Daily Digest — 2026-07-02
317 items · 3 research labs, 299 arxiv papers, 15 industry media
🏛️ Research Labs (3)
Hugging Face and Cerebras bring Gemma 4 to real-time voice AI
Hugging Face and Cerebras demonstrate a real-time speech-to-speech pipeline combining Nvidia's Parakeet for speech recognition, Gemma 4 31B VLM (via Cerebras for low-latency inference), and Alibaba's Qwen3TTS for text-to-speech. The modular architecture achieves natural conversational flow by reducing median and P95 latency, addressing key bottlenecks in language-model response times. The system currently powers 9,000+ Reachy Mini robots, emphasizing stability for embodied AI applications. The open-source stack enables inspection and modification at each layer.
speech-to-speechgemma 4cerebrasparakeetqwen3tts
The latest AI news we announced in June 2026
Google announced multiple AI advancements in June 2026, focusing on unified AI integration across devices and workflows. Key releases include Gemma 4 12B, a 12B-parameter open model with vision/voice processing running locally on 16GB RAM, and Gemini 3.5 Flash for cross-platform agent automation. Android 17 introduced floating app windows and biometric security, while Gemini Omni Flash enabled multimodal video workflows. Educational tools like NotebookLM added structured research capabilities, and AI flood prediction models achieved 7-day advance warnings. Workplace studies showed 73% UK AI adoption correlated with career advancement.
gemma 4 12bgemini omni flashin-context learningmultimodal agentsparameter-efficient
New York City educators and industry leaders gathered at Google’s offices to shape the future of AI in classrooms.
The AI Summit 2026 convened 150 educators and industry leaders to align AI education with workforce needs, emphasizing human-centric skills alongside technological literacy. Hosted by Google, the New York Jobs CEO Council, and Urban Assembly, the event featured hands-on sessions exploring tools like Google AI mode and NotebookLM, demonstrating their potential for fostering AI literacy and problem-solving capabilities. Key discussions highlighted the importance of adaptability, collaboration, and critical judgment in an AI-driven future, while stressing privacy and equitable access. The summit concluded that technological innovation must integrate with educational systems to prepare students effectively for emerging careers.
ai literacynotebooklmproblem-solvingequitable accessworkforce needs
📜 arXiv Papers (299)
Introspective Coupling: Self-Explanation Training Tracks Behavioral Change Despite Fixed Supervision
This work demonstrates that language models (LMs) trained on fixed counterfactual explanations exhibit introspective coupling, where generated explanations align with their current behaviors rather than the training targets. The method involves training LMs to explain input features influencing predictions, using counterfactual explanations derived from earlier checkpoints or behaviorally similar models as supervision. Results show this coupling persists across behavioral shifts during concurrent post-training objectives, tracking changes without updated supervision, and remains robust to label noise across tasks like sycophancy and refusal. Findings suggest fixed counterfactual explanation datasets provide scalable introspection signals.
introspective couplingcounterfactual explanationslanguage modelsbehavioral shiftpost-training
QVal: Cheaply Evaluating Dense Supervision Signals for Long-Horizon LLM Agents
The paper introduces QVal, a training-free testbed for evaluating dense supervision signals in long-horizon LLM agents. QVal measures Q-alignment by comparing method scores to Q-values from a reference policy, enabling pre-training comparison across diverse supervision methods. QVal-v1.0 benchmarks 21 methods across four environments and seven families, using six open-weight models (1.2K experiments). Results show simple prompting baselines outperform recent literature methods, with performance clustering by family, consistent across model sizes and modalities. The framework supports extensibility for new environments and methods.
dense supervisionq-alignmentlong-horizon agentstraining-free evaluationreference policy
Reinforcement Learning with Metacognitive Feedback Elicits Faithful Uncertainty Expression in LLMs
The paper introduces Reinforcement Learning with Metacognitive Feedback (RLMF), a novel paradigm to enhance Large Language Models' (LLMs) metacognitive abilities by refining completion rankings during preference optimization based on self-judgments of performance. RLMF, combined with metacognitive data selection, aims to improve faithful calibration (FC) by aligning expressed uncertainty with intrinsic uncertainty. The method employs a two-stage approach: calibrating self-reported confidence scores and mapping them to context-adaptable linguistic uncertainty. Experiments demonstrate that RLMF achieves state-of-the-art FC across diverse tasks, surpassing standard RL by up to 63% while preserving accuracy, and enhances models' ability to assess and express their capability limits.
reinforcement learningmetacognitionfaithful calibrationpreference optimizationuncertainty expression
When LLMs Read Tables Carelessly: Measuring and Reducing Data Referencing Errors
This work presents the first systematic evaluation of tabular data referencing errors (DREs) in large language models (LLMs), demonstrating their prevalence across models ranging from 1.7B to 20B parameters. The authors introduce a critic-based approach incorporating data referencing, which improves answer accuracy by up to 12.0% through filtering and rejection sampling. A lightweight 4B-parameter critic model is trained, achieving an average F1 score of 78.2% in detecting both in-distribution and out-of-distribution DREs, effectively assisting inference for larger models.
data referencing errorscritic-based filteringrejection samplingin-distributionout-of-distribution
Freeform Preference Learning for Robotic Manipulation
The paper introduces Freeform Preference Learning (FPL), a method for learning robot manipulation policies from multi-axis human preferences. FPL collects natural-language preference axes (e.g., speed, safety) with pairwise comparisons, trains a language-conditioned reward model per axis, and optimizes a reward-conditioned policy across dimensions. Evaluated on four real-world and two simulated long-horizon tasks, FPL outperforms sparse-reward and binary-preference baselines by 38 percentage points, while enabling test-time behavior steering and compositional generalization without subtask segmentation.
preference learningreward modelinglanguage-conditioned policylong-horizon manipulationmulti-axis evaluation
AdaJEPA: An Adaptive Latent World Model
AdaJEPA introduces an adaptive latent world model that performs test-time adaptation within model predictive control (MPC) to address distribution shifts. The model plans and executes an initial action chunk, uses the observed state transition as a self-supervised adaptation signal, and replans with the updated model, enabling continuous recalibration without expert demonstrations. Experiments demonstrate that AdaJEPA significantly improves planning success rates, requiring as few as one gradient step per MPC replanning step across various goal-reaching tasks.
latent world modeltest-time adaptationmodel predictive controlself-supervised learninggradient step
FLORA: A deep learning approach to predict forest attributes from heterogeneous LiDAR data
FLORA introduces a deep learning framework for predicting six forest attributes from heterogeneous LiDAR data, addressing challenges posed by variable acquisition conditions in national programs. The method combines an octree-based backbone with ecological and spatiotemporal auxiliary variables via late-fusion gating, trained on 32,052 National Forest Inventory plots across France. Results show robust performance, with rRMSE of 12.3% (R²=0.88) for dominant height and 39% (R²=0.74) for total volume, outperforming season-specific models and demonstrating cross-season generalization.
lidaroctreeforest inventorydeep learningheterogeneous data
TRIAGE: Role-Typed Credit Assignment for Agentic Reinforcement Learning
TRIAGE introduces role-typed credit assignment for agentic reinforcement learning, addressing limitations of standard GRPO's uniform outcome-based credit. The method classifies action segments into semantic roles (decisive progress, useful exploration, etc.) via a structured judge and applies role-conditioned process rewards, optimally reducing advantage estimation error when the judge is reliable. Evaluated on ALFWorld, Search-QA, and WebShop, TRIAGE improves success rates over GRPO by 10.4-14.8% and reduces redundant actions, with regression detection and exploration credit being key contributors.
credit assignmentagentic reinforcement learningsemantic rolesadvantage estimationprocess rewards
AxDafny: Agentic Verified Code Generation in Dafny
The paper introduces AxDafny, a verifier-guided repair framework for agentic code generation in Dafny that iteratively produces executable code alongside proof artifacts (invariants, assertions, termination arguments). The authors also present LCB-Pro-Dafny, a benchmark of 250 competition-style programming problems with formal specifications and verifier-based evaluation. AxDafny achieves 92.7% verification success on DafnyBench (6.5pp improvement over prior proof-hint baselines) and demonstrates superior performance on LCB-Pro-Dafny compared to GPT-5.5. Results indicate verification success and runtime tests capture distinct code quality dimensions.
verified code generationdafnyagentic programmingformal verificationproof artifacts
PolicyGuard: From Organizational Policies to Neuro-SymbolicCompliance Review Engines
PolicyGuard introduces a neuro-symbolic framework for policy-grounded document compliance review, addressing the opacity of end-to-end LLM prompting. The framework converts organizational policies into executable review engines comprising typed relational logic rules and atom-level extraction questions. During review, LLMs answer local questions using document evidence, while a symbolic evaluator applies formal rules to detect non-compliance. Evaluated on company-specific NDA compliance, PolicyGuard enhances explicitness, maintainability, and systematic testability by separating policy formalization, document interpretation, and symbolic evaluation.
neuro-symbolic frameworkpolicy-grounded reviewtyped relational logicatom-level extractionsymbolic evaluator
Self-Study Reconsidered: The Hidden Fragility of Learning from Self-Generated QA
The study reveals fragility in language model training from self-generated QA pairs, identifying two key failure modes: biased question selection and instruction compliance. Analyzing generation as an implicit policy, the authors show question coverage concentrates on salient spans (saturating early) and answers overly obey instruction-like text, with larger models more compliant (88% mean injection rate). Simple interventions—fixed-question targets and instruction-span filtering—reduce bias (13% compliance) while preserving clean text utility, demonstrating these issues stem from generation choices rather than training procedures.
synthetic supervisionquestion generationinstruction complianceknowledge distillationfailure modes
Radial Suppression Accelerates Algorithmic Generalization: A Geometric Analysis of Delayed Generalization
The study identifies radial inflation of hidden representations as the primary driver of delayed generalization in neural networks performing algorithmic tasks. Through a geometric analysis of activation-space dynamics, the authors propose that penalizing radial inflation induces anisotropic weight regularization, suppresses radial gradient energy, and biases convergence toward flatter minima. Empirical validation using a norm penalty on activations demonstrates 6x faster grokking on modular arithmetic tasks with MLPs and Transformers, and 50% reduced training steps for a 10M-parameter nanoGPT on 3-digit addition.
radial inflationalgorithmic generalizationactivation-space dynamicsanisotropic regularizationgrokking
Amplifying Membership Signal Through Chained Regeneration
We introduce MADreMIA, a model-agnostic framework that amplifies membership inference attack (MIA) and dataset inference (DI) signals through iterative regeneration trajectories, addressing limitations of one-shot approaches. Inspired by Model Autophagy Disorder (MAD), MADreMIA leverages chained generations across diverse modalities, where each output serves as subsequent input, to enhance membership evidence while maintaining low false positive rates. The method demonstrates that memorized training samples exhibit higher coherence and slower degradation during iterative regeneration compared to non-member samples. Evaluations across image autoregressive models (IARs), diffusion models, and language models show improved signal richness, with preliminary results suggesting applicability to audio models.
membership inference attackmodel autophagy disorderiterative regenerationdataset inferencefalse positive rate
GR2 Technical Report
GR2 introduces an LLM-based re-ranking framework for industrial recommendation systems, addressing three gaps: underutilization of re-ranking stages, limited RL application, and non-semantic item identifiers. The method combines semantic ID mid-training (≥99% uniqueness), reasoning-trace distillation from a teacher model, and RL with verifiable rewards, supplemented by context compression and on-policy distillation for scalability. Results show +18.7% R@1, +7.1% R@3, and +9.6% N@3 improvements over baselines, with reward design identified as critical to prevent exploitation of position bias.
re-rankingsemantic idsreasoning distillationon-policy distillationverifiable rewards
LUNA: Learning Universal 3D Human Animation Beyond Skinning
LUNA introduces an LBS-free universal neural animation model that directly maps 2D controls (images, keypoints, sketches) into 3D Gaussian deformations, bypassing explicit body fitting. The model employs a transformer-based motion regressor to disentangle global rigid motion from local dynamics, enabling coherent movement and subtle non-rigid effects. Hybrid supervision distills soft structural priors from an LBS teacher, supporting training on both fitted data and unlabeled videos. Experiments demonstrate LUNA achieves competitive visual fidelity, realistic motion, and zero-shot cross-identity generalization across diverse driving modalities, marking the first end-to-end 3D animatable model supporting implicit 2D driving.
linear blend skinninggaussian deformationsmotion regressorhybrid supervisionzero-shot generalization
TreeAgent: A Generalizable Multi-Agent Framework for Automated Bias Labeling in Forestry via Compiled Expert Rules and Vision-Language Models
We introduce TreeAgent, a multi-agent framework combining expert decision trees with Vision-Language Models (VLMs) for automated bias labeling in forestry remote sensing. The framework employs a Decoupled Declarative Decision (D3) Framework to generalize across expert-defined structures without modification, using VLMs for localized semantic perception and multi-agent voting to reduce VLM stochasticity. Evaluated on tree bias classification, TreeAgent outperforms supervised ML baselines while significantly reducing expert labeling effort. Results demonstrate that agentic orchestration of VLMs with expert priors can replicate expert labeling procedures at lower cost while preserving interpretability.
multi-agent systemvision-language modelsdecision treebias labelingforestry remote sensing
MECoBench: A Systematic Study of Multimodal Agent Collaboration in Embodied Environments
The authors introduce MECoBench, a multimodal embodied cooperation benchmark for systematically evaluating collaboration capabilities of multimodal large language models (MLLMs) in visually grounded environments. The benchmark includes diverse real-world tasks, two cooperation structures, and three collaboration modes. Key findings show that (i) collaboration improves task completion but requires balancing gains with coordination complexity, (ii) communication is crucial with optimal modes depending on team size and model capability, and (iii) collaboration enhances robustness under noisy priors and exploration conditions.
multimodal large language modelsembodied agentscooperation benchmarkvisually grounded environmentscollaboration modes
LeCropFollow: Latent Space Planning for Navigation in Unstructured Crop Fields
LeCropFollow introduces a latent space planning framework for agricultural robot navigation in unstructured crop fields, avoiding explicit geometric modeling. The method combines a self-supervised semantic heatmap extractor with TD-MPC2, a Model-Based Reinforcement Learning planner, to optimize trajectories directly within a learned latent manifold. Field experiments in late-stage corn show 2.4x fewer semantic failures than keypoint-based methods in plantation gaps, with zero-shot transfer from simulation to physical deployment.
latent space planningmodel-based reinforcement learningsemantic heatmapzero-shot transferagricultural robotics
MVP-Nav: Multi-layer Value Map Planner Navigator
MVP-Nav introduces a physical-aware RGB-only navigation framework for Zero-shot Object Goal Navigation (ZSON), addressing semantic-physical misalignment without depth sensors. The method reconstructs explicit physical occupancy from monocular RGB observations using 3D foundation models, projecting 2D semantic instances into 3D oriented bounding boxes to form a global spatial semantic representation. A Multi-layer Value Map (MVM) integrates semantic priorities and reconstructed geometry into a shared cost space, enabling physically grounded geometric planning. Experiments on zero-shot object navigation benchmarks show MVP-Nav outperforms existing depth-free methods, achieving state-of-the-art performance by effectively compensating for the absence of active depth sensors.
zero-shot object navigationmonocular rgb3d foundation modelsmulti-layer value mapphysical-aware navigation
Attend, Transform, or Silence: Operator-Level Visual Skipping for Efficient Multimodal LLM Inference
We introduce an operator-level visual-token skipping framework to enhance multimodal large language model (MLLM) inference efficiency by selectively bypassing redundant attention, FFN, or both operators while preserving the full visual-token sequence. Our approach decomposes Transformer layers into attention and FFN operators, leveraging the observation that late visual-token updates often have minimal impact on answer-token representations. Evaluated across three MLLM architectures and 10 VQA benchmarks, the method reduces 33.7% TFLOPs in Qwen3-VL while maintaining 99.5% of vanilla model performance.
multimodal large language modelsvisual-token skippingtransformer layersattention operatorsffn operators
Better Understanding, Understanding Better
The paper introduces a comparative epistemic logic framework for modeling graded understanding among agents, addressing a gap in epistemic logic's treatment of understanding as a degree-based phenomenon. The method enriches multi-agent epistemic models with agent-indexed graded explanation structures and a justification term algebra, enabling representation of minimal to ideal understanding and inter-agent comparisons. Results include soundness and strong completeness proofs for both finitary bounded-level and infinitary full-language systems, with decidability established for finite-level fragments.
epistemic logicgraded understandingexplanation structuresterm algebradecidability
Modal CEGAR-tableaux with RECAR and resolution-based SAT-shortcuts
This work introduces a novel integration of SAT, tableaux, and resolution methods for modal satisfiability, outperforming individual components. The authors extend CEGAR-tableaux with two SAT-shortcut approaches: RECAR and a new method leveraging the modal resolution theorem prover KSP as an oracle. Experiments conducted with CEGARBox++, a C++ implementation of CEGAR-tableaux, demonstrate that while RECAR-based shortcuts are ineffective, KSP-based shortcuts significantly enhance performance, particularly on large satisfiable problems. This marks the first successful unification of these techniques in modal logic.
cegar-tableauxsat-shortcutsmodal resolutionksprecar
Harnessing Textual Refusal Directions for Multimodal Safety
The paper introduces Modality-Agnostic Refusal Steering (MARS), a training-free method for enhancing multimodal safety in Multimodal Large Language Models (MLLMs) without requiring unsafe multimodal data. MARS leverages textual refusal directions from the LLM backbone, generalizing them across modalities (image, video) through activation re-centering, adaptive steering strength scaling, and optimal layer selection. Evaluated on five state-of-the-art MLLMs across safety, utility, and video jailbreak benchmarks, MARS demonstrates consistent safety improvements while maintaining utility. The findings suggest that safety-relevant structures are shared across modalities and that textual refusal directions are a powerful foundation for multimodal alignment.
multimodal safetyrefusal directionsactivation re-centeringtraining-freemodality-agnostic
Belief Contraction in Dynamic Epistemic Logic
We introduce a mechanism for belief contraction in dynamic epistemic logic (DEL) that operates directly on standard Kripke models without requiring plausibility orderings or constraints on the doxastic accessibility relation. This approach addresses expressive limitations of prior methods, particularly their inability to model belief violating positive introspection and contraction dynamics in response to hedged public announcements. The proposed mechanism satisfies some standard properties of belief contraction, though not all, and we analyze conditions under which contraction may fail. We provide sound and complete axiomatizations via reduction axioms for both the base logic and an extended DEL accommodating contractions induced by private or semi-private announcements.
belief contractiondynamic epistemic logickripke modelsreduction axiomsdoxastic accessibility
Z-1: Efficient Reinforcement Learning for Vision-Language-Action Models
Z-1 introduces a reinforcement learning framework for enhancing Vision-Language-Action (VLA) models in robotic manipulation, addressing limitations of behavior cloning and supervised fine-tuning. The method applies Group Relative Policy Optimization (GRPO) across 24 RoboCasa tasks, utilizing shared-prefix rollout construction, tree-structured trajectory branching, completion-aware reward calibration, and selective joint training of VLM and Action Expert. Z-1 achieves an average success rate of 80.6%, marking a 13.2% improvement over supervised fine-tuning initialization and surpassing state-of-the-art models. This demonstrates GRPO's efficacy in advancing flow-based VLA policies without requiring additional private demonstrations.
vision-language-actiongroup relative policy optimizationrobocasasupervised fine-tuningreinforcement learning
Bridging Local Observation and Global Simulation in Closed-Loop Traffic Modeling
The paper proposes CRAFT, a framework addressing the local-to-global context mismatch in autoregressive traffic simulators trained on ego-centric driving logs. CRAFT employs self-supervised failure discovery through what-if rollouts in a globally observable sandbox, then aligns behaviors using human-aligned driving priors via a Contextual Preference Evaluator (CPE) that reweights autoregressive decoding. The method reduces collisions by 31.2% and traffic violations by 33.2% without base simulator retraining.
autoregressive traffic simulationcontextual preference evaluatorego-centric driving logsclosed-loop environmentspreference-guided alignment
Real-Time Source-Free Object Detection
We introduce RT-SFOD, a real-time source-free object detection method that advances the Pareto frontier of speed, accuracy, and model size. Building on YOLOv10, our approach incorporates Dual-Head Pseudo-Label Fusion (DHF) for optimal pseudo-label generation under domain-shift and Multi-scale Adaptive Representation Diversification (MARD) loss to maintain feature discriminability. These modules operate only during training, preserving inference efficiency. RT-SFOD achieves 1.4-3.5% mAP gains, 1.3× higher throughput, and ∼2× fewer parameters compared to state-of-the-art SFOD methods. The method generalizes across YOLO- and DETR-based dual-head detectors, demonstrating its versatility.
source-free object detectiondual-head detectordomain-shiftpseudo-label generationmulti-scale feature
An Agentic AI Framework to Accelerate Scientific Discovery in Plant Phenotyping
The authors propose an agentic AI framework to accelerate scientific discovery in plant phenotyping by transforming Oak Ridge National Laboratory's Advanced Plant Phenotyping Laboratory from a data factory into an interactive autonomous platform. The framework comprises a conversational Co-Scientist Agent that translates natural-language queries into structured analysis plans and a headless Compute Agent that executes Vision Transformer-based segmentation and trait extraction on the Frontier exascale supercomputer. Agents operate in separate security domains, communicate via a secure token-authenticated streaming channel, and maintain end-to-end provenance. This reduces analysis time from days/weeks to interactive loops with real-time reasoning and recommendations.
agentic aiplant phenotypingvision transformerexascale computingprovenance
Breaking Failure Cascades: Step-Aware Reinforcement Learning for Medical Multimodal Reasoning
The paper introduces Medical Reasoning-aware Policy Optimization (MRPO), a reinforcement learning algorithm that mitigates cascading errors in medical visual question answering by assigning step-wise process rewards. MRPO penalizes early invalid reasoning steps exponentially when the final answer is incorrect, while preserving successful paths. Evaluated on three multimodal LLM backbones, MRPO outperforms standard GRPO and recent RL baselines, achieving a 2.79-point improvement over HuatuoGPT-Vision-34B on Qwen3-VL-8B-Instruct and reducing early-stage reasoning failures from 64.0% to 13.0%.
reinforcement learningmultimodal reasoningcascading errorsmedical vqaprocess rewards
Adaptive Cluster-First Route-Second Decomposition for Industrial-Scale Vehicle Routing
The paper proposes an adaptive cluster-first route-second (CFRS) decomposition system for industrial-scale capacitated vehicle routing problems (CVRPs), addressing limitations of fixed partitioning methods. The approach formulates decomposition as an iterative decision process using a large language model (LLM) as a high-level controller that dynamically applies clustering, balancing, and refinement operators based on instance characteristics. Evaluated on synthetic and benchmark-derived CVRP instances with up to 500,000 customers, the method achieves competitive performance on benchmarks while demonstrating superior scalability and routing quality on large-scale problems.
capacitated vehicle routingcluster-first route-secondlarge language modelsadaptive decompositionindustrial-scale optimization
Creating Intelligence: A Computational Foundation for AGI
The paper proposes a computational theory of mind based on set theory and hyperdimensional computing, using sparse binary data representations instead of continuous weights. Information is modeled as discrete sets, with associative memory emerging from combinatorially expanded hidden layers via topological plasticity. The framework unifies auto-associative and hetero-associative learning through subset pattern matching and exact nearest-neighbor search, operating in constant time. The author suggests this algorithm underlies both cerebellar and neocortical function, enabling energy-efficient hardware implementations due to its discrete logic foundation.
hyperdimensional computingassociative memorytopological plasticitysubset pattern matchingsparse distributed representations
Geometry-Preserving Orthonormal Initialization for Low-Rank Adaptation in RLVR
We propose geometry-preserving orthonormal initialization for low-rank adaptation (LoRA) in Reinforcement Learning with Verifiable Rewards (RLVR), addressing the instability and underperformance of structurally initialized LoRA variants like PiSSA and MiLoRA in this setting. Our theoretical analysis demonstrates that orthonormal initialization minimizes the gap between LoRA outcomes and full fine-tuning. We introduce RLPO and RLMO, two new LoRA variants leveraging this initialization. Experiments on mathematical reasoning benchmarks show that our approach stabilizes RLVR training and outperforms standard LoRA, while also explaining the suboptimal performance of PiSSA and MiLoRA in RLVR contexts.
low-rank adaptationorthonormal initializationreinforcement learningverifiable rewardsmathematical reasoning
Large Databases Need Small, Open-Weight Language Models
We demonstrate that quantized, open-weight language models deployed locally on 16GB VRAM achieve comparable or superior accuracy to closed-source APIs at reduced latency and cost, enabling efficient LM-database integration. By optimizing system configurations and integrating these models into the BlendSQL v0.1.0 framework, we achieve a 390x cost reduction and 3.8x latency improvement over proprietary LM APIs. Our approach challenges the necessity of closed-source solutions for LM-enhanced relational operators, particularly in large database contexts where token-based API costs can exceed $10,000 per experiment.
quantizedopen-weightrelational operatorslatencyvram
RAISE: LLM-based Automated Heuristic Design with Robust Adversary Instance Search
The paper introduces RAISE, a framework for robust automated heuristic design (AHD) using large language models (LLMs) that addresses distributional shifts in real-world deployment. RAISE combines LLM-based evolutionary search with constrained worst-case instance search, where an inner loop identifies hard instances within an epsilon-ball around the training distribution via basis distribution parameterization. Experiments on Online Bin Packing, Online Job Shop Scheduling, and Online Vehicle Routing show RAISE maintains performance across distribution shifts, outperforming existing LLM-based AHD methods that degrade by up to 19x.
automated heuristic designdistributional shiftevolutionary searchworst-case instance searchbasis distribution parameterization
Evo-PI: Aligning Medical Reasoning via Evolving Principle-Guided Supervision
Evo-PI introduces a principle-centric learning framework for improving reasoning in multimodal language models (MLLMs) by evolving language-based supervision signals dynamically. The method establishes a co-evolutionary loop where principles guide model reasoning while model behaviors refine these principles, enabling adaptive supervision. Evaluated on medical visual question answering across eight benchmarks, Evo-PI achieves up to 24.6% accuracy gains, demonstrating its effectiveness for expert-aligned reasoning.
multimodal language modelsprinciple-guided supervisionco-evolutionary loopmedical visual question answeringdynamic alignment
CHERRY: Compressed Hierarchical Experts with Recurrent Representational Yield
The paper introduces CHERRY, a framework combining three techniques for compute-efficient language models: (1) Selective Ground Truth Token Training (SGT) improves per-token efficiency by supervising only 15% of semantically meaningful tokens, recovering 67% of full-sequence loss reduction via gradient coupling (γ̄=0.72). (2) Depth compression reduces a 48-layer transformer to 6 layers (227M params) with recurrent unrolling, achieving comparable loss (2.934) to a 566M dense model. (3) Mixture of Efficient Experts (MoEE) fusion with multi-token prediction yields lower loss (2.789) than individual experts. Validated on Korean CHERRY-1.8B, results are explicitly scoped to loss-based metrics.
selective supervisiongradient couplingdepth compressionmixture of expertsrecurrent unrolling
A Self-Evolving Agentic System for Automated Generation and Execution of Biological Protocols
The paper introduces ProtoPilot, a self-evolving multi-agent system for autonomous biological experimentation, alongside an expert-grounded evaluation framework. The system integrates layer-wise verifiability, multi-agent orchestration, and a runtime-updated skill library to generate protocols, synthesize SDK-compliant code, and revise workflows based on wet-lab feedback. Evaluated on 294 synthetic- and molecular-biology tasks, ProtoPilot achieved 90.2% Top@3 expert-preference rate, 89.5% protocol-to-code gate pass rate, and 88.24% Opentrons pass rate, significantly outperforming OpenTrons-AI (32.35%). Wet-lab validation confirmed successful DNA assembly and feedback-correction, demonstrating verifiable autonomous experimentation.
autonomous experimentationmulti-agent systemprotocol generationwet-lab automationfeedback-correction
A Technical Typology of AI Systems in Public Administration
The paper introduces a technical typology of AI systems in public administration, categorizing them into five types: hand-coded, glass-box, black-box, general-purpose, and agentic systems, calibrated by their implications for public values. It evaluates technical precision in 91 highly-cited public administration papers (2019-2025) using this typology, finding widespread imprecision: 55% underspecify systems, 31% motivate with different systems than studied, and 41% overgeneralize conclusions. Practical recommendations are provided, including diagnostic questions to locate systems within the typology without requiring specialist knowledge.
typologyimprecisionpublicadministrationsystems
JL1-CC&QA: Extending the JL1-CD Benchmark with Change Captioning and Question Answering
JL1-CC&QA extends the JL1-CD benchmark by introducing change captioning (CC) and question answering (QA) tasks to enhance semantic understanding in remote sensing change detection. The dataset comprises 5,000 bi-temporal image pairs from the Jilin-1 satellite, annotated via a three-stage pipeline involving multi-modal large language model (LLM) generation, vision-grounded LLM judging, and human expert verification. It includes 17,021 captions describing land-cover transformations and 20,060 QA pairs across eight question types, enabling fine-grained interrogation of surface changes. This unified benchmark aims to advance multi-task change understanding in remote sensing.
change captioningquestion answeringremote sensingmulti-modal llmbi-temporal images
FedXDS: Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning
FedXDS introduces a novel federated learning approach that leverages explainable AI (XAI) methods to counteract data heterogeneity. The method uses propagation-based attribution to identify task-relevant features through a single backward pass, enabling selective data sharing between clients while incorporating metric privacy for formal guarantees. Experiments show FedXDS achieves higher accuracy and faster convergence than existing methods across varying client numbers and heterogeneity settings, with demonstrated robustness against membership inference and feature inversion attacks.
federated learningexplainable aidata heterogeneitymetric privacyfeature attribution
STEB: Style Text Embedding Benchmark
The authors introduce the Style Text Embedding Benchmark (STEB), an open-source framework for standardized evaluation of style embeddings across diverse tasks. STEB aggregates 96 datasets spanning 7 languages, covering authorship verification, retrieval, AI-text detection, and linguistic feature probing. Results demonstrate that semantic embeddings underperform in stylistic tasks, and no single style embedding dominates across all evaluated tasks. The benchmark is publicly available at https://github.com/rrivera1849/STEB.
style embeddingsbenchmark evaluationauthorship verificationai-text detectionlinguistic probing
Seeing Is Not Sharing: Some Vision-Language Models Overestimate Common Ground in Asymmetric Dialogue
The study investigates whether vision-language models (VLMs) can distinguish potential from established common ground in asymmetric dialogue, using an interpretation-matching task on 13,077 annotated expressions from HCRC MapTask. Evaluating VLMs under controlled manipulations of dialogue context and map-access, results show task-relevant map content (visual or textual) biases models toward over-predicting alignment, degrading accuracy on non-aligned cases. Analysis reveals models rely on static referential cues rather than tracking grounding dynamics, with Qwen3-VL-8B-Instruct and four others exhibiting this pattern to varying degrees.
vision-language modelscommon groundinterpretation-matchingreferential cuesdialogue grounding
Cross-lingual Relation Extraction with Large Language Models: Zero-Shot, Few-Shot, and Fine-Tuned Evaluation on Romanian
This work evaluates cross-lingual relation extraction (RE) for Romanian using large language models (LLMs), addressing the lack of annotated corpora in low-resource languages. The authors translate the SemEval-2010 Task 8 benchmark from English to Romanian via an LLM-based pipeline and assess Gemma 4 31B under zero-shot, few-shot, and QLoRA fine-tuned configurations against four encoder baselines. Results show a 3-5pp performance drop for Romanian versus English in prompt-only settings, marginal gains from few-shot prompting, and a 22pp macro F1-score improvement from QLoRA fine-tuning, reducing the cross-lingual gap from 3.3 to 1.4pp. Despite being 50-250 times smaller, encoder baselines perform within 1-4pp of QLoRA Gemma, with Romanian BERT matching multilingual XLM-R.
relation extractioncross-lingualqlorafew-shotsemantic evaluation
Arena-T2I Hard: Benchmarking and Improving Faithfulness with Dependency-Aware Checklist
We introduce Arena-T2I Hard, a 310-prompt benchmark for evaluating text-to-image (T2I) model faithfulness across six categories, including text rendering and spatial relationships. The benchmark decomposes each prompt into ~30 yes/no constraints, enabling fine-grained failure analysis. We propose a dependency-aware checklist reward that structures constraints as a DAG and propagates failures to descendant nodes, combined with a group-decoupled normalization (GDPO) aesthetic reward. This approach achieves a superior faithfulness-aesthetics trade-off on SD3.5-Medium and FLUX.1-dev under MMRB2 pairwise comparisons compared to single-reward and ensemble baselines. The strongest closed-source system scored 0.855, with a 33pp performance gap across 11 systems.
text-to-imagefaithfulnessdependency-awaregroup-decoupled normalizationmmrb2
Look But Don't Touch with Sparse Autoencoders for Unlearning in Diffusion Models
The study evaluates sparse autoencoders (SAEs) for concept manipulation in diffusion models, revealing a detection-intervention gap. While SAEs effectively localize semantic concepts in diffusion model activations, direct latent space interventions cause out-of-distribution artifacts. The authors propose using SAEs solely as detectors to identify target object regions, replacing those patches with non-target embeddings to maintain activation statistics and improve erasure quality. Results demonstrate that monosemantic features are unsuitable for direct steering, positioning SAEs as interpretability tools rather than control mechanisms for unlearning.
sparse autoencodersdiffusion modelsconcept detectionlatent interventionunlearning
RCT: A Robot-Collected Touch-Vision-Language Dataset for Tactile Generalization
The paper introduces RCT (Robotic Contact Tactile), a robot-collected dataset for tactile generalization, comprising 29,279 tactile frames from 122 industrial materials across 7 categories, recorded using three DIGIT sensors. The dataset preserves contact sequences, enabling evaluation across materials, categories, sensors, and contact positions. Results show that removing contact-sequence overlap reduces tactile-to-text Recall@1 by 17.7 percentage points, and performance drops sharply when materials are held out (Recall@1: 25.1 +/- 6.1%). RCT-trained embeddings improve category probes on unseen materials, highlighting novel-material generalization as a key challenge.
tactile perceptionrobot-collected datasetcontact sequencesgeneralizationcontrastive training
ShopX: A Foundation Model for Intent-to-Item Fulfillment in Agentic Shopping
ShopX introduces a foundation model for intent-to-item fulfillment in agentic shopping, addressing the gap between language understanding and item-space operations. The model unifies intent understanding, execution planning, and semantic ID (SID)-based item-space operations into a single framework, enabling direct item-space interfaces through SIDs. It employs a training recipe to retain LLM knowledge while specializing in multi-turn fulfillment. Evaluations on Taobao production logs show ShopX outperforms tool-mediated systems, particularly on complex or ambiguous requests, by reducing lossy hand-offs between agent orchestration and item-space execution.
semantic idsagentic shoppingintent-to-item fulfillmentllm agentsgenerative recommendation
When to Truncate a Feature Ranking: A Residual-Overlap Stopping Rule for Subset Selection
The paper introduces a distributional framework for determining optimal stopping points in supervised feature selection by transforming feature rankings into class-independent subsets. It proposes a risk-calibrated stopping rule based on the Bhattacharyya coefficient to measure marginal separation between class-conditional distributions, selecting the shortest ranking prefix where residual product overlap falls below a threshold. The method derives binary and multiclass Bayes-risk bounds and provides prior-dependent/free threshold calibrations. Empirical evaluations on high-dimensional genomic datasets demonstrate the rule's ability to reduce tens of thousands of features to dozens while maintaining predictive performance comparable to using all features, offering an interpretable solution for high-dimensional settings.
feature selectionbhattacharyya coefficientrisk-calibrated stopping ruleclass-conditional distributionsresidual product overlap
Histogram-constrained Image Generation
The paper introduces Histogram-constrained Image Generation (HIG), a novel control mechanism for diffusion models that enforces user-specified distributional constraints with exact precision. The method formulates histogram control as an optimal transport problem and applies explicit guidance transformations during sampling to align the diffusion trajectory with desired histograms. Results demonstrate HIG's versatility in constrained generation via color/latent histograms and high-capacity information embedding, offering a flexible and interpretable control scheme compatible with existing mechanisms.
diffusion modelsoptimal transporthistogram constraintscontrollable generationexplicit guidance
WorldRoamBench: An Open-World Benchmark for Long-Horizon Stability of Interactive World Models
WorldRoamBench introduces a comprehensive benchmark for evaluating interactive world models (IWMs) across four dimensions: action, vision, physics, and memory. It employs tailored metrics including per-frame action evaluation, segment-based drift detection, controllability-gated physics plausibility, and action-decoupled memory protocols. The benchmark comprises 600+ test cases across Nature, Urban, and Indoor scenes with continuous interaction durations of 10-60s. Evaluation of 10+ models reveals none reliably satisfies all dimensions, with the best achieving only moderate scores. This benchmark advances the development of stable, physically grounded, and memory-faithful IWMs for real-world applications.
interactive world modelssegment-based driftcontrollability-gated evaluationaction-decoupled protocol3d point-cloud reconstruction
Sparsity-Inducing Divergence Losses for Biometric Verification
The paper introduces Q-Margin, a novel α-divergence loss function for biometric verification that incorporates a probabilistic margin into the reference measure. Unlike margin-penalty softmax losses (e.g., ArcFace, CosFace), Q-Margin preserves sparsity (when α>1) while enhancing discriminative embedding learning. Evaluated on IJB-B, IJB-C (face verification) and VoxCeleb (speaker verification), Q-Margin outperforms baselines at low False Acceptance Rates (FARs) and enables memory-efficient training via sparse posteriors.
α-divergencebiometric verificationsparse solutionsprobabilistic margindiscriminative embeddings
Improving Certified Robustness via Adversarial Distillation
The paper introduces AD-CERT, a certified training method combining adversarial distillation with Interval Bound Propagation (IBP) to improve the standard-certified accuracy trade-off. The approach distills adversarial information from a robust teacher model at the logit level, providing an effective lower bound surrogate for certified training. Results demonstrate state-of-the-art certified performance on multiple benchmarks, with logit-level distillation improving certified accuracy by up to 5.40 percentage points compared to feature-space distillation.
certified robustnessadversarial distillationinterval bound propagationlogit-space distillationneural network verification
FARS: A Fully Automated Research System Deployed at Scale
FARS (Fully Automated Research System) introduces a fully automated AI-for-AI research system capable of autonomously generating and advancing research projects across diverse AI/ML topics. The system employs stage-specific agents coordinated through a shared workspace, handling ideation, planning, experimentation, and manuscript writing while preserving intermediate artifacts. In its initial deployment, FARS produced 166 complete research papers spanning 67 fine-grained AI/ML topics, evaluated through 282 structured reviews from volunteer reviewers. Results indicate FARS can generate review-worthy and occasionally strong AI/ML research artifacts at scale, though recurring limitations in experimental scope, methodology, and integrity were noted.
automated research systemai-for-aistage-specific agentsshared workspacestructured reviews
ECHO: Prune to act, trace to learn with selective turn memory in agentic RL
ECHO introduces a selective turn-memory framework for long-horizon language agents, addressing history collapse and traceable learning through source-indexed reconstruction. The method compresses completed environment turns into compact memory records, reconstructs policy contexts by selecting from these records, and routes outcome credit to supporting evidence via source indices. On BrowseComp-Plus, ECHO achieves 43.4% accuracy, outperforming GRPO (28.9%) and SUPO (36.1%), while demonstrating improved zero-shot generalization across QA, code generation, and information-seeking tasks.
selective turn-memorysource-indexed reconstructionhistory collapsetraceable learninglong-horizon agents
Think in English, Answer in Korean: Efficient Adaptation of Multilingual Tool-Using Agents
LuckyStar 111B introduces a 111B-parameter multilingual tool-using agent adapted from Cohere's Command A model for Korean-English enterprise applications under memory constraints. The method combines multilingual supervised fine-tuning, reinforcement learning with verifiable rewards, language-consistency rewards for Korean responses, and 4-bit quantization for single-GPU serving. Results show improved mathematical reasoning, function calling, and NL2SQL performance while maintaining Korean and English instruction-following quality, providing a practical framework for memory-efficient agent adaptation.
multilingual fine-tuningverifiable rewardslanguage-consistency4-bit quantizationnl2sql
A Lifecycle and Application-Stack Survey of Large Language Model Vulnerabilities: Attacks, Risks, Defenses, and Open Problems
The paper presents a systematic survey of vulnerabilities in large language model (LLM) systems, organized through a lifecycle and application-stack lens. It identifies eight stages of potential attacks: data collection, pretraining, post-training alignment, model packaging and supply chain, retrieval and memory, prompting and inference, tool/agent execution, and deployment/maintenance. For each stage, the authors analyze attacker capabilities, security objectives, representative attacks, practical risks, evaluation practices, and defenses. The study emphasizes the failure of trust boundaries, the transformation of untrusted data into executable instructions, and the amplification of model errors through delegated authority. The paper concludes with a research agenda addressing compositional security, provenance-aware retrieval, tool-call containment, and other critical areas.
large language modelstrust boundariestool-call containmentprovenance-aware retrievalcompositional security
Intrinsic decomposition and editing of 3D Gaussian splats
The paper introduces intrinsic decomposition for 3D Gaussian splatting, enabling separate editing of diffuse albedo and shading in radiance fields. The method models decomposition as independent Gaussian primitive sets, optimizes them using data-driven predictions from multi-view images, and provides an editing workflow where planar surface textures are modified via albedo adjustments in a single image. This allows re-rendering of edited scenes with consistent lighting from arbitrary viewpoints.
intrinsic decompositiongaussian splattingradiance fieldsalbedo editingmulti-view optimization
A Tutorial on Autonomous Fault-Tolerant Control Using Knowledge-Grounded LLM Agents
The paper proposes a framework for autonomous fault-tolerant control using knowledge-grounded LLM agents as constrained supervisory planners. The method integrates plant-specific knowledge for recovery action proposals, validated by external symbolic or simulation-based checks before actuation. Three key design dimensions are identified: recovery patterns, validation strategies, and deployment constraints. The framework is implemented in two open Python environments simulating a modular mixing module and continuous stirred-tank reactor with configurable faults and custom recovery interfaces.
fault-tolerant controlllm agentsknowledge groundingsupervisory planningprocess plants
Scientific Explanations in Health Sciences: Causality, Trust, and Epistemic Adequacy
The paper bridges philosophy of science and explainable AI (XAI) in medicine by critically reviewing foundational accounts of scientific explanation and assessing their adequacy for medical AI. It integrates philosophical analysis with current XAI developments to identify three key axes: causality in medical reasoning, epistemic dimensions of medical trust, and pragmatic criteria for explanatory adequacy. The study proposes principles for designing epistemically robust XAI systems aligned with clinical decision-making needs, offering a philosophically grounded approach to medical explainability.
explainable aimedical artificial intelligencecausalityepistemic adequacyclinical decision-making
Automating Cause-Effect Specification with Knowledge Graphs and Large Language Models
The paper introduces a semantic-AI framework for automating cause-and-effect (C&E) logic generation in process control and safety. The method combines a knowledge graph (KG) based on a modular alignment ontology with a constrained large language model (LLM) layer. The KG represents process structure, faults, symptoms, causes, and mitigation actions, while the LLM generates operator-ready safety narratives and Semantic Web Rule Language (SWRL) rules under strict ontology constraints. Demonstrated on a modular process plant, the framework reduces manual effort by unifying engineering semantics, diagnostic relations, and machine-verifiable specifications.
knowledge graphlarge language modelsemantic web rule languagecause-and-effect logicmodular alignment ontology
Learning Structurally Consistent Representations for Multi-View Radar Semantic Segmentation
A unified higher-order structural alignment framework is proposed for multi-view radar semantic segmentation, addressing challenges from sparse, noisy, and weakly semantic radar measurements. The method employs learnable hypergraphs to capture higher-order dependencies among radar responses, aligns view-specific features using Unbalanced Optimal Transport (UOT) for consistency across heterogeneous projections, and fuses complementary views via an adaptive attention mechanism. Supervised segmentation and cross-view consistency regularization are used for training. Evaluations on CARRADA and RADIal benchmarks show improvements of +1.7 and +2.3 mIoU, achieving 63.8% and 83.4% mIoU, respectively, demonstrating the efficacy of higher-order relational modeling in radar perception.
hypergraphsunbalanced optimal transportmulti-view segmentationradar perceptionhigher-order dependencies
Preserve the Hard, Regenerate the Rest: Uncertainty-Guided Synthetic Training Data Augmentation with Diffusion Models
We propose an uncertainty-guided synthetic context augmentation strategy for semantic segmentation that preserves label validity while maximizing pixel informativeness. The method identifies uncertain semantic regions using a baseline segmenter's predictive entropy, inpaints complementary visual context via diffusion models, and computes loss only over original pixels during fine-tuning. This focuses learning on uncertain regions presented in novel contexts. Experiments on Cityscapes, UAVID, and BDD100K demonstrate substantial mIoU gains, particularly for rare and difficult classes like buses, trains, and cars.
semantic segmentationuncertainty-guided augmentationdiffusion modelspredictive entropycontext inpainting
Token-Sparse Medical Multimodal Reasoning via Dual-Stream Reinforcement Learning
The paper introduces ViToS, a dual-stream reinforcement learning framework for token-sparse medical multimodal reasoning. The method jointly optimizes visual token pruning (VTP) and question answering through cross-feedback sequential optimization, addressing gradient conflicts in shared policy learning. Evaluated on seven benchmarks, ViToS reduces visual tokens to 77% of original length while achieving 108.27% and 104.16% relative performance on Lingshu-7B and HuatuoGPT-Vision-7B, respectively, demonstrating improved efficiency and accuracy.
visual token pruningdual-stream reinforcement learningmedical multimodal reasoningcross-feedback optimizationtoken-sparse inference
Comparative Analysis of Machine Learning based Intrusion Detection in Realistic IoT Networks
This paper presents a comparative analysis of machine learning-based intrusion detection systems for IoT networks, evaluating five algorithms on the Gotham2025 dataset. The dataset, generated using the Gotham testbed, comprises 78 emulated IoT devices employing MQTT, CoAP, and RTSP protocols. The study compares Random Forest, XGBoost, Logistic Regression, Naive Bayes, and Deep Neural Network models for attack classification. Results show that Random Forest outperforms other models, achieving an F1-score of 0.99 in detecting attacks.
intrusion detectioniot networksgotham2025random forestf1-score
Evil Spectra: How Optimisers can Amplify or Suppress Emergent Misalignment
The study systematically characterizes how optimiser choice affects emergent misalignment (EM) in LLMs, finding a 7x spread in misalignment rates across optimisers while model scale (1B-235B) and family show negligible impact. Through sweeps across Qwen3 models and 12 models from three families using Adam, the authors identify final log training loss as a strong alignment predictor, with optimiser trajectories in loss-alignment space becoming dominant post-training. Muon optimiser preserves alignment best by implicitly regularising LoRA adapter singular values; spectral regularisation mitigates EM in Adam and Lion with minimal loss impact.
emergent misalignmentoptimiser sensitivityspectral regularisationlora adaptertraining dynamics
ZEBRA: Zero-Shot Entropy-Regularized Prompt Learning for Base-to-Novel Generalization in Audio-Language Models
The paper introduces ZEBRA, a zero-shot entropy-regularized prompt learning framework that addresses the base-to-novel generalization gap in Audio-Language Models (ALMs). ZEBRA combines zero-shot and prompt-learning logits while applying self-entropy regularization to mitigate overfitting to base classes. Evaluations across multiple audio classification datasets demonstrate that ZEBRA consistently enhances novel-class performance without compromising base-class accuracy, significantly narrowing the base-to-novel performance gap compared to standard prompt learning.
audio-language modelszero-shot learningprompt learningentropy regularizationbase-to-novel generalization
DPPE: Rethinking Camera-Based Positional Encoding for Scaling Multi-View Transformers
The paper introduces Decoupled Pose Positional Encoding (DPPE), a novel camera-based positional encoding method for multi-view Transformers that addresses performance stagnation in scaled-up training. DPPE explicitly decouples rotation and translation components, resolving indeterminacy issues when these parameters share dimensions in value vectors. Evaluations on novel view synthesis tasks demonstrate DPPE's stability in long-term training and superior generalization to extrapolation scenarios, including increased viewpoints and zoom-in settings.
positional encodingmulti-view transformersnovel view synthesiscamera parametersscalability
Which Tokens Matter? Adaptive Token Selection for RLVR with the Relative Surprisal Index
The paper introduces the Relative Surprisal Index (RSI), an information-theoretic metric combining token entropy and selected-token probability to optimize policy dynamics in Reinforcement Learning with Verifiable Rewards (RLVR). RSI-S, an entropy-adaptive token filtering method based on RSI, reconciles conflicting paradigms by filtering redundant low-surprisal and unstable high-surprisal tokens. Evaluations on Qwen2.5 models (1.5B–7B parameters) show RSI-S improves avg@32 accuracy by 2–3 percentage points over GRPO on AIME and AMC benchmarks.
relative surprisal indexrlvrtoken selectionpolicy optimizationinformation-theoretic metric
Temperature Field Reconstruction of Tungsten Monoblock Divertor on EAST using Physics-aware Neural Operator Transformer
The paper proposes a Physics-aware Neural Operator Transformer (PNOT) for real-time temperature field reconstruction in tungsten monoblock divertors on EAST fusion devices. The method models boundary heat-flux relations as a structured graph using graph attention to capture spatial dependencies, incorporates a physics-aware neural operator module for heat diffusion modeling, and employs a gradient-constrained Sobolev regularization loss for physical consistency. Experiments demonstrate improved prediction accuracy while maintaining physical constraints compared to conventional FEM approaches.
neural operatorgraph attentionheat-flux modelingsobolev regularizationfusion devices
Mitigating Positional Leakage in 3D Masked Autoencoders for Robust Representation Learning
We propose MPL-MAE, a masked point learning framework that mitigates positional leakage in 3D masked autoencoders by addressing decoder over-reliance on positional information. The method introduces a recalibrated positional embedding module to suppress metric-dominant coordinate signals while preserving geometric topology, and a gated positional interface module to dynamically regulate positional injection during reconstruction. These components promote balanced interaction between spatial priors and semantic features. Experiments demonstrate MPL-MAE's competitive performance across downstream tasks, validating its effectiveness in learning robust and informative representations.
masked autoencodingpositional leakagesemantic representationgeometric topologyreconstruction
FLARE-AI: Flaw Reporting for AI
FLARE-AI introduces an open-source flaw reporting system for AI to address fragmentation in the reporting ecosystem. The authors audit 12 existing systems, identifying five design challenges (discoverability, scope, information collection, coordination, strict-liability guidance), and develop FLARE-AI through expert feedback from 49 professionals across 32 organizations. The system uses conditional logic for triage-relevant data collection and enables machine-readable report dissemination to multiple stakeholders, improving interoperability and remediation efficiency.
flaw reportinginteroperabilityconditional logicmachine-readablestrict-liability
ACE: Pluggable Adaptive Context Elasticizer across Agents
We propose Adaptive Context Elasticizer (ACE), a plug-and-play module addressing inflexibility in context management for LLM-based agents by elastically orchestrating historical step information. ACE employs a lossless message maintenance layer storing raw messages and compressed abstractions, and a context orchestration layer adaptively assigning elastic types (raw, abstract, drop) to each step based on task state. Evaluated across ReAct, DeepAgent, WebThinker, and MiroFlow frameworks without training or architectural changes, ACE consistently outperforms truncation and summarization baselines, delivering performance gains in all frameworks.
adaptive context elasticizerlossless message maintenancecontext orchestrationelastic typeplug-and-play module
CVE-TTP KG: Knowledge Graph Linking Software Vulnerabilities to Attack Behaviors
The work introduces CVE-TTP Knowledge Graph, a novel framework linking Common Vulnerabilities and Exposures (CVE) to attacker behaviors from the MITRE ATT&CK framework. Transformer-based models, including CySecBERT, are developed for behavior identification, achieving macro F1-scores of 87.71% for techniques and 96.16% for tactics. A pipeline-based approach yields macro F1-scores of 0.86 for entity extraction and 0.99 for relation extraction, while a span-based joint model achieves 0.78. The framework integrates 24,820 entities and 43,608 relations into a Neo4j-based Cyber Threat Knowledge Graph, enabling structured visualization of vulnerabilities and enhancing threat interpretation.
cve-ttp knowledge graphmitre att&ckcysecbertneo4jentity extraction
Improving multichannel speech enhancement through accurate room-acoustic simulations
This work demonstrates that high-fidelity room-acoustic simulations significantly improve multichannel speech enhancement performance compared to conventional geometric acoustics. The authors train SpatialNet on datasets augmented with different simulation methods: geometric acoustics (low-fidelity) versus wave-based and hybrid approaches (high-fidelity). Evaluation on measured data shows a 38% relative reduction in median word error rate when using high-fidelity simulations, establishing a direct link between simulation accuracy and downstream task performance.
speech enhancementroom-acoustic simulationgeometric acousticswave-based modelingmultichannel processing
Modality-Driven Search with Holistic Trace Judging for ARC-AGI-2
The paper introduces a solver for ARC-AGI-2, a few-shot visual reasoning benchmark, addressing the challenge of selecting correct reasoning traces among fluent but often incorrect LLM outputs. The method employs (i) modality-driven search, generating diverse candidates across text, image, and code channels, and (ii) holistic trace judging, where a judge model compares all candidates within a single long-context prompt. The solver achieves 72.9% accuracy on the ARC Prize semi-private evaluation set, outperforming GPT-5.2 Pro (54.2%) and Gemini 3 Pro (54.0%) by +18.7 percentage points, at $38.99 per task.
arc-agi-2modality-driven searchholistic trace judgingfew-shot reasoninglong-context prompting
A time-series classification framework for individual-level absenteeism prediction under severe class imbalance
A time series classification framework is proposed for individual-level absenteeism prediction, addressing severe class imbalance and enabling proactive workforce planning. The method separates historical attendance sequences from future absence labels, utilizing Binary Focal Loss (BFL) and Geometric Mean (G-Mean) loss calibrated by imbalance ratio ρ. Three deep learning architectures—LSTM, CNN, and LSTM-FCN—are evaluated, with LSTM-FCN achieving strong precision and specificity. Experiments demonstrate stable performance with batch sizes ≥64 and window sizes between 40-80 days, yielding balanced accuracy of approximately 80% on test data.
time series classificationclass imbalancebinary focal losslstm-fcnbalanced accuracy
On the Convergence of Self-Improving Online LLM Alignment
We propose SAIL-RevKL, a regularized variant of the Self-Improving Alignment (SAIL) algorithm, to address distribution shift in online LLM alignment. By incorporating a reverse Kullback-Leibler divergence penalty, SAIL-RevKL improves the optimization landscape, satisfying the Polyak-Lojasiewicz condition within bounded parameter spaces. Theoretical analysis establishes global convergence guarantees with near-linear sample complexity. Empirical evaluations demonstrate SAIL-RevKL's superiority over vanilla SAIL, achieving improved performance on both MuJoCo benchmarks and LLM alignment tasks.
self-improving alignmentdistribution shiftreverse kl divergencepolyak-lojasiewicz conditionsample complexity
FinPersona-Bench: A Benchmark for Longitudinal Psychometric Stability of Autonomous Financial Agents
The paper introduces FinPersona-Bench, a benchmark for evaluating Mandate Salience Decay (MSD) in autonomous financial agents based on LLMs. The benchmark uses synthetic markets to decouple price from fundamental value, testing three failure modes: irrational trading in calm markets, panic-selling during crashes, and value-ignoring during bubbles. Evaluation of 18 LLMs across three behavioral profiles shows MSD compounds over time, with a 4.4x widening gap between static and periodically re-grounded agents in crash scenarios. Re-grounding benefits conservative agents but harms aggressive ones in low-signal markets, suggesting profile-aware intervention strategies.
mandate salience decayautonomous financial agentssynthetic market simulationbehavioral profilellm evaluation
Design and Implementation of Agentic Orchestrations and Orchestration of Agents
The paper proposes a classification framework for agentic orchestration in business process management, focusing on balancing LLM-based agent autonomy with robustness and traceability. It introduces qualitative decision criteria and quantitative metrics across dimensions like task specificity, tractability, and correctness assurance. The framework is evaluated through multiple agentic implementations of a predictive light sensing scenario, demonstrating practical applicability for designing and assessing agent orchestrations.
agentic orchestrationllm-based agentsbusiness process managementtraceability metricsautonomy balancing
Surprise as a Signal for Plasticity and Metacognition
The paper proposes using prediction-error signals from a small predictor over frozen encoder latents for dual roles: gating plasticity and enabling metacognition. Two systems demonstrate this: (1) a non-parametric episodic memory that writes new concepts during high surprise, achieving 17.7-51.3 point retention gains on ImageNet classes with DINOv2/I-JEPA backbones and 91.6% 5-way 1-shot accuracy; (2) a vision-language model where surprise modulates response assertiveness (AUROC=0.966 for novelty detection), learning concepts from single utterances with 99.2% post-sleep recall. Both systems highlight limitations while advancing episodic memory and personalized VLMs.
prediction-error signalepisodic memorymetacognitionvision-language modelconsolidation phase
Robustness of Robotic Manipulation: Foundations and Frontiers
The paper presents a systematic framework for analyzing manipulation robustness in robotics, defined as a system's ability to achieve goals under uncertainty. It offers probabilistic and control-theoretic formulations, surveys robustness mechanisms across perception, planning, control, policy learning, and hardware, and reviews evaluation metrics. The work synthesizes principles from foundational and recent studies, providing design guidelines and identifying open challenges toward human-level robustness.
robotic manipulationrobustness metricsuncertainty quantificationcontrol-theoretic formulationpolicy learning
One Reflection Is Not Enough: Self-Correcting Autonomous Research via Multi-Hypothesis Failure Attribution
The paper introduces SAGE, a Self-correcting, Autonomous, Grounded Experimenter, which addresses the failure-recovery bottleneck in autonomous research agents via Multi-Hypothesis Failure Attribution (MHFA). MHFA performs structured causal diagnosis by generating multiple evidence-grounded failure explanations, evaluating their severity, and routing root causes to appropriate intervention levels. SAGE also employs grounded reporting to prevent numerical hallucinations. Evaluated on a 12-topic, 5-domain benchmark, SAGE improves metrics-bearing outputs from 42% to 92%, artifact quality from 5.00 to 6.75/10, and outperforms AI-Scientist-v2 (52.0 vs. 48.2), particularly in code development and execution.
autonomous researchmulti-hypothesis failure attributionstructured causal diagnosisgrounded reportingfailure recovery
Von Mises Based Uncertainty Quantification for Closely Spaced Automotive Radar Targets
The study contributes a comparative analysis of uncertainty-aware deep learning methods for direction of arrival (DOA) estimation in automotive radar, focusing on geometric consistency versus statistical generality. It evaluates a von Mises (VM) ensemble (ENS) framework, parameterized by (mu, kappa), against an evidential deep learning (EDL) approach based on normal inverse gamma formulation. Performance is assessed under in-distribution and out-of-distribution conditions using risk coverage and ROC/AUROC metrics. Results show ENS achieves lower uncertainty under nominal conditions and greater sensitivity to perturbations, while EDL offers smoother uncertainty variation and improved ranking consistency. The VM-based ENS enables direct probabilistic integration into association modules via closed-form likelihoods.
direction of arrivalvon mises ensembleevidential deep learningnormal inverse gammaprobabilistic modeling
CLOUDADV: Decision-Aligned Instance Sizing with Zero-Shot Foundation Models under Drift
(No summary returned.)
Team MKC at CLPsych 2026: Capturing and Characterizing Mental Health Changes through Social Media Timeline Dynamics
The paper presents an LLM-based pipeline for comprehensive mental health analysis using sequentially ordered social media posts, addressing the CLPsych shared task. The method integrates post-level assessment with user-level temporal modeling, leveraging recent advances in Large Language Models for scalable mental health monitoring. The unified framework aims to support early detection and continuous tracking of psychological well-being through social media timeline dynamics.
large language modelsmental health analysistemporal modelingsocial mediaclpsych
CSTrader: A Testbed for Language-Grounded Trading in a Community-Driven Virtual Asset Market
CSTrader introduces a multi-agent framework for language-grounded trading in the volatile Counter-Strike 2 (CS2) weapon skin market, integrating heterogeneous signals from unstructured text. The system employs specialized agents for technical analysis, liquidity, events, and sentiment, combined with risk control and transaction friction modules. Evaluated on real CS2 data during high volatility, CSTrader achieves a 7.58% cumulative return, outperforming a falling market index (-15.62%) and single-prompt LLM baselines, with liquidity and reversed sentiment agents proving critical for stable profits.
multi-agent frameworklanguage-grounded tradingvolatile asset marketssentiment analysistransaction friction
UniTac: A Unified Multimodal Model for Cross-Sensor Tactile Understanding and Generation
UniTac introduces the first unified multimodal model for tactile understanding and generation, addressing the gap in integrating cross-sensor tactile data. It models tactile processes through a dual-level representation encoding both sensor and object attributes, enabling reasoning over physical and cross-sensor information. For understanding, UniTac performs object property description and sensor identification; for generation, it employs a two-stage training paradigm with reconstruction, alignment, and sensor-prior-based sampling. Trained on large-scale multi-sensor datasets, UniTac achieves state-of-the-art performance in tactile understanding and generates realistic tactile signals across diverse sensors.
unified multimodal modeltactile understandingdual-level representationsensor-prior-based samplingcross-sensor
Who Determines the Meaning of an Emotion? Affective Sovereignty as an Epistemic Consequence of Measurement Limits
The study introduces the concept of affective sovereignty as a normative principle for emotion-sensing AI, arguing that interpretive authority over emotional meaning must reside with the experiencing subject due to epistemic limitations in measurement. The authors define a meaning distribution as the distribution of labels assigned by annotators under a fixed protocol, decomposing its uncertainty into reducible and irreducible components. They demonstrate that while emotion AI can assign high-confidence labels and detect aggregate differences, the irreducible component of meaning distributions for individual instances cannot be adequately estimated, creating an epistemic gap. This gap necessitates reserving interpretive authority for the subject, shifting design priorities from accuracy maximization to explicit authority allocation.
affective sovereigntymeaning distributionepistemic gapemotion-sensing aiinterpretive authority
CDR-Bench: Evaluating Faithful Execution of Compositional, Order-Sensitive Data Refinement Recipes
The paper introduces CDR-Bench, a benchmark with 3,462 tasks across four domains to evaluate LLMs' ability to faithfully execute compositional, order-sensitive data refinement recipes. It tests models in atomic, order-agnostic, and order-sensitive settings using deterministic reference outputs. Experiments on 10+ state-of-the-art LLMs show performance degradation in compositional settings and collapse in order-sensitive recipe success, revealing current models' lack of procedural faithfulness for reliable data refinement.
data refinementcompositional executionorder-sensitive processingllm evaluationprocedural faithfulness
Ask the World Before Acting: Budgeted Environment Probing for World-Model Calibration
The paper introduces Budgeted Environment Probing (BEP), a mechanism for long-horizon language agents to calibrate their world models by querying the environment before committing to task actions. BEP treats environment interaction as a scarce resource, allowing agents to probe one belief field and update their world model accordingly. The method distinguishes between procedural and spatial beliefs, showing that targeted checks for procedural beliefs are effective but costly, while spatial beliefs rely more on structural cues. Controlled experiments demonstrate that mid-planning environment evidence reduces terminal world-model error when probe policies align with task structure.
budgeted environment probingworld-model calibrationprocedural beliefsspatial beliefsmid-planning evidence
DA-Studio: An Agentic System for End-to-End Data Analysis
DA-Studio introduces an agentic system for end-to-end data analysis, addressing limitations of existing LLM-based tools by autonomously organizing multi-step workflows, executing code in a sandboxed environment, and maintaining inspectability through visible action traces and intermediate artifacts. The system integrates an action-structured analysis backend, a sandboxed execution workspace, and a browser interface for task setup, streamed action traces, artifact preview, code editing, and report export. It incrementally constructs executable analysis steps from raw files and natural-language requests via iterative action generation, code execution, and feedback incorporation, exposing intermediate results throughout the process.
multi-step workflowssandboxed executionaction tracesintermediate artifactsnatural-language requests
Temporal Preservation over Processing: Diagnosing and Designing Spatiotemporal Single-Stage Video Detectors
The paper introduces TemporalLens, a diagnostic framework for assessing temporal dependence in single-stage video object detectors, and proposes YOLO-3D, a spatiotemporal detector. TemporalLens employs perturbations, occlusions, and shuffling to reveal whether models genuinely use temporal context, showing stacked 2D detectors fail without target frames while spatiotemporal models recover predictions. YOLO-3D, built on YOLOv8, achieves +3.7 pp mAP@50 at 32 frames by preserving temporal depth in the backbone, demonstrating temporal reasoning is measurable and actionable.
temporal dependencevideo object detectionspatiotemporal modelingdiagnostic frameworkyolo-3d
BP-TTA: Balanced and Prototype-Guided Test-Time Adaptation in Dynamic Scenarios
BP-TTA introduces a balanced and prototype-guided test-time adaptation method addressing class imbalance and continual domain shifts in dynamic scenarios. The approach combines batch-balanced sampling, integrating current samples with high-confidence historical instances, and maintains evolving class prototypes during inference. Prototype similarity serves as a constraint for model adaptation, enhancing pseudo-label reliability and update stability. Extensive experiments show BP-TTA outperforms state-of-the-art TTA methods in dynamic test-time streaming settings.
test-time adaptationclass imbalancedomain shiftsprototype-guidedonline updates
Learning to Select, Not Relearn: Hard-Routed Mixtures of Reasoning LoRAs
The paper introduces Hard-Routed MoR-LoRA, a two-stage framework for composing frozen reasoning LoRA experts via unit-scale hard selection. First, domain-specific LoRA adapters are independently trained using reinforcement learning from verifiable feedback. Then, a lightweight shared router and small attention LoRA are trained to integrate frozen experts through hard top-1 routing per token, enabled by a straight-through estimator. Experiments across five benchmarks show the method preserves expert behavior while requiring fewer parameters than soft-routing baselines, with analysis revealing soft mixtures often concentrate routing mass on single experts.
lora adaptershard routingreasoning expertsstraight-through estimatorparameter efficiency
Xiaomi-GUI-0 Technical Report
The paper introduces Xiaomi-GUI-0, a native multimodal GUI agent for real mobile environments, addressing the gap between benchmark performance and real-world usability by operating within a real-device closed loop. The method employs a hybrid infrastructure combining physical devices and sandboxes, multi-source training data (head tasks, long-tail intents, capability enhancement), and an error-driven data flywheel, trained via supervised fine-tuning, step-level RL, and agentic RL. Evaluations show 72.0% success on RealMobile and 78.9% on AndroidWorld, with improved stability and abnormal-state recognition.
gui agentreal-device closed loopmultimodalerror-driven data flywheelagentic reinforcement learning
Visual Semantic Entropy: Do Vision Language Models Recognize Visual Ambiguity?
The paper introduces Visual Semantic Entropy (VSE), a method for quantifying uncertainty in vision-language models (VLMs) by isolating visual ambiguity from textual prompt sensitivity. VSE perturbs only the image input while keeping the text query fixed, then clusters generated answers into semantic prototypes to compute mass-weighted dispersion. Evaluations across five VLMs (e.g., CLIP, Flamingo) and five VQA benchmarks show VSE outperforms existing entropy-based and perturbation methods in capturing visual ambiguity. The approach addresses limitations of Semantic Entropy (SE) and joint text-image perturbations, which often conflate visual uncertainty with textual variability.
visual semantic entropyvision-language modelsuncertainty estimationsemantic prototypesvisual ambiguity
Wisdom Of The (AI) Crowd: Investigating Artificial Swarm Intelligence In Large Language Models
The study demonstrates that large language models (LLMs) can approximate human swarm intelligence through artificial aggregation, achieving error reductions up to 37 percentage points (MAPE). Using 960 manual prompts across GPT-5, Gemini 2.5 Pro, and Claude Sonnet 4.5, the authors tested intra-model sampling and inter-model aggregation on eight estimation tasks. Results show significant error reduction and metacognitive awareness (Spearman's ρ=0.242-0.568) via confidence interval analysis, suggesting practical applications for LLM swarms in decision-making.
artificial swarm intelligencelarge language modelserror reductionmetacognitive awarenessaggregation strategies
World-Model Collapse as a Phase Transition
The study identifies world-model collapse as a measurable bottleneck in long-horizon language agents, analogous to phase transitions in physical systems. Through a grid search across parameters including state cardinality, dependency density, horizon, branching, observation mode, and mutation rate, the authors map a phase diagram comprising a solved plateau, transition band, and collapse floor. Per-step trace analysis reveals that world-state fidelity degrades before action validity, indicating agents act from corrupted internal representations rather than merely selecting suboptimal actions. While stronger models shift the critical boundary, the qualitative transition persists. This work establishes world-model collapse as a fundamental limitation in sequential decision-making agents.
world-model collapsephase transitionlong-horizon agentsstate fidelitygrid search
Mixture-of-Control: State-Aware Fine-Tuning for Transformer-based Models
Mixture-of-Control (MoC) introduces a state-aware fine-tuning framework for transformers that combines local and global control signals via sparse mixture-of-experts. Unlike per-block state updates, MoC enables efficient cross-block communication while maintaining memory and computational efficiency. Evaluations across multiple transformer benchmarks show MoC outperforms existing state-based methods without significant overhead.
state-based fine-tuningtransformer adaptationmixture-of-expertscontrol signalsparameter efficiency
Resolving superposition in AI for interpretability and cross-modal alignment in patient-neuronal images
The study resolves superposition-induced interpretability challenges in neural networks analyzing high-dimensional biological data by employing sparse autoencoders (SAEs) on 100,000+ multiplexed images of Parkinson's disease and healthy neurons. The method theoretically and empirically demonstrates that SAEs restore geometric fidelity in latent spaces corrupted by superposition. By adapting single-cell RNA sequencing (scRNA-seq) methodologies to image data and introducing GW-map for Gromov-Wasserstein optimal transport alignment, the approach reconstructs neuronal pathology pathways like Calcium-AIS scaffold without spatial transcriptomics references.
sparse autoencoderssuperpositiongromov-wasserstein optimal transportlatent space geometrysingle-cell rna sequencing
ReGRPO: Reflection-Augmented Policy Optimization for Tool-Using Agents
ReGRPO (Reflection-augmented Group Relative Policy Optimization) enhances tool-using vision-language models by integrating reflection-guided correction. The framework employs a structured reflective data engine to collect failure observations and generate Reflection-of-Thought triplets (ErrorType, Evidence, FixPlan) paired with corrected actions for supervised fine-tuning. It jointly optimizes reflection tokens and corrective actions using group-relative advantages, incorporating a reflection-cost term to minimize unnecessary reflection. Evaluations on GTA and GAIA benchmarks demonstrate ReGRPO's superiority over open-source baselines, achieving state-of-the-art performance with the same backbone and tool suite.
reflection-augmentedvision-language modelsgroup-relative advantagessupervised fine-tuningreflection-of-thought
Stage-Transition Dense Reward Modeling for Reinforcement Learning
Stage-Transition Dense Reward (STDR) introduces a visual reward-learning framework for long-horizon robotic manipulation tasks, addressing sparse reward challenges by converting unstructured expert videos into dense rewards. STDR employs semantic understanding to infer task stage structures from demonstrations, providing stage-transition feedback and within-stage progress feedback during training. The framework integrates out-of-distribution detection and grasping regulation modules to enhance robustness and prevent reward hacking. Evaluations across 14 tasks in MetaWorld, ManiSkill, and Franka Kitchen demonstrate improved sample efficiency and success rates, matching or surpassing handcrafted dense rewards. Real-robot experiments confirm STDR's robustness to visual noise and calibrated reward assignment.
stage-transition dense rewardsemantic understandingout-of-distribution detectionreward hackingrobotic manipulation
Calibrating the Evaluator: Does Probability Calibration Mitigate Preference Coupling in LLM Agent Feedback Loops?
This work investigates probability calibration as a mitigation strategy for evaluator preference coupling (EPC) in LLM agent feedback loops. The study compares standard binary TTRL with confidence-calibrated TTRL using DeepSeek-V4-Pro as executor and GLM5.2 as evaluator, finding calibration reduces the coupling coefficient γ by 20-49% and Jensen-Shannon divergence by 45-67%. A symmetric-LR control confirms the effect stems from calibration rather than update asymmetry. The authors release the calibrated TTRL protocol as a lightweight solution for LLM-as-judge pipelines.
preference couplingprobability calibrationttrlllm-as-judgejensen-shannon divergence
From Materials Database to Materials Bank: Assetizing Data for AI Driven Materials Innovation
The authors propose a Materials Bank framework to bridge the gap between materials data accumulation and industrial innovation by transforming conventional databases into value-filtering systems. The framework introduces a multi-dimensional BankCard system that evaluates materials based on scientific validity, synthesis feasibility, application readiness, and industrial value, creating standardized, upgradable materials assets. This approach integrates databases, AI models, automated experimentation, and multi-criteria assessment into a closed-loop ecosystem, facilitating the transition from raw data to industrial products. The Materials Bank aims to accelerate AI-driven materials innovation by providing a scalable decision infrastructure that aligns academic discovery with industrial demands.
materials bankbankcard frameworkautomated experimentationmulti-criteria assessmentclosed-loop ecosystem
PGUDA: Pressure-Guided Unsupervised Domain Adaptation with Cross-Modal Knowledge Distillation for sEMG-Based Gesture Recognition
The paper proposes PGUDA, a pressure-guided unsupervised domain adaptation framework for sEMG-based gesture recognition, addressing performance degradation from cross-subject/session distribution shifts. The method employs cross-modal knowledge distillation, where a teacher network trained on robust pressure signals guides an sEMG student network on unlabeled target domains, enforcing modality-invariant representations. Experiments on an 11-subject multimodal dataset show PGUDA achieves 58.08% average accuracy in cross-domain tasks, outperforming existing DA methods while requiring only 5% labeled data for teacher training.
domain adaptationknowledge distillationsemgcross-modal learninggesture recognition
Smart charging of large fleets of Electric Vehicles: Independent Multi-Agent Reinforcement Learning approaches
The paper compares two independent multi-agent reinforcement learning approaches for decentralized electric vehicle charging optimization: contextual combinatorial bandits and policy gradient algorithms. Using a realistic simulation with autonomous agents making decisions based on local environmental information (price signals, state-of-charge, temporal constraints), the methods are evaluated under varying congestion levels and mixed-strategy configurations with heterogeneous agent groups. The evaluation employs dynamic electricity pricing derived from real photovoltaic production data to assess performance in minimizing user costs while avoiding network overloads.
multi-agent reinforcement learningelectric vehicle chargingcontextual banditspolicy gradientdecentralized optimization
Beyond Binary Instrument QA: Probing Instrument Grounding in Music Audio-Language Models
The paper introduces a diagnostic benchmark for evaluating instrument grounding in music audio-language models, moving beyond binary instrument-presence QA. The benchmark tests genre-prior reduction, confusable instrument discrimination, long-context understanding, and temporal localization. Results reveal that high binary QA accuracy masks critical failures: models exhibit option-position bias, confusable-instrument errors, and temporal response bias. The findings advocate for multi-axis evaluation over aggregate accuracy metrics in assessing audio grounding capabilities.
audio-language modelsinstrument groundingdiagnostic benchmarktemporal localizationconfusable instruments
Optimization Algorithms for Joint OFDM Waveform Design and RIS Configuration in 6G Networks: From Convex Relaxation to Foundation Models
The survey analyzes 78 joint OFDM-RIS optimization works (2021-2026) for 6G networks, classifying them into four paradigms: convex relaxation, metaheuristics, deep reinforcement/unsupervised learning, and emerging methods (foundation models, diffusion AI, quantum optimization). ML-based methods achieve 95-99% of model-based spectral efficiency at 10^2-10^4x faster inference speeds, with GPU-based neural networks showing N-invariant runtime scaling versus polynomial scaling for iterative solvers. Key challenges include benchmark standardization, hardware constraints, doubly-dispersive channels, multi-objective PAPR trade-offs, and LLM safety in network control.
ofdm-ris optimization6g networksmixed-integer nonlinear programmingfoundation modelsspectral efficiency
CryoACE: An Atom-centric Framework for Accurate and Automated Model Building in Cryo-EM
CryoACE introduces an atom-centric framework for automated protein model building from cryo-EM density maps, addressing challenges in physicochemical validity and conformational heterogeneity. The method employs direct atomic coordinate sampling with iterative feature recycling for efficient multimodal fusion, replacing voxel convolutions, and incorporates a training-free guidance mechanism using local resolution priors to resolve dynamic ambiguity. Evaluated on a high-quality dataset, CryoACE outperforms baselines on static benchmarks and reveals atomic-level dynamics on complex datasets like EMPIAR-10345 without pre-built structures.
cryo-ematom-centricmultimodal fusionconformational heterogeneitylocal resolution priors
3D HAMSTER: Bridging Planning and Control in Hierarchical Vision Language Action Models through 3D Trajectory Guidance
3D HAMSTER introduces a hierarchical vision-language-action framework that bridges planning and control by generating metrically reliable 3D trajectories for robot manipulation. The method augments a Vision-Language Model (VLM) with a depth encoder and dense depth reconstruction objective to predict 3D waypoint sequences, which are integrated into a pointcloud-based low-level policy. Evaluations across 3D trajectory prediction, simulation, and real-world tasks demonstrate superior performance over proprietary VLMs and 2D-guided baselines, particularly under appearance-altering shifts and unseen language, spatial, and visual conditions.
hierarchical vla3d trajectory guidancedepth encoderpointcloud-based policymetric reliability
HistoriQA-ThirdRepublic: Multi-Hop Question Answering Corpus for Historical Research, Parliamentary Debates from the French Third Republic (1870-1940)
HistoriQA-ThirdRepublic introduces a French-language multi-hop QA dataset (1,782 questions) derived from parliamentary debates and newspapers of the French Third Republic (1870-1940). Developed with historian input, it captures complex reasoning patterns like cross-source synthesis and temporal reasoning across heterogeneous documents. The corpus supports evaluation of retrieval-augmented and LLM systems in historical contexts, with methodology adaptable to other languages and national corpora.
multi-hop qahistorical reasoningretrieval-augmented systemstemporal reasoningdomain-specific evaluation
From Idea to Prototype in an Afternoon: Scaffolded, AI-Assisted Rapid VA Prototyping
The authors demonstrate a rapid visual-analytics prototyping method that reduces development time from months to an afternoon by combining the Artifact–Transform Workflow Language (ATWL) scaffold with AI assistance. Their case study tested a novel Pareto frontier relaxation technique called 'soft sky constellations'. Controlled experiments revealed that ATWL alone produced naive workflows, while combining it with expert knowledge injection achieved state-of-the-art quality. Results show that scaffold timing matters: introducing ATWL after initial unconstrained design outperforms simultaneous provision of language definitions and example libraries. The authors advocate for developing a typology of human knowledge injection that balances machine accessibility with human editability.
artifact-transform workflow languagepareto frontiervisual-analyticsknowledge injectionrapid prototyping
CSO-LLM: Class Subspace Orthogonalization for Post-Training Backdoor Detection and Trigger Inversion in LLMs
The paper introduces Class Subspace Orthogonalization (CSO), a novel framework for post-training backdoor detection and trigger inversion in LLMs. The method addresses challenges in discrete input spaces (up to 150,000^k k-tuples) and lack of comprehensive token blacklists by penalizing trigger tokens that induce perturbations toward the target class. Two variants are proposed: continuous optimization in embedding space and greedy accretion in discrete token space. Results demonstrate strong detection performance and accurate trigger inversion across multiple LLM architectures and classification domains.
backdoor detectiontrigger inversionclass subspace orthogonalizationllm securitydiscrete optimization
Benchmarking Large Language Models on Floating-Point Error Classification
The paper introduces InterFLOPBench, a benchmark of 90 C kernels with 1,130 test samples, to evaluate Large Language Models (LLMs) in detecting and classifying six categories of floating-point errors: cancellation, comparison, division by zero, overflow, underflow, and NaN. The evaluation treats error detection as a multi-label classification problem, using F1-score as the performance metric. Results show that state-of-the-art models (Qwen 3 32b, Gemini 2.5 Flash, Phi 4 Reasoning, DeepSeek R1T2, gpt-oss 20b, and 120b) achieve an overall F1-score greater than 0.88. Performance varies across error categories, with explicit operations like division by zero (F1: 0.8479) outperforming subtle phenomena like underflow (F1: 0.6059) and cancellation (F1: 0.6164).
floating-point errorsmulti-label classificationf1-scoreinterflopbenchc kernels
Minimizing Quantized Semantic Age of Information (QSAoI) in Foundation Model-Based Semantic Communications
The paper introduces Quantized Semantic Age of Information (QSAoI), a novel metric for semantic-aware resource allocation in 6G networks, addressing the gap between semantic and physical layers under finite blocklength constraints. A foundation model-based framework jointly optimizes mixed-precision quantization (MPQ) and physical blocklength via fixpoint inspection and bisection search. Simulations demonstrate adaptive quantization precision under varying channel conditions, reducing expected QSAoI by 12-18% versus baselines.
quantized semantic age of informationfinite blocklengthmixed-precision quantizationsemantic communication6g networks
Spatial Reasoning via Modality Switching Between Language and Symbolic Representation
The paper introduces a modality-switching framework for spatial reasoning that alternates between natural language and structured symbolic representations (e.g., grids) based on trustworthiness and complexity signals. It demonstrates that grounding multi-hop textual-spatial stories into geometry-aware modalities improves reasoning accuracy by up to 42% compared to pure language-based inference in LLMs. The method proposes a switching metric to dynamically select optimal modalities during problem-solving, advancing principled modality selection in multimodal reasoning systems.
modality switchingspatial reasoningsymbolic representationmulti-hop inferencegeometry-aware
CLIMB: Centroid-Based Hierarchical Memory for Online Continual Self-Supervised Learning
The paper introduces CLIMB (Continual Learning with Intelligent Memory Bank), a method for Online Continual Self-Supervised Learning (OCSSL) that combines centroid-based hierarchical memory with knowledge distillation. CLIMB groups similar images into centroids within a bounded memory bank, providing hard negatives for contrastive learning while maintaining distribution diversity, and applies distillation to mitigate representation drift. Evaluations on Split CIFAR-100 and Split ImageNet-100, including a novel irregular task distribution protocol, demonstrate CLIMB's superiority over state-of-the-art OCSSL approaches.
online continual learningself-supervised learningcentroid-based memoryknowledge distillationcontrastive learning
Learning from Failure: Inference-Time Self-Improvement for Computer-Use Agents
The paper introduces a failure-driven self-improvement loop for computer-use agents, complementing standard success-based fine-tuning by leveraging failed trajectories. The method employs an LLM to diagnose failure modes, generate inference-time solutions, and produce human-verified code patches, upgrading the agent without additional training. Evaluated on OpenCUA-72B with OSWorld, the approach improves success rates from 42.3% to 48.9% (6.6pp gain) with minimal inference overhead, demonstrating efficient agent improvement.
computer-use agentsfailure-driven learninginference-time improvementmultimodal llmsosworld benchmark
TDGT: A Tabular Data Generation Toolkit supporting adaptive GPU-accelerated Bayesian mixture models, diffusion-based models, and latent-space generative modeling
TDGT introduces a web-based toolkit for synthetic tabular data generation, featuring three key innovations: (1) Adaptive Bayesian Mixture Synthesizer (ABMS) for automatic mixture component selection, (2) VAE-ABMS hybrid architecture combining variational autoencoders with adaptive mixtures for nonlinear distributions, and (3) GPU-accelerated ABMS via CUDA-optimized k-means and Gaussian mixture fitting. The system evaluates fidelity using 11 statistical metrics and privacy risks via k-anonymity scoring. Experiments across healthcare, socioeconomic, and cybersecurity domains demonstrate consistent statistical coherence and generation quality for heterogeneous data types and scales.
bayesian mixture modelsvariational autoencodertabular data generationgpu-accelerationprivacy-preserving
SwiftAudio: Data-Efficient Caption-Only Distillation for One-Step Text-to-Audio Diffusion-based Generation
SwiftAudio introduces a one-step text-to-audio (TTA) generation framework that performs audio-free distillation from a pretrained diffusion teacher using only text captions, eliminating the need for paired text--audio data. The method adapts Variational Score Distillation (VSD) to the audio domain and incorporates temporal smoothness regularization to ensure coherent latent audio representations. Trained on approximately 45K captions, SwiftAudio achieves state-of-the-art performance among strict one-step methods on AudioCaps and Clotho benchmarks, significantly reducing the gap to multi-step diffusion systems.
text-to-audiodiffusion modelsvariational score distillationtemporal smoothnesslatent representations
Embodied CAD: Solver-Grounded LLM Agents for Parametric B-Rep Assembly Modeling
Embodied CAD introduces solver-grounded LLM agents for parametric boundary representation (B-Rep) assembly modeling, addressing the limitations of single-pass CAD script generation. The framework employs iterative action selection from a stratified CAD skill library, resolves actions into geometric operations, executes them in a CAD backend, and uses solver feedback for planning, repair, and learning. It combines action grammar constraints, deterministic parameter resolution, and solver-derived rewards for supervised warm-up and GRPO-style refinement. Evaluation on mechanical, industrial equipment, and mold-oriented assembly tasks demonstrates high executable rates and exposes gaps between valid tool calls and exact long-horizon policy prediction.
parametric b-repgeometric kernelsolver-groundedcad skill librarygrpo-style refinement
Probing Stylistic Appropriation using Large Language Models: An Evaluation Framework for Copyright Infringement under EU Law
We introduce PSALM, a framework for evaluating copyright infringement risks of large language models (LLMs) under EU law, addressing stylistic appropriation beyond verbatim memorization. PSALM operationalizes EU copyright doctrine through ten evaluators assessing computational overlap, stylistic dimensions, content dimensions, and statutory exceptions. Applied to Llama~3.2 models fine-tuned on historical Dutch literature, results show instruction-tuned models exhibit baseline stylistic similarity, fine-tuning induces systematic stylistic appropriation across infringement-relevant dimensions, and Negative Preference Optimisation unlearning reduces but does not eliminate residual stylistic patterns. PSALM bridges qualitative legal standards and quantitative technical measurement, highlighting tensions between generative AI and EU intellectual property law.
stylistic appropriationnegative preference optimisationllm-as-a-judgesubstantial similaritycomputational overlap
Delta-JEPA: Learning Action-Sensitive World Models via Latent Difference Decoding
Delta-JEPA introduces a reconstruction-free world model that learns action-sensitive latent dynamics via Latent Difference Action Decoder (LDAD), which reconstructs actions from latent displacements between consecutive observations. This displacement-level supervision prevents embedding collapse and ensures distinguishable latent changes for planning. The method avoids pixel reconstruction and distribution-matching regularizers, relying solely on latent prediction and action reconstruction. Evaluated on four visual continuous-control tasks, Delta-JEPA outperforms JEPA-based and representation-learning baselines in planning. Ablations confirm the superiority of displacement-based action decoding over endpoint concatenation, with clearer action-conditioned latent responses observed.
latent difference action decoderreconstruction-freeaction-sensitiveembedding collapsecontinuous-control
Agentic-Ideation: Sample Efficient Agentic Trajectories Synthesis for Scientific Ideation Agents
The paper introduces Agentic-Ideation, a framework for training scientific ideation agents via sample-efficient trajectory synthesis. The method combines an automated pipeline with Oracle-Guided Data Synthesis, using reference ideas to direct multi-agent systems in reconstructing reasoning paths and tool invocations (3 external + 3 cognitive tools). Training employs masked tool execution results to focus on decision logic. Experiments show 11.91% quality improvement over workflow-based baselines and 10× better sample efficiency for high-quality data synthesis.
agentic llmstrajectory synthesisoracle-guidedscientific ideationtool utilization
Thinking Before Retrieving: Robust Zero-Shot Composed Image Retrieval via Strategic Planning and Self-Criticism
We introduce PEC-CIR, a training-free framework for zero-shot composed image retrieval that improves retrieval precision through structured query construction. The method employs a Planner–Executor–Critic architecture: the Planner extracts explicit constraints from the reference image and textual modification, the Executor generates multiple candidate target descriptions, and the Critic evaluates these candidates for constraint compliance. By reframing query construction as a multi-stage reasoning pipeline, PEC-CIR reduces generative error propagation and enhances retrieval stability compared to single-pass generation strategies. The approach operates entirely within a frozen vision–language embedding space, requiring no additional training.
composed image retrievalzero-shot learningvision-language embeddingmulti-stage reasoningtraining-free framework
Information-Aided DVL Calibration
The paper proposes information-aided calibration (IAC) for Doppler velocity log (DVL) sensors in autonomous underwater vehicles (AUVs), addressing two scenarios: GNSS-enabled and GNSS-denied environments. The method enhances conventional Kalman filter-based calibration by incorporating additional information sources, enabling both improved GNSS-assisted calibration and novel GNSS-free self-calibration. Experimental results on real-world AUV datasets demonstrate 20% average accuracy improvement in GNSS-enabled cases and 35% better velocity vector estimation in GNSS-free operation, reducing navigation drift and improving mission reliability.
doppler velocity logautonomous underwater vehiclekalman filtergnss-denied navigationsensor calibration
Can LLMs Imagine Moral Alternatives Beyond Binary Dilemmas?
The study introduces MoralAltDataset, a novel dataset of 307 moral dilemmas with compromise and reframed alternatives, to evaluate LLMs' capacity for moral imagination beyond binary choices. It compares human and LLM judgments across 15 models, finding that compromise alternatives are frequently preferred, altering moral decision landscapes. LLM-generated alternatives often outperform human-authored ones in pairwise preferences and expert evaluations, though they exhibit trade-offs between structural quality and practical feasibility.
moral dilemmaslarge language modelsmoral cognitioncompromise alternativespairwise preference
Long-term Traffic Simulation via Structured Autoregressive Modeling
The paper introduces RosettaSim, a unified framework for long-term traffic simulation that leverages frozen LLM components to model multi-agent interactions with dynamic token cardinality. By projecting scene topology, agent states, and spawning intents into a structured autoregressive stream, the method achieves SOTA performance on the Waymo Open Sim Agent Challenge (WOSAC) in both short-term accuracy and long-horizon fidelity. The proposed Retrieval-based Traffic Evaluation (RTE) metric shows stronger correlation (r=0.83) with simulation quality than existing approaches (r=0.74).
traffic simulationautoregressive modelingmulti-agent interactionretrieval-based evaluationlong-horizon fidelity
Towards Inclusive Mobility Modeling: Characterizing and Evaluating Elderly Trajectory Patterns in Urban Systems
This study quantifies demographic bias in urban mobility modeling by analyzing elderly trajectory patterns in Citi Bike System Data (2016-2020 Jersey City subset). Using synthetic trajectory generation with Markov chains and Qwen3-4B (QLoRA-fine-tuned), it reveals elderly exhibit distinct signatures: 958m activity spaces (vs. 1,189m young), 1.82 mobility entropy (vs. 4.15), and asymmetric off-peak patterns. Models trained on majority data misrepresent elderly behavior (4.5% step length, 8.9% dwell time errors), with LLMs not improving fidelity under data scarcity. Results highlight representation gaps in smart city applications.
trajectory data miningmobility entropysynthetic trajectory generationdemographic biasactivity space
Agentic RAG-VLM: Affordance-Aware Retrieval-Augmented Generation with Self-Reflective Planning for Robotic Grasping
Agentic RAG-VLM introduces a unified framework for robotic grasping that integrates retrieval-augmented generation (RAG) with vision-language models (VLMs) and agentic self-reflective planning. The framework comprises three components: Hierarchical Affordance-Aware RAG (HAA-RAG) for affordance-based strategy retrieval, Scene Graph Constraint Reasoner for spatial relationship analysis, and Agentic Self-Reflective Pipeline for failure recovery. Evaluated on a 12-task benchmark with 360 trials per configuration, Agentic RAG-VLM achieves 78.3% overall success, a 53.3 percentage-point improvement over VLM-only baselines, demonstrating the efficacy of affordance-aware retrieval, scene graph reasoning, and agentic recovery in robust manipulation.
retrieval-augmented generationvision-language modelsaffordance-awarescene graphself-reflective planning
Distilling Temporal Coherence into 2D Networks for Transrectal Ultrasound Prostate Video Segmentation
The paper introduces a Temporally Consistent Learning Framework for real-time prostate segmentation in Transrectal Ultrasound (TRUS) videos, addressing inter-frame inconsistency in 2D networks without 3D computational overhead. The method employs a Confidence-Weighted Temporal Consistency objective based on optical flow warping residuals to mitigate gradient propagation from unstable regions, alongside a Dual-scale Prototype Alignment Module for semantic coherence via contrastive optimization. It also uses geometric equivariance-based pseudo-labeling to reduce annotation dependency. Evaluated on SUN-SEG and TRUS-V (2,679 frames), the approach achieves state-of-the-art accuracy and temporal consistency at real-time speeds.
temporal consistencyoptical flowcontrastive optimizationpseudo-labelingtransrectal ultrasound
Gated Multi-Graph Fusion via Graph Attention Networks for Alzheimer's Disease Detection
The paper proposes a Multi-View Gated Graph Attention Network for Alzheimer's Disease (AD) detection from spontaneous speech, addressing clinical heterogeneity and non-linear structural disruptions. The method constructs semantic, dependency, and PMI-based co-occurrence graphs from ASR transcripts, integrating them via adaptive gated fusion to model content, structure, and narrative flow. Evaluated on ADReSSo, the model achieves 90.00% accuracy, with ablations confirming the necessity of PMI-based graphs and heterogeneity-aware gating for robust performance across diverse populations.
graph attention networkspointwise mutual informationadaptive gated fusionalzheimer's disease detectionmulti-view learning
Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
The paper introduces Bayesian in-context experimenters, transformer-based policies trained to imitate a Bayesian posterior Neyman teacher for efficient average treatment effect (ATE) estimation. The method employs attention-based sufficient statistics and projected gradient descent to approximate Bayesian updating with Gaussian-series priors, addressing unknown outcome smoothness via a mixture-of-experts transformer. Theoretical analysis bounds the transformer class complexity, enabling empirical risk minimization through supervised pretraining. Experiments demonstrate accurate teacher imitation, adaptive allocation, and enhanced ATE precision compared to baselines.
average treatment effectbayesian posteriormixture-of-expertsattention-based statisticsempirical risk minimization
AI-Assisted Discovery of Convex Relaxations via Dual Agents
We present an AI-assisted framework for discovering convex relaxations that yield certified lower bounds in optimization problems. The method employs a dual-agent architecture: a coding agent proposes tightening constraints, while a theory agent verifies validity and searches for counterexamples. Each reported bound is certified through explicit dual-feasible points verified using rigorous interval arithmetic. Applied to two optimization constants - the first autocorrelation inequality ($C_{6.2}$) and the Erdős minimum-overlap constant ($C_{6.5}$) - the approach improves certified lower bounds from 1.28 to 1.2937 and from 0.379005 to 0.37912, respectively.
convex relaxationsdual-feasible pointsinterval arithmeticoptimization constantsdual-agent architecture
HealthAgentBench: A Unified Benchmark Suite of Realistic Agentic Healthcare Environments for Challenging Frontier AI Agents
The paper introduces HealthAgentBench, a unified benchmark suite comprising 54 agentic healthcare tasks across 7 categories, designed to evaluate AI agents in realistic clinical workflows. Each task requires multi-step reasoning over raw healthcare data, with success rate as the primary metric. Evaluation of frontier models reveals low overall performance (42% for Codex GPT-5.5), with notable weaknesses in medical imaging and compositional reasoning tasks. The benchmark highlights significant challenges in agentic healthcare applications while providing a standardized evaluation framework.
healthcare aiagentic benchmarkingmulti-modal reasoningclinical workflowstask success rate
AETDICE: Unified Framework and Offline Optimization for Nonlinear Multi-Objective RL
We introduce AETDICE, a unified offline optimization framework for nonlinear multi-objective reinforcement learning (MORL) that bridges the Scalarized Expected Return (SER) and Expected Scalarized Return (ESR) paradigms. Our Aggregation-Expectation-Transformation (AET) framework decomposes scalarization into three components, enabling principled optimization of complex trade-offs like risk aversion and fairness. AETDICE leverages DICE-style density-ratio estimation in an augmented state space to perform sample-based optimization from static datasets. This approach resolves fragmentation in existing MORL methods and captures trade-offs induced by the AET framework, addressing long-standing barriers in nonlinear MORL optimization.
multi-objective reinforcement learningscalarized expected returnexpected scalarized returndensity-ratio estimationoffline optimization
ClawArena-Team: Benchmarking Subagent Orchestration and Dynamic Workflows in Language-Model Agents
ClawArena-Team introduces a benchmark for evaluating the management capabilities of large language-model (LLM) agents in orchestrating subagents through dynamic workflows. The benchmark comprises 41 multi-turn, multimodal scenarios with 258 evaluation rounds and 72 staged updates, focusing on the main agent's ability to delegate tasks and manage subagents under constrained conditions. Experiments across twelve models reveal that privilege granting is a significant bottleneck, with no model exceeding 50% workspace-permission precision. Cost and management quality are decoupled, with open models achieving competitive performance at lower costs. The Subagent-Management Score (SMS) highlights divergent orchestration behaviors despite clustered leaderboard scores.
llm agentssubagent orchestrationdynamic workflowsprivilege grantingexecution-based scoring
Cross-Domain Feature Expansion for Tabular Medical Data via Knowledge Graphs Injection
MedKGTab introduces a knowledge-injected framework for cross-domain feature expansion in tabular medical data, addressing data scarcity by inferring uncollected biomedical features. The method employs a row-column dual-attention mechanism to process raw structured tabular data directly, integrating data-driven statistical priors with the SPOKE biomedical knowledge graph. This synergy ensures generated data are grounded in empirical medical research. Experimental results show MedKGTab outperforms state-of-the-art medical large models and specialized tabular models, achieving high data fidelity and realistic representation across various data generation scenarios.
cross-domain feature expansiontabular medical datarow-column dual-attentionspoke knowledge graphdata-driven statistical priors
MIRTH: Mutual-Information Reasoning with Temporal Hubs for Vision-Language-Action Agents
MIRTH introduces a vision-language-action (VLA) framework addressing temporal myopia, reasoning gaps, and inference inefficiency in robotic control. The method augments pretrained VLAs with dual-scale temporal memory hubs for scene dynamics, latent reasoning tokens optimized via mutual information for semantic-action alignment, and parallel action decoding for control throughput. Evaluations on LIBERO and LeRobot show state-of-the-art performance with emergent error recovery. Code and datasets are publicly released.
vision-language-actiontemporal memory hubsmutual-information optimizationparallel action decodingemergent error recovery
ComplianceGate: Classifier-Gated Multi-Tier LLM Routing for Inference in Regulated Industries
The paper introduces ComplianceGate, a classifier-gated multi-tier routing system for LLM inference in regulated industries that enforces compliance and optimizes cost. The method employs a trained encoder classifier to pre-screen queries for complexity and data sensitivity before routing them to appropriately sized dense models in compliant geographic locations, preventing PII leaks by design. Evaluated on 600 queries, the system reduces median latency by 39%, achieves 33-52% cost savings, and maintains 99.2% classifier accuracy with 7ms overhead while generating 122-200 tokens/second versus 50-64 for baselines.
llm routingcompliance enforcementpii detectionmulti-tier inferencegeographic routing
LLM-Powered Interactive Robotic Action Synthesis from Multimodal Speech, Gestures, and Music
This paper presents a novel framework for synthesizing complex robotic actions from multimodal human inputs using Large Language Models (LLMs). The system integrates speech transcription, gesture recognition, and beat detection modules, whose outputs are contextualized via prompt templates and processed by an LLM. The LLM reasons over these inputs, constrained by a predefined robot action space, to generate coherent action sequences executed on a quadruped robot via ROS. This approach enables interpretation and fusion of semantic commands, deictic information, and rhythmic cues, advancing fluid and context-aware human-robot interaction.
large language modelsmultimodal fusiongesture recognitionaction synthesishuman-robot interaction
One Retrieval to Cover Them All: Co-occurrence-Aware Knowledge Base Reorganization for Session-Level RAG
The paper introduces a co-occurrence-aware knowledge base (KB) reorganization method for session-level retrieval-augmented generation (RAG), addressing the limitation that standard RAG retrieves documents optimized for single queries rather than coherent question sessions. The approach clusters KB articles offline based on co-occurrence patterns and expands retrieval candidates through cluster neighborhoods at query time. Evaluated on WixQA (6,221 enterprise support articles), the method improves single-query session coverage from 41% to 58% (+17% absolute), reduces retrieval calls for 70% coverage by 34%, and compresses the KB to 20% of original size, demonstrating consistent gains across four embedding models and six domains.
retrieval-augmented generationknowledge base clusteringsession-level retrievalco-occurrence patternsenterprise qa
PPT-Eval: A Benchmark for Computer-Use Agents on PowerPoint Tasks
PPT-Eval introduces a benchmark of 120 PowerPoint tasks across 12 files, addressing content creation and presentation editing scenarios with varying difficulty. The benchmark employs a robust evaluation framework featuring task-specific rubrics that award partial credit for intermediate steps, penalize unnecessary changes and poor aesthetics, and provide natural language feedback. This nuanced approach achieves a Kendall's τ-b correlation of 0.77 with human judgments. Evaluation reveals that frontier agents like Claude-4.5-Opus struggle, achieving only a 45% success rate and an average partial score of 57%. The benchmark is available at https://microsoft.github.io/ppteval.
ppt-evalrubric-based evaluationmultimodal taskspartial creditkendall's τ-b
PruneGround: Plug-and-play Spatial Pruning for 3D Visual Grounding
PruneGround introduces a plug-and-play framework for efficient 3D Visual Grounding (3DVG) by leveraging local spatial context. The method comprises three components: Language-Guided Spatial Pruning (LGSP), which uses a frozen Vision Language Model to identify language-relevant regions; MultiView-Conditioned Description Reformulation (MCDR), which simplifies complex expressions and augments spatial cues through multi-view reasoning; and LLM-Grounder, which repurposes a detection-pretrained spatial LLM for language-conditioned grounding within pruned regions. Evaluations on ScanRefer, Nr3D, and Sr3D benchmarks demonstrate state-of-the-art performance, achieving top results on all ScanRefer settings and 9 out of 10 Nr3D/Sr3D settings.
3d visual groundingspatial pruningvision language modelpoint cloudlanguage-conditioned grounding
A Modular Vision-Language-Action Robotics Framework for Indoor Environments
The paper introduces a modular vision-language-action framework for autonomous robotics in indoor environments, addressing the CMU Vision-Language-Action Challenge. The system integrates two parallel pipelines: a perception module leveraging OwlViT embeddings to construct semantic voxel maps from real-time camera feeds, and a language module utilizing a Vision-Language Model to classify user commands. Mapping operates under a 500-second time constraint, allowing partial map generation if exceeded. Commands are grounded in the semantic and geometric context of the map to produce actionable outputs via the VLM. This approach effectively bridges natural language instructions with robotic task execution.
semantic voxel mapowlvit embeddingsvision-language modelmodular architectureautonomous robotics
Beyond the Library: An Agentic Framework for Autoformalizing Research Mathematics
We present an agentic autoformalization framework that translates natural language mathematics into verifiable Lean 4 code using general-purpose LLMs. The system employs an orchestrator to manage a multi-agent pipeline, dynamically extending type definitions and validating them via a novel Auxiliary Lemma technique. Evaluated on PutnamBench and five ACM STOC papers, the framework successfully formalized main theorems and proofs, achieving machine-checked Lean proofs for 32 Putnam problems and two STOC proofs with no axioms beyond Lean's kernel. All formalizations are publicly available.
autoformalizationlean 4auxiliary lemmaputnambenchstoc
Scenario Generation for Testing of Autonomous Driving Systems Using Real-World Failure Records
The paper proposes a pipeline for generating diverse test scenarios for Autonomous Driving Systems (ADS) using historical failure records in natural language format. The method employs modular LLM-based synthetic scenario generation, incorporating categorical and contextual information to produce scenarios compatible with system-specific testing constraints. Evaluated on the Metadrive simulator using NHTSA ADS crash records, the approach generates scenarios with 4 road types, 3 non-ego vehicle movement types, and on-road anomalies like work zones, revealing system failures within a 20-scenario testing budget.
autonomous driving systemsscenario generationllm-based synthesismetadrive simulatornhtsa crash records
UniSAE: Unified Speech Attribute Editing on Speaker, Emotion and Low-Level Content via Discrete Phonetic Posteriorgram Modelling
UniSAE introduces a unified framework for composable speech attribute editing across speaker, emotion, and content at sub-phoneme to word levels. The method employs Discrete Phonetic PosteriorGram (DPPG) representations to factorize speech into discrete tokens encoding phoneme identity, pronunciation variants, and duration, enabling precise phoneme-level editing. An autoregressive content transformer predicts edited DPPG sequences for word-level modifications, while a diffusion-based acoustic decoder renders speech conditioned on disentangled speaker and emotion representations. Experiments demonstrate precise control over speaker and emotion attributes, multi-granular content editing, and joint modification of all three attributes within a single architecture.
discrete phonetic posteriorgramautoregressive content transformerdiffusion-based acoustic decoderspeech attribute editingphoneme-level editing
SkillSpotter: Pose-Aware Multi-View Skilled Action Detection and Grading in Ego-Exo Videos
The paper introduces SkillSpotter, a pose-aware multi-view architecture for skilled action detection and grading in ego-exo videos. The method employs three novel modules: adaptive temporal suppression for varying action densities, gated 3D body pose fusion to integrate kinematic signals, and bidirectional cross-view attention for effective ego-exo view combination. Evaluated on Ego-Exo4D's proficiency benchmark, SkillSpotter achieves 76% higher class-specific mAP (21.82 vs. 12.40) and 4.41% better balanced accuracy (60.40% vs. 55.99%) over baselines. The modules demonstrate transferability to other action detection models and generalize to the HoloAssist dataset.
ego-exo videotemporal action detection3d pose fusioncross-view attentionskill grading
The Past Is Prologue: A Plug-in Controller for Selective Updates in Sequentially Evolving LLM Memory
The paper introduces Janus, a plug-in controller for selective memory updates in sequentially evolving LLM memory systems. Janus employs a Memory Momentum Trigger to detect anomalous update trajectories and evaluates candidate updates using a compact hybrid task set (coverage, boundary, fresh tasks) rather than full history replay. The method-agnostic approach improves accuracy by +2.7 to +4.6 points across six datasets, two backbone LLMs (unspecified), and two memory updaters, while preserving existing update rules.
memory controllersequential evolutionmomentum triggerhybrid evaluationin-context learning
Revealing Safety-Critical Scenarios for UTM via Transformer
The paper proposes a transformer-based reinforcement learning framework for discovering safety-critical scenarios in Unmanned Traffic Management (UTM) systems. The method employs a Policy Model to generate targeted test scenarios and an Action Sampler to enforce domain constraints, using attention mechanisms to model system state relationships and a risk-based reward function for exploration. Evaluation on 700-hour simulations shows an 8× improvement in vulnerability discovery efficiency over expert-guided testing, uncovering previously missed edge cases.
unmanned traffic managementtransformer-based rlattention mechanismsrisk-based rewardvulnerability discovery
What Probing Reveals about Autonomous Driving: Linking Internal Prediction Errors to Ego Planning
The study investigates the predictive and planning capabilities of autonomous driving policies to assess their robustness beyond nominal performance. Using linear probing and targeted perturbations, the authors analyze imitation learning and reinforcement learning models across varying scales of datasets and simulation training. Results reveal that despite strong closed-loop performance, policies often fail to predict surrounding vehicle movements during near-collision events, limiting ego planning. Causal interventions demonstrate that correcting prediction errors enhances trajectory safety, highlighting the importance of internal predictive signals for robust planning.
autonomous drivinglinear probingego planningimitation learningreinforcement learning
Seeing Through Multiple Views: Parameter-Efficient Fine-Tuning via Selective Neurons for Consistent Radiology Report Generation
The paper introduces View-PNDF, a parameter-efficient framework for view-consistent radiology report generation (RRG) that addresses clinical inconsistencies in multi-view X-ray image processing. The method combines (i) view-specific neuron detection, (ii) neuron verification, and (iii) selective fine-tuning to strengthen view-specific neurons while preserving view-agnostic representations, updating only 1-2% of parameters. Experiments on two RRG benchmarks show View-PNDF improves view-specific report quality (measured by NLG metrics and GPT-4o assessment) while maintaining general-view performance, with reduced computational costs.
radiology report generationparameter-efficient fine-tuningview-specific neuronsmulti-view x-rayselective fine-tuning
When Reranking Hurts: Uncertainty-Based Gating for Few-Shot Reranking
The paper introduces Training-Free Gated Reranking, a method that selectively applies few-shot example reranking based on model uncertainty to avoid performance degradation. The approach gates the reranking step using the model's uncertainty estimates, reducing unnecessary computations. Experiments across 8 LLMs, 7 NLU datasets, and 9 MT domain-language combinations show 15%-80% lower computational costs and up to 2% average performance improvement, demonstrating that reranking is most effective for high-uncertainty cases.
few-shot learningrerankingmodel uncertaintycomputational efficiencynatural language understanding
DDIAgents: Mechanism-Conditioned Context Flow for Drug-Drug Interaction Prediction
We propose DDIAgents, a mechanism-conditioned multi-agent framework for drug-drug interaction (DDI) prediction that dynamically orchestrates heterogeneous biomedical knowledge. The system employs a planner agent to instantiate specialized expert agents, route mechanism-relevant knowledge sources, and aggregate analyses through a conclusion agent, adapting context flow to inferred interaction mechanisms. This approach reduces irrelevant information, supports complementary expert reasoning, and provides interpretable agent-level rationales. Experiments on realistic DDI benchmarks demonstrate that DDIAgents consistently outperforms feature-based, graph-based, LLM-based, and agent-based baselines, showcasing adaptive and interpretable AI4Science reasoning through multi-agent systems.
drug-drug interactionmulti-agent frameworkmechanism-conditionedcontext flowknowledge orchestration
Beyond But-for Test: Counterfactual Explanation in Abstract Argumentation via Actual Causality (Extended Version)
The paper introduces an intervention-based counterfactual reasoning framework for abstract argumentation, overcoming limitations of prior but-for test approaches. The method encodes argument acceptance conditions as equations and defines an intervention operator supporting simultaneous changes to argument sets and fixation of witness arguments to actual labels. By aligning with the Halpern-Pearl definition of actual causality, the framework correctly handles complex cases like Preemption and Overdetermination, demonstrating superior expressiveness and reliability compared to existing methods.
counterfactual explanationabstract argumentationactual causalityintervention operatorhalpern-pearl definition
Triospect: A Three-Dimensional Framework for Robust Statistical AI-Generated Text Detection Against Diverse Attacks
The study introduces Triospect, a three-dimensional framework for robust AI-generated text detection that incorporates content (core ideas) and expression (stylistic elements) alongside existing textual characteristics. This approach significantly improves detection robustness against diverse attacks. Evaluated on two benchmarks with 17 attacks, 12 domains, and 17 source models, Triospect outperforms strong baselines by 22.3% (AUROC) and 13% (TPR01) on Humanize-16K, and by 9.1% (AUROC) and 22% (TPR01) on adversarial RAID. The framework represents a pioneering statistical method for enhancing detection reliability under adversarial conditions.
ai-generated text detectionstatistical frameworkadversarial robustnesscontent-expression analysisauroc improvement
MultiUAV-Plat: An LLM-Oriented Platform, Benchmark and Framework for Multi-UAV Collaborative Task Planning
The authors introduce MultiUAV-Plat, a lightweight simulation platform and benchmark for evaluating LLM-driven multi-UAV collaborative task planning, addressing gaps in existing UAV simulators and LLM-agent benchmarks. The platform features RESTful APIs, agent-facing observations, and validation logic, while the benchmark includes 75 mission sessions with 1500 tasks and 9396 validation checks across three scenarios. Their proposed Agent4Drone framework achieves 57.9% task pass rate, outperforming ReAct by 27.3 percentage points, demonstrating effective multi-UAV coordination under realistic constraints.
multi-uav collaborationllm-agent benchmarkingtask planningrestful api simulationpartial observability
ADAPT: Attention Dynamics Alignment with Preference Tuning for Faithful MLLMs
The paper introduces ADAPT, an attention-based framework for reducing hallucinations in Multimodal Large Language Models (MLLMs) by aligning text-to-image cross-attention dynamics. The method combines three components: a cross-attention visual anchor for stable spatial grounding, an attention-supervised inference mechanism for online correction of attention drift, and Visual Attention Guidance DPO to preference visually grounded responses. Experiments demonstrate a 40%-60% reduction in hallucination rates across multiple benchmarks while maintaining general multimodal capabilities, establishing new state-of-the-art performance.
multimodal large language modelscross-attention dynamicshallucination mitigationpreference tuningvisual grounding
Learning Video Dynamics with Predictive Differentiable Rendering
The authors propose Predictive Differentiable Rendering (PDR), a novel video prediction paradigm combining discrete and continuous representations via a 2D Gaussian adapter (PredGS) integrated with existing predictors. PredGS uses 5 + C learnable parameters per Gaussian and a CUDA-accelerated renderer (predgsplat) for 10x faster rendering, optimized by L1 + SSIM loss to avoid MSE-induced blurring. Experiments on TaxiBJ, WeatherBench, KTH, and Human3.6M show PDR outperforms prior methods in detail preservation, visual fidelity, and predictive accuracy.
differentiable renderinggaussian splattingvideo predictionssim losscuda acceleration
Knowledge Distillation from Large Reasoning Models to Compact Student Models: A Case Study on the John O Bryan Mathematics Competition
The study demonstrates knowledge distillation from DeepSeek-R1 to Qwen2.5-7B for mathematical reasoning, using Chain-of-Thought (CoT) training on John O'Bryan Mathematics Competition problems. A dual-agent framework constructs the dataset, with LoRA fine-tuning on Apple Silicon via MLX. The distilled model achieves 69.43% accuracy (4.76pp improvement over base) and generalizes to 73.1% on MATH-500. Analysis reveals accuracy declines from 69.43% (R1, 220 words) to 41.9% (R6, 31.2 words), showing response length critically impacts reasoning quality.
knowledge distillationchain-of-thoughtlow-rank adaptationmathematical reasoningresponse length
OpenLife: Toward Open-World Artificial Life with Autonomous LLM Agents
The paper introduces OpenLife, a paradigm for open-world Artificial Life (ALIFE) using autonomous LLM agents with persistent memory, tool use, and economic interactions. The method employs a society of asynchronous processes—memory, perception, evaluation, and budget-based metabolism—to enable life-like behavior in open social and economic environments. Results from running six agents for twelve weeks show emergent dynamics: reactive-to-spontaneous activity shifts, individuation, social structure, and self-earned income, demonstrating the viability of open-world ALIFE as an experimental platform.
open-world alifellm agentsasynchronous processesbudget-based metabolismemergent dynamics
LabGuard: Grounding Natural-Language Laboratory Rules into Runtime Guards for Embodied Laboratory Agents
LabGuard introduces a safety framework for grounding natural-language laboratory rules into executable runtime guards for embodied agents. The system comprises LabGuard-IR (typed executable representation), LabGuard-Bench (812 annotated rules from 203 seeds), and LabGuard-Grounder (NL-to-IR mapper), compiled into runtime monitors via a pipeline. Evaluations demonstrate 79.4 F1 on unseen rules, reducing unsafe events from 39.5% to 23.8%, with ACT integration maintaining <0.5% intervention rates in LabUtopia while preserving task success.
embodied agentsruntime guardsexecutable specificationssafety rulesnatural-language grounding
LLM-Driven Personalities for Decision Making in Emergency Simulations
The study demonstrates that LLM-driven personality profiles significantly influence decision-making in virtual humans during evacuation simulations. Using the OCEAN personality framework encoded in language prompts, the authors implemented heterogeneous agent behaviors in a crowd simulation scenario. Results show distinct behavioral patterns emerge from different trait configurations, suggesting LLM-guided agents can enhance realism compared to rule-based approaches. The work provides empirical evidence for language-model-based personality expression in multi-agent systems.
large language modelsocean personality traitsagent-based simulationdecision-makingevacuation scenario
OTCache: Optimal Transport for Geometry-Aware Caching in Diffusion Models
OTCache introduces a training-free framework for accelerating diffusion model sampling via optimal transport-based caching schedule prediction. The method addresses limitations of graph-based caching by modeling schedule evolution across inference budgets through three stages: high-fidelity reference schedule generation, anchor search via Optuna optimization, and quantile interpolation between policies. Evaluations on FLUX.1, Qwen-Image, and HunyuanVideo demonstrate 3.66x-4.7x acceleration while improving generation fidelity over state-of-the-art baselines.
optimal transportdiffusion modelscaching schedulequantile interpolationperceptual objective
Beyond Compilation: Evaluating Faithful Natural-Language-to-Lean Statement Formalization
The paper introduces a protocol for evaluating faithful natural-language-to-Lean formalization, distinguishing between compilation success and semantic faithfulness. Using a 400-entry benchmark spanning advanced mathematics, the method combines Lean compilation, cross-model semantic judging, and human expert calibration. Results reveal a 29.0-point gap between compilation rate (89.5%) and consensus faithfulness (60.5%), with human audits confirming 96.0% of consensus-positive outputs as faithful. The study also decomposes three pipeline interventions (parametric drafting, Mathlib search, Lean feedback), finding feedback most impactful for validity but prone to semantic failures, while search improves grounding.
lean formalizationsemantic faithfulnesstheorem provingmathlibelaboration feedback
A Three-Phase Foundation Model for Tax-Aware Personalized Portfolio Management
The paper introduces a three-phase deep reinforcement learning system for personalized portfolio management, addressing limitations of prior financial RL work. Phase 1 employs a self-supervised cross-asset encoder with a Chronos (T5-based) time series foundation model, generalizing to new assets via metadata. Phase 2 fine-tunes a Mixture of Experts (MoE) actor-critic with PPO, handling six distinct investment objectives through specialized expert heads. Phase 3 adds a LoRA-based personalization layer adapted to individual brokerage histories. The system is the first to integrate a time series foundation model into portfolio RL and uses natural language parsing for goal specification.
deep reinforcement learningmixture of expertstime series foundation modelportfolio managementlora module
Wait, am I Being Fair? Characterizing Deductive Stereotyping and Mitigating It with Fair-GCG
This work introduces Fair-GCG, a reasoning-time injection framework to mitigate deductive stereotyping in large language models (LLMs), where models apply population-level statistical regularities to individual cases, producing logically coherent yet socially biased inferences. Fair-GCG systematically discovers effective injection phrases to steer models toward fairness-aware reasoning. Results demonstrate that Fair-GCG improves performance across multiple fairness benchmarks, generalizes from smaller to larger LLMs, enhances reasoning-level fairness, reduces bias in open-ended generation, and transfers to real-world fairness-sensitive tasks.
deductive stereotypingfair-gcgreasoning-time injectionfairness benchmarksopen-ended generation
When Regulation Has Memory: Hysteresis and Control Burden in Artificial Agency
The study demonstrates that adaptive artificial agents exhibit history-dependent regulatory burden, quantified through hysteresis loops in adaptive gain. Using a computational model of uncertainty regulation, agents were driven through continuous uncertainty target changes and reversals without resetting. Results show that identical uncertainty targets require different control levels depending on the agent's trajectory (toward or returning from demanding regimes). Anticipatory regulation before disturbance exposure reduces adaptive gain compared to post-disturbance recovery, though state-level coherence remains path-dependent. This highlights the importance of evaluating agents not just by stability but by the regulatory effort required to maintain it.
adaptive gainhysteresis loopuncertainty regulationcontrol burdenanticipatory regulation
AgentBound: Verifiable Behavioral Governance for Autonomous AI Agents
AgentBound introduces a verifiable behavioral governance framework for autonomous AI agents, addressing the gap between authorization and context-aware execution. The system combines three decision authorities (delegated authorization, behavioral constitutions, site contracts) through formal composition, generating cryptographically verifiable receipts for accountability. Key innovations include standing delegation for long-running agents and the AgentBound-Bench evaluation framework. The approach provides deterministic governance between authorization and execution, enabling independent verification through governance receipts rather than relying on trusted processes.
behavioral governanceautonomous agentsverifiable receiptsdecision compositionaccountability framework
HyPOLE: Hyperproperty-Guided Multi-Agent Reinforcement Learning under Partial Observation
HyPOLE introduces hyperproperty-guided MARL under partial observability, leveraging HyperLTL for formal specification to overcome reward shaping limitations. The framework combines Centralized Training for Decentralized Execution (CTDE) with hyperproperties to synthesize decentralized policies. Evaluations on SMAC, MessySMAC, and WildFire benchmarks demonstrate superior performance compared to baselines, showcasing the benefits of formal methods in MARL.
multi-agent reinforcement learninghyperpropertiespartial observabilityhyperltlctde
Loc2Repair: A Framework for Evaluating the Impact of File-Level Issue Localization in Repo-Level LLM Repair
Loc2Repair introduces a modular framework for evaluating repository-grounded repair pipelines by isolating file-level issue localization as an independent variable. The framework decouples localization and repair, enabling controlled comparisons across different localization models and repair backends under shared conditions. Experiments on SWE-bench Verified with three repair backbones show that explicit localization improves resolved rates (44.7% to 52.4% with gold localization) and reduces mean elapsed time (up to 154.45s faster), though token effects vary across models.
repository-grounded repairfile-level localizationmodular evaluationswe-benchpatch synthesis
Neuro-Bayesian-Symbolic Residual Attention Shallow Network: Explainable Deep Learning for Cybersecurity Risk Assessment
The authors propose Neuro-Bayesian-Symbolic Residual Attention Shallow Network (NBS-RASN), a hybrid neural architecture for explainable cybersecurity risk assessment. The 12-layer shallow network (80 neurons) incorporates domain knowledge as differentiable components, enforcing five epistemological axioms via a gatekeeper mechanism. It combines residual attention with symbolic reasoning to achieve deep-learning capabilities while maintaining interpretability, decomposing scores into deterministic weights and traceable expert adjustments. Evaluation on 20 OWASP Top 10:2025 projects shows confidence scores of 0.79-0.97, demonstrating that shallow networks with deep reasoning can outperform opaque models in high-stakes domains requiring interpretability.
explainable airesidual attentioncybersecurity riskshallow networksymbolic reasoning
Learning Where to Look: A Reinforcement Learning Framework for Robust Micro-Ultrasound Prostate Cancer Detection
Prost-RL introduces a reinforcement learning framework for robust prostate cancer detection in micro-ultrasound ($μ$US) imaging, addressing sparse supervision and class imbalance. The method integrates a lightweight RL policy into a foundation-model encoder-decoder to generate interpretable spatial attention maps, using Adaptive Policy Optimization (APO) for training stability and a noise-robust objective combining symmetric cross-entropy with negative-entropy regularization. Evaluated on 6,607 biopsy cores from 693 patients across five sites, Prost-RL achieves 79.0±3.5 AUROC for core-level detection (+2.1 AUROC over baselines) and 79.3±5.8 AUROC for clinically significant cancer classification.
reinforcement learningmicro-ultrasoundadaptive policy optimizationnoise-robust objectivespatial attention maps
AgRefactor: Self-Evolving Agentic Workflow for HLS Compatibility and Performance
AgRefactor introduces a self-evolving, multi-agent LLM workflow for refactoring software into High-Level Synthesis (HLS)-compatible programs, addressing limitations of existing automated and LLM-based approaches. The system combines a memory mechanism for accumulating factual and strategic knowledge with hybrid LLM/tool-based transformations to enhance robustness and scalability. Evaluated on 11 real-world benchmarks (5-10x longer than prior work), it matches or outperforms state-of-the-art tools, achieving a 6.51x geometric mean speedup over the best pragma tuning tool and 1.20x over optimized designs with <20% extra resource overhead.
high-level synthesisllm-based refactoringmulti-agent workflowself-evolving memorypragma tuning
Motion Planning in Compressed Representation Spaces
The paper introduces a generative framework unifying learned motion priors with model-based planning through latent space search. The method first trains a hierarchical discrete token autoencoder for high-compression trajectory representation, then performs optimization-based planning in this latent space while preserving realism. Evaluated on nuPlan and Waymo Open Motion Dataset, the approach demonstrates strong performance in closed-loop planning and multi-agent scenario synthesis, achieving flexibility via test-time objective specification without task-specific training.
motion planninglatent space searchdiscrete autoencoderhierarchical representationgenerative framework
Physics-informed Conditional Normalizing Flows for Angles-only Cislunar Orbit Determination
The work advances generative astrodynamics by applying conditional normalizing flows to angles-only cislunar orbit determination. The method formulates initial state estimation as conditional density estimation, training a normalizing flow on perturbed topocentric observations from Near Rectilinear Halo Orbits to model potentially multimodal posteriors. Generated physics-informed state hypotheses are refined via nonlinear least-squares, providing warm starts for classical algorithms. Results demonstrate statistically consistent state estimation from short observation arcs.
generative astrodynamicsconditional normalizing flowscislunar orbit determinationnear rectilinear halo orbitsnonlinear least-squares
RoPoLL: Robust Panel of LLM Judges
The paper introduces RoPoLL, a robust aggregation method for LLM evaluation panels that addresses bias vulnerabilities in conventional consensus scoring. The method replaces mean aggregation with geometric median estimation, achieving optimal breakdown point 1/2 and parametric convergence rate σ√(d/N). Experiments across 13 models (4B-675B parameters) show RoPoLL outperforms standard panels by 19% on cross-dimensional attacks and maintains superiority under 30-50% corruption, with a 38B 3-judge panel surpassing 675B Mistral-Large-3 by 1.31x on HelpSteer-2.
llm juryrobust mean estimationgeometric medianbreakdown pointbyzantine adversaries
Behavior Cloning is Not All You Need: The Optimality of On-Policy Distillation for Noisy Expert Feedback
The paper analyzes imitation learning with noisy expert feedback, explaining why on-policy distillation (OPD) outperforms offline methods like supervised fine-tuning (SFT) in practice. By modeling expert noise and analyzing sample complexity, the authors prove offline learning requires exponential horizon dependence, while their proposed OPD variant achieves polynomial dependence. Theoretical results include necessary conditions for horizon-free learning, a novel loss function for noisy settings, and extensions to unknown corruption with deterministic experts, providing foundations for OPD's empirical success in language model training.
imitation learningon-policy distillationnoisy expertsample complexityhorizon dependence
Budget-Adaptive Routing: Skipping the Weak When the Strong Answers Anyway
The paper introduces budget-adaptive routing for edge-cloud inference, optimizing computation by skipping weak edge models when offloading to stronger cloud models. The method combines a lightweight weak-skipping estimator (0.153 GFLOPs, 29x lighter than the weak detector) with weak-conditioned routing, dynamically selecting between them via offline-tuned thresholds based on offload budget. On PASCAL VOC, it achieves up to 19.1 ms latency reduction (30% at ρ=0.9) and surpasses the strong model's peak mAP by +1.7 pp at certain operating points. The approach outperforms SOTA methods while reducing compute overhead.
edge-cloud inferencerouting estimatorweak-skippingbudget-adaptiveoffload budget
Why Solve It Twice? Hierarchical Accumulation of Skills for Transfer-Efficient ML Engineering
The paper introduces HASTE, a hierarchical multi-agent system for efficient knowledge transfer in ML engineering competitions. HASTE organizes skills into three tiers (global, domain, competition-specific) with matching agent levels, coordinated by an orchestrator using LLM-driven abstraction. Experiments on MLE-Bench Lite (22 Kaggle competitions) show tiered loading achieves a 77.3% medal rate with Claude Sonnet 4.6, outperforming flat loading (62.5%) and reducing refinement iterations by 52% in warm starts. Skill retention improves from 42% to 85% with 50+ skills, demonstrating hierarchical organization reduces redundant computation.
hierarchical multi-agent systemknowledge transferllm-driven abstractionml engineeringskill accumulation
Investigating Multi-Agent Deliberation in Law
The paper introduces two novel multi-agent deliberation (MAD) frameworks for legal reasoning tasks, inspired by courtroom procedures and legal argumentation. Using Large Language Models (LLMs), the authors compare these multi-agent approaches against monolithic baselines on legal and non-legal benchmarks. Results show comparable overall performance but distinct answer distributions, with multi-agent systems excelling in cases requiring critical thinking from multiple perspectives, while baselines solve different subsets of problems.
multi-agent deliberationlegal reasoninglarge language modelscourtroom procedureslegal argumentation
How Human Feedback Shapes AI-generated Community Notes
This work analyzes 19,146 Collaborative Notes and 211,850 instances of human feedback to understand how AI-generated drafts are refined in X's Community Notes system. Using a taxonomy of human suggestions, the study finds factual corrections and contextual additions are most frequently incorporated, while subjective policy judgments are rarely adopted. Human feedback significantly improves note helpfulness, particularly when challenging previous drafts' main claims. Despite iterative improvements, Collaborative Notes achieve helpful status and platform visibility at lower rates than human-only or AI-only notes, with limited human participation identified as a bottleneck. The system serves a complementary role by targeting posts not addressed by other methods.
community notescollaborative noteshuman feedbackcontent moderationfact-checking
Curvature-Guided Module Localization for Low-Rank Detoxification of Backdoored Large Language Models
The paper introduces a curvature-guided framework for localized detoxification of backdoored large language models (LLMs) without full retraining. The method combines activation patching with Fisher/K-FAC curvature analysis to identify critical modules, then applies targeted low-rank repair to suppress trigger-induced malicious behavior. Evaluations on poisoned Llama-3.2-1B-Instruct show effective mitigation of trigger-conditioned responses (beginning/middle/end of prompts) while maintaining benign performance, suggesting backdoor removal as a structural repair problem rather than behavioral alignment.
backdoor attackslow-rank repairfisher curvatureactivation patchingllm detoxification
Training Therapeutic Judges and Multi-Agent Systems for Human-Aligned Mental Health Support
The paper introduces a two-stage framework for improving therapeutic quality in mental health support LLMs by using human-aligned evaluation as a control signal. Stage I develops TheraJudge, an open-source evaluator trained via preference-based optimization to assess responses across 7 psychological dimensions. Stage II introduces TheraAgent, a multi-agent system with specialized roles (Critic, Coach, Therapist) that refine responses based on TheraJudge's evaluations. TheraJudge achieves ICC = 0.87-0.95 agreement with clinicians, while TheraAgent improves human-rated quality by +0.43 points (5-point scale), with 94% recovery rate for low-quality responses.
therapeutic evaluationpreference-based optimizationmulti-agent systemhuman-alignedmental health llms
The Label Imitation Game: Turing Test Network for Zero-Shot Pseudo-Label Pruning
The paper introduces the Label Imitation Game (LIG), a Turing-inspired framework for pruning hallucinated pseudo-labels from foundation models via adversarial interrogation. It trains a task-agnostic Turing Test Network (TTN) to evaluate pseudo-labels contextually across datasets, eliminating the need for supervision or retraining. Experiments on four datasets show TTN improves label accuracy for three vision-language models, enabling zero-shot task transfer (e.g., image classification-trained TTN prunes object detection labels) with F1-score gains up to 44%. TTN also facilitates Category Revival, recovering zero-recall classes in downstream models.
pseudo-label pruningzero-shot learningturing test networkfoundation modelsadversarial interrogation
Beyond expert users: agents should help users construct preferences, not just elicit them
The paper challenges the assumption of expert users in agent design, proposing that agents should help users construct preferences through dialog rather than merely eliciting them. It introduces CoPref, a preference construction model based on the Search-Experience-Credence framework, and CoShop, an interactive benchmark for evaluating agentic recommender systems. Evaluating five frontier models on CoShop reveals ≤56% accuracy, with failures attributed to insufficient user knowledge expansion during interaction.
preference constructionagentic recommender systemsdialog systemssearch-experience-credenceinteractive benchmarking
When Does Learning to Stop Help? A Cost-Aware Study of Early Exits in Reasoning Models
The paper introduces LearnStop, a hidden-state-free checkpoint stopper for reasoning language models that predicts prefix correctness using online features like answer confidence and stability. Evaluated across 18 task-model settings including GSM8K and MATH-500, results show task-dependent utility: learned multi-feature stopping improves fixed-budget frontiers in free-form math (+0.157 peak gain on GSM8K with Qwen3-32B), while scalar rules suffice for multiple-choice tasks. The study provides validation-selected operating points, cost analyses under different serving regimes, and establishes that learned stopping adds value primarily when no single scalar signal reliably indicates correctness.
early exitingreasoning modelscost-aware inferencecheckpoint schedulingkv-fork
Test-Time Verification for Text-to-SQL via Outcome Reward Models
The paper introduces GradeSQL, a framework for training Outcome Reward Models (ORMs) as semantic verifiers in Text-to-SQL tasks, eliminating the need for manual annotation via automated candidate generation and execution-based labeling. ORMs are integrated into a Best-of-N pipeline, outperforming heuristic methods like execution-based selection and Majority Voting by up to +4.33% on BIRD and +2.10% on Spider benchmarks. Results demonstrate ORMs' effectiveness with larger candidate sets and complex queries, offering a scalable alternative to traditional test-time verification.
text-to-sqloutcome reward modelstest-time verificationbest-of-n samplingmajority voting
BayesBench: Evaluating LLM Belief Trajectories Under Multi-Turn Evidence Accumulation
The paper introduces BayesBench, a novel evaluation suite for assessing LLMs' Bayesian reasoning capabilities during multi-turn evidence accumulation. The benchmark comprises three tasks of increasing complexity: Bayesian estimation, prediction, and latent-framed prediction, comparing model belief updates against rational Bayesian baselines. Results across seven LLMs (3B-70B parameters) show that while scaling improves latent inference, downstream prediction accuracy often fails to match Bayesian optimality, revealing a disconnect between inference and belief utilization.
bayesian reasoningevidence accumulationmulti-turn evaluationlatent inferencebelief trajectories
How Can AI Find My Model? A Model-Finding Experimental Study Considering Data Formats, Embeddings, and Retrieval Strategies
This study investigates AI-driven approaches for discovering reusable simulation models by analyzing the impact of data representation, transformer-based embeddings, and retrieval strategies. The authors evaluate performance using natural language queries across multiple query types, measuring recall@5 and nDCG@5 metrics. Results indicate that data representation significantly affects retrieval accuracy, open-source embedding models achieve competitive performance, and reranking methods become increasingly important as query complexity grows. The work establishes a baseline for AI-driven model discovery and highlights its potential for advancing composability and interoperability in Modeling and Simulation.
transformer-based embeddingsretrieval strategiesrecall@5ndcg@5model discovery
Contrastive Reflection for Iterative Prompt Optimization
The paper introduces Contrastive Reflection, an iterative prompt-optimization framework for LLM agents in information retrieval (IR) workflows. The method leverages structured traces (retrieval/reasoning traces, dimension-level scores) to identify error-anchored behavioral slices, contrast them with successful examples, and generate targeted prompt edits via a Teacher LLM. Edits are validated for performance improvements and regression checks. On HotpotQA, one contrastive repair improved exact-match accuracy from 51.4% to 60.4%, outperforming failure-only (51.4%) and random-evidence variants. The framework aligns with modern optimizers (MIPROv2: 59.4%, GEPA: 57.0%) while offering interpretable prompt repair.
prompt optimizationcontrastive reflectionllm agentsinformation retrievalbehavioral slices
A Stationary-Distribution Theory for Triplet-Based Plateau Search in Random Forest Ensemble-Size Selection
The paper develops a stationary-distribution theory for plateau-based tuning of ensemble size in Random Forests, where the central triplet point fluctuates stochastically rather than converging deterministically. By modeling the central ensemble size $B_t$ as a birth-death Markov chain on a geometric grid, the authors derive its stationary distribution via local balance and centered folded-normal approximation. Results show the stationary center scales as $B_*=O(\varepsilon^{-2})$, with spread $\sigma_{B,*}=O(\varepsilon^{-2})$ and variance $O(\varepsilon^{-4})$, while relative spread remains $\varepsilon$-independent. The analysis reinterprets plateau-based tuning as a stochastic process rather than deterministic stopping.
random forestsstationary distributionbirth-death chainensemble-size selectionplateau search
AI-Generated PowerShell Malware: An Experimental Framework and Dataset
The study introduces an experimental framework for evaluating LLM-generated PowerShell malware, featuring a novel sandbox for dynamic analysis and a manually curated dataset of real-world malware annotated in natural language. The framework assesses open-weight LLMs adapted for malware generation, comparing their output to real malware through dynamic OS event analysis. Results indicate high similarity, with a median Jaccard index of 84.5% and 48.4% of instances showing complete event overlap.
powershell malwaredynamic analysisjaccard indexopen-weight llmsmalicious events
When transformers learn "impossible" languages, what do they learn?
The study investigates transformer language models' performance on theoretically 'impossible' languages, proposing two linking hypotheses: grammatical sensitivity and generative deficiency. Using GPT-2-style models trained on perturbed English variants, the authors evaluate grammatical sensitivity via BLiMP minimal pairs and assess generative quality through sentence production. Results show gradual degradation in grammatical sensitivity based on information locality, while generation exhibits pronounced failures at longer sequences, suggesting generative deficiency as a key factor in non-attestation of impossible languages.
transformer language modelsimpossible languagesgrammatical sensitivitygenerative deficiencyblimp
Gradient Smoothing: Coupling Layer-wise Updates for Improved Optimization
The paper introduces Depth-wise Gradient Augmentation, an optimization framework that transforms block-wise updates along network depth to exploit cross-layer structure. It proposes Gradient Smoothing, a family of methods using local Window Smoothing operators to couple layer-wise updates from base optimizers (SGD/Adam/Muon) with minimal overhead. Evaluations across language model pretraining, RL fine-tuning, diffusion models, and Vision Transformers demonstrate consistent improvements in optimization and generalization, while promoting structured representation evolution without architectural changes.
depth-wise gradient augmentationgradient smoothingwindow smoothingstructured optimizationcross-layer updates
Indi-RomCoM: Code-Mixed Benchmark for Evaluating LLMs on Romanized Indic-English Instructions
The Indi-RomCoM benchmark evaluates Large Language Models (LLMs) on Romanized Code-Mixed (RCM) instructions across seven tasks, four Indic languages, and three code-mixing intensity levels. Proprietary, open-weight, and Indic-focused LLMs were assessed under zero- and few-shot settings. Results indicate consistent underperformance on RCM instructions, with degradation increasing alongside code-mixing density. Reasoning tasks exhibited less degradation than detection tasks, attributed to contextual explanations generated by LLMs. This benchmark aims to advance the development of inclusive multilingual systems by addressing gaps in LLM performance on RCM content.
romanized code-mixinglarge language modelsindic languagesinstruction-following taskscode-mixing density
Security--Fidelity Tradeoffs: The Hidden Cost of Prompt Injection Defense
We introduce SecFid, a benchmark for evaluating security-fidelity tradeoffs in defending large language models (LLMs) against indirect prompt injection. SecFid distinguishes between executing, processing, and ignoring injected instructions, enabling fidelity measurement alongside security. Across 1,168 examples and 48 configurations, no model or defense achieves both high security and fidelity: the highest-fidelity model reaches 96.5% fidelity at 47.8% security, while the most secure defenses achieve 99.3% security but only 71.0%-73.9% fidelity. Defenses differ in how they achieve security, either repairing hijacks or suppressing content. Decision-theoretic analysis shows that optimal behavior depends on deployment-specific costs, highlighting that security alone inadequately measures robustness.
prompt injectionsecurity-fidelity tradeofflarge language modelsbenchmarkdecision-theoretic analysis
Detecting Audio Deepfakes on the Edge:Lightweight SSL-Based Detection in a Browser Plugin
The authors propose a lightweight, privacy-preserving solution for audio deepfake detection via a browser plugin, addressing concerns with cloud-based processing. Their method combines a truncated self-supervised learning (SSL) backbone with a logistic regression classifier, optimizing for both accuracy and speed. Evaluations show a 10% accuracy improvement over AASIST and 40% faster inference. The plugin enables secure, on-device detection for journalists and fact-checkers. Code is publicly available.
audio deepfakesself-supervised learningon-device detectionlogistic classifierbrowser plugin
A Single Rewrite Suffices: Empirical Lessons from Production Skill Description Optimization
The study introduces an automated pipeline for optimizing skill descriptions in enterprise AI agents, addressing skill collision issues where overlapping descriptions cause query misrouting. The method employs LLM rewrites using false-positive and false-negative cases, achieving 79.2% F1 on a production system (9 skills, 372 cases), comparable to manual tuning (79.4% F1) but with a 32x reduction in engineering effort (3.8 vs. 120 minutes per skill). Ablation studies on ToolBench (16k tools) reveal that a single LLM rewrite captures most improvements, with other design choices affecting F1 by less than 0.5%. A diagnostic for genuine skill scope overlap is also proposed.
skill collisionllm rewritedescription optimizationf1 scoreenterprise ai agents
SemRF: A Semantic Reference Frame for Residual-Stream Dynamics in Language Models
SemRF introduces Semantic Reference Frames (SemRF) to analyze residual-stream dynamics in language models, separating semantic measurement from residual computation. The method fixes semantic anchors and measures states against them using pseudo-inverse tying, ensuring stable semantic-basis coordinates and distortion bounds. SemRF induces a semantic Voronoi diagram, enabling layerwise steps, contribution profiles, and imbalance diagnostics. The canonical trace, defined as the minimum-action path within a margin-relaxed tube, obeys a discrete spline equation and links parameter efficiency to trace complexity. Results show that lower-action and lower-complexity traces use fewer semantic degrees of freedom, with guarantees requiring controlled interface error and small projection residual.
semantic reference framesresidual-stream dynamicsvoronoi diagramminimum-action pathparameter efficiency
Automated Background Swapping for Robustness against Spurious Backgrounds
The paper introduces Automated Background Swapping (AutoBackSwap), a method to mitigate spurious background correlations in deep neural network classifiers. The approach employs a secondary network for foreground-background disentanglement, infilling for background synthesis, and data augmentation via foreground-background recombination. Requiring only patch-wise labeling of hundreds of samples, AutoBackSwap outperforms prior methods across multiple image classification tasks, even without training samples that break spurious correlations.
spurious correlationsforeground-background disentanglementdata augmentationimage classificationdeep neural networks
FedLAB: Traceable Semantic Codebooks for Federated Multimodal Graph Foundation Learning
FedLAB introduces traceable semantic codebooks for federated multimodal graph foundation learning, addressing privacy constraints in decentralized multimodal graphs. The framework organizes knowledge into typed hierarchical codebooks for modality evidence, node semantics, and topology context, refined via federated semantic barycenter pre-training while keeping raw data local. Experiments on 10 benchmarks and 6 tasks demonstrate a 7.53% improvement over baselines while maintaining semantic traceability.
federated learningmultimodal graphssemantic codebookstraceabilityfoundation models
CoMet: Context and Multiplicity Decomposition for Multimodal Uncertainty Estimation
CoMet introduces a multimodal uncertainty estimation method for MLLMs by decomposing uncertainty into context-specific and multiplicity-specific terms. The context-specific term captures ambiguity from the given context (e.g., task or prompt), while the multiplicity-specific term quantifies the number of plausible answers compatible with the input. A lightweight post-hoc uncertainty module is trained to estimate these terms efficiently, avoiding autoregressive answer generation or repeated sampling. Experiments across open-ended multimodal benchmarks, hallucination detection, and visual question answering demonstrate CoMet's consistent improvement over baselines in uncertainty estimation while maintaining practical efficiency.
multimodal uncertaintycontext-specific termmultiplicity-specific termpost-hoc modulehallucination detection
Surrogate Fidelity: When Can Open LLMs Explain Closed Ones?
The paper investigates surrogate fidelity in mechanistic interpretability (MI), examining when open LLMs can reliably explain closed models. The authors evaluate fidelity at prediction, attribution, and representation levels using log-odds and leave-one-out attributions across eleven models (Llama, Qwen, GPT, Gemini). Key findings show prediction fidelity overstates attribution fidelity, with models agreeing on answers but not explanations, and reveal an access-validity inversion where white-box signals are stable but weakly predictive of causal attributions. Results indicate prediction-level agreement is insufficient for mechanistic transfer.
mechanistic interpretabilitysurrogate fidelitylog-oddsinput ablationsaccess-validity inversion
Random Reshuffling Dominates Stochastic Gradient Descent
This work resolves a longstanding open question by proving that Random Reshuffling (RR) dominates Stochastic Gradient Descent (SGD) in smooth convex optimization under any reasonable stepsize after any finite number of epochs. The analysis addresses two key limitations in existing theory: the requirement for stepsizes smaller than 1/n and the suboptimal convergence rate when epochs are fewer than n. The results demonstrate RR's superiority over SGD, aligning theoretical guarantees with empirical observations in practical implementations of Shuffling SGD.
random reshufflingstochastic gradient descentsmooth convex optimizationstepsizeepochs
Evaluation of Population Initialization Methods for Genetic Programming-based Symbolic Regression
This study evaluates population initialization methods for genetic programming-based symbolic regression, comparing random initialization with optimized solutions from exhaustive symbolic regression (ESR). Using NSGA-II on twelve synthetic and one real-world dataset, the authors analyze accuracy and model complexity. Results show no significant differences in final Pareto fronts, with ESR's initial advantage vanishing within few generations. The initialization method's impact on outcomes is negligible when initial population diversity is comparable.
genetic programmingsymbolic regressionpopulation initializationnsga-iipareto front
Semantic Leakage and Privacy Preservation in Relay-Assisted Semantic Communications
The paper identifies a privacy vulnerability in relay-assisted semantic communication (SemCom) systems, demonstrating that relays can infer semantic meaning and reconstruct signals nearly as effectively as legitimate receivers. To mitigate this semantic leakage, the authors propose an iterative adversarial training framework that jointly optimizes the relay's eavesdropping function and the legitimate system, degrading the relay's inference while preserving receiver performance. Results show a significant semantic accuracy gap between legitimate and adversarial decoding across channel conditions, achieved without compromising reconstruction fidelity.
semantic communicationprivacy preservationadversarial trainingsemantic leakagerelay-assisted systems
Signed-Permutation Coordinate Transport for RMSNorm Transformers
The paper introduces signed-permutation coordinate transport for RMSNorm transformers, addressing gauge symmetry incompleteness in permutation-only alignment. The method employs sign-marginalized Hungarian matching to overcome structural accuracy ceilings in signed-correlation matching, achieving 91.1% coordinate recovery versus 60.3% for endpoint matching. Results demonstrate improved tool transfer: TinyLlama SAE reconstruction NMSE drops to 0.004 (vs 1.08 under permutation), Qwen sentiment steering preserves 95.8% effect (vs 17.2%), and AdamW state transport maintains trajectory fidelity.
rmsnormgauge symmetryhungarian matchingcoordinate transportsigned-permutation
Making Sense of Touch from the Child's View for Contrastive Learning
The study proposes a structured coding system for infant touch events to investigate tactile contributions to visual concept learning. Researchers collected 264k two-second touch event clips from child-centric perspectives, then pretrained developmentally inspired models on this dataset. Results demonstrate the system's viability for quantifying tactile-visual learning relationships and provide insights into developmental learning mechanisms.
developmental learningtactile-visual learningstructured coding systeminfant touch eventscontrastive pretraining
FlexViT: A Flexible FPGA-based Accelerator for Edge Vision Transformers
FlexViT introduces a reconfigurable FPGA accelerator for efficient Vision Transformer (ViT) inference on edge devices, addressing architectural heterogeneity in hybrid ViT models. The accelerator employs a hardware-software co-design approach, mapping both fully connected and convolutional layers onto a unified INT8 GEMM engine via runtime im2col transformation. A dual-mode dataflow dynamically switches between input and weight reuse, while a depth-first tiling strategy eliminates off-chip partial-sum transfers. Implemented on a PYNQ-Z2 FPGA, FlexViT achieves up to 2.74x speedup on accelerator-executed layers and 1.40x end-to-end speedup compared to CPU-only execution.
vision transformerfpga acceleratorint8 gemmdual-mode dataflowdepth-first tiling
Interface-Aware Neural Newton Preconditioning for Robust Cohesive Zone Model Simulations
The paper introduces Interface-Aware Neural Newton Preconditioning (IA-NNP), a learned method to improve convergence in Cohesive Zone Model (CZM) simulations suffering from negative tangents and solution jumps. IA-NNP generalizes manual Newton-Raphson modifications through state-dependent interface corrections, operating only on active interface variables while preserving original traction-separation laws and dissipation checks. Two variants are proposed: IA-NNP-Init for initial-guess lifting and IA-NNP-NL for nonlinear right preconditioning. Evaluations on horizontal, circular, and multi-interface benchmarks demonstrate improved convergence rates (82% reduction in failures) and branch recovery compared to standard Newton-Raphson, without altering force-displacement responses.
cohesive zone modelsneural newton preconditioninginterface fracturenonlinear solverfinite element analysis
Accelerating Conformal Prediction via Approximate Leave-One-Out
The paper introduces an accelerated conformal prediction framework using approximate leave-one-out (ALO) estimators, addressing computational bottlenecks in uncertainty quantification. The method leverages ALO cross-validation risk estimators, adapted for conformal prediction where leave-i-out residuals are required for predictions at new points. Theoretical analysis establishes asymptotic coverage and efficiency, while simulations demonstrate comparable performance to exact methods with significantly reduced runtime. This approach builds on Jackknife+ and Jackknife-minmax but avoids costly leave-one-out refits for all observations.
conformal predictionapproximate leave-one-outuncertainty quantificationcross-validationasymptotic coverage
Sequential RC-TGAN: Generating Relational Time Series with Spectral Envelope Loss
Sequential RC-TGAN introduces a temporal extension of RC-TGAN for generating relational time series, featuring a novel spectral envelope loss to preserve latent periodic structures via backpropagation. The method employs Variational Gaussian Mixture Model discretization for continuous time series and establishes a theoretical benchmark using simulated categorical sequences with known spectral envelopes. Experiments on real-world and simulated datasets show superior performance in reproducing cyclic patterns and seasonality compared to state-of-the-art systems, validated by proposed metrics Spectral Density Divergence and Spectral Envelope Divergence.
relational time seriesspectral envelope lossvariational gaussian mixture modelcategorical sequencesfrequency-domain regularization
Review Residuals: Update-Conditioned Residual Gating for Transformers
The paper introduces Review Residuals, a novel gating mechanism for Transformer residual connections that scales updates by a learned, input-dependent gate conditioned on both the current state and the proposed update: h_l = h_{l-1} + r_l * u_l with r_l = sigmoid(W[RMSNorm(h_{l-1}), RMSNorm(u_l)]). Unlike prior gated residuals, this method conditions the gate on the update itself. Experiments reveal two key findings: (1) an additive, identity-preserving gate form enables stable training at all depths, unlike convex Highway-style gates that suffer from vanishing gradients beyond ~20 layers; (2) Review Residuals exhibit emergent benefits with scale, outperforming parameter-matched Highway gates and standard residuals at 590M parameters (p<0.05), with advantages growing further at 1B parameters.
residual connectionsgating mechanismvanishing gradientsrmsnormtransformers
Low-dimensional topology of deep neural networks
This work analyzes the topological expressivity of neural architectures by restricting layer width to d=3, isolating activation and depth effects from width. Using linking number as a topological invariant, the study proves that ResNets and transformers exhibit equivalent topological transformation power, both surpassing feedforward networks with monotonic activations. Nonmonotonic activations elevate feedforward networks to ResNet/transformer expressivity levels, while invertible and flow-based models show lower expressivity. Empirical experiments validate these theoretical insights, demonstrating that low-dimensional topology provides architectural design guidance. The results generalize to arbitrary d>3, maintaining the core hierarchy of architectural expressivity.
linking numbertopological invariantsnonmonotonic activationsexpressivitylow-dimensional topology
Explicit Fuzzy Logic in the Feed-Forward Layer: Self-Forgetting Quantifiers Discover Legible Grammatical-Licensing Detectors
The paper introduces a negation-capable feed-forward network (NC-FFN) that replaces standard transformer FFN layers with explicit fuzzy logic operations (intersection and set-difference) on sigmoid-bounded memberships. This parameter-neutral modification matches GELU baseline perplexity on OpenWebText (125M params) while enabling logical interpretability. Key innovations include sequence quantifiers with learned forgetting rates, which resolve grammatical licensing deficits and maintain interpretability across layers. Results show improved LAMBADA performance and legible grammatical detectors (e.g., licensing comparatives or negations) without dictionary learning, though full Boolean FFNs diverge during training.
negation-capable ffnfuzzy logic operationsgrammatical licensinglearned forgetting rateinterpretable transformers
Relational and Sequential Conformal Inference for Energy Time Series over Graphs via Foundation Models
We propose STOIC, a novel framework for uncertainty quantification in energy demand forecasting that integrates spatial-temporal graph neural networks (STGNNs) with tabular foundation models for zero-shot calibration. STOIC generates point forecasts using an STGNN, reformulates spatial-temporal residuals into a tabular format, and leverages in-context learning to calibrate prediction intervals without task-specific retraining. Evaluated on five benchmarks including synthetic simulations and real-world electricity and district heating networks, STOIC outperforms existing conformal prediction baselines, providing more reliable uncertainty estimates for graph-structured energy time series.
spatial-temporal graph neural networksconformal predictionin-context learninguncertainty quantificationtabular foundation models
Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers
The paper introduces LOTUS (Looped Transformers with parallel supervision on latents), a latent chain-of-thought (CoT) method that bridges the performance gap with explicit CoT at the 3B parameter scale. LOTUS employs a looped Transformer architecture processing K latent blocks in parallel for R iterations, supervised via cross-entropy loss on gold CoT-step tokens at each latent position. Results show 2.5x-6.9x latency reduction over explicit CoT while matching performance, with interpretable latent spaces that recover gold reasoning steps and alternative valid intermediates. Ablations confirm the necessity of both looped architecture and parallel supervision.
latent chain-of-thoughtlooped transformersparallel supervisioninterpretable latent spacereasoning efficiency
Policy Optimization Achieves Data-Dependent Regret Bounds in MDPs with Unknown Transitions
The paper introduces a policy optimization algorithm for online episodic tabular Markov decision processes with unknown transition kernels, achieving data-dependent regret bounds in both adversarial and stochastic regimes. The method employs optimistic follow-the-regularized-leader with novel $Q$-function estimators and a data-dependent transition bonus to control estimator bias via loss-prediction error. Results include first-order, second-order, and path-length bounds with a transition-dependent complexity term, alongside gap-dependent $\mathrm{polylog}(T)$ regret in stochastic settings.
policy optimizationmarkov decision processesdata-dependent regretq-function estimatorstransition bonus
Addressing Over-Refusal in LLMs with Competing Rewards
The paper introduces SEAR, a method to mitigate over-refusal in safety-trained LLMs by leveraging unsafe reasoning as an exploratory signal. The approach frames safety training as an adversarial optimization problem with competing rewards: a reasoning player explores unsafe responses while an answer player ensures safe outputs, both implemented within a single model via dense process rewards. Experiments show SEAR reduces over-refusal while maintaining safety, demonstrating robustness against reasoning-based attacks.
over-refusalsafety trainingadversarial optimizationchain-of-thoughtprocess rewards
Is Natural Always Appropriate? Investigating Naturalness and Appropriateness Across Different Domains for TTS Evaluation
This work investigates how perceived naturalness and appropriateness in text-to-speech (TTS) systems vary across different application domains. The study evaluates five state-of-the-art TTS systems across five domains (AI assistant, reader, actor, animated character, spontaneous speaker) using human judgments. Results reveal domain-dependent appropriateness that diverges from naturalness ratings, with systems excelling in reading tasks but struggling with expressive domains. Naturalness scores were found to favor spontaneous speech while penalizing stylized outputs. The findings demonstrate the limitations of universal TTS evaluation metrics and emphasize the need for context-aware assessment.
text-to-speechnaturalnessappropriatenessevaluation metricsdomain adaptation
Nonlinearity-Aware LoRA: Structured Gate Adaptation under Low-Rank Constraints
NA-LoRA introduces a nonlinearity-aware principle for parameter-efficient fine-tuning in self-gated Transformer feed-forward networks, addressing selection misalignment caused by low-rank updates. The method employs two lightweight mechanisms: a derivative-based temporal-importance mask for gate-related LoRA updates and an activation-specific step-scaling rule when effective-homogeneity partitions are available. NA-LoRA adds no auxiliary loss or inference-time overhead. Evaluations on language-model fine-tuning and vision-language transfer benchmarks demonstrate consistent improvements over vanilla LoRA and competitive performance against strong PEFT variants.
low-rank adaptationself-gated transformerselection misalignmentparameter-efficient fine-tuningeffective-homogeneity
WIDER-FAIR: An Annotated Version of the WIDER-FACE Dataset for Fairness Evaluation
The authors introduce WIDER-FAIR, an annotated version of WIDER-FACE containing 16,256 images labeled with perceived ethnicity (Asian, Black, Indian, White) and sex to evaluate fairness in face detection. Annotation consistency is validated through face embeddings, K-Nearest Neighbors classification, and t-SNE visualization. Experiments with YOLOv5 reveal significant performance disparities, particularly lower detection accuracy for Black individuals, and demonstrate that excluding this group during training exacerbates fairness gaps more than other ethnicities.
face detectionfairness evaluationdemographic biasdataset annotationperformance disparity
Diffusing Blame: Task-Dependent Credit Assignment in Biologically Plausible Dual-Stream Networks
The study extends Error Diffusion (ED), a biologically plausible learning rule for dual-stream excitatory/inhibitory networks adhering to Dale's principle, to achieve competitive performance in both supervised and reinforcement learning. The method introduces modulo error routing, layer-specific sigmoid widths, batch-centered class error signals, and asymmetric initialization, achieving 96.7% accuracy on MNIST and 61.7% on CIFAR-10. In reinforcement learning, ED-PPO matches Direct Feedback Alignment on continuous-control tasks in Google Brax and Craftax. Ablation studies reveal task-dependent credit-assignment bottlenecks.
dale's principleerror diffusioncredit assignmentproximal policy optimizationbiologically plausible learning
Calibration, Not Compilation: Detecting and Repairing Misspecified Probabilistic Programs Written by Language Models
The paper introduces a calibration-based approach for detecting and repairing statistical misspecifications in probabilistic programs generated by language models, surpassing traditional unit-test verification. It evaluates detection via posterior predictive checks and simulation-based calibration on 200 instances across 10 model families, achieving 0.97 AUC when using a reference program and 62-78% reference-free. Repair experiments with LLM feedback show calibration outperforms unit tests (GPT-5.1: 33→92%, Claude: 75→100%). Real-world tests reveal 15-47% of runnable LLM-generated programs are misspecified, with calibration-guided repair surpassing alternative methods.
probabilistic programmingbayesian workflowposterior predictive checkssimulation-based calibrationmisspecification detection
On Optimal Data Splitting for Split Conformal Prediction
The paper develops a theoretical framework for optimal data splitting in split conformal prediction, addressing the unresolved issue of minimizing prediction interval length while maintaining coverage. It derives analytical characterizations of the length-optimal split ratio under symmetric and asymmetric regimes, and specializes these results to common regression settings including linear regression, nonparametric regression, and neural networks. A data-based method for selecting the optimal proportion is also described. Experiments on synthetic and real-world datasets demonstrate the framework's applicability across various scenarios, providing principled guidance for constructing shorter prediction intervals.
split conformal predictionprediction intervalsoptimal data splittingnonparametric regressioncoverage guarantees
From Failure to Alignment: A Requirements Engineering Framework for Machine Learning Systems
The paper proposes REAL (Requirements Engineering for mAchines that Learn - and Fail), a model-based framework for aligning machine learning systems (MLS) with stakeholder needs through requirements engineering. REAL integrates three principles: weaving data, model, and system requirements; using failure modes to explore alternatives; and iterative requirement refinement with traceability. The framework is demonstrated via an autonomous driving case study, showing improved alignment with stakeholder requirements. A replication package is provided for validation.
requirements engineeringmachine learning systemsstakeholder alignmentfailure analysismodel-based framework
Robustness of neural networks to random noise perturbations of their inputs
The authors introduce a computationally efficient robustness measure for neural networks, providing an upper bound on mean squared error under input perturbations with high probability. Treating networks as black boxes, the method evaluates robustness by analyzing the interplay between accuracy and perturbation sensitivity. Experimental validation on multiple real-world datasets demonstrates the measure's efficacy. Additionally, robustness curves are proposed to enable comparative analysis of robustness across and within datasets, offering insights into network behavior under noise.
robustness measuremean squared errorinput perturbationsblack box analysisrobustness curves
Localized Conformal Prediction for Image Classification with Vision-Language Models
The paper introduces a non-linear transformation of cosine similarities to improve localized conformal prediction for image classification with vision-language models (VLMs). While standard approaches using visual feature similarities fail to outperform non-local baselines, the proposed method maintains marginal coverage guarantees and significantly reduces mean prediction set sizes. Extensive benchmarking on natural image tasks demonstrates statistically significant improvements over conventional methods, with open-source implementation provided.
conformal predictionvision-language modelsuncertainty quantificationcosine similarityimage classification
Introduction to Stochastic Differential Equations for Generative Machine Learning: A Variational Perspective
The paper introduces stochastic differential equations (SDEs) and their probabilistic framework for generative machine learning, focusing on applications in image, video, and biomolecule generation. It derives the variational lower bound (ELBO) as a unifying foundation for diffusion models, score matching, and flow matching, framing these approaches as specific parameterizations of a general variational method. The Fokker--Planck equation is discussed to describe the temporal evolution of marginal distributions in SDEs. A one-dimensional density modeling problem illustrates the comparative analysis of these parameterizations.
stochastic differential equationsvariational lower bounddiffusion modelsscore matchingflow matching
Beyond the Expressivity-Trainability Paradox: A Dynamical Lie Algebra Perspective on Navigating Barren Plateaus in Quantum Machine Learning
The study addresses the expressivity-trainability paradox in Quantum Machine Learning (QML), demonstrating that high Hilbert space capacity in Parameterized Quantum Circuits (PQCs) induces Barren Plateaus (BPs) with exponentially flat gradients. By integrating Dynamical Lie Algebras (DLAs) and Geometric QML, the authors establish a framework connecting algebraic dimension of circuit generators to optimization dynamics. Empirical validation on binary classification reveals that symmetry-preserving geometric priors enable scalable training by restricting DLA growth, trading memorization for gradient-rich landscapes.
quantum machine learningbarren plateausdynamical lie algebrasparameterized quantum circuitsgeometric qml
RaBitQCache: Rotated Binary Quantization for KVCache in Long Context LLM Inference
RaBitQCache introduces a sparse attention framework for KV-cache optimization in long-context LLM inference, addressing limitations of static retrieval and biased proxy scores. The method employs randomized rotated binary quantization and binary-INT4 arithmetic to efficiently estimate attention weights, enabling adaptive Top-p retrieval with proven error bounds. Evaluations show significant inference acceleration and memory I/O reduction while maintaining generation quality, outperforming state-of-the-art baselines.
kv-cachesparse attentionbinary quantizationtop-p retrievallong-context inference
Falsification, Not Exposure: An Internally Preregistered Placebo-Controlled Decomposition of Self-Repair Feedback in Frozen Small Code Models
The study introduces a falsification-centered methodology to evaluate self-repair feedback in frozen small code models, contrasting with conventional retry mechanisms. Using placebo-controlled decomposition (packet decomposition, placebo mirroring, matched-budget tests), it assesses feedback's efficacy by comparing blind resampling and content-free placebos. Experiments on six HumanEval+/MBPP+ tasks with 0.5B-1.5B models (7,000 generations) show blind resampling outperforms bare-code retry (+18 net unlocks, p=0.0021), while code-plus-facts feedback matches resampling (26 unlocks each). No significant instruction-only effect was found (p=0.36).
frozen code modelsself-repair feedbackplacebo-controlled decompositionblind resamplingexecutable counterexamples
Fork-Think with Confidence
The paper introduces Fork-think with confidence, a decide-first-then-think paradigm for efficient LLM reasoning that identifies forking points using model confidence in a single seeding path before sampling multiple continuations. This method reduces token consumption by up to 30% and runtime by 57% while matching or outperforming parallel thinking on three models and reasoning benchmarks. Analysis shows meaningful forking point identification and improved generations from later sampling, with further gains when combined with early stopping and weighted voting.
llm reasoningforking pointsmodel confidenceparallel thinkingearly stopping
Constrained Online Convex Optimization without Slater's Condition
We propose an anytime primal-dual framework for constrained online convex optimization that eliminates the need for Slater's condition while maintaining optimal regret and constraint violation bounds. The method incorporates an adaptive regularizer into the dual update, stabilizing the dual process without relying on negative drift assumptions. For stochastic constraints and convex losses, the algorithm achieves $O(\sqrt{T})$ expected regret and $O(\sqrt{T}\log T)$ expected cumulative constraint violation, with high-probability bounds of the same order. For strongly convex losses, regret improves to $O(\log T)$ with matching violation bounds. The framework also extends to adversarial constraints with guarantees for hard constraint violation.
online convex optimizationprimal-dual frameworkslater's conditionadaptive regularizerconstraint violation
TabPATE: Differentially Private Tabular In-Context Learning Without Public Data
TabPATE introduces a differentially private PATE-style defense for tabular in-context learning (ICL) that eliminates the need for public in-distribution data. The method partitions private context across teacher models, privately aggregates their labels on synthetic tabular queries, and releases the labeled queries as a student context. TabPATE leverages bounded, low-dimensional tabular features to generate useful queries from feature ranges or lightly privatized marginals. Evaluated across tabular benchmarks, TabPATE maintains competitive utility while reducing membership inference attack success to near-random levels, offering a practical solution for private tabular ICL.
differential privacyin-context learningtabular datamembership inferencepate-style defense
Zero-Shot Quantization for Object Detectors using Off-the-Shelf Generative Models
GoodQ introduces a Zero-Shot Quantization-Aware Training (QAT) pipeline for Object Detection (OD) models using off-the-shelf generative models to construct training sets when original data is inaccessible. It addresses three key challenges: dense multi-instance images, imbalanced class distributions, and noisy pseudo-labels, through Information-Dense Prompting, Intrinsic Distribution-Aware Selection, and Teacher-guided Adaptive Noise Reduction. The framework achieves state-of-the-art performance in low-bit (W4A4) and extreme bit-width (W3A3) quantization, while providing extensive analysis of its efficacy.
zero-shot quantizationobject detectionquantization-aware traininggenerative modelslow-bit quantization
Contextual Slate GLM Bandits with Limited Adaptivity
The authors propose two algorithms for contextual slate bandits with generalized linear rewards under limited adaptivity: B-SlateGLinCB for batched settings and RS-SlateGLinCB for rarely-switching settings. B-SlateGLinCB partitions the time horizon into O(log log T) batches, while RS-SlateGLinCB performs O(Nd log T) parameter updates. Both algorithms achieve regret bounds of O(Nd^(3/2)√T) and O(Nd√T) respectively, independent of the GLM non-linearity parameter κ. The methods are computationally efficient (poly(N) time per round) and outperform baselines in simulations, with B-SlateGLinCB matching the fully adaptive Slate-GLM-OFU. Practical validation includes in-context example selection for language models.
contextual banditsgeneralized linear modelslate recommendationlimited adaptivityregret bounds
Linguistic Bias Mitigation for Spoofing Detection via Gradient Reversal and A Variational Information Bottleneck
The paper proposes a linguistic-invariant spoofing detection framework to address performance degradation in out-of-domain settings caused by linguistic bias. The method combines teacher-student adversarial learning—where a linguistic-aware teacher guides the student via gradient reversal—with a Variational Information Bottleneck to suppress principal linguistic cues while preserving non-linguistic spoofing indicators. Evaluated across nine DF Arena datasets, the approach achieves up to 36.2% relative reduction in equal error rate compared to baselines.
spoofing detectionlinguistic biasgradient reversalvariational information bottleneckteacher-student learning
Direction-Magnitude Decomposition for Low-Rank Matrix Optimization: Faster Convergence and Saddle-to-saddle Dynamics
The paper introduces Direction-Magnitude Decomposition (DMD), a unified framework for low-rank matrix optimization that improves efficiency when the target rank is unknown. DMD decomposes the optimization variable into direction and magnitude components, enabling two approaches: overparameterized DMD, which uses a higher rank for faster convergence, and recursive DMD, which reduces memory and computational costs by leveraging saddle-to-saddle dynamics. Both methods exhibit exponential speedup over gradient descent on the Burer-Monteiro formulation. Empirical validation on matrix factorization, sensing, and completion tasks confirms the theoretical advantages and practical effectiveness of DMD.
low-rank matrix optimizationdirection-magnitude decompositionburer-monteiro formulationsaddle-to-saddle dynamicsmatrix factorization
Dualformer: Efficient Feature Extractor for Complex-valued Blind Communication Signal Analysis
The paper proposes Dualformer, a Transformer-based dual-channel neural network (DualNN) for complex-valued blind signal analysis tasks like automatic modulation recognition (AMR) and signal structure parsing (SSP). Dualformer shares parameters across IQ channels while processing complex signals, theoretically reducing generalization error without sacrificing expressivity. It employs patch-level tokenization and multi-granularity feature extraction. Experiments show consistent improvements over three Transformer baselines and four conventional DL approaches on AMR, SSR, and SSP tasks, with demonstrated generalizability to blind source separation and low-SNR spectrum sensing.
dual-channel neural networkcomplex-valued signalsautomatic modulation recognitiontransformer architectureparameter sharing
Domain-Decomposed Randomized Neural Networks for Partial Differential Equations in Unbounded Domains
A domain-decomposed randomized neural network framework is proposed for solving partial differential equations (PDEs) on unbounded domains, addressing challenges of truncation error and localized structures. The method assigns distinct randomized subnetworks to near-field and far-field regimes, coupled via boundary and interface conditions, with output-layer coefficients solved through linear least-squares systems derived from Petrov--Galerkin or collocation formulations. Theoretical analysis includes a conditional bounded-parameter approximation result and error decomposition. Numerical experiments on Poisson and time-dependent Schrödinger equations demonstrate the method's accuracy and flexibility.
partial differential equationsdomain decompositionrandomized neural networkspetrov--galerkin methodcollocation formulation
Expected Gain-based Escalation in Vertical Federated Learning
The paper introduces an expected-gain-based selective escalation method for vertical federated learning (VFL) to optimize the communication-accuracy trade-off. The proposed two-round inference protocol first uses client posteriors for low-cost predictions, then invokes embedding fusion only when expected to improve correctness, formulated as an analytical score combining calibrated pooled posteriors and classwise reliability estimates. Experiments on multi-view classification benchmarks demonstrate superior communication-accuracy performance compared to confidence-, learned-gain-, and deferral-based baselines, particularly under test-time view degradation.
vertical federated learningselective escalationcollaborative inferencemulti-view classificationexpected-gain score
Safe Online Learning via Smooth Safety-Structured Policy Composition
AutoSafe introduces a safety-aware policy architecture for online reinforcement learning that integrates structured safety monitoring and intervention into action generation. This approach enables smooth, risk-dependent transitions between performance-driven and safety-preserving behaviors, maintaining continuous interaction and learning dynamics. Empirical evaluations on continuous-control benchmarks demonstrate robust safety enforcement without compromising learning smoothness. Practical validation on a physical cart-pole system confirms AutoSafe's effectiveness for real-world safe online learning.
reinforcement learningsafety constraintspolicy architecturecontinuous-controlonline learning
Deep Reinforcement Learning for Spacecraft Attitude Control During Atmospheric Re-Entry
The study demonstrates that deep reinforcement learning (RL) can outperform traditional proportional-integral-derivative controllers for spacecraft attitude control during atmospheric re-entry. Using continuous, off-policy RL with dynamics randomization to enhance generalization, the authors develop hybrid controllers combining RL and traditional methods. Results show superior performance in tracking angle of attack and robustness to variations in mass, inertia tensor, and actuator bandwidth within a predefined operational envelope, though out-of-distribution generalization remains limited.
deep reinforcement learningattitude controldynamics randomizationproportional-integral-derivativespacecraft re-entry
Patch-PODiff-ViT: Structured Latent Diffusion with Patchwise POD for Super-Resolution and Uncertainty Quantification
Patch-PODiff-ViT introduces a structured latent diffusion framework using patchwise Proper Orthogonal Decomposition (POD) for super-resolution and uncertainty quantification. The method replaces learned nonlinear autoencoders with a fixed linear orthonormal basis over local patches, yielding low-dimensional, variance-ordered tokens that preserve spatial structure. This enables efficient diffusion in a structured latent space via Vision Transformers, with analytic uncertainty propagation through the linear decoder. Evaluated on sea surface temperature, medical imaging, and natural images, it achieves strong reconstruction with reduced parameters and memory while producing well-calibrated spatial uncertainty matching empirical ensembles.
latent diffusionproper orthogonal decompositionvision transformeruncertainty quantificationsuper-resolution
Probabilistic Inversion with Flow Matching
The paper adapts Flow Matching, a generative AI technique, for probabilistic inversion in geophysical applications like Full-Waveform inversion. The method leverages the mathematical framework of Flow Matching to address inverse problems, demonstrating its utility through two case studies: a simplified 2D velocity model and the complex OpenFWI dataset. Results show the approach's effectiveness in probabilistic inversion for seismic velocity models.
flow matchingprobabilistic inversionfull-waveform inversiongenerative aiseismic velocity models
Sequential sparse Gaussian process quantile regression
The paper introduces a sequential sparse Gaussian process framework for Bayesian quantile regression, addressing computational challenges via inducing variables and Laplace approximation. The method decomposes predictive uncertainty into conditional-prior and posterior-induced variance to drive adaptive inducing-input infilling and data acquisition. Experiments on benchmarks show the Laplace approximation's accuracy, benefits of variance-based inducing-input placement, and superior performance of the sequential enrichment strategy over predefined approaches.
quantile regressiongaussian processlaplace approximationinducing variablessequential learning
Revisiting the Volume Hypothesis
The study reconciles contradictory evidence regarding the volume hypothesis in deep learning by examining different dataset size regimes. Using the Replica Exchange Wang-Landau algorithm, the authors estimate the joint density of states over training and test accuracies in binary networks. Results across multiple architectures and datasets show that gradient learning's generalization advantage over random sampling diminishes with increased training data size, resolving the apparent paradox.
volume hypothesisimplicit biasstochastic gradient descentdensity of statesgeneralization
The Calibration Turn in AI-Assisted Research: A Conceptual and Methodological Framework for Evidence-Licensed Claims
The paper introduces a conceptual framework for evidence-licensed claims in AI-assisted research, emphasizing calibration as a mechanism for managing scientific assertion rights. It identifies five operators: hypothesis generation, model-mediated consequence derivation, external validation, belief update, and claim calibration, and distinguishes linguistic, consequence-based, interventional, and evidence-licensed semantics. The framework defines the claim-evidence gap and epistemic debt, proposing minimal structural reconstruction as a form of claim calibration. Principles include no claim without license, validation not determining claim level, and automation amplifying calibration needs. Reliable AI-assisted research is evaluated as a loop generating hypotheses, deriving testable consequences, accepting independent adjudication, updating beliefs, and outputting evidence-licensed claims.
calibrationevidence-licensed claimsepistemic debtclaim-evidence gapmodel-mediated consequence derivation
The Decomposition Is the Fingerprint: Per-Component Identity for Agent Skills
The paper introduces a locality-sensitive fingerprint for AI agent skills, decomposing them into per-component triples (prompt, code, tools) to enable stable identity tracking across paraphrases, refactoring, and controlled translations. The method employs a multi-bank SimHash to project each component into a compact 120-byte signature, compared via Hamming distance, preserving ranking with 77x fewer bits than full embeddings. Evaluation on 4,950 pairwise comparisons shows AUC 0.974 (95% CI [0.956, 0.994]), with successful localization of tampered components in a 906-skill benchmark, though it remains complementary to behavioral verification.
locality-sensitive fingerprintsimhashhamming distanceagent skillscomponent decomposition
MNAR-$k$-means: A $k$-means Clustering for Data Missing Not at Random with Magnitude-Decaying Probability
The paper proposes MNAR-$k$-means, a novel $k$-means clustering method for data missing not at random (MNAR) with magnitude-decaying probability, where missingness correlates with smaller absolute values. The method constrains imputation values to mitigate distortion of cluster centers, supported by statistical consistency guarantees for estimated centers versus true fully observed data. An alternating minimization algorithm optimizes the proposed loss function. Simulations demonstrate improved clustering accuracy and reduced center bias, with real-world applications validating practical utility.
k-means clusteringmissing not at randommagnitude-decaying probabilitystatistical consistencyalternating minimization
Scaling Storm-Resolving Atmospheric AI Simulation to the Entire Planet
The authors introduce STRATA (Storm-resolving Tile-based autoRegressive Atmosphere Transformer Architecture), the first autoregressive AI emulator for global storm-resolving atmospheric dynamics at 4.9-km resolution. The method employs 3D patch embedding, local 3D neighborhood attention, Stereographic Rotary Position Embedding (StereoRoPE), and a pixel-space de-aliasing decoder, trained on 17 days of SCREAM physics-model output. STRATA achieves stable 24-hour rollouts with realistic km-scale dynamics, 48 simulation days per megawatt-hour (50× more efficient than SCREAM), and 741 simulated days per wall-clock day on 512 H100 GPUs, though large-scale biases emerge with longer lead times.
storm-resolving simulationautoregressive transformer3d neighborhood attentionstereographic rotary embeddingkm-scale emulation
Learning Gaussian Graphical Models from a Glauber Trajectory Without Mixing
The authors present a polynomial-time algorithm for learning the structure of a $d$-sparse Gaussian graphical model from a single Glauber dynamics trajectory, without requiring mixing-time assumptions. The method involves three key components: (1) conditional variance estimation and trajectory rescaling to unit-diagonal form, (2) a local edge test isolating pairwise influence from short update windows, and (3) robust median-based aggregation of local statistics to handle temporal dependence. The algorithm achieves sublinear-in-$n$ sample guarantees under general sparsity and minimum edge-strength conditions, addressing a gap in prior i.i.d. sample approaches.
gaussian graphical modelsglauber dynamicsstructure learningconditional variancerobust estimation
Probing Memorization of Tabular In-Context Learning
The study investigates parametric memorization in large tabular models (LTMs) using in-context learning (ICL), introducing ICLMEM, a probing framework to isolate memorization from contextual predictions. The method employs a zero-information multiple-choice context and controlled fine-tuning to address distribution shift, feature contamination, and base-rate fallacy. Results on a real-world LTM show moderate memorization signals in 8/10 tasks (AUC ≤ 0.67, TPR > 0.1 at 1% FPR), strongest for low-cardinality and binary tasks, but largely absent under realistic training conditions.
tabular foundation modelsin-context learningparametric memorizationmembership inferencedistribution shift
Machine Learning-based Feedback Linearization Control of Quadrotor Subject to Unmodeled Dynamics
The paper presents a machine learning-based feedback-linearization controller for quadrotors with unmodeled dynamics, using a Gaussian RBF neural network for real-time compensation. The method employs online weight adaptation without pretraining and guarantees stability via Lyapunov theory, ensuring asymptotic convergence in trajectory tracking. Experimental results on a Crazyflie 2.1 quadrotor demonstrate 7.13% and 49.27% RMSE reductions in position and yaw tracking respectively compared to baseline, despite unmodeled drag and disturbances.
feedback linearizationradial basis functionlyapunov stabilityunmodeled dynamicstrajectory tracking
ISM:Self-Improving Strategy Memory for Continual Mathematical Reasoning
We introduce Intelligent Schema Memory (ISM), a self-evolving memory-augmented system that enhances mathematical reasoning for frozen large language models (LLMs) under continual learning with episodic resets. ISM maintains a compact, self-refined bank of strategy schemas, leveraging symbolic tools to verify intermediate steps and certify answers. Without updating model parameters, ISM outperforms passive, retrieval, and reflection baselines on MATH-Hard and OlympiadBench, using 64% and 86% fewer schemas respectively than the strongest passive baseline. These results demonstrate that small, actively maintained, and verified strategy memories enable reliable continual mathematical reasoning under strict episodic isolation.
intelligent schema memorycontinual learningepisodic resetssymbolic toolsmathematical reasoning
Probe Choice Changes Canary-Memorization Verdicts: Three Post-Hoc Disagreement Case Studies in a Text-Dominant LoRA-Tuned Autoregressive Testbed
The study identifies discrepancies in memorization detection when using a fixed prefix-window mean-NLL probe (K=20) on a Qwen2.5-VL-7B model, revealing three post-hoc cases of disagreement with full-span secret NLL and greedy exact-recall metrics. Case C3 demonstrates a false negative due to window truncation, C4 a false positive from non-secret drift, and C5 an ambiguous in-window drop despite positive full-span hex. The authors recommend reporting full-span secret NLL, span-localized decomposition, behavioral exact-recall at k>=4, and decoy probes to ensure secret-specificity. Findings are based on controlled canaries within a single backbone, with magnitudes specific to the testbed.
memorization probeprefix-windownllexact-recallcanary testbed
TAG-DLM: Diffusion Language Models for Text-Attributed Graph Learning
TAG-DLM introduces a diffusion language model for text-attributed graph (TAG) learning that unifies textual reasoning and graph message passing. The method linearizes sampled local neighborhoods into token sequences and injects graph structure via a topology attention mask, enabling bidirectional attention and generative decoding. This approach supports node classification, link prediction, and cross-dataset transfer without task-specific fine-tuning. Experiments demonstrate TAG-DLM outperforms graph neural networks, graph transformers, and LLM-based baselines on three TAG benchmarks, achieving up to 3.9-point improvements over the strongest baseline.
diffusion language modeltext-attributed graphtopology attention maskbidirectional attentiongenerative decoding
An Empirical Study of Security Calibration in Large Language Models for Code
This work presents the first large-scale empirical study of security calibration in LLM-generated code, evaluating GPT-4o-mini, Gemini-2.0-Flash, and Qwen3-Coder-Next across temperature settings on security tasks and repository-level contexts. Results demonstrate prevalent overconfidence, with security calibration outperforming functional calibration, suggesting models better estimate security outcomes than functional correctness. Calibration-guided repair yields limited improvements while introducing functional regressions, and architectural gating improves calibration on controlled benchmarks but deteriorates in realistic settings, increasing high-confidence vulnerable outputs.
security calibrationllm-generated codefunctional regressionarchitectural gatingfalse trust
A Bayesian Filtering Approach for Learning Lagrangian Dynamics from Noisy Measurements
The paper introduces a Bayesian filtering approach for learning Lagrangian dynamics from partial, noisy measurements. The method models system dynamics using Lagrangian mechanics, parameterizing kinetic and potential energies with neural networks and treating unknown external forces as white Gaussian noise. This yields a continuous-time stochastic state-space model, with neural network parameters and system states jointly learned via maximum-likelihood estimation using Gaussian-approximation-based Bayesian filters. The approach is validated on pendulum and Duffing oscillator examples, demonstrating superior performance compared to conventional Lagrangian neural networks and approximate Bayesian filters with known system models.
bayesian filteringlagrangian mechanicsstochastic state-space modelmaximum-likelihood estimationneural networks
MSNN-LINet: Cross-Modal Learning via Continuous Linear Integration
LINet introduces a Multi-Stream Neural Network (MSNN) for RGB-D scene classification, addressing limitations in multi-modal feature fusion via continuous cross-modal learning. The architecture employs three parallel streams (RGB, depth, integration) with a novel Linear Integration Convolution (LIConv2d) operator, enabling integration before nonlinear activation thresholds. A 1/N constant initialization mitigates gradient flow issues, while progressive modality dropout prevents pathway collapse. Trained on SUN RGB-D, LINet achieves 45.2% mean class accuracy at ResNet18 scale, surpassing prior from-scratch results, and reaches 49.6% with ScanNet pretraining.
multi-stream neural networklinear integration convolutionrgb-d scene classificationprogressive modality dropoutsomatic integration
Can Tabular In-Context Learners Generalize to Biomolecular Property Prediction?
The study demonstrates that tabular foundation models (TabPFN3, TabICL), despite being pretrained on synthetic causal graphs, effectively transfer to biomolecular property prediction tasks. Using fixed ESMC representations for proteins and ECFP/RDKit descriptors for small molecules, the authors evaluate in-context learning across ProteinGym, esterase datasets, TDC ADMET, MoleculeNet, FS-Mol, and DrugOOD. Results show competitive performance for protein fitness regression and variable outcomes for small-molecule classification, where representation choice dominates predictor performance. The work establishes tabular models as viable few-shot learners for biomolecular prediction.
tabular foundation modelsin-context learningbiomolecular property predictionfew-shot regressionmolecular representation
Visualizing High-Dimensional Graph Embeddings via Informed Multi-View Projections
The paper introduces a method for visualizing high-dimensional graph embeddings through optimized 2D projections that preserve structural patterns. The approach employs a differentiable surrogate for edge crossings to search for viewpoints maximizing aesthetic metrics (e.g., edge crossings, angular resolution). Experiments demonstrate that these projections outperform standard 2D layouts and even specialized metric-optimization methods. The authors also present DataFly, an interactive system for exploring multiple viewpoints, with a usability study confirming improved pattern discovery over conventional visualizations.
graph embeddingsdifferentiable surrogateaesthetic metricsmulti-view projectionsinteractive visualization
Explaining Machine Learning and Memorization with Statistical Mechanics
The thesis advances theoretical understanding of neural networks and machine learning by analyzing adversarial attacks and low-dimensional learning dynamics using statistical mechanics. It examines dense associative memory (DAM) and restricted Boltzmann machines (RBM) as model classes exhibiting varying degrees of learning versus memorization. Analytical connections between model variants are established to improve computational efficiency in theoretical investigations.
statistical mechanicsneural networksadversarial attackslow-dimensional learningdense associative memory
Fora: From Weight-Space to Function-Space Protection in Capability-Preserving Fine-Tuning
The paper introduces FORA (Function-space Orthogonal Residual Adaptation), a method for capability-preserving fine-tuning that protects activation subspaces rather than weight-space geometry. FORA estimates principal directions of input-activation covariance via label-free calibration inputs, constructing orthogonal projectors to constrain updates while allowing controlled plasticity through a spectral channel. Evaluated on Qwen3-1.7B across COGS, GSM8K, and translation tasks, FORA outperforms weight-space projection and standard regularization in capability preservation, with minimal new-task trade-offs. Ablations confirm the advantage stems from capability-derived, not weight-derived, projection directions.
fine-tuningactivation subspaceorthogonal projectioncapability preservationspectral channel
Dynamic Gaussian Processes and the Vanilla-SPDE Exchange
The authors propose Vanilla-SPDE Exchange, a hybrid Gaussian process inference method that combines standard and SPDE formulations to reduce computational costs in spatio-temporal settings. By exploiting an equivalence between these formulations, the method achieves linear complexity in time while mitigating the cubic spatial complexity of exact inference, particularly when observation and prediction locations are disjoint. Complexity analysis and numerical experiments demonstrate improved computational efficiency compared to existing approaches.
gaussian processesspatio-temporal inferencecomputational complexityspde formulationhybrid scheme
Online TT-ALS for Streaming Tensor Decomposition with Incremental Orthogonalization
The authors propose Online TT-ALS, a streaming tensor decomposition algorithm that incrementally enforces orthogonality constraints for efficient core tensor updates. The method combines alternating least squares with sequential orthogonalization, theoretically guaranteeing monotonic objective decrease and temporal smoothness while reducing rank dependence from quadratic to linear (O(I^{n-1}r) complexity). Experiments show superior approximation accuracy and video quality metrics over existing online methods, with orders-of-magnitude speedups compared to deep learning approaches, enabling low-latency real-time processing.
tensor train decompositionalternating least squaresorthogonal gauge constraintsstreaming datalow-latency processing
Warp RL: Reshaping Base Policy Distributions for Dynamics Adaptation
Warp RL introduces a policy adaptation method that replaces additive residual corrections with state-conditioned invertible transformations of base policy distributions, addressing limitations in variance, confidence calibration, and non-uniform corrections under dynamics shifts. The method employs monotonic rational-quadratic spline flows for identity-preserving initialization and structured adaptation, generalizing additive residuals. Evaluated on ManiSkill3 manipulation tasks with controlled dynamics shifts, Warp RL matches residual correction when translation suffices and outperforms it by 30% in task completion speed for real-robot peg-insertion when distributional reshaping is required.
residual reinforcement learningdynamics adaptationinvertible transformationsrational-quadratic spline flowspolicy-gradient optimization
Teaching LLMs to Recommend and Defer in Underrepresented Epilepsy Care
The paper introduces MANANA, a non-parametric prompt-learning framework for adapting LLMs to recommend anti-seizure medications in Ugandan pediatric epilepsy care while learning when to defer. MANANA converts prescription errors into auditable prompt memories, with single-agent and multi-agent variants, and employs Bayesian prompt averaging for uncertainty-based deferral. Evaluated on two Ugandan cohorts, it improves top-3 prescription accuracy by 4-8 percentage points over baselines and enables selective prediction with 95-99% precision for confident cases.
prompt-learninganti-seizure medicationbayesian prompt averagingselective predictionlongitudinal treatment
Offline Reinforcement Learning for Fluid Controls: Data-based Multi-observational Policy Extraction
The authors propose an offline reinforcement learning framework for fluid control that enables policy adaptation to multiple sensor configurations without retraining. Their method employs a sensor position-conditioned architecture with Point Attention layers to model spatial relationships, allowing a single policy network to generalize across varying sensor placements. Evaluations on Kuramoto-Sivashinsky equation and Navier-Stokes flow control demonstrate the approach's flexibility for sensor optimization, reducing computational costs compared to online RL methods.
offline reinforcement learningfluid controlpoint attentionsensor optimizationpolicy extraction
Certified Speculative Execution for Untrusted AI Agents
The paper introduces Certificate-Gated Prefix Acceptance (CGPA), a certified speculative-execution framework for untrusted AI agents in hard-constrained sequential decision systems. CGPA combines a trusted verifier for exact constraint violation rejection, a conformally calibrated value boundary for prefix gating, and solver deferral to decouple safety, regret, and speed. Evaluations show zero constraint violations across adversarial drafters and six frozen LLMs (including a 12B model), with mean regret reduced by three orders of magnitude versus unguarded acceptance. Deployment on unit-commitment demonstrates a 2.96x speedup using an 8B LLM at 2.1% regret, outperforming domain heuristics (1.79x) and safe baselines (1.07x).
certified speculative executionconstraint violation rejectionconformal calibrationsequential decision systemsunit-commitment
Estimating Supply Incrementality in Two-sided Marketplaces: A Causal Machine Learning Approach
The paper proposes a causal machine learning method for estimating supply incrementality effects in two-sided marketplaces, using Airbnb as a case study. The approach combines double/debiased machine learning with hierarchical Bayesian modeling, incorporating geospatial similarity measures as informative features. Results demonstrate plausible supply elasticity estimates and strong out-of-sample predictive performance across heterogeneous product segments.
supply incrementalitytwo-sided marketplacesdouble/debiased machine learninghierarchical bayesiangeospatial similarity
Multistage Defer Trees for Hybrid Interpretability: If at First You Can't Succeed, Tree Again
The paper introduces Multistage Defer Trees, a sequence of sparse decision trees that defer uncertain predictions to subsequent trees or a black box model, balancing interpretability and accuracy. The method trains a cascade where each tree classifies most samples, routing only challenging cases downstream, thus maintaining interpretability for the majority of inputs while matching ensemble performance. Experiments show the approach expands the accuracy-interpretability frontier, achieving state-of-the-art results with predominantly single-tree predictions.
decision treesinterpretabilityensemble learningdeferral learningsparse models
Hierarchical Clustering As a Novel Solution to the Notorious Multicollinearity Problem in Observational Causal Inference
The paper introduces hierarchical clustering as a novel solution to multicollinearity in observational causal inference, addressing limitations of existing methods like shrinkage estimators that obscure original causal relationships. The method aggregates correlated variables by clustering geographic units based on normalized, demeaned marketing spend data, then applies Bayesian Marketing Mix Modeling to cluster-level data. Empirical results demonstrate reduced collinearity and improved identification of individual marketing channel impacts compared to traditional approaches.
multicollinearityhierarchical clusteringcausal inferencebayesian marketing mix modelobservational data
ElemeNet: Multiscale Molecular Machine Learning with Uncertainty Quantification Across the Periodic Table
ElemeNet introduces a unified molecular machine learning framework supporting elements 1-100, extending beyond organic chemistry to organometallic and biological systems. The package implements E(3)-equivariant and transformer architectures alongside classical 2D models, with built-in uncertainty quantification via deterministic and statistical measures. It enables atom-, bond-, molecule-, and moiety-level predictions with optional charge/spin state conditioning. Benchmarks demonstrate state-of-the-art performance across organic, inorganic, coordination, and biological chemistry datasets, scaling to millions of molecules. The command-line interface lowers adoption barriers for non-experts.
molecular machine learninguncertainty quantificatione(3)-equivariantperiodic table coveragecommand-line interface
ShardNet: Training Neural Controllers with Hard, Non-Convex Constraints
ShardNet introduces a neural network architecture that strictly enforces unions of polyhedral constraints by construction, using a differentiable projection layer parameterized by a classification network. The method embeds safety into the network's structure, enabling independent optimization of performance while maintaining formal safety guarantees. It supports nonconvex unions of polyhedras and learned value function level sets, with a novel technique for verifying and training ReLU networks. On double integrator benchmarks, ShardNet maintains 100% safety on verified sets, achieves lower objective loss than existing methods, and produces safe sets 3 times larger than prior verification approaches.
neural controllerspolyhedral constraintsforward-invariantnonconvex optimizationrelu networks
Quality-Aware Modulation for Diffusion Transformers
The paper introduces the Quality Representation Module (QRM), a lightweight transformer module that enhances image quality in diffusion transformers (DiT) by learning quality-aware representations. QRM generates vectors $M_{qrm}$ that modulate adaptive LayerNorm within DiT transformer blocks, injecting quality-sensitive signals into denoising parameters without altering the sampling schedule or diffusion backbone. Experiments demonstrate consistent improvements in image quality over baseline DiT models, supported by ablations on QRM training losses and architectures.
diffusion transformersquality representation moduleadaptive layernormdenoising parametersimage quality
Personalizing Marketplace Policies with Competing Objectives and Constrained Experiments: Evidence from a Job Marketplace
We present an integrated framework for personalizing free-value thresholds in a two-sided job marketplace, addressing competing objectives and constrained experimentation. The framework combines ensemble-based hybrid ranking models for separate target and guardrail metrics, treatment effect extrapolation under monotonicity assumptions, and production deployment validation. This approach reduces guardrail risk by over 10% compared to single-objective methods while maintaining target gains, despite limitations imposed by cluster-level randomization and few discrete treatment levels. Post-launch analysis confirms extrapolation accuracy and guardrail compliance, demonstrating effective personalization in a marketplace with millions of employers and job seekers.
two-sided marketplaceensemble-based hybrid rankingtreatment effect extrapolationcluster-level randomizationguardrail constraints
SGD at the Edge of Stability: Stochastic Stabilization with Large Learning Rates
This work provides sharp convergence guarantees for Stochastic Gradient Descent (SGD) applied to multiclass cross-entropy loss, addressing the edge-of-stability phenomenon in deep learning. The analysis focuses on linear classifiers and two-layer neural networks, demonstrating that SGD dynamics alternate between curvature-driven oscillations in the edge-of-stability regime and controlled loss reduction in the stable regime. Theoretical results prove SGD self-stabilizes, ensuring iterates return to stability within fixed iterations and enabling best-iterate convergence with large learning rates. Experimental validation confirms the theoretical findings and highlights SGD's benefits in the large-stepsize regime.
stochastic gradient descentedge-of-stabilitymulticlass cross-entropyself-stabilizationlarge learning rates
Conditional Tropical Cyclogenesis Rates via Rare-Event Sampling in a Neural Weather Emulator
The study introduces a method coupling Forward Flux Sampling (FFS) with SDL-WXFormer, a 1°-resolution neural weather emulator, to estimate conditional tropical cyclogenesis rates without altering model dynamics. FFS decomposes cyclone intensification into flux through pressure interfaces, enabling rare-event sampling with O(10^4) trajectories per initial condition. Applied to 98 Atlantic basin cases, the method resolves genesis rates spanning three orders of magnitude, showing seasonal consistency (mean FFS-to-direct-sampling ratio: 1.03 ± 0.15) and computational enhancements (geometric mean: 14X). Case studies identify rate-limiting steps in cyclogenesis, with diagnostics varying by environmental conditions.
forward flux samplingneural weather emulatortropical cyclogenesisrare-event samplingstochastic layers
Structure-Regularized Interpretable TCR-Epitope Prediction
TCR-SRIM introduces a structure-regularized interpretable model for TCR-epitope binding prediction, combining protein language model embeddings with contact prototypes to capture residue-level interactions. The method achieves state-of-the-art performance on the TCR-XAI benchmark while providing improved interpretability. Analysis reveals that AlphaFold3-, TCRModel2-, and tFold-TCR-generated structures yield competitive predictive accuracy but less precise interaction patterns and reduced binding-site diversity compared to experimental structures.
tcr-epitope bindinginterpretable-by-designprotein language modelcontact prototypesstructure prediction
Dynamic Prediction of Alternating Recurrent Events via Neural Network
The authors propose a neural network-based framework for dynamically predicting alternating recurrent events, addressing statistical challenges including correlated observations and censored outcomes. The method incorporates inverse probability weighted pseudo-observations and develops neural network theory tailored for statistical applications. Simulations demonstrate strong predictive performance, and the model excels in predicting low mood periods among first-year medical residents, showcasing its practical utility in behavioral science and biostatistics contexts.
alternating recurrent eventsdynamic predictionneural networkinverse probability weightingcensored outcomes
A Systematic Approach to Multi-Agent AI from Advanced Regulatory Control Theory: Safe and Auditable LLM Operator Agents for Process Control
The paper proposes a multi-agent LLM framework for process control derived from Advanced Regulatory Control (ARC) theory, addressing LLMs' poor performance on domain-specific tasks through structural decomposition. Each feedback loop in the ARC chain is mapped to a specialized Qwen 2.5 7B Instruct operator agent with control-theoretic context, while an orchestrator agent (either deterministic or Claude-based) manages interactions via MIN/MAX selectors and override paths. Evaluated on a dairy-barn ventilation scenario, the system produced auditable trajectories with operator-voice rationales while maintaining ARC's safety properties for constraint conflict resolution.
multi-agent systemsadvanced regulatory controlllm operator agentsmin/max selectorsprocess control
A Transferable Learned Temporal Prior for Transmission Reconstruction and Decision-Relevant Uncertainty in Real Outbreak Labels
The authors introduce a transferable learned temporal prior for outbreak transmission reconstruction, trained on eleven disease families and applied without refitting to an Andes virus benchmark. The locked prior achieved mean reciprocal rank (MRR) 0.571 and Top-1 accuracy 37.9%, significantly outperforming the best source-trained parametric baseline (MRR 0.274, Top-1 13.8%). A phylogenetic concordance audit of 75 NYC mpox inter-host pairs revealed 54.67% were genomically unresolved or unsupported, demonstrating measurable transmission-label uncertainty in outbreak evidence modules.
temporal priortransmission reconstructionphylogenetic concordancereciprocal rankoutbreak evidence
Partition-Guided Distance Saliency: Bridging Decision and Objective Spaces in Many-Objective Optimization
The Partition-Guided Distance Saliency (PGDS) framework addresses interpretability challenges in Many-Objective Optimization (MaO) by automating explanation generation through a three-stage pipeline. First, a surrogate model maps geometric distances in decision space to objective space proximity. Second, the framework partitions the objective landscape into regions and identifies Dominating Points as automated improvement targets. Third, sensitivity analysis quantifies decision variable impact, categorizing them as Drivers or Blockers. Evaluated on 10-objective benchmarks and the Welded Beam engineering problem, PGDS outperforms traditional visualization and rule-based XAI methods in providing actionable insights.
many-objective optimizationpartition-guided distance saliencysurrogate modeldominating pointssensitivity analysis
Separation Capacity of Scattering Networks
The paper extends Cover's function-counting theory to analyze the separation capacity of convolutional neural networks (CNNs) as feature extractors, focusing on scattering networks. By formulating separation capacity as a combinatorial quantity counting realizable dichotomies, the work identifies architectural factors governing this capacity in scattering networks. The results provide practical design insights for such networks, linking their building blocks to theoretical separation performance.
separation capacityscattering networksfunction-counting theoryrealizable dichotomiesfeature extractors
Mind the Residual Gap: Probabilistic Downscaling under Real-World Bias
ReMatch introduces a novel probabilistic downscaling method addressing residual target misspecification in mean–residual frameworks, a key limitation in atmospheric science and climate modeling. The approach aligns training and test-time residual distributions via optimal transport in PCA space, preserving statistical benefits while reducing train–test mismatch. Evaluated on synthetic benchmarks with varying bias levels and real-world HRRR–ERA5 wind field downscaling, ReMatch significantly reduces under-dispersion, improves calibration (SSR and CRPS), and outperforms standard mean–residual models and state-of-the-art super-resolution approaches.
probabilistic downscalingoptimal transportresidual distributionmean–residual frameworktrain–test mismatch
Using AI Agents to Automate Black-Box Audits of Personalization Algorithms at Scale
We propose a framework for black-box audits of personalization algorithms using generative AI agents as behavioral engines for synthetic accounts. Agents are instantiated with fixed personas grounded in demographic and political survey data, enabling counterfactual analysis by perturbing platform-visible signals while maintaining consistent behavior. Deploying 1,120 agents on X (formerly Twitter) across 14 personas and three counterfactual conditions, we collected over 200,000 content exposures. Results show X's algorithmic feed amplifies toxic, polarizing, political, and right-leaning content relative to the chronological feed, with demographic signals affecting content delivery in persona-dependent ways. This establishes GenAI-based agents as a scalable tool for algorithmic auditing.
black-box auditspersonalization algorithmsgenerative ai agentscounterfactual analysisalgorithmic feed
Predictable GRPO: A Closed-Form Model of Training Dynamics
The authors develop a first-principles reduced-order model to describe the training dynamics of Group Relative Policy Optimization (GRPO), a method for enhancing large language models' reasoning. The model subsumes empirical single-exponential saturation laws, predicts group-size invariance, and identifies stability thresholds and dynamical transitions. It also provides diagnostics to distinguish failure modes like reward hacking and policy concentration. Validated across three models and two group sizes, the closed-form trajectory achieves R² ≥ 0.91 in fitting training rewards and demonstrates group-size invariance on eight math benchmarks. Predictions are further confirmed in a softmax-bandit reduction.
group relative policy optimizationtraining dynamicsreward hackingpolicy concentrationsoftmax-bandit
Revocable Learned State via Process Sidecars
(No summary returned.)
ReactionAtlas: Ab origine exploration of chemical reaction networks with machine learning
ReactionAtlas introduces a machine learning framework for constructing chemical reaction networks ab origine, eliminating the need for hand-crafted rules or reactant-product pairs. The method combines a generative model proposing reactions from kinetically sampled compounds with a DFT-trained machine-learned force field (MLFF) to validate transition states (TS). Starting from eight pre-biotic seed molecules, ReactionAtlas discovers ~47,000 reactions among ~12,000 compounds, achieving 85% RMSD accuracy within 0.5 Å for MLFF TSs compared to PBE0 references. The framework maps small carbohydrate chemistry up to C$_4$H$_8$O$_4$ with unprecedented scale and accuracy, enabling novel insights into reaction pathways like the formose cycle.
chemical reaction networktransition statesmachine-learned force fieldformose cyclekinetically sampled compounds
Diffusion-warm sampling of the XY model enables fast thermalization at scale
The paper introduces a temperature-conditioned diffusion model for scalable sampling of the XY model, addressing Markov chain Monte Carlo (MCMC) limitations in generalizing across system sizes. By training on small lattices, the method generates accurate samples for larger systems, leveraging diffusion sampling followed by brief MCMC refinement. Experiments show an order-of-magnitude reduction in thermalization time compared to random-initialized MCMC, validated through spin correlation measurements. The work demonstrates generative models' potential for studying continuous-state condensed matter systems efficiently.
diffusion modelsxy modelmarkov chain monte carlospin correlationsthermalization time
TraceLab: Characterizing Coding Agent Workloads for LLM Serving
The paper introduces TraceLab, a dataset of 4,300 coding-agent sessions (350K LLM steps, 430K tool calls) from daily use of Claude Code and Codex, addressing the lack of real-world workload data for LLM serving optimization. The trace collection reveals key workload characteristics: long autonomous loops, short outputs in long contexts, heavy-tailed tool call distributions, and high but imperfect KV-cache hit rates. These findings suggest optimizations like append-length-aware prefill, semantic tool-latency prediction, and improved KV-cache management during human interaction gaps. The dataset and analysis tools are publicly released.
coding agentsllm servingkv-cacheworkload characterizationtool calls
Hierarchical Global Attention (HGA)
Hierarchical Global Attention (HGA) proposes a parameter-preserving replacement for dense causal attention in long-context transformers, enabling efficient inference without retraining. The method employs two-level hierarchical routing: first retrieving relevant chunks via RoPE-aware summaries, then refining selections for exact token-level attention, drastically reducing fetched tokens while maintaining attention quality. Implemented on Qwen3-30B-A3B-Instruct-2507-FP8, HGA achieves 64K-token contexts on a single RTX 5090 (32GB) with only 3% sparsity, staying within 0.01–0.02 nats of dense attention performance across 4K–64K contexts.
hierarchical attentionlong-context transformersrope-aware routingkv storage offloadingsparse attention
HSAP: A Hierarchical Sequence-aware Parallelism for Hybrid-Context Generative Models
The paper proposes Hierarchical Sequence-Aware Parallelism (HSAP), a framework combining existing sequence parallelism paradigms while addressing cross-contamination in attention computation for hybrid-context packed sequences. HSAP introduces a Sequence-Aware Parallelism algorithm using JIT compilation to optimize NCCL-level communication and manages memory/communication overhead hierarchically. Experiments show HSAP outperforms state-of-the-art sequence parallelism approaches across multiple metrics.
sequence parallelismhybrid-contextattention computationjit compilationnccl
Arko-T: A Foundation Model for Text-to-Structured 3D Generation
Arko-T introduces a 4B-parameter foundation model for text-to-structured 3D generation, producing editable parametric CAD programs rather than static renderings. The method aligns pipeline stages (data curation, code normalization, execution-grounded supervision) with a formal design state to preserve editability features. Evaluated against seven frontier LLMs on 12 metrics, Arko-T achieves top performance on 8 metrics and second-best on 3, at 10% benchmark cost, demonstrating targeted design-level training matches general-purpose models for structured CAD generation.
text-to-3dparametric caddesign stateexecution-grounded supervisioncode normalization
Why Do Few-Step Text Latents Fail When Image Latents Work? Non-Commitment at Sharp Categorical Readouts
The paper identifies geometric constraints as the primary cause of deterministic few-step generation failures in continuous text latents, contrasting with successful image latent generation. Through theoretical analysis, it demonstrates that decoder sharpness, not transport accuracy, governs this failure, with posterior-mean terminal steps flipping tokens at a rate proportional to latent mass near decision boundaries. Empirical diagnostics (DABI and CCI) reveal that continuous-text decoders amplify boundary-aligned perturbations significantly more than isotropic ones, unlike image decoders. The study proposes two escape mechanisms: categorical commitment and stochastic re-injection, supported by theoretical transport laws and a dimension phase diagram.
deterministic generationdecoder sharpnesscategorical commitmentstochastic re-injectiondimension phase diagram
📰 Industry Media (15)
LLMs are stuck in a groupthink groove. This startup is trying to get them out.
Springboards introduces Flint, an LLM trained on Qwen 3 to mitigate groupthink in generative AI by selectively increasing response diversity at key output points. Unlike mainstream models (ChatGPT, Claude) that converge on high-probability responses (e.g., '7' for random numbers, 'Toyota' for cars), Flint employs targeted randomness injection while maintaining coherence. Empirical tests show Flint produces divergent outputs (e.g., '3.7916', 'Ford F-150') where conventional models exhibit homogeneity, as documented in the NeurIPS-best-paper 'Artificial Hivemind'. The approach balances creativity and reliability for open-ended tasks like marketing ideation.
in-context learningresponse diversitytemperature parameteropen-source llmrandomness injection
Claude Science is Anthropic’s newest flagship product
Anthropic launched Claude Science, a flagship AI product for scientific research, extending its toolset beyond Claude Code and Claude Cowork. The system autonomously executes tasks in computational biology and drug development, interfacing with specialized tools while emphasizing reproducibility. Demonstrated capabilities include identifying drug candidates for phenylketonuria, with Anthropic leveraging it for internal neglected disease research. The Opus 4.5 model reportedly matches second-year graduate student proficiency in scientific project execution, per external evaluations.
autonomous agentscomputational biologydrug discoveryreproducibilityopus 4.5
Anthropic Redeploys Claude Fable 5 on July 1 After US Export Controls Lift, Adds New Cybersecurity Classifier
Anthropic redeployed Claude Fable 5 globally on July 1, 2026, following the lifting of US export controls imposed due to a safeguard bypass reported by Amazon researchers. The model, a Mythos-class variant optimized for general use, incorporates a new cybersecurity classifier that blocks the reported bypass technique in over 99% of cases, routing blocked requests to Claude Opus 4.8. Anthropic also proposed a four-criteria framework for scoring jailbreaks, developed in collaboration with Amazon, Microsoft, and Google. Independent benchmarks show Fable 5 outperforms rivals like GLM-5.2 on tasks such as financial analysis and codebase migrations, though GLM-5.2 offers lower cost due to its open-weight design.
claude fable 5cybersecurity classifierjailbreak frameworkglm-5.2export controls
NVIDIA Releases Nemotron-Labs-TwoTower: an Open-Weight Diffusion Language Model Built on a Frozen Autoregressive Nemotron-3-Nano-30B-A3B Backbone
NVIDIA introduces Nemotron-Labs-TwoTower, an open-weight diffusion language model addressing autoregressive generation throughput bottlenecks. The architecture employs a frozen Nemotron-3-Nano-30B-A3B autoregressive backbone (context tower) paired with a trained denoiser tower (60B total params), using layer-aligned cross-attention and Mamba-2 state propagation. At γ=0.8 and block size 16, it achieves 2.42× wall-clock throughput versus AR baseline while retaining 98.7% aggregate benchmark quality (MMLU 78.24 vs 78.56, HumanEval 75.58 vs 79.27). The denoiser trained on 2.1T tokens supports three decoding modes: diffusion, mock-AR, and AR.
diffusion language modelautoregressive backbonemamba-2kv-cacheblock-wise decoding
Google AI Introduces TabFM: A Hybrid-Attention Tabular Foundation Model for Zero-Shot Classification and Regression
Google Research introduces TabFM, a hybrid-attention foundation model for zero-shot classification and regression on tabular data. TabFM reframes tabular prediction as an in-context learning problem, processing entire datasets as unified prompts without dataset-specific training. The architecture combines TabPFN-style alternating row/column attention with TabICL-style in-context learning, trained on hundreds of millions of synthetic datasets generated via structural causal models. Evaluated on TabArena across 38 classification and 13 regression datasets, TabFM outperforms traditional supervised methods like XGBoost, achieving robust generalization without hyperparameter tuning or feature engineering.
tabular datain-context learninghybrid-attentionstructural causal modelszero-shot
CUP (Common Useful Python): Building Reliable Python Workflows with Baidu’s Utility Toolkit
Baidu's CUP (Common Useful Python) library provides a comprehensive utility toolkit for enhancing Python workflow reliability, featuring modules for logging, configuration management, caching, and concurrency. The tutorial demonstrates CUP's implementation through practical examples, including structured logging with rotation, singleton decorators, nested configuration parsing, in-memory KV caching with TTL, and thread pool management with job callbacks. Key functionalities tested include platform checks (Linux/Mac/Windows), unique ID generation, interruptible threads, cron-style task scheduling, and system resource monitoring. The library's modular design integrates seamlessly into development tasks requiring monitoring, automation, and configuration management.
python utility toolkitstructured loggingnested configurationin-memory cachingthread pool management
Linq’s iMessage Apps Bring Payments, Tickets, Flights, and Games Into the iMessage Bubble Through the imessage_app Part
Linq introduces iMessage Apps, enabling developers to embed interactive mini-apps directly within iMessage threads via the `imessage_app` message part. These apps support full workflows such as payments, ticket booking, and gaming without requiring external browser redirection. The system uses a `team_id` and `bundle_id` to identify the rendering extension, with fallback to static captions if the app is not installed. Cards can be dynamically updated in-place using the `/messages/{id}/update` endpoint, allowing stateful interactions. However, functionality is limited to iMessage and requires recipient app installation, restricting global reach.
imessage_appteam_idbundle_idin-place updatesmessage part
Anthropic Claude Sonnet 5 vs Sonnet 4.6 vs Opus 4.8: Agentic Coding Benchmarks, API Pricing, and Cost-Performance Tradeoffs Compared
Anthropic released Claude Sonnet 5, a mid-tier model with enhanced agentic capabilities for long-horizon tasks like coding and tool use. Evaluations show Sonnet 5 outperforms Sonnet 4.6 across benchmarks: 63.2% on SWE-bench Pro (vs 58.1%), 81.2% on OSWorld-Verified (vs 78.5%), and 57.4% on Humanity’s Last Exam (vs 46.8%). It narrows the gap to Opus 4.8 while offering lower API costs ($2/$10 per MTok during intro pricing). The model introduces effort levels (low-xhigh) for cost-quality tradeoffs and uses an updated tokenizer (1.0-1.35× expansion factor).
agentic codingswe-bench protokenizer expansioneffort levelsgdpval-aa
Meta AI Releases Brain2Qwerty v2: A Non-Invasive MEG Brain-to-Text Pipeline Decoding Typed Sentences at 61% Word Accuracy
Meta AI introduces Brain2Qwerty v2, a non-invasive brain-to-text decoder achieving 61% average word accuracy (39% WER) from magnetoencephalography (MEG) signals during typing. The system combines a convolutional encoder (processing raw MEG data), transformer (modeling temporal structure), and character-level language model (ensuring semantic coherence), trained end-to-end on 22,000 sentences from 9 participants. Performance scales log-linearly with data volume, with top participants reaching 78% accuracy. This represents a 7.6× improvement over prior non-invasive methods (8% accuracy). The architecture and training code (CC BY-NC 4.0) demonstrate transferable techniques for biosignal decoding.
magnetoencephalographybrain-to-textconvolutional encoderword error ratenon-invasive decoding
OpenClaw Releases iOS and Android Companion Node Apps That Connect a Phone to a Self-Hosted AI Agent Gateway
OpenClaw introduces iOS and Android companion node apps that enable smartphones to function as peripherals in a self-hosted AI agent network. The architecture separates the AI assistant, running on a Gateway (macOS, Linux, or Windows via WSL2), from mobile nodes, which provide hardware capabilities like camera, location, and voice. Nodes connect via WebSocket on port 18789, requiring explicit pairing approval, and expose command surfaces through node.invoke. Privacy-sensitive commands are opt-in, and the system supports persistent memory, community plugins, and integration with chat apps like WhatsApp and Telegram. Both apps report no data collection and require a running Gateway for functionality.
websocketgatewaynode.invokepersistent memoryself-hosted
PyGraphistry Implementation Workflow for Interactive Graph Intelligence Pipelines in Security Analytics and Risk Investigation
The tutorial presents a PyGraphistry-based workflow for interactive graph intelligence in security analytics, demonstrating how to transform enterprise access logs into enriched graph structures with risk scoring and anomaly detection. The method involves generating synthetic access data, constructing node/edge tables with NetworkX, computing centrality metrics (PageRank, betweenness) and community detection, and visualizing results via PyGraphistry's interactive interface. Results include identification of suspicious patterns (compromised users, risky devices) through graph embeddings and visual encodings of risk scores, enabling practical investigation of security threats.
graph analyticssecurity analyticsnetworkxpagerankcommunity detection
Deploying retail AI to scale personalisation and customer insight
Retail AI systems are transitioning from static interfaces to dynamic Generative UIs that personalize layouts in real-time using predictive models analyzing clickstreams, purchase history, and inferred intent. Multi-modal social listening platforms process video/audio to identify unbranded mentions, achieving 76% ROI versus 60% for text-based systems. Synthetic user simulations leverage LLMs to test campaigns, while edge-based computer vision optimizes physical retail spaces. The Model Context Protocol standardizes integration with legacy systems, reducing latency and token costs in multi-step interactions.
generative uismulti-modal social listeningsynthetic user simulationsmodel context protocoledge computing
Japan’s answer to its worker shortage: An AI model for 10 million robots
Japan has formalized a national strategy to deploy 10 million AI-powered robots across 18 industries by 2040, backed by a $6.1 billion public funding initiative. The project, commissioned by METI and NEDO, involves developing a multimodal foundation model by Noetra and AIST, capable of integrating language, images, video, and sensor data for physical AI applications. Initial versions are slated for release this fiscal year, with annual upgrades thereafter. Funding is contingent on milestone achievements, with a stage-gate review process ensuring accountability. The consortium includes SoftBank, NEC, Sony, and Honda, leveraging Japan's existing robotics expertise to address labor shortages and enhance export potential.
multimodal foundation modelphysical aistage-gate processsensor datarobotics expertise
Bank of England reviews AI rules for agentic AI in finance
The Bank of England is evaluating regulatory frameworks for agentic AI in finance, highlighting gaps in current rules designed for human-supervised systems. Deputy Governor Sarah Breeden emphasized that agentic AI systems, which autonomously execute tasks in payments, trading, and cybersecurity, require new safeguards due to their potential to amplify risks like cyberattacks and market volatility. A 2026 Cambridge Centre for Alternative Finance report found 52% of financial firms actively adopting agentic AI, primarily for internal functions. Proposed measures include circuit breakers, kill switches, and enhanced recovery requirements for core systems to mitigate systemic risks. The Financial Stability Board also outlined 12 sound practices for responsible AI adoption in finance.
agentic aicyber resiliencecircuit breakerskill switchesfinancial stability
Anthropic deploys Claude Sonnet 5, Fable and Mythos restored
Anthropic deployed Claude Sonnet 5 and reinstated access to Fable and Mythos models after addressing a safety vulnerability identified by Amazon researchers. The fix involved an automated classifier blocking malicious prompts with 99% efficacy, though it increased false positives. Performance metrics show Sonnet 5 achieves 63.2% on SWE-bench and 80.4% on Terminal-Bench 2.1, with cost parity to Sonnet 4.6. Safety audits confirmed reduced non-compliant behavior versus predecessors, and the model demonstrated autonomous task execution in deployments by Rakuten, Zapier, and Zed. A new industry framework classifies exploit severity across capability gain, breadth, weaponization ease, and discoverability.
claude sonnet 5automated classifierswe-benchterminal-bench 2.1exploit severity
Generated automatically at 2026-07-01 21:17 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.