivityrelation complexity
Active Tabular Augmentation via Policy-Guided Diffusion Inpainting
The paper introduces TAP (Tabular Augmentation Policy), a method combining diffusion inpainting with a learner-conditioned policy to address the fidelity-utility gap in generative tabular augmentation. TAP dynamically steers generation toward high-utility regions and controls injection via gating and windowed commitment. Evaluated on seven real-world datasets under severe data scarcity, TAP outperforms generative baselines, improving classification accuracy by up to 15.6 percentage points and reducing regression RMSE by up to 32%.
tabular augmentationdiffusion inpaintingfidelity-utility gaplearner-conditioned policywindowed commitment
Positive Alignment: Artificial Intelligence for Human Flourishing
The paper introduces Positive Alignment, a paradigm shift in AI alignment research that extends beyond safety and harm prevention to actively promote human and ecological flourishing. It critiques current alignment approaches for being reactive and proposes methods like virtue cultivation, polycentric governance, and context-sensitive design. Technical challenges include data filtering, upsampling, and collaborative value collection across the LLM lifecycle. The framework emphasizes pluralism, user authorship, and decentralized oversight to address issues like engagement hacking and epistemic humility.
positive alignmentpolycentric governancevirtue cultivationcontext-sensitiveepistemic humility
Qwen Goes Brrr: Off-the-Shelf RAG for Ukrainian Multi-Domain Document Understanding
The authors present a retrieval-augmented pipeline for Ukrainian multi-domain document understanding, achieving top performance in the Fifth UNLP shared task. Their method combines contextual PDF chunking, question-aware dense retrieval with Qwen3-Embedding-8B, and answer-conditioned reranking using Qwen3-Reranker-8B, followed by constrained answer generation from top passages with Qwen3-32B. Reranking improved Recall@1 from 0.6957 to 0.7935, while using top-2 passages increased answer accuracy from 0.9348 to 0.9674. The system scored 0.9452 and 0.9598 on public and private leaderboards, demonstrating the effectiveness of preserving document structure and answer-aware relevance estimation over complex heuristics.
retrieval-augmented pipelinecontextual chunkingdense retrievalanswer-conditioned rerankingconstrained answer generation
Robust Probabilistic Shielding for Safe Offline Reinforcement Learning
The paper introduces robust probabilistic shielding for safe offline reinforcement learning, integrating safe policy improvement (SPI) with shielding techniques to ensure both performance and safety guarantees. The method restricts the action space to provably safe actions based on a dataset and known safe/unsafe states, shielding policy improvement steps to produce safe policies with high probability. Experimental results show that shielded SPI outperforms unshielded SPI, particularly in low-data regimes, improving both average and worst-case performance.
offline reinforcement learningsafe policy improvementshieldingprobabilistic guaranteeslow-data regimes
LeapTS: Rethinking Time Series Forecasting as Adaptive Multi-Horizon Scheduling
LeapTS introduces a novel framework reformulating time series forecasting as adaptive multi-horizon scheduling, addressing temporal decoupling in traditional models. The method employs a hierarchical controller for dynamic prediction scale selection and advancement length, coupled with neural controlled differential equations for continuous-time state evolution. This controlled update mechanism integrates irregular temporal dynamics with discrete scheduling feedback. Evaluations on real-world and synthetic datasets show LeapTS achieves a 7.4% performance improvement and 2.6× to 5.3× inference speedup over Transformer-based models, while autonomously adapting to non-stationary dynamics through explicit scheduling trajectory tracing.
time series forecastingadaptive schedulingneural controlled differential equationshierarchical controllernon-stationary dynamics
Generative AI Fuels Solo Entrepreneurship, but Teams Still Lead at the Top
The study examines how generative AI impacts entrepreneurial dynamics using Product Hunt launch data (n=160,000+). It finds ChatGPT-3.5's release increased solo entrepreneurship entry by 45% in traditionally team-dominated categories, but team-based ventures maintain a 2.3x higher representation in top-ranked outcomes. Methodologically, the analysis employs difference-in-differences to isolate AI's causal effect, revealing AI lowers entry barriers for solos while teams retain quality advantages in product development.
generative aientrepreneurshipproduct huntdifference-in-differencessolo ventures
AgentRx: A Benchmark Study of LLM Agents for Multimodal Clinical Prediction Tasks
This work introduces AgentRx, a benchmark evaluating LLM-based agents for multimodal clinical prediction tasks using real-world data. The study systematically assesses unimodal and multimodal performance, comparing single-agent frameworks against multi-agent systems. Results demonstrate that single-agent systems outperform naive multi-agent approaches in handling multimodal data, achieving better calibration and prediction accuracy. The findings emphasize the need for improved multi-agent collaboration strategies to address heterogeneous healthcare data. The authors open-source their evaluation framework to facilitate future research in agentic healthcare systems.
multimodal predictionclinical decision supportllm agentscalibrationheterogeneous data
Drum Synthesis from Expressive Drum Grids via Neural Audio Codecs
The paper presents a Transformer-based system for generating realistic drum audio from expressive drum grids (time-aligned MIDI with microtiming and velocity data) via neural audio codec token prediction. The method maps drum grids to discrete tokens of pre-trained codecs (EnCodec, DAC, X-Codec), subsequently decoded to waveform audio. Evaluated on the Expanded Groove MIDI Dataset (E-GMD), the approach demonstrates effective grid-to-audio conversion, with comparative analysis of codec performance on percussive synthesis tasks. Results validate codec-token prediction as a viable strategy and offer practical guidance for audio tokenizer selection in drum synthesis.
drum synthesisneural audio codecexpressive drum gridtransformermicrotiming
DP-LAC: Lightweight Adaptive Clipping for Differentially Private Federated Fine-tuning of Language Models
DP-LAC introduces a lightweight adaptive clipping method for differentially private federated fine-tuning of language models, eliminating hyperparameter tuning while preserving privacy budget. The approach first estimates an initial clipping threshold via private histogram estimation, then dynamically adjusts it during training. Evaluations demonstrate DP-LAC's superiority over state-of-the-art adaptive clipping and vanilla DP-SGD, yielding a 6.6% average accuracy improvement.
federated learningdifferential privacyadaptive clippinglanguage modelsgradient estimation
MemReread: Enhancing Agentic Long-Context Reasoning via Memory-Guided Rereading
MemReread enhances agentic long-context reasoning by introducing memory-guided rereading, avoiding quadratic attention complexity and retrieval-based recall limitations. The method combines streaming reading with question decomposition and adaptive rereading triggered by insufficient final memory, preserving document flow while enabling non-linear reasoning. A reinforcement learning framework optimizes rereading passes dynamically. Experiments show MemReread outperforms baselines on long-context tasks with linear time complexity relative to context length.
long-context reasoningmemory-guided rereadingquestion decompositionreinforcement learninglinear complexity
IndustryBench: Probing the Industrial Knowledge Boundaries of LLMs
We introduce IndustryBench, a 2,049-item multilingual benchmark for evaluating LLMs in industrial procurement QA, grounded in Chinese national standards (GB/T) and structured product records. The benchmark spans seven capability dimensions, ten industry categories, and difficulty tiers, with item-aligned translations in English, Russian, and Vietnamese. Our pipeline rejects 70.3% of LLM-generated candidates via external verification, highlighting the unreliability of LLM-only filtering. Evaluation of 17 Chinese and 8 multilingual models reveals that the top system scores only 2.083 on a 0–3 rubric, with Standards & Terminology as the most persistent weakness. Extended reasoning introduces safety violations, reshuffling leaderboards post-SV adjustment. We release IndustryBench with prompts, scoring scripts, and documentation.
industrial procurementmultilingual benchmarksafety-violationexternal verificationcapability dimensions
E-TCAV: Formalizing Penultimate Proxies for Efficient Concept Based Interpretability
E-TCAV introduces an efficient framework for approximating TCAV (Testing with Concept Activation Vectors) scores, addressing computational overhead, inter-layer disagreement, and statistical instability. The method leverages three key insights: the impact of latent classifiers on TCAV score stability, inter-layer agreement of TCAV scores, and the use of the penultimate layer as a fast proxy for earlier layers. Evaluations across four architectures and five datasets demonstrate strong inter-layer agreement in the final block and identify latent classifier choice as a source of TCAV score variance. E-TCAV achieves linearly scaling speed-ups with network size and evaluation samples, enabling efficient model debugging and concept-guided training.
tcavpenultimate layerlatent classifierinter-layer agreementconcept activation vectors
Towards Autonomous Railway Operations: A Semi-Hierarchical Deep Reinforcement Learning Approach to the Vehicle Rescheduling Problem
The paper introduces a semi-hierarchical deep reinforcement learning (RL) approach to address the Vehicle Routing and Scheduling Problem (VRSP) in railway operations, focusing on disruption management. The method separates dispatching and routing into distinct action and observation spaces, enabling specialized policies for each decision scope and addressing the imbalance between infrequent dispatch decisions and frequent routing updates. Evaluated on the Flatland-RL simulator across five difficulty levels and 50 random seeds with 7 to 80 trains, the approach demonstrates improved coordination, resource utilization, and robustness. It nearly doubles the number of trains reaching destinations while maintaining deadlock rates below 5% and adaptively managing congestion through sequencing, delaying, or canceling trains.
vehicle routingreinforcement learningdispatchingrailway operationsflatland-rl
A Cold Diffusion Approach for Percussive Dereverberation
We introduce a cold diffusion framework for percussive dereverberation, addressing the understudied domain of drum signals in audio processing. The method models reverberation as a deterministic degradation process, employing two reverse-process parameterizations: Direct (next-state) prediction and Delta-normalized residual (velocity-style) prediction. Implementations utilize both UNet and diffusion Transformer architectures, trained on curated datasets of acoustic and electronic drum recordings with synthetic and real room impulse responses. Extensive evaluations on in-domain and out-of-domain test sets demonstrate superior performance over score-based and conditional diffusion baselines, validated by signal-based and perceptual metrics specific to percussive audio.
cold diffusiondereverberationpercussive audiodelta-normalized residualdiffusion transformer
Knowledge Poisoning Attacks on Medical Multi-Modal Retrieval-Augmented Generation
We propose M3Att, a knowledge-poisoning framework for medical multi-modal retrieval-augmented generation (RAG) systems, assuming limited distribution knowledge of the underlying database. The method injects covert misinformation into textual data while using paired visual data as a query-agnostic trigger to manipulate retrieval probabilities. It leverages the inherent ambiguity of medical diagnosis to degrade diagnostic accuracy while evading model self-correction. Experiments on five LLMs and datasets show that M3Att consistently produces clinically plausible yet incorrect generations, demonstrating its effectiveness in undermining system reliability.
knowledge poisoningretrieval-augmented generationmultimodal retrievalquery-agnostic triggerself-correction
SciIntegrity-Bench: A Benchmark for Evaluating Academic Integrity in AI Scientist Systems
The authors introduce SCIINTEGRITY-BENCH, the first benchmark evaluating academic integrity in AI scientist systems through dilemmatic scenarios where honest failure acknowledgment is the only correct response. The benchmark comprises 33 scenarios across 11 trap categories, tested on 7 state-of-the-art LLMs (231 runs total), revealing a 34.2% overall integrity problem rate with no model achieving zero failures. Key findings include universal synthetic data generation in missing-data scenarios (100% of models) and prompt ablation showing undisclosed fabrication drops from 20.6% to 3.2% when completion pressure is removed, indicating an intrinsic completion bias.
academic integrityai scientist systemsdilemmatic evaluationsynthetic data generationcompletion bias
When Normality Shifts: Risk-Aware Test-Time Adaptation for Unsupervised Tabular Anomaly Detection
The paper introduces RTTAD, a risk-aware test-time adaptation method for unsupervised tabular anomaly detection, addressing normality shifts and anomaly contamination. The method employs a two-stage mechanism: collaborative dual-task learning during training to capture multi-level representations, and a Test-Time Contrastive Learning (TTCL) module during testing that selectively updates using high-confidence pseudo-normal samples. TTCL also uses a k-nearest neighbor-based contrastive objective to refine embeddings. Experiments on 15 datasets show RTTAD achieves state-of-the-art performance.
tabular anomaly detectiontest-time adaptationcontrastive learningnormality shiftsrisk-aware
When Does Non-Uniform Replay Matter in Reinforcement Learning?
The study identifies three key factors governing the effectiveness of non-uniform replay in off-policy reinforcement learning: replay volume, expected recency, and sampling distribution entropy. Through empirical analysis across diverse RL settings, the authors demonstrate that non-uniform replay benefits low-volume regimes and requires high-entropy sampling. They propose Truncated Geometric replay, a computationally efficient method that biases toward recent experiences while maintaining entropy. Evaluations on five benchmark suites with three modern algorithms show improved sample efficiency in low-volume scenarios without compromising high-volume performance.
off-policy reinforcement learningreplay buffersample efficiencynon-uniform samplingentropy preservation
Hypothesis-Driven Deep Research with Large Language Models: A Structured Methodology for Automated Knowledge Discovery
The paper introduces Hypothesis-Driven Deep Research (HDRI), a structured methodology for automated knowledge discovery using large language models. HDRI formalizes six core principles and an eight-stage pipeline, featuring a gap-driven iterative research mechanism for targeted supplementary investigation, a fact reasoning framework with traceable reasoning chains, a subject locking mechanism, and multi-dimensional quality assessment. Implemented in the INFOMINER system, HDRI improves fact density by 22.4%, subject matching accuracy by 90%, multi-source verification confidence by 0.92, and completeness by 14%. Five case studies validate its practical applicability, achieving an average quality rating of 4.46/5.0.
hypothesis-drivengap-driventraceable reasoningsubject lockingmulti-dimensional assessment
Beyond Autonomy: A Dynamic Tiered AgentRunner Framework for Governable and Resilient Enterprise AI Execution
The paper proposes Dynamic Tiered AgentRunner, a framework addressing governability gaps in enterprise AI agent deployment. It introduces three mechanisms: (1) Risk-Adaptive Tiering for dynamic resource allocation based on task risk profiles, (2) Separation of Powers architecture with independent proposal/review/execution/verification agents, and (3) Resilience-by-Design via Verifier-Recovery closed loops. The approach is derived from a production multi-tenant SaaS platform, achieving Pareto-optimal safety-efficiency tradeoffs through formalized tier selection processes.
agent frameworksrisk-adaptive tieringseparation of powersresilience-by-designpareto-optimal tradeoffs
To Redact, or not to Redact? A Local LLM Approach to Deliberative Process Privilege Classification
The study introduces a local LLM approach for classifying deliberative process privilege in FOIA documents, addressing legal constraints on cloud-based processing. Using Qwen3.5 9B on consumer hardware, the authors evaluate eight prompting variants, finding that Chain-of-Thought with few-shot error-based examples achieves superior recall and F2 scores, rivaling Gemini 2.5 Flash. Linguistic analysis reveals deliberative sentences feature opinion verbs, first-person phrasing, and multi-indicator combinations.
deliberative process privilegefew-shot promptingchain-of-thoughtqwen3.5foia exemption 5
HeteroGenManip: Generalizable Manipulation For Heterogeneous Object Interactions
HeteroGenManip introduces a task-conditioned, two-stage framework for generalizable manipulation of heterogeneous objects, addressing contact point localization and interaction trajectory planning separately. The Foundation-Correspondence-Guided Grasp module aligns initial contact states using structural priors, while the Multi-Foundation-Model Diffusion Policy routes objects to category-specialized foundation models via dual-stream cross-attention. This approach integrates fine-grained geometric information with part-specific features, reducing pose uncertainty and improving generalization. Evaluations show a 31% performance improvement in simulation tasks and a 36.7% gain in real-world tasks across diverse interaction types.
generalizable manipulationfoundation-correspondence-guided graspmulti-foundation-model diffusion policydual-stream cross-attentionheterogeneous objects
Empty SPACE: Cross-Attention Sparsity for Concept Erasure in Diffusion Models
SParse cross-Attention-based Concept Erasure (SPACE) introduces a closed-form method for erasing specific concepts in text-to-image diffusion models, addressing limitations of existing techniques in scaling to larger architectures like Stable Diffusion XL. SPACE iteratively modifies cross-attention parameters to induce sparsity and map concepts to a lower-dimensional subspace, enhancing erasure efficacy and robustness against adversarial prompts. Experiments demonstrate SPACE achieves 80%-90% cross-attention sparsity, reducing storage requirements for modified parameters by 70%, while maintaining superior erasure effectiveness compared to dense baselines.
cross-attentionconcept erasuresparsitydiffusion modelsclosed-form
TRACE: Distilling Where It Matters via Token-Routed Self On-Policy Alignment
We propose Token-Routed Alignment for Critical rEasoning (TRACE), a method for on-policy self-distillation that selectively distills knowledge on annotator-marked critical spans rather than full responses. TRACE employs forward KL on key spans of correct rollouts, optional reverse KL on localized error spans, and GRPO on remaining tokens, with KL channel annealing after warm-up. This approach mitigates privileged-information leakage and entropy rise observed in all-token distillation. On four math benchmarks and GPQA-Diamond, TRACE outperforms GRPO by 2.76 percentage points on average and maintains Qwen3-8B's OOD score, where baselines degrade. Gains persist under online self-annotation (+1.90 percentage points), demonstrating robustness to external annotator reliance.
on-policy self-distillationtoken-routed alignmentforward klreverse klgrpo
ProteinOPD: Towards Effective and Efficient Preference Alignment for Protein Design
ProteinOPD introduces a multi-objective preference alignment framework for protein design that maintains designability while balancing competing objectives. The method adapts pretrained protein language models (PLMs) into preference-specific teachers and distills their knowledge into a shared student via token-level On-Policy Distillation (OPD) on the student's trajectories. This approach ensures bounded optimization under conflicts and achieves a normalized geometric consensus of weighted teachers. Experiments demonstrate substantial gains in target preference objectives without compromising designability, with an 8x training speedup over RL-based alignment methods.
protein language modelson-policy distillationmulti-objective alignmentpreference-specific teachersdesignability
LegalCiteBench: Evaluating Citation Reliability in Legal Language Models
LegalCiteBench introduces a benchmark for evaluating citation reliability in legal language models, addressing the critical failure mode of incorrect or fabricated case citations in closed-book settings. The benchmark comprises 24K evaluation instances derived from 1,000 U.S. judicial opinions, focusing on five tasks: citation retrieval, completion, error detection, case matching, and verification/correction. Evaluation across 21 LLMs reveals poor performance, with exact citation recovery scores below 7/100 and Misleading Answer Rates exceeding 94% for retrieval-heavy tasks. Scale and legal-domain pretraining offer limited improvements, and explicit uncertainty instructions reduce fabrication but do not enhance correctness.
legal language modelscitation reliabilityclosed-book settingmisleading answer ratescase matching
DynGhost: Temporally-Modelled Transformer for Dynamic Ghost Imaging with Quantum Detectors
DynGhost introduces a transformer architecture for dynamic ghost imaging that models temporal coherence and addresses Poissonian noise in quantum detectors. The method employs alternating spatial-temporal attention blocks and a quantum-aware training framework with Anscombe normalization, using simulated single-photon detectors (SNSPDs, SPADs, SiPMs). Experiments show DynGhost outperforms traditional and deep learning baselines, particularly in dynamic scenes and low-photon regimes.
ghost imagingtransformerpoissonian noisesingle-photon detectorsanscombe normalization
Developing a foundation model for high-resolution remote sensing data of the Netherlands
We introduce a foundation model for high-resolution (1.2m) remote sensing data of the Netherlands, combining Convolutional Neural Networks and Vision Transformers to capture both low- and high-frequency landscape features. The model leverages temporal data to exploit dependencies in topographic features, land-cover changes, and seasonal dynamics, reducing feature ambiguity and improving representation learning with fewer labeled samples. Evaluated on downstream tasks including Dutch vegetation monitoring and global benchmarks, the model demonstrates competitive performance despite using fewer parameters and limited pretraining data. This indicates its ability to learn generalizable representations from constrained data. Scripts and models are publicly available for reproducibility.
foundation modelconvolutional neural networkvision transformertemporal dependenciesrepresentation learning
A Comparative Study of Machine Learning and Deep Learning for Out-of-Distribution Detection
This study demonstrates that traditional machine learning (ML) achieves comparable out-of-distribution (OOD) detection performance to deep learning (DL) on medical imaging tasks while offering superior computational efficiency. The authors evaluate ML and DL approaches on a dataset of over 60,000 fundus and non-fundus images acquired under standardized protocols, assessing performance on internal and external validation sets. Both methods achieved perfect AUROC scores (1.000) and near-perfect accuracies (0.999-1.000), but ML exhibited significantly lower end-to-end latency. These findings suggest that lightweight ML approaches are viable alternatives to DL for OOD detection in visually constrained domains, enabling practical deployment with reduced computational costs.
out-of-distribution detectionmedical imagingcomputational efficiencyaurocend-to-end latency
One-Step Graph-Structured Neural Flows for Irregular Multivariate Time Series Classification
Proposes Graph-Structured Neural Flows (GSNF), a novel approach for irregular multivariate time series classification that enhances inter-variable interaction modeling through one-step neural ODEs. GSNF introduces two self-supervision strategies: interaction-aware trajectory generation via re-initialization, which ensures trajectory divergence to expose graph-induced interactions, and reverse-time trajectory generation, leveraging flow invertibility for forward-backward consistency. Evaluated on five real-world datasets, GSNF achieves state-of-the-art classification performance while maintaining competitive training time and memory efficiency.
neural flowsmultivariate time seriesode trajectoriesself-supervisiongraph-structured
MTA-RL: Robust Urban Driving via Multi-modal Transformer-based 3D Affordances and Reinforcement Learning
MTA-RL introduces a novel framework combining Multi-modal Transformer-based 3D Affordances and Reinforcement Learning for robust urban autonomous driving. The method fuses RGB images and LiDAR point clouds via a transformer architecture to predict geometry-aware affordances, serving as a structured observation space for RL policies. Evaluations in CARLA Town01-03 demonstrate superior performance, with 9.0% higher Route Completion, 11.0% increased Total Distance, and 83.7% improved Distance Per Violation versus baselines. The approach shows strong zero-shot generalization and benefits from multi-modal fusion and reward shaping.
autonomous drivingtransformeraffordancesreinforcement learningmulti-modal fusion
When Prompts Become Payloads: A Framework for Mitigating SQL Injection Attacks in Large Language Model-Driven Applications
The study introduces a multi-layered security framework to mitigate SQL injection attacks in large language model (LLM)-driven database applications. The framework combines prompt sanitization, advanced threat detection for behavioral and semantic anomalies, and signature-based controls for known attack patterns. It was evaluated under diverse attack scenarios, including prompt injection, obfuscated SQL payloads, and context manipulation, using a curated benchmark dataset of adversarial prompts. Experimental results demonstrate high detection accuracy and low false-positive rates, enhancing the secure deployment of LLM-powered database interfaces.
sql injectionprompt sanitizationthreat detectionadversarial promptsllm-driven applications
When Reviews Disagree: Fine-Grained Contradiction Analysis in Scientific Peer Reviews
The paper introduces a fine-grained approach to analyzing contradictions in scientific peer reviews, moving beyond binary detection to identify evidence spans and graded intensity scores. It presents RevCI, an annotated benchmark of review pairs, and proposes IMPACT, a multi-agent framework combining aspect-conditioned evidence extraction, deliberative reasoning, and adjudication. IMPACT outperforms baselines in evidence identification and intensity agreement, while its distilled version TIDE achieves competitive performance with lower inference costs.
peer review analysiscontradiction detectionevidence extractionmulti-agent frameworkintensity scoring
Automated Approach for Solving Infinite-state Polynomial Reachability Games
We propose ranking certificates, a sound and complete proof rule for determining winning strategies in infinite-state polynomial reachability games, and develop a fully automated algorithm for computing such strategies with formal correctness witnesses. The algorithm handles turn-based games on infinite-state graphs defined by polynomial constraints over real variables, running in sub-exponential time with soundness and semi-completeness guarantees. Experimental results demonstrate the method's effectiveness, including solving the Cinderella-Stepmother game with arbitrary precision for the first time.
reachability gamesranking certificatespolynomial constraintswinning strategyinfinite-state graphs
Task-Agnostic Noisy Label Detection via Standardized Loss Aggregation
We propose Standardized Loss Aggregation (SLA), a task-agnostic framework for detecting noisy labels in large-scale datasets. SLA aggregates standardized fold-level validation losses across repeated cross-validation runs, generalizing discrete hard-counting schemes into a continuous estimator that captures both frequency and magnitude of performance deviations. Experiments on a public fundus dataset show SLA outperforms hard-counting baselines across all noise levels, converging faster especially under low noise ratios where subtle loss variations are informative. High SLA scores identify potentially ambiguous or mislabeled cases, enabling efficient re-annotation and improving dataset reliability for classification tasks.
standardized loss aggregationnoisy label detectioncross-validationtask-agnosticfundus dataset
Coarsening Linear Non-Gaussian Causal Models with Cycles
The paper introduces a method for learning low-dimensional causal directed acyclic graphs (DAGs) from high-dimensional linear non-Gaussian (LiNG) models with cycles, relaxing the acyclicity assumption required by prior work. By leveraging observational equivalence classes that differ by cycle reversals, the approach identifies a representative DAG invariant across equivalence class members. Theoretical results establish cubic-time worst-case complexity and sample complexity bounds, with synthetic experiments validating the method. An open-source implementation is provided.
causal abstractionlinear non-gaussiandirected acyclic graphobservational equivalencecycle reversal
Benchmarking Safety Risks of Knowledge-Intensive Reasoning under Malicious Knowledge Editing
EditRisk-Bench introduces a benchmark for evaluating safety risks in knowledge-intensive reasoning under malicious knowledge editing in large language models (LLMs). The framework integrates diverse malicious scenarios (misinformation, bias, safety violations) with multi-level reasoning tasks and editing strategies, measuring attack effectiveness, reasoning correctness, and side effects. Experiments on open-source and closed-source LLMs demonstrate that malicious edits reliably induce incorrect or unsafe reasoning while preserving general capabilities, making risks difficult to detect. Key influencing factors include edit scale, knowledge characteristics, and reasoning complexity. EditRisk-Bench provides an extensible testbed for understanding and mitigating safety risks in LLM knowledge editing.
knowledge editingreasoning correctnesssafety risksmalicious scenarioslarge language models
Scaling Vision Models Does Not Consistently Improve Localisation-Based Explanation Quality
This study demonstrates that scaling vision models does not consistently enhance the quality of post-hoc explanations. The authors evaluate 11 computer vision models from the ResNet, DenseNet, and Vision Transformer families, trained from scratch or pretrained, across three image datasets with ground-truth segmentation masks. Explanations are generated using five post-hoc explainable AI methods and assessed via Relevance Rank Accuracy and Dual-Polarity Precision. Results show that increasing architectural depth and parameter count does not improve explanation quality in most cases, with smaller models often matching or exceeding deeper variants. Pretraining improves predictive performance but not localisation scores, highlighting the need for explicit explainability assessment in safety-sensitive deployments.
post-hoc explanationslocalisation metricsvision transformersdual-polarity precisionsegmentation masks
FormalRewardBench: A Benchmark for Formal Theorem Proving Reward Models
We introduce FormalRewardBench, the first benchmark for evaluating reward models in formal theorem proving using Lean 4, addressing the challenge of sparse credit assignment in RLVR-based neural theorem provers. The benchmark comprises 250 preference pairs, pairing correct proofs with incorrect variants generated via five expert-curated error injection strategies. We evaluate frontier LLMs, judge LLMs, general-purpose LLMs, and specialized theorem proving models, finding that frontier LLMs achieve the highest accuracy (59.8%), while specialized theorem provers perform the worst (24.4%). Results suggest theorem proving ability does not transfer to proof evaluation. The benchmark is publicly released to advance research on reward models in formal mathematics.
formal theorem provingreward modelserror injectionlean 4reinforcement learning
Useful for Exploration, Risky for Precision: Evaluating AI Tools in Academic Research
The study evaluates AI tools in academic research, proposing a benchmarking framework combining human-centered and computer-centered metrics. It assesses Q&A and literature review tools, finding that while they enhance efficiency in exploratory tasks, they lack reliability for precise information extraction. Q&A tools provided generally accurate summaries but low explainable AI (xAI) accuracy, requiring human verification. Literature review tools supported exploratory searches but showed low reproducibility and inconsistent source quality. The findings emphasize the need for explainability features and careful integration of AI tools into research workflows.
benchmarking frameworkexplainable ailiterature reviewq&a toolshuman-centered metrics
Rethinking Constraint Awareness for Efficient State Embedding of Neural Routing Solver
The paper introduces Constraint-Aware Residual Modulation (CARM), a module enhancing Heavy-Encoder-Light-Decoder (HELD) neural routing solvers for Vehicle Routing Problems (VRPs) with complex constraints. CARM adaptively modulates context embeddings using constraint-relevant variables, preserving global observation space during attention computation. Empirical analysis shows CARM improves baseline performance across single-task and multi-task solvers, particularly in scaling to large instances and generalizing to unseen VRP variants. The method addresses the constraint-agnostic limitation of existing neural solvers, offering architectural insights for improved state embedding generation.
constraint-aware residual modulationneural routing solversvehicle routing problemsglobal observation spacestate embedding
Explainability of Recurrent Neural Networks for Enhancing P300-based Brain-Computer Interfaces
The authors introduce the Post-Recurrent Module (PRM), an additional layer integrated into Recurrent Neural Network (RNN) architectures to enhance both performance and explainability in P300-based Brain-Computer Interfaces (BCIs). The PRM enables dual spatio-temporal analysis of EEG signals through global and local explainability techniques, identifying critical brain regions and time intervals while aligning model decisions with neurophysiological descriptions of P300. Experimental results demonstrate a 9% performance improvement over state-of-the-art methods, addressing inter- and intra-subject variability. The framework's ability to identify key spatial and temporal features makes it generalizable to EEG-based tasks such as motor imagery, steady-state visual evoked potentials, and cognitive workload assessment.
post-recurrent modulerecurrent neural networkbrain-computer interfaceexplainabilityp300
MicroWorld: Empowering Multimodal Large Language Models to Bridge the Microscopic Domain Gap with Multimodal Attribute Graph
MicroWorld introduces a multimodal attributed property graph (MAPG) framework to enhance multimodal large language models (MLLMs) in microscopy without domain-specific fine-tuning. It constructs a knowledge graph (111K nodes, 346K edges) from image-caption corpora using scispaCy and Qwen3-VL-Embedding, then retrieves structured knowledge during inference. On MicroVQA, it boosts Qwen3-VL-8B-Instruct by 37.5%, surpassing GPT-5 by 13.0%, and achieves a 6.0% gain on MicroBench, demonstrating improved generalization.
multimodalknowledge graphqwen3-vlmicrovqascispacy
Think as Needed: Geometry-Driven Adaptive Perception for Autonomous Driving
Enhanced HOPE introduces a geometry-driven adaptive perception architecture for autonomous driving that dynamically allocates computation based on scene complexity. The method employs an unsupervised statistical estimator to route LiDAR frames through shallow or deep processing paths, replaces quadratic attention with linear-time subspace clustering for efficient interaction modeling, and incorporates a persistent temporal memory module for occluded object recall. Evaluated on nuScenes and CARLA, it reduces latency by 38% on simple scenes, improves rare-scenario mAP by 2.7 points, and maintains object tracking through >5s occlusions where baselines fail.
adaptive perceptionsubspace clusteringtemporal memorylidar processingocclusion handling
CFSPMNet: Cross-subject Fourier-guided Spatial-Patch Mamba Network for EEG Motor Imagery Decoding in Stroke Patients
CFSPMNet introduces a cross-subject adaptation framework for EEG motor imagery decoding in stroke patients, addressing challenges from pathological neural reorganization. The method combines a Fourier-Reorganized State Mamba Network (FRSM) for latent neural-state modeling with Shared-Private Prototype Matching (SPPM) for improved pseudo-label updating. Evaluated on two stroke MI-EEG datasets (XW-Stroke and 2019-Stroke), CFSPMNet achieves average accuracies of 68.23% and 73.33%, outperforming CNN-, Transformer-, and Mamba-based baselines by 5.63 and 8.25 percentage points, respectively. Neurophysiological analyses validate the roles of Fourier-domain token-state reorganization and calibrated pseudo-label updating.
motor imagery eegcross-subject adaptationfourier-domain reorganizationstate-space mambapseudo-label updating
Arcane: An Assertion Reduction Framework through Semantic Clustering and MCTS-Guided Rule Exploring
Arcane introduces an assertion reduction framework for hardware verification, addressing redundancy in assertion-based verification (ABV) systems. The framework employs a two-tier semantic clustering approach to classify assertions accurately and utilizes Monte Carlo Tree Search (MCTS) to optimize rule-application sequences for assertion reduction. Evaluated on Assertionbench, Arcane reduces assertion counts by up to 76.2% while maintaining formal coverage and mutation-detection capabilities. Simulation studies indicate a 2.6x to 6.1x speedup in simulation time. The framework is publicly available for further research and application.
assertion-based verificationsemantic clusteringmonte carlo tree searchassertion reductionhardware verification
ViSRA: A Video-based Spatial Reasoning Agent for Multi-modal Large Language Models
ViSRA introduces a training-free framework for enhancing spatial reasoning in Multi-modal Large Language Models (MLLMs) by leveraging explicit spatial information from expert models. The approach emphasizes modularity and extensibility, enabling a plug-and-play paradigm without post-training computational costs or manual dataset curation. ViSRA achieves human-aligned and transferable 3D understanding, avoiding task-specific overfitting. Experimental results show consistent improvements across MLLMs, with ViSRA outperforming baselines by up to 15.6% on existing benchmarks and 28.9% on unseen 3D spatial reasoning tasks.
spatial reasoningmulti-modal large language modelstraining-freeplug-and-play3d understanding
HYPERPOSE: Hyperbolic Kinematic Phase-Space Attention for 3D Human Pose Estimation
(No summary returned.)
Retrieve-then-Steer: Online Success Memory for Test-Time Adaptation of Generative VLAs
The paper introduces Retrieve-then-Steer, an online success-memory framework for test-time adaptation of Vision-Language-Action (VLA) models in robotic manipulation. The method stores successful observation-action segments in long-term memory during deployment, retrieves state-relevant action chunks at inference, and filters inconsistent candidates via trajectory-level consistency. Confidence-adaptive prior guidance integrates the elite action prior into an intermediate state of the flow-matching action sampler, adjusting guidance strength based on retrieval confidence. This approach enables lightweight, non-parametric adaptation without parameter updates. Experiments demonstrate improved task success and closed-loop stability, particularly in long-horizon and multi-stage tasks.
vision-language-actiontest-time adaptationflow-matchingtrajectory-level consistencyconfidence-adaptive guidance
PoDAR: Power-Disentangled Audio Representation for Generative Modeling
PoDAR (Power-Disentangled Audio Representation) introduces a framework to improve latent modelability in audio generative models through explicit factor disentanglement. The method employs randomized power augmentation and a latent consistency objective to decouple signal power from invariant semantic content, simplifying the latent space. Applied to Stable Audio 1.0 VAE with an F5-TTS generator, PoDAR accelerates convergence by 2×, enhances speaker similarity by 0.055, and increases UTMOS by 0.22 on LibriSpeech-PC. Additionally, isolating power into dedicated channels extends stable guidance regimes via CFG application to power-invariant content.
latent diffusion modelsfactor disentanglementpower augmentationlatent consistencystable guidance
Active Testing of Large Language Models via Approximate Neyman Allocation
The paper introduces an active testing algorithm for generative tasks in large language models (LLMs), addressing the high costs of evaluation. The method uses semantic entropy from surrogate models to stratify the evaluation pool and applies approximate Neyman allocation based on surrogate signals. Evaluated across multiple language and multimodal benchmarks with various surrogate-target model pairs, the approach achieves up to 28% MSE reduction over uniform sampling and averages 22.9% budget savings, closely tracking Oracle-Neyman performance.
active testingsemantic entropyneyman allocationgenerative taskssurrogate models
Metis: Learning to Jailbreak LLMs via Self-Evolving Metacognitive Policy Optimization
Metis introduces a novel framework for jailbreaking Large Language Models (LLMs) by reformulating the task as inference-time policy optimization within an adversarial Partially Observable Markov Decision Process (POMDP). The method employs a self-evolving metacognitive loop to diagnose target defense logic and refines its policy using structured feedback as semantic gradients, enhancing interpretability through transparent reasoning traces. Evaluations across 10 diverse models show Metis achieves an average Attack Success Rate (ASR) of 89.2%, outperforming traditional baselines, particularly on resilient frontier models (76.0% on O1, 78.0% on GPT-5-chat). Token costs are reduced by an average of 8.2x, highlighting efficiency gains.
jailbreakingpartially observable markov decision processmetacognitive loopattack success ratetoken cost
NCO: A Versatile Plug-in for Handling Negative Constraints in Decoding
The paper introduces NCO, a plug-in decoding strategy for Large Language Models (LLMs) that efficiently handles negative constraints during generation. NCO performs online pattern matching over finite hard constraints and regex patterns, avoiding state explosion while maintaining compatibility with standard inference methods like beam search and sampling. It supports both hard and soft masking for probabilistic suppression of undesirable content. Empirical evaluations demonstrate effectiveness in practical tasks such as PII and profanity suppression. The method reduces computational overhead compared to automaton-based approaches.
constrained decodingnegative constraintsregex constraintssoft maskingbeam search
MAGE: Multi-Agent Self-Evolution with Co-Evolutionary Knowledge Graphs
MAGE introduces a multi-agent framework for self-evolving language models that externalizes self-knowledge into a four-subgraph co-evolutionary knowledge graph. The framework leverages an experience subgraph storing teacher-written corrections and learner-generated success traces, retrieved as task-conditioned guidance for a frozen execution model. Evolution updates the graph and two bandits (task-level search and skill-level routing) from a shared reward stream while maintaining the learner's backbone. Evaluations across nine benchmarks demonstrate strong performance against prompt-based frozen-backbone baselines, with ablations revealing complementary benefits of success traces and corrective memories.
co-evolutionary knowledge graphfrozen execution modeltask-conditioned guidancesearch banditrouting bandit
Not-So-Strange Love: Language Models and Generative Linguistic Theories are More Compatible than They Appear
The article argues that neural language models (LMs) can instantiate formal structure-based linguistic theories, expanding the space of theories testable with LMs. This challenges the framing by Futrell and Mahowald (2025) that LMs primarily support gradient, usage-based theories. By demonstrating compatibility with generative linguistic theories, the work suggests potential reconciliations between usage-based and generative accounts. The argument leverages the theoretical flexibility of LMs to bridge traditionally opposed linguistic frameworks, offering a broader empirical basis for evaluating linguistic theories through computational modeling.
neural language modelslinguistic theoriesgenerative traditionusage-based theoriescomputational modeling
Strategic Exploitation in LLM Agent Markets: A Simulation Framework for E-Commerce Trust
The authors introduce TruthMarketTwin, a simulation framework for studying strategic behavior of LLM agents in e-commerce markets characterized by information asymmetry. The framework models bilateral trade where agents optimize seller profit and buyer utility through decisions on listing, purchasing, rating, and recourse. Experiments reveal that LLM agents autonomously exploit weaknesses in reputation-based governance in traditional markets, while warrant enforcement reduces deception and alters strategic reasoning. This work positions LLM-agent simulation as a tool for analyzing institution-governed autonomous markets.
agent-based modelinginformation asymmetrybilateral tradereputation-based governancewarrant enforcement
Route by State, Recover from Trace: STAR with Failure-Aware Markov Routing for Multi-Agent Spatiotemporal Reasoning
STAR introduces a failure-aware routing framework for multi-agent spatiotemporal reasoning, externalizing inter-agent control via a state-conditioned transition policy. The method employs an agent routing matrix combining expert-specified nominal routes with recovery transitions learned from execution traces, enabling distinct responses to various failure states. Specialists execute through a tool-grounded protocol and share intermediate results via a blackboard. Results demonstrate that training with unsuccessful traces enhances routing policy support on error states, improving performance across three benchmarks and eight backbone LLMs, particularly for queries deviating from nominal paths.
spatiotemporal reasoningmulti-agent systemsfailure-aware routingtransition policyexecution traces
Swarm Skills: A Portable, Self-Evolving Multi-Agent System Specification for Coordination Engineering
The paper introduces Swarm Skills, a portable specification extending the Anthropic Skills standard to enable multi-agent coordination engineering. The method includes roles, workflows, execution bounds, and a self-evolution mechanism, operationalized via an algorithm that autonomously refines coordination protocols based on Effectiveness, Utilization, and Freshness metrics. Results demonstrate zero-adapter cross-agent portability and framework-independent self-evolution using the JiuwenSwarm reference implementation, addressing the bottleneck in multi-agent collaboration.
swarm skillscoordination engineeringself-evolutionanthropic skillsjiuwenswarm
Guided Streaming Stochastic Interpolant Policy
The paper introduces the Streaming Stochastic Interpolant Policy (SSIP), a framework for real-time control that optimally guides generative robot policies via Stochastic Interpolants (SI) derived from Backward Kolmogorov Equation analysis. SSIP unifies deterministic Streaming Flow Policy (SFP) with a modified drift term, enabling sampling from target distributions. Two mechanisms are proposed: training-free Stochastic Trajectory Ensemble Guidance (STEG) for zero-shot adaptation and training-based Conditional Critic Guidance (CCG) for amortized inference. Empirical results show SSIP outperforms chunk-based policies in reactivity and provides physically valid guidance for dynamic environments.
stochastic interpolantsbackward kolmogorov equationstreaming flow policyzero-shot adaptationamortized inference
Rethinking Loss Reweighting for Imbalance Learning as an Inverse Problem: A Neural Collapse Point of View
The paper proposes a novel inverse-problem approach to loss reweighting in long-tailed classification, inspired by Neural Collapse (NC) geometry. By framing class weight inference as an inverse problem targeting equal per-class average loss—consistent with the ideal simplex Equiangular Tight Frame (ETF) terminal state—the method dynamically aligns loss distributions. Empirical evaluations demonstrate reduced loss imbalance coefficients and improved NC metric alignment, outperforming baseline methods across multiple datasets.
neural collapseequiangular tight frameloss reweightinglong-tailed classificationinverse problem
Adaptive Action Chunking via Multi-Chunk Q Value Estimation
We introduce Adaptive Action CHunking (ACH), an offline-to-online RL algorithm that dynamically modulates action chunk length during training and inference. ACH simultaneously estimates Q-values for all candidate chunk lengths in a single forward pass using a Transformer-based architecture, enabling adaptive selection of the optimal chunk length based on the current state. Evaluated on 34 challenging tasks, ACH consistently outperforms fixed-length baselines, demonstrating superior generalization and learning efficiency in complex environments.
adaptive action chunkingoffline-to-online rltransformer-based architectureq-value estimationaction chunking
Personalizing LLMs with Binary Feedback: A Preference-Corrected Optimization Framework
The paper introduces C-BPO, a framework for personalizing LLMs using binary feedback that explicitly models inter-user differences. The method treats target user data as positive signals and other users' data as implicit negatives, then applies a Positive-Unlabeled (PU) learning objective to subtract shared task knowledge ('positive bias') from negative signals. This preserves general helpfulness while aligning outputs with user-specific idiosyncrasies. Experiments across multiple personalization tasks and LLM backbones demonstrate C-BPO's consistent superiority over baseline methods.
llm personalizationbinary feedbackpositive-unlabeled learningpreference calibrationinter-user differences
TimeClaw: A Time-Series AI Agent with Exploratory Execution Learning
TimeClaw introduces an exploratory execution learning framework for time-series analysis, addressing limitations in execution-centric approaches by enabling reusable hierarchical distilled experience. The method employs a four-stage loop (Explore, Compare, Distill, Reinject) combining metric-supervised exploratory execution, task-aware tool dropout, and hierarchical distillation for inference-time reinjection, while keeping the base model frozen. Evaluated on 17 MTBench-aligned tasks spanning finance and weather prediction, TimeClaw demonstrates consistent performance improvements over baselines, highlighting the importance of exploratory experience comparison and reuse in scientific systems.
exploratory executionhierarchical distillationtask-aware tool dropouttime-series analysismetric-supervised learning
Bridging the Cognitive Gap: A Unified Memory Paradigm for 6G Agentic AI-RAN
The article proposes a unified memory paradigm for 6G AI-RAN to address the cognitive gap in current disaggregated architectures. By mapping biological memory hierarchies onto heterogeneous computing fabrics and leveraging coherent interconnects, the method creates a cognitive continuum enabling state sharing across microsecond reflexes, millisecond reasoning, and long-term evolution. This memory-centric approach replaces message passing with zero-copy observability, allowing AI agents to bridge real-time responsiveness and long-horizon context for autonomous 6G networks.
6g ai-ranmemory-centric architecturecognitive continuumcoherent interconnectszero-copy observability
DECO: Sparse Mixture-of-Experts with Dense-Comparable Performance on End-Side Devices
DECO introduces a sparse Mixture-of-Experts (MoE) architecture optimized for end-side deployment, achieving dense-Transformer performance under identical parameter budgets and training tokens. Key innovations include ReLU-based differentiable routing with expert-wise scaling, NormSiLU activation for stable sparsity, and non-gated MLP experts, simplifying MoE design. Experiments show DECO matches dense performance while activating only 20% of experts, with a specialized kernel yielding 3.00× speedup on hardware versus dense inference.
mixture-of-expertssparse routingend-side deploymentactivation functionparameter efficiency
Quantifying Concentration Phenomena of Mean-Field Transformers in the Low-Temperature Regime
The paper analyzes token evolution in deep encoder-only transformers at inference time, modeled as a mean-field continuity equation in the large-token limit. By leveraging convergence analysis techniques from interacting multi-particle systems, the authors prove that the token distribution rapidly concentrates onto the push-forward of the initial distribution under a projection map induced by key, query, and value matrices. They derive a Wasserstein distance scaling as $\sqrt{{\log(β+1)}/β}\exp(Ct)+\exp(-ct)$ with respect to temperature $β^{-1}\to 0$ and inference time $t\geq 0$, showing metastability for moderate times. Numerical experiments confirm the theory and reveal a terminal phase dominated by the value matrix spectrum for finite $β$ and large $t$.
mean-field transformerswasserstein distancetoken distributionmetastabilityprojection map
Dynamic Skill Lifecycle Management for Agentic Reinforcement Learning
The paper introduces SLIM, a framework for dynamic Skill LIfecycle Management in agentic reinforcement learning, addressing the limitations of static skill retention. SLIM treats the active external skill set as a dynamic optimization variable, updated jointly with policy learning. It estimates each skill's marginal contribution via leave-one-skill-out validation and applies three lifecycle operations: retaining high-value skills, retiring negligible ones, and expanding the skill bank when capability gaps are detected. Experiments on ALFWorld and SearchQA demonstrate SLIM's superiority, outperforming baselines by an average of 7.1%. Results show that policy learning and external skill retention can coexist, validating SLIM as a general paradigm for skill-based agentic RL.
agentic reinforcement learningskill lifecycle managementleave-one-skill-out validationdynamic optimizationparametric capacity
Optimal and Scalable MAPF via Multi-Marginal Optimal Transport and Schrödinger Bridges
The paper presents an optimal and scalable approach to anonymous multi-agent path finding (MAPF) by reformulating it as a multi-marginal optimal transport (MMOT) problem with Markovian structure. This reduces the exponentially large MMOT to a polynomial-sized linear program (LP) that yields min-cost, integral transports without spatiotemporal overlaps. For scalability, the authors employ Schrödinger bridges, which reduce the problem to an entropically regularized MMOT solvable via Sinkhorn iterations. Experimental results demonstrate near-optimal integral transports with significantly reduced complexity.
multi-agent path findingmulti-marginal optimal transportschrödinger bridgesentropic regularizationsinkhorn iterations
Equivariant Reinforcement Learning for Clifford Quantum Circuit Synthesis
The paper presents an equivariant reinforcement learning approach for synthesizing optimal Clifford quantum circuits under all-to-all qubit connectivity. The method formulates circuit synthesis as a Markov decision process where an agent learns to decompose symplectic matrix representations into elementary Clifford gates using a curriculum based on random walks from identity. A novel size-agnostic neural architecture preserves equivariance under qubit relabeling, enabling generalization across qubit counts. Evaluations show the agent achieves near-optimal performance (99.2% optimal circuits) on six-qubit benchmarks and outperforms Qiskit's Aaronson-Gottesman synthesizer on circuits up to 30 qubits, reducing two-qubit gate counts in thousand-gate circuits.
equivariant reinforcement learningclifford circuit synthesissymplectic matrixquantum compilationqubit relabeling
Revisiting Policy Gradients for Restricted Policy Classes: Escaping Myopic Local Optima with $k$-step Policy Gradients
The authors propose a $k$-step policy gradient method to address myopic local optima in restricted policy classes, where standard policy gradients fail due to reliance on one-step $Q$-functions. By coupling randomness within a $k$-step time window, the method escapes suboptimal critical points in MDPs. Theoretical guarantees show convergence to solutions exponentially close to the optimal deterministic policy with respect to $k$, achieved in $O(\frac{1}{T})$ iterations using projected gradient descent and mirror descent. The approach avoids distribution mismatch factors, enabling improved performance in state aggregation and partially observable cooperative multi-agent settings.
policy gradientlocal optimamarkov decision processesmirror descentstate aggregation
Beyond Red-Teaming: Formal Guarantees of LLM Guardrail Classifiers
The paper introduces formal verification methods for LLM Guardrail Classifiers by shifting analysis to pre-activation space, where harmful regions are defined as convex shapes enclosing known harmful prompts. Using sigmoid monotonicity, it derives closed-form soundness proofs in O(d) time and proposes two region constructions: SVD-aligned hyper-rectangles for exact certificates and Gaussian Mixture Models for probabilistic certificates. Evaluation on three classifiers (GPT-2, Llama-3.1-8B, BERT) reveals safety holes in all hyper-rectangle configurations and divergent stability, with BERT showing volatile coverage (55-100%) versus robust GPT-2 (90%) and Llama-3.1-8B (80%) performance.
guardrail classifiersformal verificationpre-activation spacegaussian mixture modelssigmoid monotonicity
RubricEM: Meta-RL with Rubric-guided Policy Decomposition beyond Verifiable Rewards
RubricEM introduces a rubric-guided reinforcement learning framework for training deep research agents in long-form report synthesis tasks beyond verifiable rewards. The method combines stagewise policy decomposition with reflection-based meta-policy evolution, using self-generated rubrics to condition planning, evidence gathering, review, and synthesis. It employs Stage-Structured GRPO for credit assignment and trains a shared-backbone reflection meta-policy to distill judged trajectories into reusable guidance. RubricEM-8B demonstrates strong performance across four long-form research benchmarks, outperforming comparable open models and approaching proprietary deep-research systems. Thorough analyses elucidate the framework's key components.
rubric-guided reinforcement learningstagewise policy decompositionreflection meta-policystage-structured grpolong-form research
V4FinBench: Benchmarking Tabular Foundation Models, LLMs, and Standard Methods on Corporate Bankruptcy Prediction
We introduce V4FinBench, a benchmark for corporate bankruptcy prediction comprising over one million company-year records from Visegràd Group economies (2006-2021), featuring 131 financial/non-financial features, six prediction horizons, and a composite distress criterion. The benchmark supports evaluation under severe class imbalance (0.19%-0.36% positive rates). We evaluate standard tabular methods, finetuned TabPFN, and QLoRA-finetuned Llama-3-8B. TabPFN matches or exceeds gradient boosting on $F_1$-score and ROC-AUC at longer horizons, while Llama-3-8B underperforms gradient boosting across all horizons. V4FinBench-finetuned TabPFN improves over vanilla TabPFN on external evaluation, indicating transferable financial-distress patterns. The benchmark is publicly released for further research.
tabular foundation modelsclass imbalancemulti-horizon forecastingqlora-finetuningroc-auc
Neural Weight Norm = Kolmogorov Complexity
The paper establishes that, in fixed-precision regimes, the smallest weight norm of a looped neural network outputting a binary string equals the Kolmogorov complexity of that string, up to a logarithmic factor. This equivalence implies that weight decay induces a prior matching Solomonoff's universal prior, optimal for computable functions, up to a polynomial factor. The proof leverages two tight reductions: encoding universal Turing machine programs into neural weights at unit cost per bit and describing fixed-precision networks via non-zero parameters with logarithmic overhead. The result holds for any weight norm as regularizer, collapsing to non-zero parameter count in fixed precision, but fails for infinite precision networks encoding non-computable functions.
kolmogorov complexityweight decayfixed-precisionuniversal priorneural weight norm
Compute Where it Counts: Self Optimizing Language Models
Self-Optimizing Language Models (SOL) introduce dynamic budget allocation for autoregressive decoding, optimizing computation per token based on difficulty. SOL pairs a frozen LLM with a lightweight policy network that selects efficiency actions—controlling attention sparsity, MLP activation pruning, and quantization bit-width—while preserving base model weights. Trained via group-relative policy optimization on teacher-forced episodes, SOL balances language-model quality against budget constraints. Experiments show SOL outperforms static allocation and random schedules, achieving a superior quality-efficiency Pareto front and improving MMLU accuracy by up to 7.3% over uniform budget strategies.
dynamic budget allocationautoregressive decodingattention sparsityactivation pruningquantization bit-width
Masked Generative Transformer Is What You Need for Image Editing
We propose EditMGT, the first Masked Generative Transformer (MGT)-based framework for image editing, addressing the global denoising limitations of diffusion models. EditMGT employs multi-layer attention consolidation for precise edit localization and region-hold sampling to prevent token flipping in non-target areas. Trained on CrispEdit-2M, a 2M-sample high-resolution (>1024) dataset spanning seven categories, EditMGT achieves state-of-the-art image similarity on multiple benchmarks with only 960M parameters. It delivers 6x faster editing than diffusion-based approaches, demonstrating MGTs as a compelling alternative for localized image editing.
masked generative transformermulti-layer attention consolidationregion-hold samplingcrispedit-2mimage similarity
Conditional anomaly detection methods for patient-management alert systems
The paper introduces instance-based methods for conditional anomaly detection, focusing on identifying anomalous patterns in subsets of attributes conditioned on remaining attributes. The methods utilize distance metrics and metric learning techniques to optimize anomaly detection performance. Empirical evaluation demonstrates their effectiveness on two real-world healthcare datasets: detecting unusual admission decisions for community-acquired pneumonia patients and identifying anomalous orders of HPF4 tests for Heparin-induced thrombocytopenia. The results highlight the utility of instance-based approaches in patient-management alert systems.
conditional anomaly detectioninstance-based methodsdistance metricmetric learningpatient-management alert systems
Transcoda: End-to-End Zero-Shot Optical Music Recognition via Data-Centric Synthetic Training
Transcoda introduces an end-to-end Optical Music Recognition (OMR) system addressing dataset scarcity and encoding non-uniqueness in **kern formats. It employs a synthetic data generation pipeline, **kern normalization to enforce unique representations, and grammar-based decoding for syntactic correctness. The 59M-parameter model, trained in 6 hours on a single GPU, achieves 18.46% OMR-NED on synthetic benchmarks, outperforming billion-parameter baselines like Legato (43.91%), and reduces error rates on historical Polish scans to 63.97% OMR-NED from 80.16% (SMT++).
optical music recognitionsynthetic datakern normalizationgrammar-based decodingomr-ned
Predicting 3D structure by latent posterior sampling
We propose a probabilistic 3D reconstruction method combining Neural Radiance Fields (NeRF) with diffusion models, treating 3D scenes as stochastic latent variables. The approach employs a two-stage training process: first training a reconstruction model with auto-decoded latent representations, then learning a latent prior using a diffusion model. Posterior sampling leverages score-based inference and volumetric rendering likelihoods. Experiments demonstrate accurate 3D structure prediction from diverse inputs including single-view, multi-view, noisy images, sparse pixels, and sparse depth data, effectively modeling task-specific uncertainty.
neural radiance fieldsdiffusion modelsstochastic latent variablesvolumetric renderingposterior sampling
NoRIN: Backbone-Adaptive Reversible Normalization for Time-Series Forecasting
(No summary returned.)
Benchmarking Sensor-Fault Robustness in Forecasting
SensorFault-Bench introduces a standardized protocol for evaluating sensor-fault robustness in cyber-physical system forecasting models, addressing noisy, biased, missing, or misaligned sensor data. The benchmark includes a severity model, disjoint fault-transfer splits, and metrics such as worst-scenario degradation, clean MSE, and worst-scenario fault-time MSE. Evaluations across four datasets and eight scenarios reveal that models optimized for clean MSE often degrade sharply under faults, with Chronos-2 showing significant degradation in worst-case scenarios. Robustness-improvement methods like adversarial training and fault augmentation selectively reduce degradation based on fault type. The benchmark provides open-source tools for reproducibility and extension.
sensor-fault robustnesscyber-physical systemsmean squared erroradversarial trainingfault augmentation
On periodic distributed representations using Fourier embeddings
The paper introduces periodic distributed representations using Fourier embeddings to address challenges in processing scalar angular measures. By employing real-valued, periodic embeddings in high-dimensional space, the method mitigates issues with distinguishing nearby angles when absolute differences exceed pi. The approach formalizes Dirichlet and periodic Gaussian kernels using Spatial Semantic Pointers, enabling control over dot product similarities and diverse kernel shapes. Results demonstrate the neural plausibility and versatility of these representations for modeling physical and perceptual phenomena.
fourier embeddingsperiodic representationsdirichlet kernelsspatial semantic pointersgaussian kernels
Likelihood scoring for continuations of mathematical text: a self-supervised benchmark with tests for shortcut vulnerabilities
The authors introduce an automated benchmark for evaluating mathematical text continuation prediction, using hidden text from technical papers as test cases. The method involves generating forecasts (Z) from visible context (X) and comparing next-token probabilities for hidden continuations (Y) with and without Z, using controls to detect shortcut vulnerabilities. Testing on 1363 equation continuations from physics/mathematics papers, GPT-5.5, Opus 4.7, and GPT-5.4 nano show improved likelihood scores over controls, with varying performance across models and task settings. The benchmark demonstrates utility for model comparison and shortcut vulnerability assessment prior to reinforcement learning.
mathematical text predictionlikelihood scoringshortcut vulnerabilitiesself-supervised benchmarkequation-suffix prediction
Mistake-Bounded Language Generation
(No summary returned.)
LLMs for Secure Hardware Design and Related Problems: Opportunities and Challenges
The article presents a comprehensive review of Large Language Models (LLMs) in Electronic Design Automation (EDA) and hardware security, highlighting both advancements and vulnerabilities. It examines LLM-driven methodologies such as reasoning-driven synthesis, multi-agent vulnerability extraction, and adversarial machine learning evasion. The review emphasizes critical countermeasures, including dynamic benchmarking to mitigate data memorization and aggressive red-teaming for robust security assessment. Key findings underscore the potential of LLMs in automating Register Transfer Level (RTL) code generation and bridging semantic gaps, while also addressing significant security risks. The synthesis aims to guide future research toward secure, trustworthy, and autonomous hardware design ecosystems.
electronic design automationregister transfer leveladversarial machine learningdynamic benchmarkingmulti-agent vulnerability extraction
Muown: Row-Norm Control for Muon Optimization
(No summary returned.)
Factual recall in linear associative memories: sharp asymptotics and mechanistic insights
The authors precisely characterize the storage capacity of linear associative memories for factual recall, establishing a baseline for understanding neural network memory limits. They analyze a minimal setting where p input embeddings in ℝ^d are mapped to d-dimensional targets via a single layer, requiring strict separation between mapped inputs and targets. Using statistical physics tools and a decoupled model with independent competing outputs, they demonstrate equivalence to the original model and derive an optimal storage capacity of p_c log p_c / d^2 = 1/2. The analysis reveals mechanistic insights into how optimal solutions outperform naïve Hebbian learning by raising correct scores above extreme-value thresholds.
linear associative memorystorage capacitystatistical physicshebbian learningextreme-value threshold
ConQuR: Corner Aligned Activation Quantization via Optimized Rotations for LLMs
The paper introduces ConQuR, a post-training rotation calibration method for efficient activation quantization in large language models (LLMs). By optimizing orthogonal rotations to align normalized activations with hypercube corners, the method redistributes activation energy evenly across dimensions via closed-form solutions to the orthogonal Procrustes problem. An online calibration procedure eliminates storage overhead while adapting to quantized distributions. Evaluations on Llama-2/3 models (3B-70B parameters) demonstrate competitive perplexity and reasoning performance without end-to-end training or activation storage requirements.
activation quantizationorthogonal procrustespost-training calibrationllama modelshypercube alignment
Fixed-Point Neural Optimal Transport without Implicit Differentiation
The authors propose a neural optimal transport method that avoids adversarial min-max optimization and multi-network architectures by parameterizing a single potential in the Kantorovich dual. The c-transform is reformulated as a proximal fixed-point problem, enforcing dual feasibility through proximal optimality conditions without implicit differentiation. This single-network framework efficiently computes forward and backward transport maps, including class-conditional extensions. Experiments on high-dimensional Gaussians, physical datasets, and image translation demonstrate improved accuracy, stability, and computational efficiency.
optimal transportkantorovich dualproximal fixed-pointimplicit differentiationtransport maps
Elucidating Representation Degradation Problem in Diffusion Model Training
The paper introduces Elucidated Representation Diffusion (ERD), a plug-and-play framework addressing Representation Degradation in diffusion model training. ERD dynamically reallocates optimization effort based on effective recoverability, mitigating structural distortion caused by increasing noise levels. The method leverages Neural Tangent Kernel (NTK) spectral analysis to stabilize representation learning without external supervision. Empirical results demonstrate that ERD accelerates convergence and enhances generation quality across various diffusion backbones. The framework effectively resolves optimization bottlenecks associated with mismatched target recoverability and low-rank behavior.
representation degradationdiffusion modelsneural tangent kerneleffective recoverabilityoptimization bottleneck
MASS-DPO: Multi-negative Active Sample Selection for Direct Policy Optimization
MASS-DPO introduces multi-negative active sample selection for Direct Policy Optimization (DPO), addressing redundancy in gradient updates from large negative pools under the Plackett-Luce model. The method employs a Fisher-information objective to select compact, informative negative subsets per prompt, optimizing a log-determinant criterion that prioritizes complementary gradient directions. Evaluated across four benchmarks (recommendation and QA tasks) and three model families, MASS-DPO matches or exceeds baseline accuracy, improves Recall/NDCG and optimization dynamics, and achieves stronger alignment with fewer negatives.
direct policy optimizationplackett-luce modelfisher-informationmulti-negative optimizationgradient redundancy
Rebellious Student: Reversing Teacher Signals for Reasoning Exploration with Self-Distilled RLVR
The paper introduces RLRT (RLVR with Reversed Teacher), a novel method that reverses self-distillation signals to reinforce a student LLM's successful reasoning paths when they diverge from teacher predictions. Building on GRPO, RLRT treats these divergent tokens as valuable exploration grounded in the student's own success, rather than suppressing them. Evaluated across Qwen3 base, instruction-tuned, and thinking-tuned checkpoints, RLRT outperforms both standard self-distillation and exploration-based baselines, demonstrating information asymmetry as a principled design axis for RLVR.
self-distillationrlvrexplorationinformation asymmetryreasoning paths
Locking Pretrained Weights via Deep Low-Rank Residual Distillation
The paper introduces DLR-Lock, a defense mechanism against unauthorized fine-tuning of pretrained language models by exploiting inference-training asymmetry. The method replaces each MLP with a deep low-rank residual network (DLR-Net) of comparable parameter count, increasing activation memory linearly with depth during backpropagation. DLR-Nets are trained via module-wise distillation, creating architectural mismatches that complicate optimization and disproportionately increase backward pass overhead. Experiments on LLMs demonstrate robustness against adaptive attackers while preserving model capabilities.
dlr-locklow-rank residualmodule-wise distillationinference-training asymmetryunauthorized fine-tuning
On the global convergence of gradient descent for wide shallow models with bounded nonlinearities
The paper extends global convergence guarantees for gradient descent to wide shallow neural networks with bounded nonlinearities, including multi-head attention layers and two-layer sigmoid networks with vector outputs. Building on Chizat and Bach (2018), the authors prove that all non-global minimizers are unstable under gradient descent dynamics, ensuring convergence to global minima when parameters are initialized with full support (e.g., Gaussian) in the infinite-width limit. Key technical contributions include completing the escaping active set construction for bounded nonlinearities and extending it to vector outputs. Results also establish well-posedness and discretization stability for mean field dynamics with sub-Gaussian initializations.
gradient descentglobal convergencewide neural networksbounded nonlinearitiesmean field dynamics
DynaMiCS: Fine-tuning LLMs with Performance Constraints using Dynamic Mixtures
DynaMiCS introduces a dynamic mixture optimizer for multi-domain fine-tuning of large language models (LLMs), addressing the challenge of improving target-domain performance while preserving capabilities in constrained domains. The method casts fine-tuning as a constrained optimization problem, using short domain-specific probing runs to estimate a slope matrix of local cross-domain effects. These estimates inform mixture weights computed via optimization over the probability simplex, ensuring constrained-domain losses remain below reference levels. DynaMiCS outperforms fixed-mixture baselines in target-domain improvements and constraint satisfaction, achieving this at lower computational cost without requiring reference models, per-example scoring, or manual weight tuning.
multi-domain fine-tuningconstrained optimizationslope matrixprobability simplexdynamic mixture optimizer
Reinforce Adjoint Matching: Scaling RL Post-Training of Diffusion and Flow-Matching Models
The paper introduces Reinforce Adjoint Matching (RAM), a reinforcement learning (RL) post-training method for diffusion and flow-matching models that preserves the supervised regression structure of pretraining. RAM leverages KL-regularized reward maximization, combining the adjoint-matching optimality condition with a REINFORCE identity to derive a consistency loss that corrects the pretraining target based on reward. This approach eliminates the need for SDE rollouts, backward adjoint sweeps, or reward gradients, maintaining scalability. Evaluated on Stable Diffusion 3.5M, RAM achieves superior rewards in composability, text rendering, and human preference, matching Flow-GRPO's peak reward in up to 50× fewer training steps.
reinforcement learningdiffusion modelskl-regularizationadjoint-matchingflow-matching
AdaPaD: Adaptive Parallel Deflation for PEFT with Self-Correcting Rank Discovery
AdaPaD introduces adaptive parallel deflation for parameter-efficient fine-tuning (PEFT) with self-correcting rank discovery, enabling simultaneous training of rank-1 components via deflation targets updated from predecessors' latest estimates. The method incorporates advance learning and dynamic rank discovery, making rank distribution an output rather than an input. Theoretical analysis shows exponential error decay post-warm-up, with a generalization bound separating algorithmic and statistical terms. Empirically, AdaPaD matches adaptive-rank LoRA baselines on GLUE with DeBERTaV3-base and fixed-rank LoRA on Qwen3-0.6B SQuAD/SQuAD v2, achieving a 30.7% smaller average adapter size.
adaptive parallel deflationparameter-efficient fine-tuningself-correcting rank discoverydeflation targetsdynamic rank discovery
XQCfD: Accelerating Fast Actor-Critic Algorithms with Prior Data and Prior Policies
XQCfD introduces a novel actor-critic framework that enhances sample efficiency in robotic reinforcement learning by leveraging prior data and pretrained policies. The method employs stationary policy architectures and augmented replay buffers to prevent rapid unlearning of initial policies, outperforming standard architectures through higher entropy predictions. Evaluated on Adroit, Robomimic, and MimicGen benchmarks, XQCfD achieves state-of-the-art performance in sparse-reward manipulation tasks with minimal update-to-data ratio and no ensemble networks.
actor-criticsample efficiencypretrained policiesstationary architecturesparse rewards
Kernel-Gradient Drifting Models
The authors propose kernel-gradient drifting, a one-step generative modeling framework that generalizes drifting models by replacing fixed Euclidean displacement directions with kernel-induced directions. This reformulation reveals a score-based structure for general kernels, where the drift represents the score difference between kernel-smoothed data and model distributions, ensuring identifiability for characteristic kernels and providing a smoothed-KL descent interpretation. The method extends naturally to Riemannian manifolds and discrete data via Fisher-Rao geometry. Empirical results demonstrate state-of-the-art one-step generation performance on spherical geospatial data, promoter DNA, and molecule generation tasks without distillation.
kernel-gradient driftingscore matchingfisher-rao geometryriemannian manifoldssmoothed-kl descent
On Improving Graph Neural Networks for QSAR by Pre-training on Extended-Connectivity Fingerprints
The authors propose a pre-training strategy to enhance Graph Neural Networks (GNNs) for Quantitative Structure-Activity Relationship (QSAR) tasks by predicting Extended-Connectivity Fingerprints (ECFP). They validate the approach using statistical tests and out-of-distribution (OOD) splits across six Biogen benchmarks. Results show statistically significant improvements in standard metrics for five benchmarks, though pre-trained GNNs underperformed in OOD settings for complex endpoints like binding affinity prediction. The study also examines substructure-level data leakage during pre-training, concluding that ECFP-based pre-training can improve OOD performance on diverse QSAR tasks despite some limitations.
graph neural networksquantitative structure-activity relationshipextended-connectivity fingerprintsout-of-distributionpre-training
Heteroscedastic Diffusion for Multi-Agent Trajectory Modeling
The authors propose U2Diffine, a unified diffusion model for multi-agent trajectory modeling that performs trajectory completion while providing state-wise heteroscedastic uncertainty estimates. The method augments the standard denoising loss with the negative log-likelihood of predicted noise and propagates latent space uncertainty to real state space via first-order Taylor approximation. A faster variant, U2Diff, avoids gradient computation during sampling. The approach includes a Rank Neural Network for error probability estimation per generated mode. Evaluations on four sports datasets (NBA, Basketball-U, Football-U, Soccer-U) show superior performance in trajectory completion and forecasting compared to state-of-the-art methods.
heteroscedastic uncertaintydiffusion modeltrajectory completionmulti-agent systemsrank neural network
What should post-training optimize? A test-time scaling law perspective
The authors introduce Tail-Extrapolated estimators (TEA and Prefix-TEA) to address the mismatch between post-training objectives and test-time deployment in large language models, where training budgets (m rollouts) are significantly smaller than deployment budgets (N rollouts). By extrapolating upper-tail statistics from small rollout groups, these estimators approximate the policy gradient of the best-of-N objective. Experiments on instruction-following tasks demonstrate improved best-of-N performance across diverse language models, reward models, and datasets under varying budget constraints.
tail-extrapolated estimatorspolicy gradientbest-of-nrollout budgetupper-tail statistics
Price of Quality: Sufficient Conditions for Sparse Recovery using Mixed-Quality Data
The paper establishes sufficient conditions for sparse recovery using mixed-quality data, where observations combine high-quality (low-variance) and low-quality (high-variance) measurements. It introduces the 'Price of Quality' concept, quantifying the linear trade-off between sample types: in the agnostic setting (decoder unaware of quality), one high-quality sample is never worth more than two low-quality samples, while the informed setting (decoder aware of variances) allows unbounded trade-offs. Algorithmically, LASSO analysis shows recovery thresholds match homogeneous-noise cases, depending only on average noise, demonstrating computational robustness to data heterogeneity. Results reveal a fundamental divergence between information-theoretic and algorithmic adaptation to data quality.
sparse recoverymixed-quality dataprice of qualitylassoheterogeneous noise
RelFlexformer: Efficient Attention 3D-Transformers for Integrable Relative Positional Encodings
RelFlexformer introduces a novel class of efficient 3D-Transformers utilizing universal 3D Relative Positional Encoding (RPE) methods defined by arbitrary integrable modulation functions. The method leverages Non-Uniform Fourier Transform (NU-FFT) theory to generalize existing RPE-attention techniques from structured token grids to unstructured, heterogeneous scenarios, enabling application to point clouds. Attention computation achieves O(L log L) complexity for L-length sequences. Empirical evaluations on diverse 3D datasets demonstrate quality improvements from NU-FFT-driven attention modulation in RelFlexformers.
relative positional encodingnon-uniform fourier transform3d-transformersattention mechanismspoint clouds
DANCE: Detect and Classify Events in EEG
The paper introduces DANCE, a deep learning pipeline for joint detection and classification of neural events in raw, unaligned EEG signals, framing the task as a set-prediction problem. The method eliminates reliance on known event onsets required by conventional window-based approaches. Evaluated across ten diverse datasets spanning cognitive, clinical, and BCI tasks, DANCE outperforms existing methods, achieving state-of-the-art in seizure monitoring and matching onset-informed model accuracy for BCI tasks. This advances end-to-end asynchronous neural decoding capabilities.
eeg decodingset-predictionevent detectionneural recordingsasynchronous decoding
The finite expression method for turbulent dynamics with high-order moment recovery
The paper introduces a two-stage framework combining symbolic regression (Finite Expression Method) and generative models to identify turbulent dynamical systems' governing equations and predict higher-order statistical moments. Stage I uses FEX to derive closed-form deterministic dynamics without predefined libraries, while Stage II employs generative models to correct stochastic residuals. Theoretical analysis confirms estimator consistency and quantifies error bounds. Numerical experiments on stochastic triad models demonstrate accurate recovery of interaction terms and forcing expressions, with successful prediction of moments up to order five.
symbolic regressionfinite expression methodstochastic residualshigher-order momentsturbulent dynamics
Scalable Mamba-Based Message-Passing Neural Decoder for Error-Correcting Codes
The Mamba message-passing decoder (MMPD) introduces a scalable attention-free neural decoder for binary linear codes, addressing the quadratic memory and computational limitations of attention-based approaches. MMPD combines local pairwise aggregation along Tanner-graph variable-check edges with bidirectional Mamba state-space blocks for efficient long-range information propagation. Experiments on the (1056, 880) LDPC code demonstrate a 0.45 dB gain over the CrossMPT decoder at a target bit error rate, while reducing memory consumption by 1.5×, with greater reductions for longer codes.
mamba message-passing decoderbinary linear codestanner-graphstate-space blocksneural decoder
Exact Unlearning from Proxies Induces Closeness Guarantees on Approximate Unlearning
The paper introduces a novel paradigm linking machine unlearning to data distribution structure rather than neural network parameter updates. It demonstrates that precise inference of these distributions enables distillation of exact unlearning signals. Theoretical bounds on Kullback-Leibler divergence between ideal retrained models and unlearned models are established under verifiable admissibility criteria. Experimental validation across three forgetting scenarios shows the method achieves the closest classifier to the ideal retrained model compared to competing approaches.
machine unlearningdata distributionskullback-leibler divergenceneural networksadmissibility criterion
Not Blind but Silenced: Rebalancing Vision and Language via Adversarial Counter-Commonsense Equilibrium
The authors propose Adversarial Counter-Commonsense Equilibrium (ACE), a training-free framework addressing attention imbalance in multimodal large language models (MLLMs) during decoding. ACE perturbs visual context using counter-commonsense patches, leveraging the stability of authentic visual features versus hallucinated ones under perturbation. This dynamic game decoding strategy suppresses perturbation-sensitive priors while compensating for stable visual signals, restoring equilibrium between linguistic priors and visual information. Extensive experiments show ACE enhances model trustworthiness with minimal inference overhead, functioning as an effective plug-and-play solution.
multimodal large language modelsadversarial counter-commonsense equilibriumdynamic game decodingvisual-language imbalanceperturbation-sensitive priors
Compander-Aligned Query Geometry for Quantized Zeroth-Order Optimization
The paper introduces CAQ-ZO (Compander-Aligned Queries for Zeroth-Order Optimization), a method addressing geometric misalignment in quantized zeroth-order optimization. By modeling scalar nonuniform quantization as $Q = φ^{-1} \circ U \circ φ$ and forming Rademacher stencils in transformed space $z = φ(x)$, CAQ-ZO eliminates query-time residuals that plague generic off-grid queries. Theoretical analysis decomposes estimator residuals and proves stationarity bounds, showing CAQ-ZO achieves zero residual error while generic methods retain $Δ^2/μ^2$ residuals. Experiments on NF4-quantized Qwen/Llama models demonstrate improved fine-tuning performance under identical quantization and evaluation budgets.
zeroth-order optimizationnonuniform quantizationcompandingrademacher stencilsstationarity bounds
Natural Policy Gradient as Doubly Smoothed Policy Iteration: A Bellman-Operator Framework
The authors establish that natural policy gradient (NPG) is equivalent to a doubly smoothed policy iteration (DSPI) framework, unifying policy iteration, dual-averaged policy iteration, and NPG under a Bellman-operator formalism. DSPI computes policies via regularized greedy steps on averaged past Q-functions, leveraging monotonicity and contraction properties of smoothed Bellman operators. They prove distribution-free geometric convergence for DSPI, yielding an iteration complexity of O((1-γ)^(-1)log((1-γ)^(-1)ε^(-1))) for ε-optimal policies in standard NPG and policy dual averaging. The framework extends to discounted MDPs with linear function approximation and stochastic shortest path problems.
natural policy gradientbellman operatorpolicy iterationq-functionsgeometric convergence
A Spectral Framework for Closed-Form Relative Density Estimation
The authors propose a closed-form spectral framework for relative log-density estimation in linearly parameterized probabilistic models, including unnormalized and conditional models. The method represents the Kullback-Leibler divergence as an integral of weighted chi-squared divergences, converting KL estimation into a family of least-squares problems. A spectral formula based on first- and second-order feature moments yields closed-form estimators for divergences and log-density potentials. The framework generalizes to f-divergences and integrates with kernelization or neural network feature learning. Theoretical convergence guarantees are provided, and empirical comparisons with optimization-based variational formulations demonstrate competitive performance on synthetic data.
spectral frameworkkullback-leibler divergencechi-squared divergencesleast-squares problemsf-divergences
Why Zeroth-Order Adaptation May Forget Less: A Randomized Shaping Theory
The paper introduces a theoretical framework explaining why zeroth-order (ZO) adaptation may outperform first-order (FO) methods in continual learning by reducing forgetting. Through a local randomized gradient-shaping analysis, it demonstrates that ZO methods preserve isotropic retention curvature while contracting anisotropic components, leading to a quadratic forgetting gap favoring ZO when FO directions exhibit above-average retention curvature. The proposed RISE algorithm applies ZO shaping to exact FO gradients within parameter blocks, balancing stability and plasticity. Blockwise analysis separates mean-step damage from random exposure, identifying conditions where this transfer is effective.
zeroth-order adaptationcontinual learningrandomized shapingretention curvatureblockwise analysis
BCJR-QAT: A Differentiable Relaxation of Trellis-Coded Weight Quantization
BCJR-QAT introduces a differentiable relaxation of trellis-coded weight quantization, replacing the non-differentiable Viterbi argmax with the BCJR forward-backward sum-product algorithm at temperature T. This method produces a soft codeword equivalent to the Boltzmann expectation over trellis paths, recovering the hard QTIP code as T approaches 0. The authors contribute a fused Triton kernel enabling BCJR on a single consumer GPU, a drift-budget theory for escaping the QTIP-PTQ Voronoi basin, and empirical results on Llama-3.2-1B showing a -0.084 PPL improvement on WikiText-2 under end-to-end forward-KL distillation.
trellis-coded quantizationbcjr algorithmviterbi argmaxboltzmann expectationforward-kl distillation
A Random-Matrix Criterion for Initializing Gated Recurrent Neural Networks
The authors derive a criterion for estimating the critical weight variance $g_c$ in recurrent neural networks, enabling initialization at an effective critical point that separates ordered and chaotic phases. Their method applies to a broad class of recurrent architectures, including gated-RNN reservoirs, and is validated through chaotic forecasting tasks where it tracks peak performance. The criterion is proposed as a design principle for future initialization schemes, addressing the importance of proper weight initialization in deep learning and reservoir computing.
weight initializationrecurrent neural networkscritical pointchaotic forecastingreservoir computing
A Single-Layer Model Can Do Language Modeling
We propose Grounded Prediction Networks (GPN), a single-layer recurrent architecture for language modeling that revisits one state vector per step through a shared matrix memory and FFN, contrasting with deep stacked-layer models. GPN+M achieves FineWeb-Edu perplexity of 18.06 at 130M parameters, within 13% of a 12-layer Transformer++ (16.05) and 18% of a 10-layer Gated DeltaNet (15.34); a 2-layer variant reduces the gap to 6%/11%. Analysis reveals a persistent default-token direction, content-bearing horizon of tens of tokens, and spontaneous memory head splitting into fast/slow retention pools.
grounded prediction networksfineweb-eduperplexityrecurrent architecturematrix memory
Composing diffusion priors with explicit physical context via generative Gibbs sampling
We introduce Generative Gibbs for Physics-Aware Sampling (GG-PA), a training-free framework that combines pretrained diffusion priors with explicit physical context via Gibbs sampling in an augmented state space. The method derives an asymptotically exact Gibbs sampler for joint target distributions, proven exact for quadratic interactions at finite diffusion times, and employs replica exchange over diffusion time to accelerate mixing. Experiments on double-well systems, φ4 lattice models, and atomistic peptide systems demonstrate GG-PA's ability to recover context-induced distribution shifts and emergent collective behavior without retraining. These results validate GG-PA as a practical approach for integrating generative priors with physical context.
diffusion priorsgibbs samplingreplica exchangephysical contextjoint target distribution
Hierarchical End-to-End Taylor Bounds for Complete Neural Network Verification
We present HiTaB, a novel neural network verification framework that systematically exploits second-order smoothness through Hessian matrices and their Lipschitz constants for tighter reachability analysis. The method introduces a hierarchical approach combining zeroth-, first-, and second-order bounds, with precise conditions for when higher-order approximations yield provable improvements. Key innovations include a compositional procedure for efficiently bounding curvature Lipschitz constants via layerwise propagation and extensions to ℓ₂- and ℓ∞-constrained input sets. Results demonstrate tighter safety certificates compared to state-of-the-art methods, marking the first practical framework leveraging Lipschitz continuity of curvature for smooth neural network verification.
reachability analysishessian matrixlipschitz continuityneural network verificationcurvature bounds
MulTaBench: Benchmarking Multimodal Tabular Learning with Text and Image
The paper introduces MulTaBench, a benchmark of 40 datasets (20 image-tabular, 20 text-tabular) designed to evaluate multimodal tabular learning where modalities provide complementary predictive signals. It addresses limitations in existing benchmarks by focusing on tasks requiring target-aware representations, where generic frozen embeddings lose critical information. Experiments show consistent performance gains from tuning embeddings across modalities, tabular learners, encoder scales, and embedding dimensions. MulTaBench represents the largest image-tabular benchmarking effort to date, spanning healthcare and e-commerce domains, and facilitates research into joint modeling architectures for multimodal tabular foundation models.
multimodal learningtabular datatarget-aware representationsfoundation modelsbenchmarking
Exact Fixed-Point Constraints in Neural-ODEs with Provable Universality
The paper introduces a method for constraining Neural-ODEs to exactly preserve prescribed fixed-points while maintaining universal approximation capabilities. The technique explicitly enforces zero-velocity conditions at finite collections of points in the state space through gradient-based training constraints, without compromising expressivity. Theoretical results prove universality under arbitrary local velocity constraints, and computational methods are provided for practical implementation. Experiments demonstrate the approach on two physical models, validating both theoretical guarantees and empirical performance.
neural-odesfixed-point constraintsuniversal approximationvelocity fieldgradient-based training
Reconfigurable Computing Challenge: Real-Time Graph Neural Networks for Online Event Selection in Big Science
This work presents an end-to-end demonstrator for real-time deployment of a dynamic Graph Neural Network (GNN) on the AMD Versal VCK190 platform, targeting the Belle II electromagnetic calorimeter hardware trigger. The method employs a Python-based semi-automated design flow encompassing operator fusion, partitioning, mapping, spatial parallelization, and kernel-level optimization, leveraging both FPGA fabric and AI Engine tiles. The implementation achieves a throughput of 2.94 million events per second with an end-to-end latency of 7.15 microseconds, representing a 53% throughput improvement over an FPGA-only baseline while reducing DSP utilization from 99% to 19%. Real-time monitoring is enabled via an interactive visualization pipeline.
graph neural networksfpgaai enginelatencythroughput
Amortizing Causal Sensitivity Analysis via Prior Data-Fitted Networks
The authors introduce an amortized approach to causal sensitivity analysis using prior-data fitted networks, enabling in-context learning for bounding causal effects under unobserved confounding. The method employs Lagrangian scalarization to generate training labels by balancing causal effect optimization and sensitivity model violation, applicable across generalized treatment sensitivity models. Under convexity and linearity conditions, the approach recovers the full Pareto frontier. Empirical evaluations demonstrate test-time computation speeds orders of magnitude faster than per-instance methods, establishing the first foundation model for in-context learning in causal sensitivity analysis.
causal sensitivity analysisamortized inferencelagrangian scalarizationpareto frontierprior-data fitted networks
Controllability in preference-conditioned multi-objective reinforcement learning
The article introduces controllability as a critical metric for evaluating preference-conditioned multi-objective reinforcement learning (MORL) agents, addressing the limitation of standard MORL metrics in assessing whether preference changes reliably influence agent behavior. The authors argue that existing metrics fail to capture this property, potentially leading to agents that score well but remain insensitive to preference inputs. They propose a complementary metric specifically designed to measure controllability, aiming to ensure the symbolic interface between user intent and agent behavior remains intact. The results advocate for revising evaluation protocols to better consolidate advances in preference adaptation for complex MORL problems.
controllabilitymulti-objective reinforcement learningpreference-conditionedevaluation protocolssymbolic interface
Online Sharp-Calibrated Bayesian Optimization
Online Sharp-Calibrated Bayesian Optimization (OSCBO) introduces a Bayesian optimization algorithm that adaptively balances Gaussian process sharpness and calibration by framing hyperparameter selection as a constrained online-learning problem. OSCBO addresses the challenge of miscalibrated uncertainty arising from online refitting of GP kernel hyperparameters on non-i.i.d. data, preserving sublinear regret bounds through theoretical guarantees of the underlying online learning algorithm. Empirical evaluations demonstrate OSCBO's competitive performance on synthetic and real-world benchmarks, achieving strong final simple regret while maintaining robust cumulative-regret behavior.
bayesian optimizationgaussian processonline learningregret boundshyperparameter selection
Affine Tracing: A New Paradigm for Probabilistic Linear Solvers
The paper introduces affine tracing, a framework for automatically constructing probabilistic iterative methods (PIMs) from standard iterative linear solvers. It demonstrates that Bayesian probabilistic linear solvers are a special case of non-stationary affine PIMs and proves their calibration. The method uses symbolic tracers to build affine computational graphs, enabling algebraic simplifications via equality saturation. Applied to Gaussian process approximation, the framework generates a probabilistic multigrid solver, showing practical utility in uncertainty quantification for linear systems.
affine tracingprobabilistic iterative methodsbayesian linear solversgaussian process approximationequality saturation
EnergyLens: Interpretable Closed-Form Energy Models for Multimodal LLM Inference Serving
EnergyLens introduces an interpretable closed-form energy model for optimizing multimodal LLM inference serving across heterogeneous accelerators. The method employs symbolic regression over profiling data to derive a twelve-parameter model that decouples tensor and pipeline parallelism contributions and separates prefill from decode energy. Requiring only 50 profiling measurements, EnergyLens achieves 88.2% Top-1 configuration selection accuracy, outperforming prior analytical baselines by 27.3 percentage points and matching ensemble ML methods with 10x fewer samples. The model reliably extrapolates to unseen batch sizes and hardware platforms without structural modifications, offering a practical tool for energy-optimal LLM deployment.
symbolic regressionmultimodal llmenergy optimizationheterogeneous acceleratorsinterpretable models
It's All Connected: Topology-Aware Structural Graph Encoding Improves Performance on Polymer Prediction
We propose a topology-aware graph construction method for polymer property prediction, addressing limitations of repeat-unit-only representations by sampling representative chains from molecular mass distributions and constructing large graphs encoding chain-scale topology. Our approach combines rich chemical descriptors with self-supervised pretraining of GNN encoders on 100,000 unlabeled PSMILES strings. On a dataset of 381 polymers, the method achieves 24.76 K RMSE (±3.30 K) with pretraining, a 5.1% improvement over the pretrained repeat-unit baseline (26.08 K ±4.20 K, p < 0.001). Ablation studies confirm the necessity of both chemical features and pretraining. Results are consistent across GINE and GATv2 architectures.
graph neural networkspolymer predictionself-supervised learningmolecular mass distributionmasked graph modeling
PhysEDA: Physics-Aware Learning Framework for Efficient EDA With Manhattan Distance Decay
PhysEDA introduces a physics-aware learning framework for electronic design automation (EDA) that integrates Manhattan distance decay as an inductive bias. The framework comprises Physics-Structured Linear Attention (PSLA), which reduces attention complexity from quadratic to linear by incorporating separable Manhattan decay, and Potential-Based Reward Shaping (PBRS), which provides dense reward signals in reinforcement learning while preserving optimal policies. Evaluated on decoupling-capacitor placement, macro placement, and IR-drop prediction, PhysEDA achieves a 56.8% improvement in zero-shot cross-scale transfer, a 14x inference speedup, and 98.5% memory savings on 100x100 grids, with PBRS contributing an additional 10.8% improvement in sparse-reward DPP.
electronic design automationmanhattan distancelinear attentionreward shapinginductive bias
Higher Resolution, Better Generalization: Unlocking Visual Scaling in Deep Reinforcement Learning
This work demonstrates that higher-resolution visual observations significantly enhance performance and generalization in pixel-based deep reinforcement learning, challenging the convention of aggressive input downsampling. The authors identify that the Impala encoder's quadratic parameter growth with resolution limits its effectiveness, proposing Impoola with global average pooling to decouple parameter count from resolution. Impoola achieves a 28% performance gain over Impala at optimal conditions, particularly excelling in environments requiring precise perception of small or distant objects. Gradient saliency analysis reveals that higher resolutions enable more spatially localized visual attention. The Procgen-HD benchmark is released to support further research on resolution scaling in deep RL.
deep reinforcement learningvisual scalingimpala encoderglobal average poolinggradient saliency analysis
ConfoundingSHAP: Quantifying confounding strength in causal inference
The authors introduce ConfoundingSHAP, a Shapley-based method for quantifying confounding strength in causal inference by attributing confounding effects to individual covariates. The method employs a specialized Shapley game distinct from standard SHAP applications, coupled with a scalable TabPFN-based estimation to avoid exhaustive model refitting across adjustment sets. Empirical evaluations demonstrate ConfoundingSHAP's utility in identifying key confounders across diverse datasets, enhancing interpretability in observational studies.
causal inferenceconfoundingshapley valuesobservational studiestabpfn
Priority-Driven Control and Communication in Decentralized Multi-Agent Systems via Reinforcement Learning
The authors propose a model-free, priority-driven reinforcement learning algorithm for decentralized multi-agent systems that jointly learns communication priorities and control policies from data. This approach addresses the limitations of existing event-triggered control methods that rely on accurate system models by circumventing the hybrid action space inherent in binary communication decisions. The algorithm is evaluated on benchmark tasks, demonstrating superior performance compared to baseline methods.
reinforcement learningmulti-agent systemsevent-triggered controlcommunication prioritieshybrid action space
Regret Minimization in Bilateral Trade With Perturbed Markets
The paper introduces an adaptive algorithm for regret minimization in bilateral trade under perturbed markets, where an underlying stochastic distribution faces adversarial corruption. The method bridges purely adversarial and stochastic settings by dynamically scaling with corruption level C. It achieves Õ(T^{3/4}) + O(C log(T)) regret against the best budget-balanced price distribution while maintaining worst-case Õ(T^{3/4}) regret relative to per-round budget balance. This guarantees optimal performance in both corrupted stochastic and fully adversarial environments.
regret minimizationbilateral tradeperturbed marketsbudget balanceadversarial corruption
Can Muon Fine-tune Adam-Pretrained Models?
This work investigates the optimizer mismatch problem when fine-tuning Adam-pretrained models with Muon, attributing performance degradation to their distinct implicit biases. Through controlled experiments, the authors demonstrate that the mismatch disrupts pretrained knowledge and scales with update strength. They hypothesize and validate that constraining updates via LoRA mitigates this issue, reducing the performance gap between Adam and Muon across language and vision tasks. Additional studies on LoRA rank, catastrophic forgetting, and LoRA variants confirm the correlation between mismatch severity and update strength. Code is available at https://github.com/XingyuQu/muon-finetune.
optimizer mismatchimplicit biaslorafine-tuningmuon
Self-Attention as a Covariance Readout: A Unified View of In-Context Learning and Repetition
The paper establishes a unified theoretical framework explaining in-context learning (ICL) and repetitive generation in large language models (LLMs) through the lens of self-attention as a covariance readout mechanism. Under stationary, ergodic, and elliptical input conditions, softmax attention converges to a linear function of input covariance, enabling population-level statistical summarization. For ICL, stacked attention heads with residual connections implement iterative gradient descent steps in linear regression tasks. Across transformer layers, this mechanism drives terminal hidden states toward a deterministic function of the current token, inducing first-order Markov behavior and explaining repetition phenomena. Both ICL and repetition emerge as consequences of covariance readout.
self-attentioncovariance readoutin-context learningrepetitive generationsoftmax attention
QT-Net: Rethinking Evaluation of AI Models in Atomic Chemical Space
The authors propose QT-Net, a rotationally augmented graph neural network for predicting atomic properties like electron populations and multipoles, addressing the lack of principled out-of-distribution evaluation in atomic chemical space. They introduce a held-out evaluation protocol using SOAP descriptors and 5×5 cross-validation, comparing E(3)-equivariant and non-equivariant models. QT-Net demonstrates improved performance on QM9 data, with its inferred atomic properties enhancing downstream molecular property prediction and accurately recovering ground-truth dipole moments. The work includes a JAX implementation and released code/data.
atomic propertiessoap descriptorsgraph neural networkqm9 datasetrotational augmentation
AxiomOcean: Forecasting the Three-Dimensional Structure of the Upper Ocean
AxiomOcean introduces a global AI ocean forecasting model that explicitly preserves vertical hierarchy and cross-layer dependence in the upper ocean, addressing limitations in current AI models that over-smooth subsurface features. The model employs a fully three-dimensional encoder-backbone-decoder architecture combined with surface atmospheric forcing to jointly predict upper-ocean temperature, salinity, and three-dimensional currents at 1/12° resolution down to 643 m depth. In 10-day forecasts, AxiomOcean reduces day-1 RMSE by 20-35% compared to an advanced AI model, maintains higher anomaly correlation, and better preserves eddy kinetic energy, temperature variance, and salinity variance, particularly in regions like the equatorial Pacific and Southern Ocean.
three-dimensional encoder-backbone-decodervertical hierarchyeddy kinetic energyanomaly correlationupper-ocean heat content
SlimSpec: Low-Rank Draft LM-Head for Accelerated Speculative Decoding
The paper introduces SlimSpec, a low-rank parameterization of the draft model's LM-head in speculative decoding to accelerate autoregressive generation in LLMs. Unlike prior methods that truncate vocabulary, SlimSpec compresses inner representations while maintaining full vocabulary support. Evaluated with EAGLE-3 drafter across three target models and diverse benchmarks, SlimSpec achieves 4-5× acceleration over standard LM-head architectures, with 8-9% end-to-end speedup improvement over existing methods. The approach requires minimal pipeline adjustments, making it a versatile alternative for draft LM-head architectures.
speculative decodinglow-ranklm-headautoregressive generationvocabulary compression
Don't Fix the Basis -- Learn It: Spectral Representation with Adaptive Basis Learning for PDEs
The paper introduces Adaptive Basis Learning (ABLE), a framework for learning data-dependent spectral representations in PDE learning, addressing limitations of fixed global bases in spectral neural operators. ABLE constructs spatially adaptive Parseval frames via learned ancillary densities, preserving invertibility and maintaining O(N log N) complexity through FFT-based implementation. The method enhances expressivity by shifting focus to representation rather than spectral coefficients, improving accuracy in regimes with sharp gradients and multiscale behavior. ABLE integrates as a drop-in replacement in existing architectures (e.g., U-FNO, HPM), demonstrating complementary spectral refinement. Results highlight data-driven basis choice as a key bottleneck in neural operator design.
spectral neural operatorsadaptive basis learningparseval framepde learningfft-based implementation
Beyond Spatial Compression: Interface-Centric Generative States for Open-World 3D Structure
The paper introduces interface-centric generative states as an alternative to spatial compression in 3D tokenization, addressing representation mismatches in open-world assets with intersecting components and noisy topology. The proposed Component-Conditioned Canonical Local Tokens (C2LT-3D) factorizes representation into canonical local geometry, partition-conditioned context, and relational seam variables, enabling querying, constraining, and repairing during decoding. Evaluated zero-shot on multi-component CAD assets, C2LT-3D demonstrates improved structural robustness and actionable latent variables under adversarial attachment settings, suggesting operational discrete states as a key metric for 3D generative representations.
3d tokenizationgenerative statesinterface-centriccanonical local tokensstructural robustness
DRIFT: Drift-Resilient Invariant-Feature Transformer for DGA Detection
We propose DRIFT, a drift-resilient Transformer-based framework for Domain Generation Algorithm (DGA) detection that addresses temporal degradation in deep learning-based detectors. The model employs a hybrid tokenization strategy combining character-level encoding for stochastic morphological patterns and subword-level encoding for word-based DGAs, alongside multi-task self-supervised pre-training to learn robust structural and contextual features. Evaluations across a 9-year longitudinal study (2017-2025) demonstrate that DRIFT significantly mitigates temporal degradation and outperforms state-of-the-art baselines in forward-chaining experiments. The approach provides a dependable foundation for long-term DGA defense in evolving threat landscapes.
domain generation algorithmstemporal drifthybrid tokenizationmulti-task pre-trainingforward-chaining experiments
Remember to Forget: Gated Adaptive Positional Encoding
The paper introduces Gated Adaptive Positional Encoding (GAPE), a novel augmentation for Rotary Positional Encoding (RoPE) in large language models. GAPE addresses out-of-distribution rotary phases by incorporating content-aware bias into attention logits, using query-dependent and key-dependent gates to manage token relevance. Theoretical analysis confirms that GAPE preserves access to important tokens while suppressing irrelevant distant ones. Empirical validation demonstrates improved attention sharpness and long-context robustness over rotary baselines in synthetic retrieval and long-context benchmarks.
rotary positional encodinglong-context robustnesscontent-aware biasattention logitsgated adaptive positional encoding
Equilibrium Residuals Expose Three Regimes of Matrix-Game Strategic Reasoning in Language Models
The study demonstrates that large language models exhibit a performance gap between semantically rich and procedurally generated matrix games, dropping to 34%, 18%, and 2% success on anonymous 2×2, 3×3, and 5×5 payoff matrices respectively. Using supervised fine-tuning on 2×2 and 3×3 games, the authors improve unseen 5×5–7×7 game success from 2% to 61%, while exploitability-reward training achieves 37% with high variance. The work proves the exploitability residual is 2-Lipschitz in payoff perturbations, enabling transfer learning despite formatting instability. Dominated-action padding experiments confirm trained models solve 3×3 games embedded in larger matrices, highlighting procedural evaluation as essential for measuring strategic reasoning.
matrix gamesnash equilibriumexploitability residualsupervised fine-tuningstrategic reasoning
Identified-Set Geometry of Distributional Model Extraction under Top-$K$ Censored API Access
This work characterizes the distribution-recovery limits of top-$K$ censored LLM API access, where only top-$K$ logits are revealed. The authors derive exact total-variation bounds ($U_K=(V-K)\exp(τ)/(Z_A+(V-K)\exp(τ))$) for the identified set of compatible teacher distributions and provide computable KL divergence bounds. Experiments on Qwen3 demonstrate a layered extraction hierarchy: top-$K$ distillation recovers 12% of private capability, full-logit distillation achieves 56% despite 99% KL closure, and generation-based extraction reaches 96%. Results show that top-$K$ censoring limits per-position distribution recovery but does not prevent capability extraction, separating fidelity from transfer in prompt-only logit distillation.
top-k censoringdistribution recoverykl divergencelogit distillationcapability extraction
Valid Best-Model Identification for LLM Evaluation via Low-Rank Factorization
The paper introduces a principled framework combining multi-armed bandit (MAB) algorithms with low-rank factorization for efficient large language model (LLM) evaluation. It employs doubly robust estimators to leverage predicted scores from low-rank approximations, reducing variance while maintaining statistical validity. The method constructs finite-sample confidence intervals under adaptive model selection and non-replacement sampling. Empirical results on real-world benchmarks demonstrate significant reductions in evaluation costs and accurate identification of the best-performing model, achieving meaningful compute and cost savings.
multi-armed banditlow-rank factorizationdoubly robust estimatorsconfidence intervalsadaptive model selection
Causal Explanations from the Geometric Properties of ReLU Neural Networks
The paper presents a geometric method for generating causal explanations from ReLU neural networks by leveraging their piecewise linear structure. By analyzing the input space partitioning into convex polytopes where each region applies a distinct linear transformation, the approach extracts decision rules directly from the network's geometry without performance-degrading distillation. This yields accurate 'why' and 'why not' explanations that faithfully reflect the original model's behavior, addressing interpretability challenges in safety-critical autonomous systems.
relu networkscausal explanationspiecewise linearconvex polytopesinterpretability
Sharp feature-learning transitions and Bayes-optimal neural scaling laws in extensive-width networks
(No summary returned.)
The Polynomial Counting Capabilities of Message Passing Neural Networks
This paper characterizes the polynomial counting capabilities of Message Passing Neural Networks (MPNN) by analyzing their ability to express extensions of graded modal logic with polynomial counting constraints. The authors primarily utilize local and global mean aggregations to study these capabilities. Results show that global polynomial counting constraints in node-labelled graphs can be checked using mean MPNN under mild assumptions, while local constraints require either sum/max aggregations or restriction to regular graphs. Additionally, formulas with nested modalities are captured by mean MPNN over tree-like graph structures under similar assumptions.
message passing neural networkspolynomial countinggraded modal logicmean aggregationstree-like structures
Regret Analysis of Guided Diffusion for Black-Box Optimization over Structured Inputs
The paper introduces a certificate-based expected simple-regret framework for analyzing guided-diffusion black-box optimization (BO) over structured inputs, addressing limitations of existing BO regret analyses. The framework avoids assumptions like maximum-information-gain bounds, RKHS constraints, and exact acquisition maximization, focusing instead on mass lift: the probability mass increase for near-optimal designs relative to the pretrained generator. This explains phenomena such as exponential finite-budget convergence and polynomial acceleration. Practical diagnostics for estimating search exponents and a proposal-corrected resampling construction are provided, yielding a certified sampler instance.
guided-diffusionblack-box optimizationmass liftsimple-regretcertified sampler
Multifidelity Gaussian process regression for solving nonlinear partial differential equations
The paper introduces a multifidelity Gaussian process regression method for solving nonlinear PDEs, addressing kernel selection challenges in kernel-based approaches. The method first learns a differentiable non-stationary kernel from low-fidelity simulations via cokriging, then constructs a high-fidelity kernel and mean function within a multifidelity framework. This physics-informed approach combines empirical kernel learning with Gaussian process regression, demonstrating efficacy on Burgers' equation through improved PDE solution accuracy.
multifidelity gaussian processnon-stationary kernelcokrigingphysics-informed learningburgers' equation
PC3D: Zero-Shot Cooperation Across Variable Rosters via Personalized Context Distillation
The paper introduces PC3D (Personalized Central Coordination Context Distillation), a method for zero-shot cooperation in multi-agent reinforcement learning with variable team sizes. The approach trains decentralized policies to recover and use personalized coordination context from local histories, leveraging a centralized teacher to distill agent-specific contexts during training. Evaluated on three cooperative MARL benchmarks, PC3D outperforms baselines in both seen and unseen roster sizes, with gains attributed to context distillation and adaptive context use.
multi-agent reinforcement learningzero-shot cooperationcontext distillationdecentralized policiesroster variation
DeepLévy: Learning Heavy-Tailed Uncertainty in Highly Volatile Time Series
DeepLévy introduces a neural framework for modeling heavy-tailed uncertainty in volatile time series by learning mixtures of Lévy stable distributions. It minimizes the discrepancy between empirical and parametric characteristic functions, avoiding the intractability of Lévy probability density functions. The framework employs a mixture mechanism that adaptively learns context-dependent weights and parameters across multiple Lévy components, enabling flexible multi-horizon uncertainty modeling. Evaluations on real and synthetic datasets show that DeepLévy outperforms state-of-the-art deep probabilistic forecasting methods in tail risk metrics, particularly under extreme volatility conditions.
lévy stable distributionscharacteristic functionstail risk metricsmulti-horizon uncertaintydeep probabilistic forecasting
Foundations of Reliable Inference: Reliability-Efficiency Co-Design
The thesis proposes a unified framework for reliable inference in AI models, emphasizing the co-design of reliability and computational efficiency. It addresses the challenge of maintaining trustworthy uncertainty quantification while minimizing computational overhead through advances in Bayesian learning. The work explores whether efficient reliable inference is achievable, contributing to both theoretical foundations and practical implementations.
bayesian learninguncertainty quantificationcomputational overheadreliable inferenceefficiency co-design
Fast Training of Mixture-of-Experts for Time Series Forecasting via Expert Loss Integration
The authors propose a Mixture-of-Experts (MoE) framework for time series forecasting that integrates expert-specific losses directly into training, enhancing specialization. The method combines a base forecasting loss with expert-level losses and employs partial online learning for efficient gating and expert updates, reducing computational costs. Evaluations on economic, tourism, and energy datasets show superior accuracy and efficiency compared to statistical methods and neural models like Transformers and WaveNet, with ablation studies confirming the loss integration's effectiveness.
mixture-of-expertstime series forecastingexpert-specific lossonline learningcomputational efficiency
Signature Approach for Contextual Bandits with Nonlinear and Path-dependent Rewards
(No summary returned.)
Follow the Mean: Reference-Guided Flow Matching
The paper introduces a novel approach to controllable generation in flow matching models by leveraging reference-guided adaptation rather than fine-tuning or auxiliary networks. The method exploits the deterministic interpolants' velocity field, which is governed by a conditional endpoint mean, enabling control by shifting this mean. Two instantiations are proposed: Reference-Mean Guidance, a training-free technique that computes endpoint-mean corrections from a reference bank, and Semi-Parametric Guidance, which uses an explicit mean anchor and learned residual refiner. Results demonstrate effective control over attributes like color, identity, style, and structure using a frozen FLUX.2-klein (4B) model, while maintaining unconditional DiT-B/4 quality on AFHQv2.
flow matchingdeterministic interpolantsendpoint meanreference-guided adaptationsemi-parametric guidance
Nearly-Optimal Algorithm for Adversarial Kernelized Bandits
(No summary returned.)
Set Prediction for Next-Day Active Fire Forecasting
The Wildfire Ignition Set Predictor (WISP) introduces a query-based model for next-day active fire forecasting, reformulating the task as point-set prediction on a 375 m grid. WISP leverages 48 hours of multi-source covariates—meteorology, satellite vegetation, static land, and fire history—to predict ranked sets of future fire cluster centers globally. The model employs Hungarian matching with asymmetric classification-localization weighting to address conflicting roles of classification scores. Evaluated on a globally distributed benchmark, WISP achieves 38.2% average precision, covers 53.4% of fire cluster mass by fire radiative power, and localizes 54.1% of observed clusters within 5 km, establishing sparse set prediction as a viable approach for high-resolution wildfire forecasting.
point-set predictionhungarian matchingfire radiative powerasymmetric weightingquery-based model
Characterizing the Generalization Error of Random Feature Regression with Arbitrary Data-Augmentation
The paper provides a tight characterization of the test error in mean squared error for supervised regression methods under data augmentation in the proportional regime, where covariates grow proportionally to sample size. The analysis focuses on arbitrary augmentation schemes, misspecified feature maps, and architectures with frozen or randomly initialized layers except the last readout layer. Results are derived in terms of population quantities of true data and first/second-order statistics of augmentation. The asymptotic characterization is shown to be tight for Gaussian data.
proportional regimemean squared errormisspecified feature mapsreadout layergaussian data
Sample-Mean Anchored Thompson Sampling for Offline-to-Online Learning with Distribution Shift
We propose Sample-Mean Anchored Thompson Sampling (Anchor-TS), a novel bandit algorithm addressing distribution shift in offline-to-online learning. Anchor-TS introduces a median-based anchoring rule that constructs arm indices from online posterior samples, hybrid posterior samples, and online sample means, systematically correcting bias induced by distribution shift. Theoretical guarantees demonstrate safe offline data utilization for accelerated online learning, with regret reduction quantified by shift degree and offline data size. Extensive experiments show consistent improvements over baselines, mitigating over-estimation for suboptimal arms and under-estimation for optimal arms while exploiting offline information effectively.
thompson samplingdistribution shiftbandit algorithmoffline-to-online learningregret reduction
BROS: Bias-Corrected Randomized Subspaces for Memory-Efficient Single-Loop Bilevel Optimization
(No summary returned.)
Scalable Gaussian process inference via neural feature maps
We introduce a scalable Gaussian process framework that employs neural feature maps to construct expressive kernels, enabling fast and accurate exact GP inference. The method interprets the learned feature map as an optimal low-rank approximation to a Gram matrix from an implied RKHS, ensuring GP posterior consistency. Spectral analysis of induced kernels and product feature-map kernels address oversmoothing. This approach supports regression and classification across diverse data modalities, including tabular inputs and structured domains like images. Benchmark evaluations demonstrate superior accuracy and efficiency in training and prediction compared to existing methods.
gaussian processneural feature mapsgram matrixrkhsoversmoothing
DeepLog: A Software Framework for Modular Neurosymbolic AI
DeepLog introduces a modular neurosymbolic framework that integrates logic and deep learning within PyTorch workflows, serving as a universal backend for diverse neurosymbolic systems. The framework compiles high-level neurosymbolic specifications into optimized arithmetic circuits, enabling both machine learning practitioners and neurosymbolic developers to prototype and deploy integrated systems efficiently. The implementation is available as open-source software, lowering the barrier for adopting neurosymbolic AI techniques.
neurosymbolicpytorcharithmetic circuitslogic programmingdeep learning
Predictive Radiomics for Evaluation of Cancer Immune SignaturE in Glioblastoma: the PRECISE-GBM study
The study developed radiogenomic biomarkers for non-invasive prediction of macrophage subtype M0 immune signatures in IDH-wildtype glioblastoma. Using retrospective multicenter data (TCGA-GBM, CPTAC, IvyGAP, REMBRANDT, CGGA), researchers extracted MRI-based radiomic features from auto-segmented tumor regions, selected features via nested cross-validated LASSO, and trained SVM/ensemble models on deconvoluted transcriptomic immune signatures. Evaluated on three holdout datasets (n=176), ensemble models achieved stable performance (balanced accuracy: 0.67, precision: 0.89) in predicting macrophage signatures, demonstrating potential for immunotherapy patient stratification.
radiogenomicsglioblastomaimmune signaturefeature selectionensemble learning
Generalization Error Bounds for Picard-Type Operator Learning in Nonlinear Parabolic PDEs
The paper presents generalization error bounds for Picard-type operator learning in nonlinear parabolic PDEs, focusing on discretization-invariant solution operators. The method formulates Picard iteration as an abstract state-transition model, separating implementation error from estimation error in the learning framework. Key results show that increasing Picard depth reduces truncation error without unbounded growth of entropy-based estimation error, with extensions to long-time prediction via local model rollout. Theoretical claims are validated for nonlinear heat equations using a Picard-type Fourier neural operator.
operator learningpicard iterationgeneralization errorparabolic pdesfourier neural operator
Teaching LLMs to See Graphs: Unifying Text and Structural Reasoning
The Graph Transformer Language Model (GTLM) introduces a parameter-efficient architecture enabling pretrained LLMs to process graph-structured data natively without semantic bottlenecks. GTLM achieves this by injecting graph-aware attention biases into the LLM's attention modules, adding only 0.015% additional parameters while preserving node permutation equivariance and backward compatibility. Evaluations show that a 1B-parameter GTLM matches or exceeds 7B-parameter state-of-the-art models on Text-Attributed Graph benchmarks and significantly outperforms baselines on GraphQA. Attention heads in GTLM implicitly learn message passing, facilitating superior algorithmic reasoning and enabling scalable GraphRAG and relational deep learning.
graph transformer language modelattention biasesnode permutation equivariancegraphqagraphrag
Building Korean linguistic resource for NLU data generation of banking app CS dialog system
We introduce FIAD (Financial Annotated Dataset), a Korean linguistic resource for generating NLU training data in banking customer service domains. By analyzing banking app reviews, we identified three linguistic patterns—TOPIC (ENTITY, FEATURE), EVENT, and DISCOURSE MARKER—and encoded them in Local Grammar Graphs (LGGs) to produce diverse annotated utterances. We evaluated FIAD-generated data by training DIET-based models with various Korean BERT variants (HANBERT, KoBERT, KorBERT), achieving intent recognition accuracies ranging from 0.91 to 0.95 and topic (entity+feature) extraction F1 scores from 0.83 to 0.86.
natural language understandinglocal grammar graphsbanking customer servicekorean bertannotated dataset
MARGIN: Margin-Aware Regularized Geometry for Imbalanced Vulnerability Detection
MARGIN introduces a margin-aware regularization framework for imbalanced software vulnerability detection, addressing frequency and difficulty imbalances through geometric embedding optimization. The method employs adaptive margin metric learning and hyperspherical prototype modeling, dynamically adjusting geometric regularization based on von Mises-Fisher concentration estimates to align embedding distributions with Voronoi cells. Experiments on public datasets demonstrate MARGIN's superiority over baselines, achieving significant improvements in classification and detection performance, particularly on imbalanced datasets. Analysis reveals enhanced embedding geometry structure, leading to improved robustness, interpretability, and generalization.
metric learninghyperspherical embeddingvon mises-fishervoronoi cellsimbalanced classification
The Benefits of Temporal Correlations: SGD Learns k-Juntas from Random Walks Efficiently
The paper demonstrates that temporal correlations in data enable efficient learning of Boolean k-juntas via gradient-based methods, overcoming traditional barriers under independent samples. Using a lazy random walk on the hypercube, a two-layer ReLU network trained with stylized-SGD and temporal-difference loss exploits these dependencies. The method achieves sample complexity linear in ambient dimension d for fixed k, outperforming large-batch gradient methods with standard convex losses, which fail to leverage temporal correlations.
boolean k-juntastemporal correlationsstylized-sgdtemporal-difference losssample complexity
FORGE: Fragment-Oriented Ranking and Generation for Context-Aware Molecular Optimization
FORGE introduces a fragment-oriented framework for context-aware molecular optimization, addressing limitations of language-model approaches that rely on natural language and suffer from chemical hallucinations. The method operates in two stages: Stage 1 ranks candidate fragments based on their property contributions within the full molecular context, while Stage 2 generates explicit fragment replacements. Built on a compact 0.6B language model, FORGE adapts to unseen objectives via in-context learning. Evaluated on Prompt-MolOpt, PMO-1k, and ChemCoTBench, FORGE outperforms larger language models and graph-based methods, demonstrating the effectiveness of fragment-level supervision over natural language training.
molecular optimizationfragment-orientedcontext-awarein-context learningchemical hallucinations
Stellar Age Compression Reshapes Interpretations of the Milky Way Thick-Disk Formation History
This study demonstrates that systematic compression in stellar age estimation significantly reshapes interpretations of Milky Way thick-disk formation history. Using identical stellar samples and physical covariate matching conditions, the authors compare spectroscopic (astroNN) and asteroseismic (APOKASC-3) age scales. Key findings include a flattening of the age-metallicity relation slope from -3.29 to -1.86 Gyr dex⁻¹, widening of formation timescale from 3.04 to 3.55 Gyr, and shifting of peak formation age from 9.1 to 6.0 Gyr. Transport inversion experiments reveal that compressive transformation (λ < 1) generates rapid-formation-like observables without requiring an intrinsically bursty history, highlighting the sensitivity of Galactic archaeology to stellar age definitions.
stellar age compressionage-metallicity relationtransport inversiongalactic archaeologyasteroseismic ages
Parameterized Complexity of Stationarity Testing for Piecewise-Affine Functions and Shallow CNN Losses
This paper establishes the parameterized complexity of testing approximate first-order stationarity for continuous piecewise-affine (PA) functions, a fundamental task in nonsmooth optimization. Using ambient dimension as the parameter, the authors provide XP algorithms for tractable cases and prove W[1]-hardness for complementary cases, with lower bounds under the Exponential Time Hypothesis ruling out algorithms running in time ρ(d)size^o(d). The results extend to local minimality testing of PA functions and shallow ReLU CNN training losses, demonstrating a unified parameterized-complexity landscape across these domains.
parameterized complexitypiecewise-affine functionsstationarity testingw[1]-hardnesscnn training losses
Extended Wasserstein-GAN Approach to Causal Distribution Learning: Density-Free Estimation and Minimax Optimality
The paper introduces GANICE (GAN for Interventional Conditional Estimation), a novel method for distributional causal inference that addresses limitations of existing GAN-based approaches. GANICE minimizes the averaged Wasserstein risk for estimating conditional interventional distributions, leveraging the extended Wasserstein distance and a cellwise critic in its dual formulation. Theoretical analysis establishes minimax optimality using Besov space theory. Empirical evaluations demonstrate that GANICE consistently outperforms existing methods in estimating interventional outcome distributions, including quantiles and tail risks.
distributional causal inferencewasserstein distanceminimax optimalitybesov spaceinterventional distribution
Unveiling High-Probability Generalization in Decentralized SGD
(No summary returned.)
Task-Aware Calibration: Provably Optimal Decoding in LLMs
The paper introduces task-aware calibration, a method to optimize decoding in large language models (LLMs) by aligning predictive distributions with task-induced latent structures. Building on the insight that free-form outputs can be interpreted semantically, the authors propose Minimum Bayes Risk (MBR) decoding on task-calibrated latent distributions, proving its optimality. They also introduce Task Calibration Error (TCE) to quantify miscalibration losses. Empirical results demonstrate consistent improvements in generation quality across diverse tasks and baselines, enhancing reliability in model decisions.
task-aware calibrationminimum bayes risklatent structuretask calibration errordecoding strategy
Many Needles in a Haystack: Active Hit Discovery for Perturbation Experiments
The article introduces Probability-of-Hit, a novel acquisition function for hit discovery in high-throughput gene perturbation experiments, formalized as a sequential experimental design problem. This method ranks candidates based on their posterior probability of exceeding a predefined phenotypic effect threshold, addressing the inefficiency of pure exploration and the over-exploitation of Bayesian optimization. The approach is proven asymptotically optimal and demonstrates empirical superiority, achieving up to 6.4% improvement over baselines on the Schmidt IL-2 dataset.
acquisition functionhit discoverybayesian optimizationsequential experimental designgene perturbation
Breaking the Reward Barrier: Accelerating Tree-of-Thought Reasoning via Speculative Exploration
SPEX accelerates Tree-of-Thought (ToT) reasoning by addressing the reward dependency barrier through speculative exploration. It introduces three techniques: intra-query speculative path selection to predict high-potential branches, inter-query budget allocation for dynamic resource distribution, and adaptive early termination to prune redundant branches. Implemented on SGLang, SPEX achieves 1.2∼3× speedup across various ToT algorithms and LLMs, with cumulative speedups up to 4.1× when combined with token-level speculative decoding. Ablation studies validate the individual contributions of each technique, demonstrating SPEX's effectiveness in enhancing ToT reasoning efficiency and scalability.
tree-of-thoughtspeculative explorationreward dependency barrieradaptive early terminationspeculative decoding
Fix the Loss, Not the Radius: Rethinking the Adversarial Perturbation of Sharpness-Aware Minimization
Loss-Equated SAM (LE-SAM) improves upon Sharpness-Aware Minimization (SAM) by addressing the mismatch between first-order linearized surrogates and second-order curvature notions in flat minima. Instead of fixing the perturbation radius, LE-SAM fixes a loss-space budget, thereby removing gradient-norm-dominated learning signals and emphasizing curvature-dominated terms. Extensive experiments across diverse benchmarks and tasks demonstrate that LE-SAM consistently outperforms SAM and its variants, achieving state-of-the-art generalization performance.
sharpness-aware minimizationloss-space budgetcurvature-dominated termsgeneralization performanceperturbation radius
Joint sparse coding and temporal dynamics support context reconfiguration
The study identifies joint sparse coding and temporal dynamics in the mouse medial prefrontal cortex (mPFC) and computational networks as mechanisms enabling context reconfiguration without erasing prior knowledge. Using spiking neural networks, the authors demonstrate that sparsity reduces cross-context interference, while temporal dynamics enhance context separability. These properties improve retention in lifelong learning tasks without auxiliary heuristics, offering an energy-efficient architectural principle for stable adaptation. The findings provide a mechanistic framework for understanding flexible context transitions in biological and artificial systems.
sparse codingtemporal dynamicscontext reconfigurationlifelong learningspiking neural networks
Balancing Efficiency and Fairness in Traffic Light Control through Deep Reinforcement Learning
The paper introduces a deep reinforcement learning agent for traffic light control that jointly optimizes efficiency and fairness for both vehicular and pedestrian flows. The proposed method dynamically adjusts signal timing based on real-time demand, departing from vehicle-centric approaches. Experiments show the agent successfully reduces congestion while maintaining equitable service across user types, offering a practical solution for smart city traffic management.
deep reinforcement learningtraffic light controlfairness-aware optimizationdynamic signal timingsmart city mobility
Hyperparameter Transfer for Dense Associative Memories
The authors develop hyperparameter transfer methods for Dense Associative Memory (DenseAM) architectures, addressing challenges posed by shared weights across layers and rapidly peaking activation functions uncommon in feed-forward networks. They derive explicit prescriptions for transferring hyperparameters from small to large-scale DenseAM models, leveraging theoretical insights into the architecture's energy landscape and temporal dynamics. Empirical results demonstrate strong agreement with theoretical predictions, validating the proposed transfer methods. This work fills a gap in hyperparameter optimization for DenseAMs, enabling more efficient scaling of these architectures.
dense associative memoryhyperparameter transferenergy landscapetemporal dynamicsactivation functions
OUIDecay: Adaptive Layer-wise Weight Decay for CNNs Using Online Activation Patterns
The paper introduces OUIDecay, an adaptive layer-wise weight decay scheduler for CNNs that dynamically adjusts regularization strengths based on online activation patterns. The method employs the Overfitting-Underfitting Indicator (OUI) to monitor layer-specific structural behavior without requiring validation data or gradient computations. Evaluations on EfficientNet-B0, ResNet50, DenseNet121, and MobileNetV2 across four datasets (Stanford Cars, Food101, CIFAR100, CIFAR10) show OUIDecay achieves the best mean validation loss in 7/8 settings, outperforming fixed and gradient-based adaptive decay methods.
weight decayactivation patternsonline adaptationconvolutional neural networksregularization
jNO: A JAX Library for Neural Operator and Foundation Model Training
jNO introduces a JAX-native library for training neural operators and foundation models, unifying data-driven and physics-informed approaches through a symbolic tracing system. The framework compiles domains, model calls, residuals, supervised losses, and diagnostics into a single optimization pipeline, enabling seamless transitions between operator regression, mesh-aware residual evaluation, and PDE-constrained training. It supports multi-model compositions, fine-grained parameter control, hyperparameter tuning, and JAX-native workflows for PDE foundation-model families. The implementation is available at https://github.com/FhG-IISB/jNO.
neural operatorsphysics-informed trainingsymbolic tracingpde-constrained trainingjax-native
Unsupervised Process Reward Models
The paper introduces unsupervised Process Reward Models (uPRM) that eliminate the need for costly expert annotations in step-level supervision of LLM reasoning. The method leverages a scoring function derived from LLM next-token probabilities to identify erroneous steps across reasoning trajectories without human supervision. Results show uPRM improves accuracy by 15% over LLM-as-a-Judge in error detection, matches supervised PRMs in verification, and enhances reinforcement learning policy optimization by 6.9% over majority voting baselines.
unsupervisedprocess reward modelsnext-token probabilitiesreasoning trajectoriesreinforcement learning
Stable Long-Horizon PDE Forecasting via Latent Structured Spectral Propagators
The authors propose a Structured Spectral Propagator (SSP) framework for stable long-horizon forecasting of time-dependent partial differential equations (PDEs). SSP reformulates PDE rollout by mapping physical states into a propagation-oriented latent space, isolating recurrent dynamics from spatial details, and evolving spectral modes using a frequency-conditioned linear backbone with a nonlinear spectral closure. This approach decouples reconstruction fidelity from rollout regularity, providing a strong inductive bias for coherent modal evolution. Experiments demonstrate SSP reduces relative L2 errors by up to 48.9% compared to state-of-the-art baselines and improves temporal extrapolation stability beyond the supervised horizon.
structured spectral propagatorpartial differential equationslatent spacemodal evolutiontemporal extrapolation
APEX: Audio Prototype EXplanations for Classification Tasks
APEX introduces a post-hoc explanation framework for pre-trained audio classifiers, addressing the gap in XAI for audio by avoiding vision-centric approaches. The method disentangles explanations into four perspectives: Square-based (transient events), Time-based (temporal patterns), Frequency-based (spectral bands), and Time-Frequency-based (integrated analysis). APEX preserves output invariance without fine-tuning, offering more semantically intuitive explanations than gradient-based methods by leveraging prototype reasoning tailored to acoustic properties.
audio prototype explanationspost-hoc interpretabilityoutput invariancespectrogram analysismultidimensional similarity
Learning to Sparsify Stochastic Linear Bandits
(No summary returned.)
PFN-TS: Thompson Sampling for Contextual Bandits via Prior-Data Fitted Networks
The paper introduces PFN-TS, a Thompson sampling algorithm for contextual bandits that leverages prior-data fitted networks (PFNs) like TabPFN v2+ and TabICL v2. By converting PFN posterior predictives into mean-reward samples via a subsampled predictive central limit theorem, PFN-TS reduces computational complexity from O(n) to O(log n) while reusing cached representations. Theoretical analysis proves consistency of the variance estimator and provides a Bayesian regret bound. Empirical results show PFN-TS achieves top performance on synthetic, OpenML, and mobile-health benchmarks, outperforming existing methods in nonlinear and linear reward settings.
thompson samplingcontextual banditsprior-data fitted networksbayesian regretpredictive central limit theorem
Per-Loss Adapters for Gradient Conflict in Physics-Informed Neural Networks
The paper introduces per-loss adapters to address gradient conflict in physics-informed neural networks (PINNs), identifying distinct regimes of conflict requiring different interventions. A diagnostic-first framework profiles unmodified PINN runs to determine whether scalar reweighting or lightweight architectural changes are needed, employing low-rank adapters to create loss-indexed parameter subspaces attached to a shared PINN trunk. Evaluated across 60+ PDE configurations, including forward, inverse, and multi-physics problems up to 50D, adapters combined with reweighting significantly improve performance in persistent directional conflict scenarios, while reweighting alone suffices for magnitude-dominated cases. Full-parameter-space gradient surgery often fails on heterogeneous parameter spaces.
physics-informed neural networksgradient conflictlow-rank adaptersparameter subspacesscalar reweighting
GELATO: Generative Entropy- and Lyapunov-based Adaptive Token Offloading for Device-Edge Speculative LLM Inference
The paper introduces GELATO, a framework for device-edge speculative LLM inference that optimizes token throughput under energy constraints. It combines an outer drift-plus-penalty loop for long-term energy-throughput trade-off management with an entropy-driven generation mechanism for dynamic per-token uncertainty adaptation. Theoretical analysis confirms its performance bound, while empirical results show 64.98% higher throughput and 47.47% lower energy consumption versus state-of-the-art approaches, maintaining decoding quality.
speculative decodingtoken offloadinglyapunov optimizationentropy-driven generationedge computing
Complex-Valued Phase-Coherent Transformer
The Phase-Coherent Transformer (PCT) introduces a complex-valued attention mechanism that preserves phase information across layers by replacing softmax with a real-valued, element-independent gate applied to L2-normalized complex query-key similarities. This token-non-competing design avoids traditional row-normalized competition, maintaining phase coherence. Evaluated on mid-scale benchmarks including long-range memory, hierarchical reasoning, and image classification, PCT outperforms standard softmax Transformers and complex-valued counterparts in parameter-fair comparisons. It remains competitive with strong real-valued baselines (Multiscreen) on challenging tasks like NIAH and LRA-Text. Ablation studies confirm the necessity of phase-preserving gates, with performance degrading when violating PCT conditions. The architecture shows no depth-related collapse across tested ranges.
complex-valued transformerphase-coherent attentiontoken-non-competingl2-normalizationlong-range reasoning
Scaling the Memory of Balanced Adam
The paper demonstrates that in balanced Adam (where $\beta_1=\beta_2$), the remaining hyperparameter $\beta$ should be interpreted as defining a memory horizon $H_\beta=(1-\beta)^{-1}$. By introducing the refresh count $R_\beta=(1-\beta)T_{\mathrm{ES}}$ (where $T_{\mathrm{ES}}$ is the effective learning horizon), the authors show that setting $R_\beta\approx1000$ adaptively selects $\beta$ values across 11 vision/language tasks, improving robustness. Compared to fixed $\beta=0.94377$, this method reduces the worst-case validation gap by 33.4% and keeps all runs within 1% of their oracle performance, suggesting $\beta$ should be treated as a memory-scale variable rather than a constant.
balanced adammemory horizonrefresh counteffective learning horizonoptimizer scaling
Generating Symmetric Materials using Latent Flow Matching
We introduce SymADiT, a symmetry-aware variant of the All-atom Diffusion Transformer (ADiT) for materials generation, leveraging Wyckoff positions to enforce crystal symmetry constraints. The method performs generative modeling in latent space, ensuring outputs adhere to space group symmetries and atom-specific Wyckoff positions. SymADiT generates stable, symmetric materials using a Transformer architecture, demonstrating competitive performance against both symmetry-aware and symmetry-agnostic models in materials generation benchmarks.
symmetry-awarewyckoff positionslatent spacediffusion transformermaterials generation
GLiNER-Relex: A Unified Framework for Joint Named Entity Recognition and Relation Extraction
GLiNER-Relex introduces a unified framework for joint named entity recognition (NER) and relation extraction (RE) using a shared bidirectional transformer encoder. The architecture extends GLiNER to handle both tasks in a single model, enabling zero-shot extraction of arbitrary entity and relation types through text, entity type, and relation type label representations. Evaluated on CoNLL04, DocRED, FewRel, and CrossRE benchmarks, it matches specialized RE models and LLMs while maintaining GLiNER's efficiency. The open-source package supports flexible inference with user-specified labels.
named entity recognitionrelation extractionbidirectional transformerzero-shot extractionknowledge graphs
TopoU-Net: a U-Net architecture for topological domains
We introduce TopoU-Net, a U-Net architecture generalized to topological domains using combinatorial complexes. The method leverages hierarchical encoder-decoder principles with rank-based transport maps and skip connections between matched ranks, replacing spatial scale with rank paths through nodes, edges, faces, or hyperedges. Key architectural decisions involve the bottleneck support ratio, determined by the complex and path, which governs skip connection utility. Evaluated on node classification, graph classification, hypergraph node classification, mesh classification, and image reconstruction tasks, TopoU-Net achieves superior mean accuracy on six of eight node-classification datasets and four of five hypergraph datasets, particularly excelling on heterophilic graphs. Ablations confirm skip connections are critical under severe bottleneck compression.
combinatorial complexesrank-pathbottleneck support ratioskip connectionsheterophilic graphs
Unlocking air traffic flow prediction through microscopic aircraft-state modeling
AeroSense introduces a state-to-flow modeling framework for air traffic flow prediction by directly mapping microscopic aircraft states to future traffic flow, bypassing aggregated time-series approaches. The method leverages dynamic sets of aircraft states from ADS-B trajectories to preserve fine-grained kinematics and interactions. Experimental results demonstrate superior predictive accuracy over conventional aggregation-based methods, especially during high-density traffic conditions, validating the efficacy of instantaneous airspace state modeling.
aerosenseads-b trajectoriestraffic flow predictionmicroscopic modelingaircraft kinematics
A Stability Benchmark of Generative Regularizers for Inverse Problems
The study benchmarks the stability of generative priors for inverse problems in imaging, evaluating their performance under imperfect conditions. It examines convergent regularization, robustness to out-of-distribution data, and sensitivity to inaccuracies in the forward operator or noise model, comparing generative approaches to modern optimization-based methods. Results identify scenarios where generative priors achieve state-of-the-art reconstructions and highlight limitations where they may underperform or pose risks.
generative priorsinverse problemsconvergent regularizationout-of-distribution robustnessvariational techniques
PixelFlowCast: Latent-Free Precipitation Nowcasting via Pixel Mean Flows
PixelFlowCast introduces a latent-free precipitation nowcasting framework combining deterministic coarse forecasting with Pixel Mean Flows (PMF) for high-fidelity prediction. The method employs a two-stage approach: KANCondNet extracts spatiotemporal features for conditional guidance, while PMF uses x-prediction to preserve fine-grained structures with few-step sampling. On SEVIR, it outperforms existing methods in accuracy and efficiency, particularly for long sequences, demonstrating operational potential.
precipitation nowcastingconditional flow matchingpixel mean flowsspatiotemporal forecastingsevir dataset
TrajDLM: Topology-Aware Block Diffusion Language Model for Trajectory Generation
We propose TrajDLM, a topology-aware trajectory generation framework that combines block diffusion language models with road network embeddings to generate high-fidelity GPS trajectories efficiently. TrajDLM models trajectories as sequences of discrete road segments, leveraging a block diffusion backbone for denoising, topology-aware embeddings from a road network encoder, and topology-constrained sampling for coherence. Evaluated on three city-scale datasets, TrajDLM achieves strong local similarity metrics, is up to 2.8× faster than prior methods, and demonstrates robust zero-shot transfer across domains, including unseen transportation modes. This highlights the scalability and accuracy of block-wise discrete diffusion for trajectory generation.
block diffusiontopology-aware embeddingstrajectory generationzero-shot transferdenoising
The two clocks and the innovation window: When and how generative models learn rules
This work characterizes the tension in generative models between learning rule-valid patterns ($τ_{\mathrm{rule}}$) and memorizing training samples ($τ_{\mathrm{mem}}$), defining an innovation window $[τ_{\mathrm{rule}}, τ_{\mathrm{mem}}]$. Using rule-valid synthetic tasks on parity and combinatorial puzzles, the authors analyze DiT and GPT models, showing $τ_{\mathrm{rule}}$ increases with rule complexity and decreases with model capacity, while $τ_{\mathrm{mem}}$ scales linearly with dataset size. The innovation window widens with larger datasets and narrows with complex rules, vanishing when $τ_{\mathrm{rule}} \geq τ_{\mathrm{mem}}$. Analysis of DiT's learned score reveals evolving optimization landscapes, with rule-valid basins expanding at $τ_{\mathrm{rule}}$ and training-sample basins dominating at $τ_{\mathrm{mem}}$.
generative modelsinnovation windowrule complexityoptimization landscapesscore-matching
The Value of Mechanistic Priors in Sequential Decision Making
The paper quantifies the value of mechanistic priors in sequential decision-making by introducing mechanistic information, defined as the mutual information between a model's recommended policy and the true optimal policy, measured via occupancy-weighted bias. Asymptotic analysis shows Bayesian regret scales with residual entropy, yielding a theoretical sample complexity reduction of H(μ)/H_mech compared to uninformed baselines, while burn-in regime analysis establishes penalties for incorrect priors. Empirical validation on 5-fluorouracil dosing simulations demonstrates significant sample-efficiency gains with hybrid priors. The study contrasts mechanistic priors with LLM priors, highlighting severe mechanistic information losses in LLMs and advocating for physically-grounded priors in safety-critical applications.
mechanistic priorssequential decision-makingbayesian regretresidual entropysample complexity
Differentially Private Sampling from Distributions via Wasserstein Projection
The paper introduces a differentially private (DP) sampling framework using Wasserstein distance as the utility measure, addressing limitations of prior density ratio-based approaches that ignore geometric structure and support differences. The proposed Wasserstein Projection Mechanism (WPM) achieves minimax optimality by projecting onto Wasserstein space. Efficient approximation algorithms with convergence guarantees are provided for practical implementation.
differential privacywasserstein distancesampling mechanismminimax optimalityconvergence guarantees
Anchor-guided Hypergraph Condensation with Dual-level Discrimination
The authors propose Anchor-guided Hypergraph Condensation with Dual-level Discrimination (AHGCDD) to address computational challenges in hypergraph neural network training. AHGCDD introduces three components: a Heat Kernel PageRank-based node initialization module for encoding structural knowledge, an anchor-guided hyperedge synthesis strategy for joint optimization of condensed features and structure, and a dual-level discrimination objective to preserve utility without redundant training. This method overcomes limitations of existing hypergraph condensation approaches that rely on decoupled training architectures and trajectory-based optimization. Extensive experiments demonstrate AHGCDD's superior effectiveness and efficiency in condensing large-scale hypergraphs.
hypergraph condensationheat kernel pagerankanchor-guided synthesisdual-level discriminationhypergraph neural networks
Continual Harness: Online Adaptation for Self-Improving Foundation Agents
The paper introduces Continual Harness, a reset-free self-improving framework for embodied agents that autonomously refines prompts, sub-agents, skills, and memory during online interaction. Building on observations from Gemini Plays Pokemon (GPP) - the first AI to complete Pokemon Blue, Yellow Legacy (hard mode), and Crystal undefeated - the method alternates between action and self-modification without human intervention or episode resets. Evaluations on Pokemon Red and Emerald show the approach reduces button-press costs by 47-63% compared to baselines and recovers 72% of the performance gap to hand-engineered expert harnesses. The system further demonstrates sustained progress via online process-reward co-learning, where frontier models relabel open-source agent rollouts for iterative improvement.
continual learningembodied agentsprompt optimizationonline adaptationprocess-reward co-learning
Lakestream: A Consistent and Brokerless Data Plane for Large Foundation Model Training
Lakestream introduces a brokerless, object-store-native data plane for Large Foundation Model training, addressing limitations of existing systems through three innovations. It proposes the Transactional Global Batch (TGB), extending lakehouse ACID semantics with training-specific consistency, including atomic batch visibility and checkpoint-aligned lifecycle management. Recovery and retention are implemented directly in the storage layer via inlined producer state and distributed checkpoint state. The Decentralized Adaptive Commit (DAC) algorithm ensures stable ingestion throughput without inter-producer communication. Evaluations on 64-GPU multimodal pre-training and SFT workloads demonstrate Lakestream's superiority in throughput, failure isolation, and latency over colocated dataloaders and Apache Kafka.
transactional global batchlakehousedecentralized adaptive commitfailure isolationexactly-once recovery
Learning Graph Foundation Models on Riemannian Graph-of-Graphs
R-GFM introduces a Riemannian Graph-of-Graphs (GoG) foundation model to address scale mismatch in graph learning by treating structural scale as a first-class citizen. The method constructs a multi-scale GoG from subgraphs sampled at varying hop distances and learns geometry-adaptive representations on Riemannian manifolds, theoretically reducing structural domain generalization error. Experiments show state-of-the-art performance with up to 49% relative improvement on downstream tasks compared to fixed-scale graph foundation models.
graph foundation modelsriemannian manifoldsmulti-scale learninggraph-of-graphsdomain generalization
Attention Drift: What Autoregressive Speculative Decoding Models Learn
The paper identifies attention drift in autoregressive speculative decoding models, where drafters progressively shift attention from the prompt to their own generated tokens during speculation chains. The authors trace this to un-normalized residual paths causing hidden state magnitude growth and propose two architectural fixes: post-norm on drafter hidden states and per-hidden-state RMSNorm. These changes improve acceptance length by up to 2× under template perturbation, 1.18× on long-context tasks, and 1.10× across seven benchmarks, while enabling better generalization to longer drafting sequences.
speculative decodingattention driftautoregressive modelsrmsnormhidden state normalization
Optimizer-Induced Mode Connectivity: From AdamW to Muon
(No summary returned.)
Federated Language Models Under Bandwidth Budgets: Distillation Rates and Conformal Coverage
The paper establishes statistical guarantees for federated language model training and inference under explicit bandwidth constraints. It introduces two protocols: Federated Probe-Logit Distillation (FPLD) for training with a high-probability KL-consistency rate dependent on node count, sample size, quantization budget, probe-set size, and vocabulary size; and Federated Conformal RAG (FC-RAG) for inference with a distribution-free marginal-coverage bound incorporating retrieval-bandwidth slack. Theoretical results show exponential decay of quantization error and $K^{-1/2}$ slack reduction with arithmetic aggregation. Synthetic and GPT-2 experiments validate parameter scaling and bandwidth-accuracy tradeoffs.
federated learningbandwidth budgetconformal predictionknowledge distillationretrieval-augmented generation
📰 Industry Media (6)
Tilde Research Introduces Aurora: A Leverage-Aware Optimizer That Fixes a Hidden Neuron Death Problem in Muon
Tilde Research introduces Aurora, a leverage-aware optimizer addressing neuron death in Muon, a polar factor-based optimizer widely used in large-scale MLP training. Aurora enforces uniform row norms and left semi-orthogonality simultaneously, resolving Muon's row-norm anisotropy issue that deactivates over 25% of neurons by step 500. Aurora achieves 100x data efficiency on open-source internet data, outperforms larger models on HellaSwag, and sets a new state-of-the-art on the modded-nanoGPT speedrun benchmark at 1.1B parameters. With only 6% compute overhead, Aurora scales effectively with MLP width and serves as a near-drop-in replacement for Muon.
polar factorneuron deathmlp trainingrow-norm anisotropymodded-nanogpt
A Coding Implementation to Portfolio Optimization with skfolio for Building Testing, Tuning, and Comparing Modern Investment Strategies
The tutorial demonstrates skfolio, a scikit-learn-compatible Python library for portfolio optimization, implementing modern investment strategies with rigorous backtesting. Methodologically, it covers mean-variance optimization, hierarchical risk parity (HRP), nested clusters optimization (NCO), Black-Litterman models, and factor-based approaches, evaluated via walk-forward validation and hyperparameter tuning. Results show comparative performance metrics (Sharpe ratios 0.3–0.5) across 15 strategies on S&P 500 data, with hierarchical methods and robust estimators outperforming baseline equal-weighted portfolios.
portfolio optimizationhierarchical risk parityblack-litterman modelwalk-forward validationfactor models
OpenAI Introduces Daybreak: A Cybersecurity Initiative That Puts Codex Security at the Center of Vulnerability Detection and Patch Validation
OpenAI introduces Daybreak, a cybersecurity initiative leveraging Codex Security to integrate vulnerability detection and patch validation into the software development lifecycle. The system employs GPT-5.5 models under a Trusted Access framework, enabling threat modeling, isolated validation, and patch generation with human review. Daybreak reduces vulnerability analysis time from hours to minutes and integrates with 20+ security partners across edge, endpoint, and supply chain defense. Access is tiered, with GPT-5.5-Cyber restricted to authorized workflows like red teaming. The initiative aims to make software resilient by design, addressing dual-use risks through verification and safeguards.
codex securitytrusted accessthreat modelingpatch validationdual-use risk
JBS Dev: On imperfect data and the AI last mile – from model capability to cost sustainability
Joe Rose of JBS Dev challenges the misconception that perfect data is required for generative AI systems, demonstrating that current tooling enables effective processing of imperfect inputs. Using medical billing reconciliation as a case study, he shows how LLMs extract structured data from heterogeneous sources (PDFs, images) via OCR and text extraction, achieving incremental automation (20-80%) with human-in-the-loop validation. Rose predicts a shift from model capability improvements to cost sustainability, advocating for edge deployment (laptops/phones) over data centers and in-house cloud-based agentic workloads over SaaS solutions.
generative aihuman-in-the-loopocragentic workloadsedge deployment
Hugging Face hosted malicious software masquerading as OpenAI release
HiddenLayer identified a malicious Hugging Face repository posing as OpenAI's Privacy Filter release, which delivered infostealer malware via a loader.py script. The attack leveraged a command-and-control channel using jsonkeeper.com to rotate payloads and targeted browser data, Discord storage, cryptocurrency wallets, and system information. Researchers found six additional repositories with similar malicious logic, highlighting risks in AI development workflows. The repository accrued 244,000 downloads and 667 likes before removal, though these metrics may have been artificially inflated. This incident underscores vulnerabilities in public AI model registries and the need for improved supply chain security measures.
infostealercommand-and-controlpayloadloader.pysupply chain
Laserfiche unveils AI agents for natural language workflows
Laserfiche introduces AI agents for autonomous natural language workflow automation within content management systems, operating under strict security and compliance frameworks. The agents leverage generative LLM reasoning models to perform context-aware document analysis and task execution, including legal contract review, accounts payable processing, and HR record organization. Initial deployment in Laserfiche Cloud (May 2026) supports one-time actions via Smart Chat, with planned enhancements for background process monitoring and business process integration.
generative llmcontent managementnatural language workflowcontext-aware actioncompliance framework
Generated automatically at 2026-05-12 21:01 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.