Daily Digest — 2026-07-11
257 items · 2 research labs, 248 arxiv papers, 7 industry media
🏛️ Research Labs (2)
How Deutsche Telekom is rewiring telecommunications with AI
Deutsche Telekom has implemented a large-scale AI-native transformation across its telecommunications operations, integrating generative AI into customer service, network optimization, and employee workflows. The company employs ChatGPT Enterprise and custom API tooling, achieving 50,000+ monthly active users and a 546% increase in AI adoption since 2026. Key innovations include real-time translation, in-call assistants, and dynamic network resource allocation, demonstrating measurable improvements in efficiency and customer experience.
generative aichatgpt enterprisereal-time translationnetwork optimizationai-native
Profiling in PyTorch (Part 3): Attention is all you profile
The article profiles attention mechanisms in PyTorch, comparing naive implementations with optimized backends. It analyzes GPU kernel traces for naive attention (matmul, scaling, masking, softmax), demonstrates a 1-kernel reduction via in-place operations, and evaluates PyTorch's Scaled Dot Product Attention (SDPA) backends. Results show the math backend introduces 20 kernels due to FP32 upcasting and safe softmax, while the efficient backend fuses operations into a single bf16-aligned CUTLASS kernel. Profiling reveals Tensor Core utilization disparities between implementations.
attention mechanismsgpu profilingscaled dot product attentiontensor coresin-place operations
📜 arXiv Papers (248)
OpenCoF: Learning to Reason Through Video Generation
OpenCoF introduces a framework for Chain-of-Frame (CoF) reasoning through video generation, comprising the OpenCoF-17K dataset (11 task families) and Wan-CoF, a fine-tuned video model. The method combines diverse temporal supervision with visual/textual reasoning tokens to capture spatial-temporal cues. Wan-CoF outperforms Wan2.2-I2V-A14B on four benchmarks, demonstrating that explicit intermediate state organization enhances video reasoning. Analysis reveals token contributions vary across model depth, denoising steps, and spatiotemporal dimensions.
chain-of-frame reasoningvideo generationtemporal supervisionreasoning tokensspatial-temporal cues
Ideas Have Genomes: Benchmarking Scientific Lineage Reasoning and Lineage-Grounded Idea Generation
The paper introduces IdeaGene-Bench (IG-Bench), a benchmark for evaluating scientific lineage reasoning and lineage-grounded idea generation in AI systems. IG-Bench employs the IdeaGene framework, representing papers as Idea Genome objects and tracking evolutionary dynamics via GenomeDiff records. It includes 1,961 lineage traces, 1,085 Idea Genome objects, and 920 GenomeDiff records across 10 domains, supporting evaluations IG-Exam (42 task types) and IG-Arena (lineage-conditioned generation). Experiments on 14 LLMs reveal a compositional bottleneck, with top systems achieving only 27.3% exact accuracy on lineage reasoning and lineage context reshuffling system rankings.
scientific lineage reasoningidea genomegenomedifflineage-grounded generationbenchmark evaluation
SLORR: Simple and Efficient In-Training Low-Rank Regularization
SLORR introduces a simple, stateless framework for in-training low-rank regularization that preserves model architecture and avoids costly SVD computations. The method employs GPU-friendly approximations of Hoyer sparsity and nuclear norm regularizers, with proven approximation guarantees. Evaluations on ImageNet-1K (ResNet-50, ViT-B/16, ViT-L/16) and LLM pretraining (135M/560M scales) show compressibility with <8% training overhead and better performance retention than unregularized models.
low-rank regularizationhoyer sparsitynuclear normgpu-friendly approximationsmodel compressibility
Using AI-based Learning Assistants in Higher Education: A Large-Scale Descriptive Analysis
This study provides a large-scale empirical analysis of AI-based learning assistant usage in higher education, addressing a research gap in educational chatbot studies that typically rely on small samples and self-reported data. The authors analyze objective log data from 77,543 distance education students using the Syntea learning assistant, examining usage patterns across demographic and structural variables. Results demonstrate widespread integration of Syntea into study routines, with significant variations by gender, age group, study cluster, degree program, and study mode. The findings offer quantitative evidence for optimizing AI learning support systems.
ai-based learning assistanteducational chatbotlarge-scale analysisusage patternsdistance education
Dimensionality Reduction Meets Network Science: Sensemaking on UMAP's kNN Graph
The paper demonstrates that UMAP's internal k-nearest-neighbor (kNN) graph, typically discarded in favor of its 2D embedding, provides valuable high-dimensional structural insights when analyzed with graph algorithms. Applying PageRank, k-core decomposition, and clustering coefficient to this graph enables representative point selection, density-based region identification, and tight neighborhood detection respectively. Evaluations on MNIST and Fashion MNIST show these methods match or complement specialized techniques like k-medoids and HDBSCAN while preserving original manifold structure lost in dimensionality reduction.
umapk-nearest-neighborpagerankk-core decompositionclustering coefficient
AUTOPILOT VQA: Benchmarking Vision-Language Models for Incident-Centric Dashcam Understanding
The paper introduces AUTOPILOT-VQA, a benchmark for evaluating vision-language models on incident-centric dashcam video understanding. The dataset features structured questions addressing safety-critical categories (weather, traffic, signage, etc.) to assess temporally grounded reasoning beyond object recognition. Designed for the AUTOPILOT CVPR 2026 competition, it provides standardized evaluation of autonomous driving systems' reliability in safety-relevant scenarios.
vision-language modelsautonomous drivingvisual question answeringsafety-critical reasoningdashcam video understanding
Workflow as Knowledge: Semantic Persistence for LLM-Mediated Workflows
The paper proposes a semantic persistence model for LLM-mediated workflows, treating workflow definitions, instances, and execution traces as persistent knowledge objects. Drawing from Lisp-inspired concepts (symbolic forms, object identity) without implementation constraints, it introduces a derive/infer distinction: deterministic computation (derive) versus LLM-mediated judgment under controlled policies (infer). This enables workflows to be inspectable, resumable knowledge artifacts within a shared substrate, though formal transition semantics remain undeveloped.
semantic persistencellm-mediated workflowsknowledge objectsderive/infer distinctionsymbolic forms
The Illusion of Equivalency: Statistical Characterization of Quantization Effects in LLMs
The study exposes a discrepancy between conventional evaluation metrics and actual behavioral changes in quantized large language models, introducing correctness agreement as a decision-level metric to assess prediction overlap between base and quantized models. Through statistical analysis of quantization effects on attention weights across multiple models (8-bit to 2-bit) and schemes, the authors identify non-linear breakpoints at low bit-widths and demonstrate differential sensitivity among attention projections. Results reveal that query and key projections are more vulnerable than value/output projections, challenging the assumed equivalency of quantized models despite preserved task performance.
post-training quantizationcorrectness agreementattention weightsbehavioral divergencenon-linear breakpoints
Validity of LLMs as data annotators: AMALIA on authority
The study evaluates the validity of Portugal's AMALIA-9B, a 9B-parameter national language model, as a data annotator for moral foundation coding, specifically authority. Using the recovery gap method, it decomposes holistic prompts into atomic clauses to test whether the model follows theoretical constructs or relies on surface features. Results show AMALIA-9B recovers only half its holistic performance, indicating reliance on surface correlates like moral outrage, unlike a multilingual LLM that closes the gap. The findings suggest national LLM benchmarks should test evidential routes, not just human coder agreement.
recovery gapmoral foundationholistic promptsurface correlatesnational language model
Pose-to-Biomechanics: Bridging 3D Human Pose Estimation and Biomechanical Attribute Prediction
The paper introduces BioModule, a lightweight temporal transformer that extends 3D human pose estimators to predict biomechanical attributes from 17-joint skeletons, enabling physically interpretable motion analysis. The method constructs an aligned dataset (Human3.6Mplus) pairing Human3.6M video and 3D keypoints with biomechanical labels, ensuring frame-accurate cross-modal supervision. Evaluated across seven state-of-the-art pose estimators, BioModule demonstrates estimator-agnostic performance, bridging vision-based pose estimation and biomechanical analysis without modifying upstream models.
biomechanical attributestemporal transformer3d pose estimationcross-modal supervisionmarkerless motion capture
Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents
The paper introduces a proactive memory agent to address behavioral state decay in long-horizon tasks, where decision-relevant information becomes inaccessible. The method employs a separate memory agent that updates a structured memory bank and selectively injects reminders into an unmodified action agent's context. Evaluated on Terminal-Bench 2.0 and $τ^2$-Bench, the approach improves pass@1 by +8.3 pp and +6.8 pp, respectively, outperforming passive retrieval and always-on injection. The authors also train Qwen3.5-27B on SETA using SFT and GRPO, demonstrating improved validation reward and partial transfer.
behavioral state decayproactive memory agentlong-horizon tasksstructured memory bankselective intervention
ProjAgent: Procedural Similarity Retrieval for Repository-Level Code Generation
ProjAgent introduces procedural similarity as a novel retrieval signal for repository-level code generation, addressing limitations of existing lexical and semantic methods. The system decomposes target functions into reasoning steps, retrieves repository functions with similar procedural behavior via an agentic workflow, and combines this with semantic retrieval. It employs a static-analysis feedback loop for iterative code repair. Evaluated on REPOCOD, ProjAgent achieves 41.14% Pass@1, outperforming baselines and demonstrating procedural similarity's effectiveness.
repository-level code generationprocedural similarityagentic workflowstatic-analysis feedbackrepocod
A Practical Investigation of Training-free Relaxed Speculative Decoding
This work investigates training-free relaxed speculative decoding techniques for accelerating autoregressive LLM sampling, unifying existing approaches and benchmarking them in contemporary settings. The study employs a shared framework to analyze methods that relax the lossless guarantee of standard speculative decoding, enabling potential speed-ups or capability gains. Key findings indicate that relaxation necessitates extensive capability evaluation and often requires a high-quality drafter language model, limiting applicability with lightweight multi-token-prediction drafters.
speculative decodingautoregressive llmtraining-freemulti-token-predictioncapability evaluation
SolarChain-Eval: A Physics-Constrained Benchmark for Trustworthy Economic Agents in Decentralized Energy Markets
SolarChain-Eval introduces a physics-constrained benchmark for evaluating trustworthy economic agents in decentralized energy markets, addressing both performance and trustworthiness. The benchmark frames market governance as a Gymnasium-compatible Markov Decision Process, incorporating an LLM-based Planner/Auditor layer for action bounds and auditability. Evaluations of static, RL, and RL+LLM policies reveal a utility-safety trade-off, with RL agents improving market utility but risking unsafe behavior, while LLM intervention enhances auditability but cannot fully correct reward misspecification. The work emphasizes the need for physical constraints and transparent intervention traces in agent evaluation.
physics-constrained benchmarkdecentralized energy marketsmarkov decision processllm-based plannertrustworthy ai
WebSwarm: Recursive Multi-Agent Orchestration for Deep-and-Wide Web Search
WebSwarm introduces a recursive multi-agent framework for deep-and-wide web search, addressing limitations of single-agent and parallel multi-agent systems in recursive depth and collaboration adaptability. The method dynamically instantiates agentic search nodes that decompose tasks, recursively delegate subtasks, and collaborate through evidence aggregation, guided by web structure probing and process-level experience reuse. Evaluations on BrowseComp-Plus, WideSearch, DeepWideSearch, and GISA benchmarks demonstrate consistent improvements over baselines in deep, wide, and hybrid search tasks, with ablation studies validating design choices.
multi-agent systemsrecursive delegationweb searchllm-based agentsevidence aggregation
Formal Mechanisms for Market Stability in Self-Interested Agent Societies: A Marketplace Simulation Study
The paper identifies Mediation as a robust formal mechanism for maintaining market stability among self-interested LLM agents, demonstrating resilience against adversarial attacks. Using a multi-agent marketplace simulation with 18 DeepSeek-V3 agents trading within constrained social networks, the study compares eight mechanisms under troll injection and conducts adversarial red-teaming. Results show Mediation sustains positive honest-agent utility despite optimized attacks (best attack reduces utility by 13.3%), proving it can recover under sustained pressure.
multi-agent simulationmarket stabilityadversarial robustnessmediation mechanismllm agents
Multi-Modal, Multi-Environment Machine Teaching for Robust Reward Learning
The authors propose a hierarchical machine teaching algorithm for robust reward learning across multiple Markov Decision Processes (MDPs), addressing the limitations of single-environment inverse reinforcement learning (IRL). The method strategically selects informative environments to expose complementary reward constraints and queries low-cost feedback within these environments. Theoretical analysis shows that comparisons impose stronger global constraints than other feedback modalities in the unlimited-data regime. Empirical results demonstrate that the approach achieves significantly lower regret and better generalization to held-out environments compared to uniform teaching baselines under identical feedback budgets.
inverse reinforcement learningmarkov decision processesreward learningmachine teachingfeedback modalities
UltraX: Refining Pre-Training Data at Scale with Adaptive Programmatic Editing
UltraX introduces a function-calling refinement framework for large-scale pre-training data, addressing limitations of rule-based and LLM-based approaches in quality, efficiency, and reliability. It extends the editing function space with insertion alongside deletion and modification, enabling fine-grained instance-level editing. UltraX employs dataset-adaptive prompt optimization, Line Alignment Mapping, Dynamic Context Replacement, and low-confidence example filtering to generate reliable program supervision. Experiments demonstrate that UltraX achieves superior performance across corpora with fewer training tokens, enhancing data efficiency and refinement reliability.
function-callingpre-training dataline alignment mappingdynamic context replacementlow-confidence filtering
The complexities of patient-centred conversational artificial intelligence
The study addresses limitations in health chatbot evaluation by analyzing 2,053 real patient-chatbot conversations, revealing significant variation in communication patterns and emotional expression. It introduces a patient simulator that separately models clinical content, emotional state, conversational strategy, and communication style, achieving near-indistinguishability from real conversations (human grader accuracy: 55%). Evaluating four LLMs across 1,164 clinician-graded cases with five distinct patient personae demonstrated that communication style significantly impacts triage outcomes, highlighting the need for patient-centered AI to accommodate real-world interaction diversity.
health chatbotslarge language modelspatient simulatortriage outcomescommunication diversity
When Structured Sparse Autoencoders Learn Consistent Concepts Across Modalities
The paper introduces Structured Sparse AutoEncoder ($S^2AE$), a method to improve modality-consistent concept learning in vision-language models (VLMs) by addressing fragmented visual concept coverage in vanilla sparse autoencoders (SAEs). $S^2AE$ enforces semantic and spatial consistency through structured sparsity regularization, combining exclusive sparsity for inter-group disentanglement and group sparsity for intra-group consistency. Evaluated on Qwen2.5-VL-7B-Instruct, $S^2AE$ achieves a 6.06% mIoU improvement in semantic alignment, 60.81 lower l0 norm for efficiency, and maintains 99% explained variance. Cross-modal analysis shows 3.08% better semantic consistency and 2.37% higher monosemanticity scores.
structured sparse autoencodervision-language modelsmechanistic interpretabilitysemantic alignmentmonosemanticity
Towards Precision Therapy in Hepatocellular Carcinoma: A Clinical-Reasoning LLM for Risk Stratification and Treatment Guidance
HCC-STAR, a clinically aligned large language model, introduces precision therapy for hepatocellular carcinoma by jointly predicting risk stratification, treatment recommendations, and survival estimates from electronic medical records. The model was trained on 30,000 HCC cases expanded into EMR-style narratives via clinician-validated prompt augmentation, optimized with a step-verifiable composite reward for knowledge-aligned reasoning. Evaluated on 6,668 patients across 12 Chinese hospitals, HCC-STAR outperformed clinical guidelines, GPT-5, and Gemini-2.5 Pro, achieving a median survival of 51 months versus 29-32 months for BCLC/CNLC. Clinician evaluations confirmed its trustworthiness and superiority over physicians in treatment accuracy and decision speed.
hcc-starrisk stratificationprecision therapystep-verifiable rewardemr narratives
SHAP-Weighted Cross-Modal Expert Fusion for Emotion and Sentiment Recognition: Evidence and Limits
The paper introduces SHAP-weighted adaptive fusion (XGAF), a tree-based mixture of unimodal and cross-modal experts for multimodal emotion and sentiment recognition, addressing limitations of early and late fusion. XGAF employs TreeSHAP attribution magnitudes to derive sample-level weights, focusing on the impact of SHAP attribution reduction methods (mean-abs, median-abs, sum-abs) when experts have unequal feature dimensionalities. On MELD 7-class emotion recognition, sum-abs XGAF achieves 0.5983 weighted F1-score, nearly matching early fusion (0.6018) and significantly outperforming late fusion (0.4598). On CMU-MOSEI 3-class sentiment recognition, sum-abs XGAF reaches 0.6519, slightly exceeding early fusion (0.6485) and late fusion (0.5696). Ablation studies highlight the importance of cross-modal experts, particularly the trimodal expert, over complex per-sample routing.
shap-weighted fusiontree-based mixturecross-modal expertsshap attributionmultimodal recognition
SMetric: Rethink LLM Scheduling for Serving Agents with Balanced Session-centric Scheduling
The paper introduces SMETRIC, a novel LLM scheduling system optimized for agentic workloads, where requests originate from agents rather than humans. It addresses two workload shifts: prioritizing tokens per second (TPS) over per-token latency and high KV-cache reuse (80% in agentic traces vs. 54-62% in chat). SMETRIC employs balanced session-centric scheduling, routing the first request of each session for load balance and subsequent requests cache-aware, preserving KV-cache reuse. Evaluated on real-world traces, SMETRIC improves cluster TPS by 10-16% and prefill TPS by 2-34% over state-of-the-art schedulers, while maintaining low per-token latency.
llm schedulingkv-cache reuseagentic servingload balancesession-centric scheduling
CommuniWave:A Machine Learning Model for Quantifying the Degree of Temporary Informal Behavior in Urban Communities
CommuniWave introduces a machine learning model to quantify the Degree of Informal Behavior (DIB) in urban communities, addressing the lack of metrics for resident behaviors in top-down planning. The model combines a Behavior Capture Net (BCN) based on mmaction2, a custom YOLOv10 model (YLX), and a Behavior Eval Model (BEM) using random forest. It generates DIB fluctuation charts from street videos, enabling dynamic monitoring to support urban resilience decisions.
degree of informal behaviorbehavior capture netyolov10random foresturban resilience
VocaDet: Sample-Driven Open-Vocabulary Object Detection and Segmentation via Visual Tokenization and Vector Database Retrieval
VocaDet introduces a sample-driven open-vocabulary object detection and segmentation framework that learns object concepts directly from user-provided samples without model retraining. The method employs DINOv3 for visual feature extraction, applies agglomerative clustering to generate multi-granularity visual tokens, and stores them with position-debiased representations in a scalable vector database for efficient retrieval-based recognition. Experiments on UA-DETRAC demonstrate effective open-vocabulary detection performance without conventional detector training, while supporting continuously expandable recognition capability.
open-vocabulary detectionvisual tokenizationvector databaseagglomerative clusteringdinov3
DocMaster: A Hierarchical Structure-Aware System for Document Analysis
DocMaster introduces a hierarchical structure-aware system for document analysis that preserves layout information typically lost in plain-text conversion. The system parses documents into tree structures maintaining sections, tables, and other elements, then constructs a semantic index enabling both document filtering and in-depth QA. An interactive web interface demonstrates capabilities including multi-view indexing, natural-language filtering, and structured analysis over document collections.
hierarchical document treesstructure-aware semantic indexmulti-view indexingdocument filteringlayout preservation
When the Judge Changes, So Does the Measurement: Auditing LLM-as-Judge Reliability
The study audits reliability issues in LLM-as-judge systems by examining how evaluator changes affect measurement validity. Using four judgment datasets, it compares two upgrade paths: parameter scaling in Qwen3 (1.7B to 32B) and API transitions in MiniMax (M2-M2.7). Results show non-interchangeable upgrades, with only Qwen3 1.7B to 4B yielding robust gains. Stronger judges reduce but persist position/verbosity biases, while jury sampling and structured debate show limited or unverifiable improvements. The authors recommend including dataset slices, bias probes, error-dependence estimates, and protocol audits in reports.
llm-as-judgemeasurement validityparameter scalingposition biasverbosity bias
AI-guided stimuli discovery and generation to optimize facial emotion perception studies in autism
The study introduces an AI-guided framework for optimizing facial emotion perception assays in autism research by identifying stimulus-level sparsity in group differences. Using population-specific artificial neural networks, the authors predicted autistic-neurotypical judgment disparities, selected diagnostic facial expressions maximizing separation, and generated transformed images via GANs to reduce behavioral differences. Validation showed model-selected images increased group separation by 1.5× versus random stimuli, while synthesized images reduced divergence by 30% versus originals, demonstrating stimulus optimization for neurodivergent phenotyping.
facial emotion perceptionstimulus optimizationpopulation-specific anngenerative adversarial networkbehavioral phenotyping
Cognitive-structured Multimodal Agent for Multimodal Understanding, Generation, and Editing
The Cognitive-structured Multimodal Agent introduces a novel architecture for long-horizon multimodal dialogue by externalizing visual information into an Episodic Visual Memory, enabling selective reactivation of relevant episodes during reasoning. It comprises a Perceptual Abstraction Engine for structured visual abstraction, a Cognitive Retrieval Engine for cross-turn memory retrieval, and a Multimodal Executive Controller for task inference and action planning. The agent leverages a Unified Scenario Engine for generating structured multi-turn conversations with retrieval annotations, optimizing policies via reinforcement learning. Evaluated on a long-horizon visual-dialogue benchmark, the 8B parameter agent achieves 91.4% retrieval accuracy over 20-turn sessions, surpassing 32B baselines by +8.2% while reducing inference time by nearly half (23.1s → 12.7s).
episodic visual memoryperceptual abstraction enginecognitive retrieval enginemultimodal executive controllerunified scenario engine
The Context Access Divide: Interaction-Level Architecture as a Complementary Dimension of Agentic Inequality
The paper introduces the Context Access Divide (CAD) as a novel dimension of agentic inequality, complementing Sharp et al.'s framework by focusing on interaction-level disparities in AI utility. It formalizes CAD through a probabilistic model based on cognitive psychology's fan effect, showing manual context attachment leads to combinatorial task-success collapse with growing corpus size, while dynamic retrieval architectures avoid this. The analysis examines technical foundations in Model Context Protocol (MCP) and retrieval-augmented generation (RAG), highlighting implications for knowledge-work stratification.
agentic inequalitycontext access dividedynamic context retrievalretrieval-augmented generationmodel context protocol
Drift-Aware Temporal Graph Rewiring (DATGR) for Adaptive Semantic Modeling in Biomedical Text
The paper proposes Drift-Aware Temporal Graph Rewiring (DATGR), a framework for adaptive semantic modeling in biomedical text that addresses concept evolution through dynamic co-occurrence graph updates. Instead of full retraining, DATGR performs lightweight edge-weight adjustments via a logistic update rule based on estimated semantic drift. Evaluated on the Biomedical Multi-Relation Corpus (BIOMRC), DATGR improved mean AUROC by 0.066 (0.699 vs. 0.633 static baseline) while maintaining comparable AUPRC (0.738 vs. 0.744), demonstrating effective temporal adaptation without precision loss.
temporal graphsemantic driftco-occurrence graphlink-predictionbiomedical text
VEGAS: Human-Aligned Video Caption Evaluation via Gaze
The paper introduces VEGAS (Video caption Evaluation via GAze Score), a training-free metric for evaluating video captions by aligning them with viewer attention through gaze data. VEGAS employs a cross-modal, information-theoretic approach to quantify caption-gaze alignment, enabling personalized caption selection via rejection sampling without model retraining. Evaluated on a curated dataset of egocentric activities and instructional slides with synchronized gaze annotations, VEGAS-selected captions significantly improve alignment with human focus and enhance caption-to-video retrieval performance.
video captioninggaze alignmentcross-modal metricinformation-theoreticrejection sampling
Applying JEPA-Style Predictive Learning to JA4-Derived Network Fingerprints
JA4-JEPA demonstrates that Joint-Embedding Predictive Architecture (JEPA) objectives, previously successful for images and video, generalize to network fingerprinting. The Transformer-based model processes JA4-derived subfields (JA4, JA4H, JA4S, JA4X) from JA4DB and CIC-IDS-2017 datasets, comprising 397K samples without complete view overlap. Evaluated on protocol-family classification across TLS, DNS, and SSH using frozen kNN probes, the model achieves 0.9899 cosine similarity and 0.9220 kNN accuracy on 39,416 heldout samples. These results validate JEPA-style predictive learning for generating effective embeddings from incomplete network fingerprint views.
ja4jepanetwork fingerprintingpredictive representation learningself-supervised learning
Two Axes of LLM Abstention: Answer Correctness and Question Answerability
The study identifies two distinct axes for LLM abstention: answer correctness and question answerability, which conventional single-threshold confidence scoring fails to separate. Analyzing five instruction-tuned models (2B to 14B parameters), the authors show that standard answer-confidence predicts correctness but not answerability, while a hidden-state linear probe detects answerability (0.69-0.77 AUROC on CREPE false-premise questions) but not correctness. A novel two-threshold policy combining these axes achieves 0.75 coverage of correct answers with controlled unanswerable rates, outperforming single-threshold approaches (0.31 coverage). Scale does not mitigate the answerability blind spot.
llm abstentionanswer correctnessquestion answerabilityhidden-state probethreshold policy
Spatio-Temporal Scheduling Prediction Under Backhaul Delay for Resilient Coordinated Beamforming
The paper proposes a two-stage predictive framework using a Spectral Temporal Graph Neural Network (StemGNN) to mitigate backhaul delay in coordinated beamforming for 5G networks. StemGNN predicts future UE scheduling states from delayed observations, replacing stale inputs to the CBF-SLNR precoder. Evaluated on a three-cell massive MIMO downlink with 60 UEs, StemGNN achieves 87.57% mean scheduling accuracy, outperforming baselines by up to 7.71%, and recovers 57-73% of sum rate loss caused by backhaul delay, improving fairness for cell-edge users by up to 83%.
coordinated beamformingspectral temporal graph neural networkbackhaul latencymassive mimoscheduling prediction
ADORN: Adaptive Drift handling for Open RAN using Reinforcement Learning
The paper proposes ADORN, a Q-learning-based adaptive retraining approach for Open RAN that optimizes the trade-off between forecasting accuracy and computational cost. The method formulates retraining decisions as an MDP, employing a multi-expert LSTM ensemble to prevent catastrophic forgetting under dynamic traffic conditions. Experiments demonstrate reduced retraining overhead by 37% compared to baselines while maintaining SLA compliance, with the RL agent achieving 92% policy optimality in simulated O-RAN environments.
open ranq-learningcatastrophic forgettinglstm ensembleservice level agreements
Predicting Male Fertility Using Machine Learning: A Semen Parameters Based Analysis with the VISEM Dataset
This study demonstrates that machine learning can effectively classify male fertility status using semen parameters, achieving 94.2% accuracy with the Nearest Centroid classifier. The VISEM dataset, comprising 85 semen samples categorized as Fertile, Sub-Fertile, and Infertile, was pre-processed and evaluated using the LazyPredict framework. Over 40 algorithms were tested, with Nearest Centroid outperforming alternatives like Support Vector Machines and Quadratic Discriminant Analysis. Model robustness was validated through 5-fold cross-validation and multiclass ROC-AUC analysis. The findings highlight the potential of machine learning to enhance fertility diagnostics and inform personalized treatment strategies in andrology and assisted reproductive technologies.
nearest centroid classifiersemen parameterslazypredict frameworkmulticlass roc-auc5-fold cross-validation
OmniFood-Bench: Evaluating VLMs for Nutrient Reasoning and Personalized Health Advice
OmniFood-Bench introduces a novel benchmark for evaluating Vision-Language Models (VLMs) in food-related reasoning tasks, addressing the 'Systemic Information Asymmetry' between visual appearance and nutritional composition. The benchmark, derived from MM-Food-100K, assesses three capabilities: Basic Perception, Quantitative Reasoning, and Safety-Critical Advisory. Evaluations of six VLMs (including GPT-5.1 and Gemini-3-Flash) reveal a 'Semantic-Physical Gap', with high accuracy in dish identification but poor performance in mass estimation and disease-specific recommendations, highlighting critical limitations in nutritional reasoning.
vision-language modelsnutritional reasoningbenchmark evaluationsystemic information asymmetrysafety-critical advisory
When Synthetic Speech Is All You Have: Better Call GRPO
The paper demonstrates that Group Relative Policy Optimization (GRPO), a critic-free reinforcement learning method, significantly outperforms supervised fine-tuning (SFT) for adapting LLM-based ASR systems using synthetic speech in regulated domains. GRPO rewards low-WER hypotheses, reducing WER by 40% relative to SFT (36.71%→22.09%) and by 45% when combined with SFT. Analysis reveals GRPO improves stopping calibration and speech-to-text alignment by refining attention mechanisms, without altering early-layer representations. The findings advocate for RL over SFT when synthetic speech is the primary resource.
group relative policy optimizationsynthetic speechllm-based asrsupervised fine-tuningword error rate
Track2Map: Online Deformable SLAM with Motion-Aware Pose Optimization in Robotic Surgery
Track2Map introduces an online 3D Gaussian Splatting pipeline for Simultaneous Localisation and Mapping (SLAM) in robot-assisted minimally invasive surgery, eliminating dependency on accurate camera trajectory priors. The method jointly optimizes camera trajectory and 3D deformable scene representation directly from surgical video, employing track-anchored deformation initialization and leveraging track statistics to disentangle camera motion from tissue deformation. Evaluated on the StereoMIS dataset, Track2Map outperforms competing SLAM methods and non-SLAM approaches relying on trajectory priors, demonstrating enhanced reconstruction quality and camera trajectory accuracy.
gaussian splattingsimultaneous localisation and mappingdeformable scene representationtrack-anchored deformationrobot-assisted surgery
DrugGen 2: A disease-aware language model for enhancing drug discovery
DrugGen-2 introduces a disease-aware language model for drug discovery, generating small molecules conditioned on disease ontology and target protein sequences. The model fine-tunes GPT-2 via supervised learning and reinforcement learning (group relative policy optimization), optimizing chemical validity, novelty, diversity, and binding affinity. Evaluated on five diabetic nephropathy targets, DrugGen-2 outperformed DrugGPT and DrugGen, generating molecules with higher structural similarity to approved drugs and superior predicted binding affinities (e.g., -9.917 vs. -8.283 for enalapril on ACE), validated by molecular docking.
drug discoverylanguage modelreinforcement learningmolecular dockingbinding affinity
Game Theory Driven Multi-Agent Framework Mitigates Language Model Hallucination
G-Frame, a game theory-driven multi-agent framework, mitigates LLM hallucinations in scientific domains by enforcing domain constraints through structured reasoning. The method synthesizes 363,045 chains-of-thought and 199,589 QA pairs, training a 7B parameter model (OmniChem) that matches GPT 4o mini's performance on ChemBench while reducing hallucinations by 79.46%. Results demonstrate improved molecular design and synthesis planning, offering a scalable solution for specialized scientific applications.
large language modelshallucination mitigationmulti-agent frameworkbayesian reasoningmolecular design
Swapping Faces, Saving Features: A Dual-Purpose Pipeline for Pedestrian Privacy in ITS
The paper proposes a five-stage pipeline for pedestrian privacy preservation in intelligent transportation systems (ITS), specifically tailored for the Egy-DRiVeS dataset. The method employs face-swapping techniques to conceal identities while retaining essential facial attributes required for pedestrian intention and trajectory prediction models. Two face-swapping models, Roop and Ghost-v2, are evaluated, with Roop demonstrating superior performance in balancing privacy and data usability. The pipeline addresses the dual challenge of protecting pedestrian privacy and maintaining image quality for effective AI model training in autonomous vehicles.
face-swappingprivacy preservationpedestrian trajectoryintelligent transportation systemsautonomous vehicles
TRACE: A Two-Channel Robust Attribution Watermark via Complementary Embeddings for LLM-Agent Trajectories
TRACE introduces the first distortion-free, self-synchronizing watermark for LLM-agent trajectories that remains invariant under rewriting. The method employs two complementary channels: a selection channel (content-keyed, distortion-free) and a tally channel (position-keyed, rewrite-proof), leveraging decision entropy for signal strength. Evaluations on ToolBench and ALFWorld show TRACE maintains original success rates, achieves high detection scores (z ≈ 100), resists 70% step deletion, and preserves exact tally invariance under LLM rewriting.
llm agentsbehavioral watermarktrajectory logdecision entropyself-synchronizing
Towards Mechanistically Understanding Why Memorized Knowledge Fails to Generalize in Large Language Model Finetuning
The paper identifies a Knowing--Using Gap in LLM fine-tuning, where models memorize facts but fail to generalize them for reasoning. Using self-patching, an intervention technique that relocates internal representations, the authors demonstrate that memorized knowledge exists but is misaligned with computational pathways. Their heuristic strategy recovers 58--75% of the generalization gap, supporting a knowledge-circuit misalignment hypothesis across domains.
knowing-using gapself-patchingknowledge-circuit misalignmentllm fine-tuninggeneralization failure
WCog-VLA: A Dual-Level World-Cognitive Vision-Language-Action Model for End-to-End Autonomous Driving
WCog-VLA introduces a dual-level World-Cognitive Vision-Language-Action framework for proactive autonomous driving, combining semantic world forecasting with generative world evolution. At the semantic level, it integrates 3D spatial perception, agent tokens, and Game-theoretic Chain-of-Thought reasoning. At the generative level, it employs an Aligned Decoupled Diffusion Transformer (ADDT) to synthesize physically-plausible multi-agent trajectories, reducing denoising steps for faster inference. The model is trained on a dataset with 85k Game-CoT annotations. Evaluated on the NAVSIM benchmark, WCog-VLA achieves a SOTA PDMS score of 92.9.
world-cognitive vision-language-actiongame-theoretic chain-of-thoughtaligned decoupled diffusion transformermulti-agent trajectoriesnavsim benchmark
Large-Language-Models-as-a-Judge in Theory-Agnostic Adaptive Metric-Alignment for Prototypical Networks in Personality Recognition
JAM (Judge for Adaptive Metric-Alignment) introduces a theory-agnostic framework for personality recognition, shifting from predefined psychological taxonomies to discovering latent pseudo-facets capturing shared behavioral structure. The framework employs an Attention-Pooled Graph Prototypical Network for structured representation learning via clustering and Cross-Theory Harmonization (CTH) to unify heterogeneous datasets without theory-specific labels. An LLM-as-a-Judge mechanism operates in two configurations (LLM-before-the-loop, LLM-in-the-loop) to identify ambiguous samples and guide adaptive metric learning. Experiments demonstrate improved cross-framework generalization and performance, advancing theory-agnostic personality inference and supporting low-resource theories. Code and artifacts are publicly available.
theory-agnosticprototypical networkcross-theory harmonizationllm-as-a-judgeadaptive metric learning
Self-Adaptive Anomaly Detection with Reinforcement Learning and Human Feedback in Connected Vehicles
The paper presents an online anomaly detection framework for connected vehicles that integrates adaptive model selection, statistical drift detection, and human-in-the-loop retraining. The method employs a factorized deep Q-network with self-attention for detector selection, a consensus-based ensemble for drift detection, and a prioritized replay buffer for operator-guided retraining. Evaluated on a valet parking application with seven microservices, the system achieves 0.69 F1, maintains 0.65 F1 post-update after retraining, and avoids catastrophic forgetting on prior distributions.
anomaly detectionreinforcement learningconcept drifthuman-in-the-loopconnected vehicles
On the Role of Conversational Timing in Synthetic Training Data for ASR
The study investigates how conversational timing in synthetic multi-speaker training data affects automatic speech recognition (ASR) performance. Using an exponential-tilting family to parameterize pause and overlap distributions, the authors explore the timing parameter space via Latin hypercube sampling and multi-objective Bayesian optimization. Results on a Hungarian dialogue corpus show that higher overlap exposure reduces concatenated-permutation word error rate (cpWER), while longer and more variable gaps increase cpWER, with similar but weaker trends for character error rate (cpCER). Bayesian optimization reveals an overlap-gap trade-off, suggesting that task-relevant timing diagnostics should complement realistic simulation.
conversational asrsynthetic training databayesian optimizationexponential-tiltingcpwer
FedOPAL: One-Shot Federated Learning via Analytic Visual Prompt Tuning
FedOPAL introduces one-shot federated learning via analytic visual prompt tuning to address communication bottlenecks in edge intelligence. The framework employs visual prompts as feature rectifiers, applying local proximal constraints to align heterogeneous data distributions into linearly separable spaces, satisfying analytical federated learning assumptions. Experiments demonstrate FedOPAL surpasses original analytical methods on multiple benchmarks and matches state-of-the-art iterative methods' accuracy while eliminating server-side training costs, offering an efficient paradigm for edge-based large model collaboration.
federated learningvisual prompt tuningfeature rectifiersedge intelligenceanalytical aggregation
FSD-VLN: Fast-Slow Dual-System Modeling for Aerial Long-Horizon Vision-Language Navigation
FSD-VLN introduces a fast-slow dual-system architecture for aerial Vision-Language Navigation (VLN), addressing structural misalignment between global multimodal understanding and sequential action generation. The framework employs asynchronous branches: a slow stream for semantic reasoning using pre-trained vision-language models and a Diffusion Transformer (DiT) fast stream for cross-temporal action distribution modeling. A time-aware adaptive optimizer stabilizes long-sequence training. Experiments demonstrate up to 2X higher navigation success rates on unseen scenes compared to state-of-the-art methods, with over 50% reduction in single-action inference delay and total task runtime.
vision-language navigationdiffusion transformersemantic reasoningcross-temporal modelingtime-aware optimizer
MobiDiff: Semantic-Aware Multi-Channel Discrete Diffusion for Human Mobility Data Generation
MobiDiff introduces a discrete diffusion framework for generating human mobility data by directly denoising multi-channel semantic skeletons, addressing limitations of existing diffusion-based methods that rely on continuous or latent spatio-temporal traces. The method decomposes check-in events into spatial, activity, and temporal channels, using structured masking at event-, group-, and channel-levels to capture trajectory patterns and dependencies. Evaluations on datasets from Atlanta, Boston, and Seattle show MobiDiff preserves trajectory length and temporal interval distributions, achieves competitive mobility statistics, and is 5.3× faster than GeoGen during inference.
discrete diffusionhuman mobility datasemantic skeletonsmulti-channel maskingtrajectory generation
Spectral Analysis of Dueling Q-Learning
The paper strengthens theoretical understanding of dueling Q-learning by analyzing its unregularized, unprojected constant step-size recursion. It introduces a switching linear system representation for deterministic dueling Q-learning and derives a finite-time error bound for the stochastic version. The analysis reveals how value and advantage updates differentially affect the Q-function's action-common and action-differential components. Results include exact convergence guarantees for the deterministic case and probabilistic bounds for the sampled stochastic formulation.
dueling q-learningmarkov decision processesq-function approximationadvantage functionswitching linear system
TypeProbe: Recovering Type Representations from Hidden States of Pre-trained Code Models
TypeProbe investigates type representation emergence in pretrained code models by probing residual streams across Java and Python code. Using parallel datasets, the study demonstrates that cross-lingual type representations arise even from untyped code and evaluates linear encoding of function application types via cross-language probing. Results indicate partial robustness to lexical perturbations and syntactic variations, addressing a gap in interpretability research for formal type semantics. Code and datasets are publicly released.
type representationsresidual streamscross-lingual probingpretrained code modelsinterpretability
ArtMine: Discovering and Formalizing Artistic Processes
ArtMine introduces a framework for computationally formalizing artistic processes from heterogeneous historical evidence, addressing the gap in modeling creative workflows rather than just final artifacts. The method synthesizes fragmented documentation into a structured repository, employs Peircean abduction to infer production steps, and optimizes these through self-reflective comparison between generated and reference artworks. A case study demonstrates coherent process reconstruction across multiple artists and movements, enabling interpretable workflow representations for human-AI co-creativity and cultural studies.
artistic process modelingpeircean abductionheterogeneous evidence synthesiscompositional graphself-reflective optimization
GitLake: Git-for-data for the agentic lakehouse
GitLake introduces a Git-for-data architecture for agent-centric lakehouse systems, enabling version control at the lakehouse scale. The system extends single-table Iceberg snapshots to support lakehouse-wide commits, branches, and merges, facilitating isolated agent workflows with human oversight. Pipeline executions occur on temporary branches, with results published atomically via final merges. Preliminary insights from an Alloy model of core abstractions and production experiences validate the system's correctness and operational viability.
git-for-datalakehouseiceberg snapshotsagent-firstalloy model
Blind-Spots-Bench: Evaluating Blind Spots in Multimodal Models
The authors introduce $ exttt{blind-spots-bench}$, a diagnostic benchmark designed to expose persistent failure modes in multimodal AI systems through 235 human-curated tasks that are trivial for humans but challenging for models. The benchmark features structured reference solutions, a task taxonomy, and an automated grading pipeline for evaluating language, vision-language, and image-generation models. Results show a ≈10% performance gap between closed-source and open-weight models, with no single model dominating across all task types, revealing systematic blind spots even in frontier models.
multimodal modelsdiagnostic benchmarktask taxonomyautomated gradingblind spots
INTENT: An LSTM Framework for Vehicle Intention Prediction in Intersection Scenarios with Comprehensive Ablation Analysis
The INTENT framework proposes an LSTM-based model for vehicle intention prediction at intersections, focusing on straight, left-turn, and right-turn classification 2 seconds before the event. The method emphasizes real-time applicability for autonomous vehicle safety and trajectory prediction enhancement. Comprehensive ablation studies on the InD dataset demonstrate 99.71% prediction accuracy, validating the approach's effectiveness in complex driving scenarios.
lstmintention predictionautonomous vehiclesintersection scenariosablation study
From Legacy Documentation to OSCAL: An MCP-Based Agent Pipeline for Threat-Informed Continuous Compliance in Critical Infrastructure
The paper introduces a multi-agent pipeline grounded in MCP (Model Checking Problem) for converting natural-language system descriptions into NIST OSCAL-compliant artifacts, enabling continuous automated compliance management in critical infrastructure. The architecture separates LLM-based reasoning from deterministic knowledge retrieval using authoritative threat-intelligence sources, mitigating fabricated vulnerabilities and hallucinated attack paths. In a synthetic water utility scenario, the pipeline achieves 0.90 CVE recall and perfect D3FEND recall, producing schema-valid OSCAL System Security Plans and Security Assessment Reports. While errors persist in the initial asset extraction phase, the approach makes residual risks verifiable and suitable for manual review.
mcposcalknowledge graphcve recalld3fend
Psychological Competence as a Missing Dimension in AI Evaluation
The paper introduces psychological competence as a critical but overlooked dimension in AI evaluation for human-facing systems. Defining it as an AI's capacity to appropriately support user cognition, emotional interpretation, and decision-making, the authors propose assessing interaction properties like framing, tone, and uncertainty handling. Drawing on behavioral science and human-AI interaction research, they outline a conceptual framework for evaluation through scenario-based probes, structured human evaluation, and model-assisted methods, advocating for its adoption by stakeholders.
psychological competencehuman-ai interactionbehavioral scienceevaluation frameworkscenario-based probes
Understanding Axes of Difficulty For Long Context Tasks Via PredicateLongBench
The authors introduce PredicateLongBench, a novel benchmark for systematically evaluating long-context reasoning in LLMs by scaling task difficulty along multiple axes. The benchmark requires models to identify the longest contiguous subsequence satisfying given predicates (e.g., lexicographic ordering) within long inputs, using both synthetic random strings and natural document samples. Results show frontier models struggle with increasing difficulty, revealing limitations in current long-context capabilities without requiring LLM-based generation or evaluation.
long-context reasoningbenchmark designpredicate constraintssynthetic generationmodel evaluation
Multi-Agent Firewall Architecture for Privacy Protection of Sensitive Data in Interactions with Language Models
The paper introduces an open-source, privacy-focused firewall architecture for securing interactions with Large Language Models (LLMs). The system combines a browser extension and proxy to intercept HTTP(S) and WebSocket traffic, employing a multi-agent pipeline with deterministic detectors and LLM-driven semantic analysis for data leakage prevention. It features proprietary code protection and extensible components for future security enhancements like prompt injection evasion. The layered architecture supports deployment in heterogeneous environments, balancing computational cost, detection depth, and latency. Evaluations show the system achieves F1 scores up to 94.93% in optimal configurations.
firewall architecturedata leakage preventionsemantic analysisprompt injection evasionmulti-agent pipeline
PolyUQuest: Verifiable Structure-Aware Web RAG over Heterogeneous Graphs
PolyUQuest introduces a verifiable, structure-aware web RAG framework that models web content as a heterogeneous graph combining hyperlink topology, DOM hierarchy, and entity-relation knowledge. The system employs a two-tier router to select among three retrieval modes (direct block retrieval, cross-page graph traversal, multi-hop entity reasoning) based on query structure. Evaluated on 4,240 pages from Hong Kong Polytechnic University's website, it outperforms existing RAG systems in correctness (unspecified %), coverage, and faithfulness while reducing LLM token usage. Each answer includes verifiable citations with source metadata.
retrieval-augmented generationheterogeneous graphdom hierarchymulti-hop reasoningverifiable citations
Different Teachers, Different Capabilities: Sub-1B On-Device Distillation for Structured Text Enrichment
The study demonstrates task-specific distillation benefits by comparing a 0.6B student model (Qwen3-0.6B) distilled from an 8B reasoning teacher (deepseek-r1:8b) against same-size non-reasoning and managed pipeline teachers. Using QLoRA with three seeds, the student achieves 58% of the teacher's summary quality gap at 0.8s/article (vs. 39s), outperforming constrained decoding (+16.8 points) and few-shot prompting (+4.9). Teacher capabilities diverge: reasoning teachers transfer writing quality while managed pipelines improve label diversity, with grounding differences observed on short-article subsets. Results inform per-field routing for on-device structured text enrichment.
distillationstructured extractionqwen3-0.6bqloraon-device
MentalHospital: A Virtual Environment for Evaluating Psychiatric Clinical Encounters
The paper introduces MentalHospital, a virtual environment for evaluating LLM-based psychiatric clinical encounters, implementing the S.O.A.P. workflow with skill-augmented standardized patients derived from 1,193 de-identified EHR cases across 76 ICD-11 disorders. The method combines objective EHR comparisons and subjective clinical quality assessments, scaled via MentalEval, five domain-specific evaluators trained with rubric-grounded SFT and expert-guided DPO. Results show MentalHospital achieves 3.88/5 clinical fidelity from clinician surveys, with MentalEval demonstrating strong expert alignment (QWK=0.944), while top LLMs trail clinicians by 37.28 percentage points in psychiatric competence.
large language modelspsychiatric clinical encounterselectronic health recordsskill-augmented standardized patientsdomain-specific evaluators
Best-of-$N$ TTS Evaluation is Confounded by ASR Family Alignment
The paper reveals that Best-of-$N$ (BoN) text-to-speech evaluation is confounded by automatic speech recognition (ASR) family alignment, where verifier rankings reverse across Whisper, wav2vec 2.0, and HuBERT evaluators. It proposes cross-family rank ensembles (rank-averaging and conjunctive max-rank) to mitigate this bias, achieving a 12% relative reduction in word error rate (WER) to 1.61% at $N=10$ on LibriSpeech-PC test-clean, without degrading SIM-o/UTMOS metrics. Results show same-family verifier-evaluator pairs recover 2-3x more oracle headroom than cross-family pairs, suggesting lineage-level coupling despite high representational similarity (linear CKA 0.978).
best-of-nasr family alignmentcross-family rank ensemblesword error ratelibrispeech-pc
Compete Then Collaborate: Frontier AI Teachers Build a Verifiable Curriculum to Improve a Coding Student Beyond Imitation
The paper introduces a compete-then-collaborate framework where four frontier AI teachers (Claude, Codex-GPT, Grok, Gemini) are ranked via execution-based verification (unit tests, stdin-stdout checks) and then collaborate to build a verifiable curriculum for a student model (Qwen2.5-Coder). Key findings: (1) Teachers achieve 99-100% accuracy on standard problems after self-correction but show divergence on harder tasks (Gemini 77% > others); (2) Supervised fine-tuning on verified solutions degrades student performance (e.g., MBPP-test drops from 76.7% to 72.7%); (3) Reinforcement learning with verifiable rewards (RLVR) improves student performance (+49% on competition problems). The framework emphasizes learning via verifiable environments over imitation.
knowledge distillationexecution verificationreinforcement learningself-correctionverifiable curriculum
AutoPersonas: A Multi-Timescale Loop Engine for Open-Ended Persona Evolution
AutoPersonas introduces a multi-timescale loop engine for open-ended persona evolution, addressing self-locking failures in long-term persona agents. The system separates environment-side Occurrences, Observations, and persona State (OSO loop) to enable divergent future-facing material while requiring evidence-governed absorption. In a 40-day stress test with eight models generating 1,600 events, the method reduced macro-theme repetition from 61.8% to 36.3% using context-slice masking and divergence targeting, while preserving identity continuity.
self-lockingmulti-timescale loopevidence-governed absorptioncontext-slice maskingmacro-theme repetition
RhyMix: A Lightweight Adaptive Multi-Rhythm Network for Long-Term Time Series Forecasting
The paper introduces RhyMix, a lightweight hybrid neural architecture for long-term time series forecasting that combines parallel dual-path modeling with adaptive gating. The model integrates a Cyclic Path with learnable seasonal embeddings and a Multi-Scale Temporal Convolutional Network with Channel Attention Path, dynamically balanced via path and hybrid gates. Evaluated on 12 datasets, RhyMix achieves state-of-the-art performance on 10, with ~40K parameters and linear complexity, enabling efficient deployment on edge devices.
time series forecastingadaptive gatingmulti-scale convolutioncyclic embeddingslinear complexity
Playing ZendoWorld: Challenging AI Agents on Active Visual Concept Induction
The paper introduces ZendoWorld, an interactive environment for evaluating AI agents' ability to induce visual concepts through perception, hypothesis formation, and active experimentation. Agents are tested across methodologies including pure vision-language model (VLM) reasoning, Bayesian particle filtering, dynamic concept discovery, and neuro-symbolic approaches. Key findings reveal: (1) high label prediction accuracy does not guarantee rule recovery; (2) perception and induction present distinct bottlenecks; and (3) VLM-based agents generate uninformative experiments, failing to reduce hypothesis uncertainty. Human performance comparisons highlight gaps in complex rule induction.
zendoworldvisual concept inductionbayesian particle filteringneuro-symbolic methodsactive experimentation
TMI: Text-to-Image Meets Image-to-Image for Complementary Data Synthesis to Boost Long-Tailed Instance Segmentation
The paper introduces TMI, a hybrid pipeline combining text-to-image (T2I) generation and image-to-image (I2I) editing to address long-tailed instance segmentation. The method uses T2I for broad scene diversity and a teacher-student scheme for label reliability, while VRAIN (Verified Rare-class Augmentation via INstructed editing) performs context-aware I2I edits to insert rare-class instances. On LVIS, TMI improves overall AP by +4.0 and rare-class AP by +9.5, scaling effectively with backbone capacity.
instance segmentationlong-tailed distributiontext-to-imageimage-to-imagedata synthesis
A First-Principles Theory of Slow Thinking and Active Perception
The paper proposes 'active lifting', a first-principles mathematical theory for modeling slow thinking and active perception in cognitive systems. The framework lifts and projects probability distributions between observable and latent spaces, using neural networks to represent complex data distributions. It derives a design space for slow thinking models, introduces an inference process with internal time dynamics, and suggests a training objective resembling minimum-length coding. Key outcomes include a three-stage model improvement pathway, unified encoder-generator construction for multimodal data, and potential solutions to policy collapse.
active liftingslow thinkinglatent spacesminimum-length codingpolicy collapse
Open-ended Multi-agent Autocurricula via Visual Inspection of Policies with Multi-modal LLMs
The paper introduces Visual Inspection of Policies (VIP), a method for generating open-ended RL curricula by analyzing agent behavior videos with Video Language Models (VLMs). VIP leverages VideoLLaMA2-7B to process policy videos and recommend tasks, outperforming text-only approaches and scalar score-based methods on the StarCraft Multi-Agent Challenge (SMAC). Empirical results demonstrate VIP's effectiveness in creating curricula that facilitate learning complex multi-agent skills through visual policy assessment.
reinforcement learningopen-ended curriculavideo language modelsmulti-agent systemspolicy inspection
Leveraging Color Naming for Image Enhancement
NamedCurves+ introduces a novel image enhancement framework leveraging Color Naming for interpretable and user-adjustable retouching. The method integrates universal color names into a learning-based system, enabling global adjustments via tone curves for each named color, while a transformer block captures spatial dependencies for context-aware local edits. This approach enhances interpretability through explicit tone curve representations and supports interactive customization of retouching parameters. Extensive experiments on image retouching, tone mapping, and exposure correction tasks demonstrate NamedCurves+ outperforms state-of-the-art methods, offering both explainable and interactive image enhancement capabilities.
color namingtone curvestransformer blockimage retouchingexposure correction
LEEVLA: Seeing What Matters in Latent Environment Evolution for Vision-Language-Action
LEEVLA introduces a vision-language-action (VLA) architecture that explicitly guides attention to task-critical regions while preserving structured latent world representations. The method combines drift-guided dynamic prioritization (DGDP) for identifying salient regions and structured feature flow generation (SFFG) for modeling latent space evolution, using prototype-to-periphery prediction and mutual-neighborhood contrastive loss. Experiments on VLA benchmarks demonstrate consistent outperformance over prior methods, highlighting the importance of task-evidence guidance and structured latent reasoning.
vision-language-actionlatent space evolutiondynamic prioritizationcontrastive lossstructured reasoning
Out of Sight: Compression-Aware Content Protection against Agentic Crawlers
CAPE introduces a compression-aware defense against LLM-based agentic crawlers by exploiting their context compression vulnerability. The framework injects imperceptible perturbations into text, optimized via surrogate compressor extraction and prior-guided evolution, to induce severe information loss during agent compression. Evaluations across three content types and four compression settings demonstrate 75.8% higher information loss than baselines while preserving human readability, with successful transfer to LangGraph and GitHub Copilot. This work establishes context compression as a novel defense layer in content protection.
llm-based agentscontext compressioninvisible perturbationsprior-guided evolutioncontent protection
ASMR: Agentic Schema Generation for Ship Maintenance Report Writing
The paper introduces ASMR, an agentic framework for automatic schema generation from ship maintenance reports. The system employs two specialized agents: a Field Generation Agent that extracts semantic concepts via multi-granularity clustering, and a Structural Optimizer Agent that uses reinforcement learning to produce compact, non-redundant schemas. Preliminary results show the approach improves report completeness and consistency, while highlighting challenges in agentic AI and human-centered data management.
schema generationagentic frameworkmulti-granularity clusteringreinforcement learningship maintenance reports
Overthinking: Amplifying Reasoning Weights to Extract Learned Secrets
The paper introduces overthinking, a method to amplify reasoning in language models for improved auditing of hidden information. The approach constructs an overthinking model by interpolating parameters between a base instruction model M and a reasoning-distilled model R, with a scaling factor α>1: θ_Oα = θ_M + α(θ_R - θ_M). Layer-wise attenuation strategies maintain output quality while enhancing reasoning. Experiments on 2B-32B models show overthinking surfaces hidden information 10× more frequently than standard reasoning models, with effectiveness varying by secret type and perturbation direction.
overthinkingreasoning amplificationparameter interpolationlayer-wise attenuationmodel auditing
ProsMAE: Multi-Source MAE Pretraining for ISUP Grade Classification
Proposes ProsMAE, a multi-source Masked Autoencoder framework for histopathology representation learning, addressing challenges in whole slide image analysis including gigapixel scale and stain variation. The method pretrains on tiles from PANDA, CAMELYON17, and BRACS datasets to expose the encoder to diverse tissue morphology and acquisition conditions. Transfer learning with a frozen encoder and linear classifier (ProsCLS) for ISUP grade classification achieved higher mean validation quadratic weighted kappa (QWK) than vanilla MAE baselines, though robustness across data splits requires further evaluation.
masked autoencoderwhole slide imagescomputational pathologytransfer learningisup grade classification
LEXIC: Lightweight Eye-tracking eXtension via Injected Complexity
LEXIC introduces lightweight conditioning mechanisms for gaze-only eye-tracking models to bridge the performance gap with text-aware models on reading comprehension prediction. The method extends the EyeBench AhnCNN baseline (LEXIC-Base) by injecting word-level difficulty signals (GPT-2 surprisal, word frequency, word length) via concatenation (LEXIC-Concat) or a residual mechanism (LEXIC-Res). On OneStop's Unseen Text task, both variants yield +1.8 to +2.2 percentage point AUROC gains (p ≤ 0.065), with LEXIC-Concat additionally improving Unseen Reader performance by +2.9 points (p = 0.010).
eye-trackingreading comprehensionaurocresidual mechanismword-level difficulty
Prismata: Confining Cross-Site Prompt Injection in Web Agents
Prismata introduces a defense mechanism against cross-site prompt injection attacks in autonomous web agents by enforcing contextual least privilege. The method dynamically derives trust-based permission labels for page content, leveraging structural confinement inspired by classical integrity models to ensure mislabelings are bounded and privilege decreases. Mechanical confinement enforces these labels through content redaction and capability restriction, requiring no developer annotations. Evaluated across recent web agent attacks, including adaptive variants, Prismata significantly reduces attack success while maintaining benign task utility.
prompt injectionleast privilegestructural confinementpermission labelsweb agents
ICDAR 2026 HIPE-OCRepair Competition on LLM-Assisted OCR Post-Correction for Historical Documents
The ICDAR 2026 HIPE-OCRepair competition evaluates LLM-assisted OCR post-correction for historical documents, introducing a harmonized multilingual dataset (English, French, German) spanning 17th-20th century newspapers and printed works. Systems employed strategies from zero-shot prompting to continued pre-training, assessed via retrieval-oriented metrics on paragraph/article-level transcriptions without image access. Results demonstrate significant OCR quality improvements but reveal performance variability across languages, noise levels, and a recurring over-correction challenge on low-noise inputs, underscoring the need for nuanced evaluation beyond character error reduction.
ocr post-correctionlarge language modelshistorical documentsmultilingual evaluationretrieval-oriented scoring
Answer Set Programming Energised! End-to-End Neurosymbolic Reasoning and Learning with ASP and Energy Based Models
The paper introduces a neurosymbolic framework combining answer set programming (ASP) with energy-based models, enabling joint optimization in continuous latent spaces while preserving ASP's declarative semantics. The method integrates background knowledge, constraints, and non-monotonic reasoning, extending prior work on ASP-modulo theories. Implemented and evaluated on MNIST, Clevr (visual QA), and MOT (multi-object tracking), it demonstrates end-to-end training for dynamic domains requiring perception and interaction.
neurosymbolic reasoninganswer set programmingenergy-based modelsnon-monotonic inferencedeclarative semantics
PS4: Proxy-Supervised Joint Training for Real Target Speaker Extraction
The paper introduces PS4, a proxy-supervised joint training framework for target speaker extraction (TSE) in real conversational mixtures, addressing the lack of large-scale training data and clean supervision. The method constructs a 71,771-sample corpus from four public datasets and fine-tunes a BSRNN-based TSE model using four differentiable objectives: ASR cross-entropy, speaker similarity, voice activity detection, and perceptual quality. Starting from a pre-trained checkpoint, PS4 achieves 2nd place on the REAL-T challenge leaderboard, excelling in speaker similarity and timing F1.
target speaker extractionproxy-supervised trainingbsrnnvoice activity detectionreal-t challenge
CausalDS: Benchmarking Causal Reasoning in Data-Science Agents
The paper introduces CausalDS, a benchmark for evaluating causal reasoning in data-science agents, addressing gaps in existing benchmarks that lack realistic data analysis or principled causal structures. CausalDS generates synthetic structural causal models (SCMs) with observational data and natural-language stories, optionally grounded in real-world distributions to mitigate 'causal parrot' risks. It spans Pearl's three rungs of causality, incorporating data-science tasks, imperfect observations, and abstention as a scored outcome. The benchmark assesses symbolic reasoning, tool use, coding, and uncertainty quantification.
structural causal modelcausal reasoningdata-science agentspearl's rungsuncertainty quantification
Deep Learning Method for Stationary Distribution of Reflected Brownian Motion
The authors propose a deep learning method to learn the Laplace transform of high-dimensional reflected Brownian motion (RBM) stationary distributions, addressing the intractability of closed-form solutions. Their approach leverages the basic adjoint relationship (BAR) through a carefully designed loss function, training data sampling strategy, and neural network architecture. Evaluations on RBM instances with known ground-truth tail probabilities demonstrate near-perfect prediction accuracy in high dimensions, suggesting broad applicability for stochastic system analysis beyond analytically tractable cases.
reflected brownian motionlaplace transformstationary distributiondeep learningbasic adjoint relationship
PARA-PV: Physics-Aware Retrieval-Augmented PV Prediction Based on Frozen Foundation Model and Distribution Shift Correction
The paper introduces PARA-PV, a Physics-Aware Retrieval-Augmented framework for photovoltaic (PV) power forecasting. The method combines patch-level representations of PV observations with physics-aware retrieval of historical analog trajectories, followed by calibration using a frozen Chronos time-series foundation model. A distribution shift correction module adjusts forecasts based on weather and diurnal conditions, while a physics-constrained loss function prioritizes critical operational states. The approach integrates physical knowledge throughout the forecasting pipeline to improve accuracy in diverse PV generation regimes.
photovoltaic forecastingretrieval-augmented learningdistribution shift correctionphysics-aware modelingtime-series foundation model
LDFE: Laplacian Decoupled Feature Enhancement Block for Dual-Stream CNN-based RGB-IR Object Detection
The paper introduces Laplacian Decoupled Feature Enhancement (LDFE), a novel block for RGB-IR object detection that enhances dual-stream CNN backbones through modality-aware feature fusion. LDFE employs global-local decomposition via Laplacian Pyramid, followed by denoising and fusion using Global State Space Enhancement (GS2E) and Local Convolutional Correlation Enhancement (LC2E) modules. GS2E uses cross-modal attention and State Space Models for noise suppression, while LC2E extracts fine-grained details. Evaluations show mAP improvements of 2.0-6.2% over SOTA across six datasets (M3FD, DroneVehicle, LLVIP, FLIR-Aligned, KAIST, VEDAI).
laplacian pyramidstate space modelcross-modal attentiondual-stream cnnfeature fusion
COBART: Controlled, Optimized, Bidirectional and Auto-Regressive Transformer for Ad Headline Generation
We introduce COBART, a Controlled, Optimized, Bidirectional and Auto-Regressive Transformer for ad headline generation, addressing evolving ad formats and customization needs. The method integrates prefix control tokens with BART fine-tuning, enabling user control over headline length and optimization for click-through-rate (CTR). COBART demonstrates flexibility across architectures and optimization criteria. Experimental results show a 25.82% improvement in Rouge-L and a 5.82% increase in estimated CTR compared to existing baselines, establishing its effectiveness in generating optimized and customizable ad headlines.
prefix control tokensbart fine-tuningclick-through-raterouge-lad headline generation
Persuasion Attacks Can Decrease Effectiveness of CoT Monitoring
We demonstrate that chain-of-thought (CoT) monitoring, a safety mechanism for AI agents, becomes less effective against adversarial persuasion attacks, increasing approval of policy-violating actions by 9.5% on average. Through an evaluation framework with 40 tasks analyzing thousands of agent-monitor interactions, we show that CoT reasoning provides an additional channel for persuasion. We propose a model-diverse fact-checking framework, pairing monitors and fact-checkers from different model families (e.g., Claude 3.7 Sonnet with GPT-4.1), which reduces approval of harmful actions by up to 45%, compared to 6% when using the same model for both roles.
chain-of-thought monitoringpersuasion attacksfact-checking frameworkmodel-diverse mitigationpolicy-violating actions
When LLMs Agree, Are They Right? Auditing Self-Consistency and Cross-Model Agreement as Confidence Signals
The study challenges the assumption that self-consistency or cross-model agreement in LLM-as-judge systems reliably indicates correctness, demonstrating it is an unreliable proxy for accuracy. Through a large-scale experiment involving 53 runners, 265,000 samples across models (GPQA Diamond, AIME), and hierarchical bootstrapping, agreement shows weak predictive power (rho 0.20-0.59) and regime-dependent utility. Frontier models exhibit over-confidence (≥0.8 agreement on 77% of GPQA cases, 48% wrong), while mid-tier models perform better. The findings caution against using agreement as a standalone confidence score, releasing per-run data for further analysis.
llm-as-judgeself-consistencycross-model agreementconfidence signalhierarchical bootstrap
When Thinking Hurts: Epistemic Signals in the Reasoning Chains of Visual Language Models
This work provides the first empirical characterization of uncertainty signals in thinking-mode visual language models (VLMs), identifying three distinct entropy patterns across Qwen3-VL-8B-Thinking (complete collapse), GLM-4.1V-9B-Thinking (no collapse), and InternVL3-8B (selective thinking). Analyzing four models on POPE adversarial samples and VQAv2, the study demonstrates that reasoning chain entropy consistently outperforms answer token entropy as a reliability signal (0.647-0.759 vs 0.492-0.716 AUROC), particularly for free-form answers. Additional findings include structured abstention affecting 12-22% of queries and a practical abstention gate improving accuracy from 71.0% to 93.8% at 62.7% coverage.
visual language modelsuncertainty quantificationreasoning chainsentropy collapseadversarial samples
Towards Efficient Large Language Model Serving: A Survey on System-Aware KV Cache Optimization
This survey systematizes research on system-aware KV cache optimization (sKis) for efficient large language model serving, addressing memory-intensive challenges in autoregressive decoding. The authors categorize existing approaches along three dimensions: temporal (execution/scheduling), spatial (placement/migration), and structural (representation/retention), while analyzing cross-behavior co-design opportunities. The work provides a foundational taxonomy for understanding KV cache optimization in modern LLM serving infrastructure, highlighting future research directions.
kv cachellm servingautoregressive decodingmemory optimizationsystem-aware
Reinforcing the Generation Order of Multimodal Masked Diffusion Models
The paper introduces a learnable control module trained via Group Relative Policy Optimization (GRPO) to optimize generation order in multimodal masked diffusion models, addressing the insufficiency of model logits for determining optimal sequences in text-to-image synthesis and multimodal understanding. The method enhances text-to-image alignment and multimodal reasoning by improving fine-grained spatial relationship capture. Evaluations on GenEval show 4.08% relative improvement in text-to-image alignment, while VLMEvalKit results demonstrate 4.85% relative improvement in multimodal understanding.
diffusion language modelsgeneration ordergroup relative policy optimizationmultimodal understandingtext-to-image synthesis
Who Analyses the Analyser? Self-Validating LLM Hazard Analysis with Constitutional Meta-STPA
The paper introduces Constitutional Meta-STPA, a method for self-validating LLM-assisted safety analysis tools by applying Systems-Theoretic Process Analysis (STPA) to the tool itself. The approach derives a governance constitution (21 Tool Principles, 8 Meta-Safety Principles) from a meta-STPA of AI-assisted safety tools, formalized via a constitution-marginal coverage operator. Results show that a frontier LLM ensemble (claude-opus-4.8 + claude-sonnet-4) recovers 18/21 canonical and 8/8 governance principles, demonstrating model-limited but constitution-robust self-derivation.
constitutional meta-stpasystems-theoretic process analysisllm-assisted safety analysisgovernance constitutionmeta-safety principles
What LLM Forecasters Know but Don't Say: Probing Internal Representations for Calibration and Faithfulness
The study demonstrates that probing internal representations in large language models (LLMs) improves forecast calibration and reveals reasoning faithfulness. Using representation-pooling probes on intermediate activations of Eternis-Forecaster 8B, GLM-4.7-Flash, and GLM-4.5-Air, the authors achieve better calibration and detect behavioral shifts more effectively than chain-of-thought (CoT) traces. Evidence ablation and diversionary injection show CoT often remains unchanged despite forecast alterations, while probes predict direction of change in 84% of cases. Pre-reasoning activations largely determine forecasts, enabling token-efficient routing with 30-47% savings and no accuracy loss.
large language modelschain-of-thoughtrepresentation-pooling probesevidence ablationforecast calibration
Aleena: Alignment Agent for Research Software Engineering Collaborations
Aleena, an open-source lifecycle alignment agent, addresses divergent mental models in research software engineering collaborations by transforming multi-modal stakeholder interactions into structured project records. The system uses GitHub as a shared collaboration surface to surface risks, track open questions, and preserve decision continuity across meetings, Slack threads, pull requests, and GitHub issues. Grounded in university-based research software engineering center experiences, Aleena supports stakeholder alignment and project-state tracking without replacing human decision-making. The paper presents the motivating problem, system design, prototype, and illustrative lifecycle scenarios for this agentic AI solution.
lifecycle alignmentresearch software engineeringmulti-modal interactionsstructured project recordsagentic ai
A safety-oriented hypothetico-deductive framework for AI-assisted differential diagnosis
The study introduces AegisDx, a safety-oriented hypothetico-deductive framework for AI-assisted differential diagnosis, addressing limitations of one-shot LLM predictions. The framework employs specialized LLM components with role-specific contracts, structured outputs, evidence retrieval, and verification gates to ensure comprehensive differential diagnoses and explicit screening for high-risk conditions. Evaluated on case reports from NEJM, JAMA, and Annals of Emergency Medicine, AegisDx achieved Top-3 accuracies of 59.9%, 62.7%, and 85.7%, respectively, outperforming standalone LLMs (52.1%, 51.4%, 68.6%). Physician evaluations on real-world ED notes showed improved safety scores (4.55 vs. 4.31, p=2.1x10^-4), particularly in must-not-miss condition identification.
hypothetico-deductive reasoningdifferential diagnosislarge language modelsclinical decision supportsafety verification
PLURAL: A Global Dataset for Value Alignment
The authors introduce PLURAL, a global dataset for value alignment containing ~500,000 synthetic preference triplets derived from the Integrated Values Survey (IVS) across 20 countries. Using a two-stage generation pipeline, they transform nationally representative survey responses into preference triplets that preserve cross-country value differences and within-country diversity. Evaluations show PLURAL reduces mean absolute error by up to 27.7% in aligning LLM outputs with target cultural profiles, with blind human evaluations confirming improved representativeness across India, Brazil, and Japan.
value alignmentpreference tripletspluralistic alignmentintegrated values surveycultural profiles
DKDNet: Dual Knowledge and Data-Driven Network for Cross-Domain Automatic Modulation Classification
The paper proposes DKDNet, a dual knowledge and data-driven network for cross-domain automatic modulation classification (AMC) that leverages signal prior knowledge to enhance generalization. The method combines in-phase/quadrature (IQ), amplitude-phase (AP), and autocorrelation function (ACF) representations via a multi-representation feature encoder (MRFE) and dynamic lightweight fusion unit (DLFU), optimized with classification and adversarial domain alignment objectives. Experiments on simulated and public datasets show superior performance, validating the selected signal priors and the proposed approach.
automatic modulation classificationcross-domain learningsignal prior knowledgeadversarial domain alignmentmulti-representation fusion
From Prompts to Contracts: Harness Engineering for Auditable Enterprise LLM Agents
The paper introduces a harness-engineering approach for transforming enterprise LLM prototypes into auditable applications with deterministic behavior. The method reconstructs LLM-agent architecture using code, manifests, schemas, and validation artifacts around a replaceable composition boundary, ensuring traceability and source-grounding. Evaluated on a dataset of 25 Korean corporate entities, the approach demonstrated three key results: (1) preserved source-grounding and contract validation across scenarios, (2) maintained checks under model substitution across 270 runs, and (3) ensured safety and utility through code-owned guarantees, outperforming prompt-only and external guardrail methods. The approach enables reproducible traces and versioned artifacts for enterprise LLM applications.
llm-agentcomposition-boundarysource-groundingvalidation-artifactstraceability
Structured Pruning of Large Language Models via Power Transformation and Sign-Preserving Score Aggregation with Adaptive Feature Retention
The paper introduces an improved structured pruning method for large language models (LLMs) that adapts Adaptive Feature Retention (AFR) to structured pruning while addressing distribution mismatch, sign information loss, and outlier effects. The proposed approach combines power transformation for nonlinear distribution alignment, sign-preserving score aggregation, and percentile-based outlier removal. Evaluations on Llama-3-8B, Vicuna-v1.5-13B, and LLaVA-v1.5-13B show the method maintains accuracy comparable to unstructured pruning while achieving practical inference speedup.
structured pruningadaptive feature retentionpower transformationsign-preserving aggregationoutlier removal
APIVOT: Adaptive Planning with Interleaved Vision-Language Thoughts
APIVOT introduces an adaptive vision-language model (VLM) planner that interleaves language and visual reasoning for long-horizon robot tasks. The method combines semantic decomposition via language with visual imagination of future states to verify geometric feasibility, addressing both task structure and spatial constraints. Evaluated on kitchen tasks, APIVOT outperforms general-purpose VLMs and prior planners, particularly in spatially constrained scenarios, while learning meaningful modality selection. Results show improved planning success and reasoning efficiency through adaptive vision-language interleaving.
vision-language modellong-horizon planningsemantic reasoninggeometric feasibilitymodality interleaving
Concretized Proposition Prompting Resolves Composition-Knowledge Dichotomy in Large Language Models
The paper introduces Concretized Proposition Prompting (CPP) to address the composition-knowledge dichotomy in LLMs, where models struggle to balance compositional reasoning with factual knowledge. CPP explicitly concretizes propositions relevant to queries, enabling logically organized and factually grounded reasoning. Evaluations show CPP significantly improves performance on medical benchmarks requiring precise knowledge while remaining competitive on math-focused tasks. The method scales across foundation models and parameter sizes, demonstrating broad applicability.
composition-knowledge dichotomyconcretized proposition promptinglarge language modelsdeductive reasoningmedical benchmarks
Can We Trust LLM's Logic? Quantifying Uncertainty, Coherence, and Robustness via a Graph-Based Framework
The authors introduce GRAPHEVAL, a graph-based framework addressing uncertainty quantification (UQ) and reasoning fidelity in Large-Language Models (LLMs). They propose Graph Reasoning Coherence Score (GRCS), a novel UQ metric evaluating semantic-structural consensus, and Graph Self-Consistency (GSC), a medoid-based decoding strategy. GRCS consistently correlates negatively with reasoning unfaithfulness across model sizes, while GSC improves reasoning fidelity in larger models and exposes inflated accuracy from lucky guesses in smaller ones. Adversarial medoid ablation demonstrates GSC-selected paths are critical for maintaining reasoning faithfulness and accuracy.
uncertainty quantificationgraph-based reasoningsemantic-structural consensusmedoid-based decodingreasoning fidelity
Provably Optimal Learning Algorithms for Assistance Games
The paper introduces provably efficient learning algorithms for online assistance games, where an informed human and uninformed assistant interact over T timesteps to optimize a shared reward. The authors propose decentralized algorithms achieving a (1-1/e)-approximate assistance regret of Õ(T^{3/4}), with polynomial runtime in action and state space sizes, and prove computational intractability of better approximation factors. A pseudo-decentralized variant using shared randomness achieves optimal Õ(T^{1/2}) regret. The framework accommodates any no-regret algorithm for the assistant.
assistance gamesonline learningregret minimizationdecentralized algorithmscomputational intractability
Beware What You Autocomplete: Forensic Attribution of Backdoored Code Completions
The paper introduces CodeTracer, a forensic framework for attributing backdoored code completions to malicious fine-tuning data in large language models. The method extracts behavioral fingerprints from compromised outputs, narrows candidates via semantic search, and uses LLM-based reasoning for attribution. Evaluations across 3 vulnerability cases, 10 backdoor attacks, and 16 baselines show CodeTracer achieves high forensic accuracy (quantified in source), low false positives, and robustness against adaptive attacks.
code completionbackdoor attackforensic attributionbehavioral fingerprintllm-based reasoning
Reaction-network reasoning with frontier models for experimentally confirmed catalyst-selectivity hypotheses
The study introduces a human-AI co-thinking framework leveraging frontier language models for catalyst discovery, constrained to reason over explicit reaction networks to identify physical levers governing pathway competition. The method enforces network invariance to extract testable hypotheses from complex chemical graphs, focusing on electrochemical CO2 reduction. Results show the framework predicted novel pathways (ketene desorption, hydroxide capture) and guided synthesis of a copper-iron oxide catalyst with threefold increased acetate selectivity over baselines, demonstrating mechanism-guided materials discovery.
reaction-network reasoningfrontier language modelscatalyst-selectivityelectrochemical co2 reductionmechanism-guided discovery
SpO$_2$ Predictor-Guided Stage-Wise Time-Frequency Reconstruction of Low-Quality Dual-Wavelength PPG for Oxygen Saturation Estimation
The paper proposes a SpO$_2$ predictor-guided stage-wise time-frequency reconstruction framework for enhancing low-quality dual-wavelength PPG signals. The method pretrains a SpO$_2$ predictor on high-quality PPG segments, then trains a masked reconstruction model with joint time-domain waveform and frequency-domain STFT losses, incorporating the SpO$_2$ predictor as a constraint. Four-stage optimization ensures physiological relevance. Evaluations on the OpenOximetry Repository and a private dataset show subject-level MAEs of 2.882% and 2.359%, respectively, outperforming prior methods.
ppg reconstructionoxygen saturationtime-frequency analysisshort-time fourier transformwearable monitoring
Who Broke the System? Failure Localization in LLM-Based Multi-Agent Systems
AgentLocate is introduced as a framework for failure localization in LLM-based multi-agent systems, addressing challenges in diagnosing system-level failures due to distributed reasoning. The method combines an LLM-based judging mechanism with multi-perspective verification by independent evaluators, aggregated via a confidence-aware strategy, and adapts the judge through lightweight fine-tuning. Evaluated on two benchmarks with diverse tasks and configurations, AgentLocate outperforms existing methods in identifying responsible agents and failure steps while maintaining efficiency in token usage and runtime.
failure localizationmulti-agent systemsllm-based judgingconfidence-aware aggregationlightweight fine-tuning
A Reliability Assessment of LALM Audio Judges for Full-Duplex Voice Agents
The study evaluates the reliability of Gemini family models (2.5 Flash, 3.5 Flash, 3.1 Pro) as audio judges for scoring full-duplex voice agent conversations from stereo waveforms. Using Gemini 2.5 Flash as ground truth, the authors compare model performance against three human raters across 209 sessions (152 conversations, 57 adversarial clips) on 8 production dimensions. Results show Gemini 2.5 Flash achieves human-level agreement (Spearman ρ within 0.07 on 5/8 dimensions, 60-92% 1-point agreement on 6/8 dimensions) and detects defects comparably to humans (45/48 cases). Model performance varies across the Gemini family, with 3.5 Flash showing improved agreement but 3.1 Pro rating lower than humans despite comparable rank correlation.
full-duplexaudio judgesspearman rhogemini modelsadversarial clips
Agentic Neural Architecture Search
The paper introduces Agentic Neural Architecture Search (AgentNAS), a hybrid method combining large language model (LLM) generation with neural architecture search (NAS). The approach uses an LLM to produce a seed architecture decomposed into a 'slotted architecture', which defines a bounded search space for NAS optimization without manual engineering. Evaluated on 17 tasks across NAS-Bench-360 and Unseen NAS benchmarks, AgentNAS achieves state-of-the-art performance on 11 tasks, outperforming task-specific expert designs. Ablations confirm complementary benefits: LLM seeds surpass baselines alone, while NAS provides additional gains through combinatorial slot recombination.
neural architecture searchlarge language modelslotted architecturecombinatorial searchautoml
3100 Opinions on Code Review in an AI World: Building Causal Theory from Practitioner Discourse
The study constructs a causal theory of AI's impact on code review by synthesizing practitioner discourse, addressing conflicting observations about agent-authored pull requests (reviewed less, merged faster, discussed less). Analyzing 38,709 grey-literature documents, the authors code a stratified sample of 3,100 using an LLM-assisted pipeline, identifying 26 constructs and 67 relationships. The resulting model posits that review acts as a control point for AI's effects, determined by team expertise and process structure. The method offers a scalable LLM-assisted template for software-engineering research, with public implementation.
code reviewcausal theorygrey-literaturellm-assisted pipelinepull requests
When Implausible Tokens Get Reinforced: Tail-Aware Credit Calibration for LLM Reinforcement Learning
We propose Tail-Aware Credit calibratiOn (TACO), a method addressing Positive-Credit Contamination in critic-free reinforcement learning for LLMs, where uniform credit assignment indiscriminately reinforces low-probability tail tokens. TACO computes a tail-risk score incorporating local generation context to assess token reliability, then tunes positive credit for risky tokens without eliminating gradients, allowing useful rare patterns to accumulate reinforcement while dampening noise. Experiments across three LLMs and eight benchmarks demonstrate TACO's consistent outperformance of GRPO-style baselines, with improved training stability and sustained performance gains in long-horizon RL.
reinforcement learningcredit assignmenttail-risk scorepositive-credit contaminationlong-horizon rl
A Multi-cluster Boundary Learning Method for Out-of-Scope Intent Detection via MiniLM Embedding
The paper proposes a multi-cluster boundary learning method for out-of-scope (OOS) intent detection using MiniLM embeddings (all-MiniLM-L6-v2) in a one-class classification framework. The approach learns boundaries from multi-cluster embeddings of training utterances to reject OOS intents, addressing limitations of traditional multi-class classification and large LLM embeddings. Experiments on CLINC150, StackOverflow, and Banking77 datasets demonstrate state-of-the-art OOS detection performance, with ablation studies confirming MiniLM's suitability for utterance embedding requirements.
oos intent detectionminilm embeddingmulti-cluster boundary learningone-class classificationutterance embedding
Beyond Thermal Imaging: Inferring Thermophysical Properties from Time-Resolved Thermal Observations
ThermoField introduces a unified framework for thermal scene reconstruction and thermophysical parameter estimation via differentiable heat-transfer simulation. The method represents geometry and thermophysical properties as spatially varying neural fields, constrained by scene geometry, heat-transfer physics, and temporal thermal observations. Results demonstrate joint reconstruction of geometry, estimation of spatially varying thermal diffusivity, and prediction of thermal evolution under novel environmental conditions. This approach bridges thermal scene reconstruction and inverse heat-transfer analysis, enabling physically interpretable parameter inference in complex 3D scenes.
thermal imagingheat-transfer simulationthermophysical propertiesneural fieldsthermal diffusivity
Evaluating the Effect of Frame Rate in Sequence-Based Classification of Autism-Related Self-Stimulatory Hand Idiosyncrasies
The study identifies optimal neural architectures and temporal sampling rates for video-based autism behavior classification, while evaluating data augmentation strategies for small datasets. Using LSTM and GRU models on pose features from the SSBD dataset at varying frame intervals (1-90 frames), peak accuracies reached 97.5% (LSTM) and 98.75% (GRU) at 15-frame intervals. An ablation study of 10 augmentation techniques showed horizontal flip most effective (48.78% accuracy), with upsampling being critical for performance. Personalized models achieved consistent predictions (mean loss 1.84).
lstmgrudata augmentationtemporal samplingpose-derived features
Linear Attention Architectures: Mechanisms, Trade-offs, and Cross-Layer Routing
The paper presents a systematic comparison of recurrent linear-attention architectures (DeltaNet, Gated DeltaNet, Kimi Delta Attention, Gated DeltaNet-2) expressed through a unified recurrent-memory framework, analyzing trade-offs in expressivity, memory dynamics, and computational efficiency. Experiments on 350M-parameter models (scaled to 3B) trained for 15B tokens evaluate optimizer choices, hybrid architectures, and propose cross-layer routing mechanisms. Key findings show Kimi Delta Attention with Muon optimizer achieves lowest validation loss, Gated DeltaNet with AdamW maximizes throughput, and Cross-Layer Value Routing (CLVR) modestly improves performance by propagating hidden states across layers.
linear attentiondelta rulerecurrent memorycross-layer routingtraining throughput
path_boost: A Python Package for Interpretable Graph-Level Prediction using Path-Based Gradient Boosting
The paper introduces path_boost, a Python package for interpretable graph-level prediction using PathBoost, a gradient boosting algorithm that discovers predictive labeled paths in graphs. Unlike graph neural networks, PathBoost produces additive models with explicit path-based features, enabling interpretability. The algorithm avoids exhaustive path enumeration by iteratively selecting and extending paths based on predictive power, supporting regression and binary classification. Benchmarks on six molecular datasets show competitive performance against graph neural networks and graph kernel methods. The package integrates with scikit-learn and offers parallel training, custom base learners, and variable importance computation.
gradient boostinginterpretable machine learninggraph-structured datapath-based featuresmolecular property prediction
Adversarial Decoys: Misdirecting Attention-Based Defenses in ViT
The paper introduces adversarial decoys, a method to circumvent attention-based defenses in Vision Transformers (ViTs) by redirecting attention away from true adversarial regions. The approach independently optimizes decoy patches to manipulate attention rankings while preserving attack effectiveness, decoupling misclassification and defense evasion objectives. Experiments on ImageNet demonstrate that decoys successfully divert attention from adversarial tokens across multiple ViT architectures and attacks, revealing limitations of attention magnitude as an adversarial indicator.
adversarial decoysvision transformersattention-based defensesadversarial patchesimagenet
Efficient Safety Alignment of Language Models via Latent Personality Traits
The paper introduces Latent Personality Alignment (LPA), a lightweight safety alignment method for language models that replaces explicit harm refusal with adversarial training on 66 harm-agnostic personality statements. LPA hypothesizes that personality-anchored representations share latent structure with harm avoidance, implicitly constraining jailbreak attack vectors. Results show near-zero attack success rates on HarmBench across direct requests and five jailbreak methods, with no utility degradation on standard benchmarks. The method requires only minutes on a single GPU and 75x fewer examples than Latent Adversarial Training (LAT). Ablations confirm robustness and efficiency.
latent personality alignmentsafety alignmentjailbreak attacksadversarial trainingharmbench
Persona Cartography: Charting Language Model Personality Traits in Weight Space
The paper introduces a method for analyzing and controlling persona traits in large language models (4B-32B parameters) using the OCEAN framework (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism). By training low-rank adapters to modulate these traits, the authors demonstrate monotonic scaling effects, additive trait combinations, and preserved model capabilities at moderate scales. Evaluations using an LLM-judge and human-validated benchmarks show that trait manipulation affects safety-relevant behaviors (e.g., neuroticism influences frustration). An unsupervised psychometric pipeline identifies four additional behavioral factors (tone, initiative, didacticism, epistemic caution), linking personality measurement to model editing and safety.
persona cartographylow-rank adaptersocean frameworkllm-judgepsychometric pipeline
Multimodal Unlearning Across Vision, Language, Video, and Audio: Survey of Methods, Datasets, and Benchmarks
This survey presents a unified framework for multimodal unlearning across vision, language, audio, and video, addressing the challenge of selectively removing sensitive or unwanted cross-modal associations in foundation models (VLMs, DMs, LLMs, AFMs) without full retraining. It systematizes recent advances through a taxonomy comparing model architectures and modalities, evaluating trade-offs in deletion strength, retention, efficiency, reversibility, and robustness. The work highlights open problems and practical deployment considerations, accompanied by a curated repository of resources.
multimodal unlearningfoundation modelscross-modal associationsselective removalrepresentation learning
Mechanistic Interpretability of LLM Jailbreaks via Internal Attribution Graphs
The paper introduces a mechanistic framework for analyzing LLM vulnerabilities through paired internal computation graphs, revealing how adversarial prompts systematically alter model reasoning. The method constructs and aligns computation graphs for clean and attacked prompts, decomposing them into invariant, suppressed, and emergent structures, and performs causal interventions on nodes and paths. Experiments on multiple LLMs and jailbreak benchmarks show structural deviations correlate with unsafe behaviors, and targeted interventions improve robustness.
internal computation graphsmechanistic interpretabilityjailbreak attackscausal interventionsmodel robustness
Closed-Loop Dynamic Validator Node Scaling in Private Substrate Blockchains Using Takagi-Sugeno Fuzzy Inference
The paper introduces a Takagi-Sugeno fuzzy inference system for dynamic validator node scaling in private Substrate blockchains. The method uses triangular membership functions and a 27-rule base to process live blockchain parameters (block production time, size, node count), outputting an efficiency score and scaling recommendation. Evaluated on a 10-node network handling smart water meter data, the system demonstrates stable autonomous scaling, outperforming threshold-based baselines with fewer oscillations while maintaining comparable block production times.
takagi-sugenosubstrate blockchainvalidator scalingfuzzy inferenceclosed-loop control
How Do I Know What to Say Next? Barenholtz's Autogenerative Theory as an Enrichment of Harrisean Integrationism
The paper proposes Elan Barenholtz's autogenerative theory as a framework to address three gaps in Roy Harris's Integrationist linguistics: prospective openness in communication, semiotic continuity between linguistic and non-linguistic activity, and structural properties of past integrations. By analyzing Large Language Models (LLMs), the autogenerative account provides computational mechanisms for these phenomena while maintaining Harris's core tenets. The synthesis offers NLP researchers a principled explanation of LLMs' statistical structures and their inherent limitations.
integrationist linguisticsautogenerative theorylarge language modelssemiotic continuityprospective openness
Time-to-Collision Based Dynamic Obstacle Avoidance Using Pretrained Vision Models for Robots in Unstructured Environments
A data-efficient, interpretable method for vision-based dynamic obstacle avoidance in unstructured outdoor environments is proposed, eliminating the need for robot-specific training data and sim-to-real transfer. The approach leverages UniDepth for monocular depth estimation and extends the SuperPoint-SuperGlue pipeline to track keypoints, compute time-to-collision (TTC), and select evasive motion primitives. Evaluated on the M3ED dataset, the method achieves 0.49 precision and 0.38 recall for TTC <1s detection, correctly generates evasive motion in 84% of true positives, and detects TTC <1s for 20/22 obstacles. It requires only 74s of data for hyperparameter tuning, demonstrating exceptional data efficiency and generalizability.
monocular depth estimationtime-to-collisiondynamic obstacle avoidancesuperpoint-supergluemotion primitives
Nigeria Machinery: A Low-Resource Industrial Dataset with a Domain-Grounded Reasoning Layer
The authors introduce the Nigeria Machinery Usage and Failures Dataset, comprising 89 machine-level records with 28 indicators from Nigeria's manufacturing and oil/gas sectors (2006-2025), alongside a method for generating domain-grounded chain-of-thought reasoning examples. They address a common issue where language model-generated prompts lack domain relevance, improving domain-grounded prompts from 1/78 to 94/94 and ensuring 84/84 retrieval answers match source values. The dataset includes provenance files and is released under CC-BY-4.0, though its small size (17 indicators with single observations) limits it to reference/seed use.
industrial machinery datasetchain-of-thought reasoningdomain-grounded promptslow-resource settingnumeric task grounding
Feedback Manipulation Regularization: Enabling Offline Agent Alignment for Imitation Learning
The paper introduces Feedback Manipulation Regularization (FMR), an algorithm-agnostic method for improving agent alignment in offline imitation learning by using evaluative feedback as a corrective signal. FMR integrates human demonstrations and feedback into a single-stage training process for sequential decision-making, contrasting with existing multi-stage pipelines designed for language generation. Experiments on adapted Safety Gymnasium environments show FMR reduces misalignment by up to 98% across various imitation learning algorithms and remains effective in low-data regimes with noisy demonstrations.
imitation learningagent alignmentfeedback manipulationoffline trainingsequential decision-making
Agentic AI and Retrieval-Augmented Models in Straight-Through Underwriting
The paper proposes an agentic AI framework for straight-through underwriting of Business Owner Policies (BOPs), comparing three pipelines: a single-LLM baseline, naive RAG, and multi-agent Agentic RAG. The method involves constructing a synthetic experimental environment to evaluate transparency, auditability, and decision accuracy. Results show the agentic system outperforms others, particularly in multi-step and missing-information scenarios, due to structured retrieval and reflection mechanisms.
agentic airetrieval-augmented generationstraight-through underwritinglarge language modelsmulti-agent systems
Multi-agent Autoformalization of Tensor Network Theory
The authors introduce a multi-agent system for autoformalizing theoretical physics concepts, demonstrated through the formalization of matrix-product states' fundamental theorem. The workflow employs specialized large language-model agents coordinated via a structured mathematical blueprint and periodic human review, enabling autonomous execution and exploration of novel proof routes. The agents generated extensive tensor-network and quantum-information libraries, extending applications to symmetry-protected topological phases in one dimension. The primary bottleneck identified is enforcing mathematical intent. The codebase, TNLean, and a detailed blueprint of the formalization effort are publicly released.
autoformalizationmatrix-product statestensor-networksymmetry-protected topological phasesmathlib
Kime-Representation Formulations of Three Open Problems in the Foundations of Classical Mechanics: Uncertainty, Invariant Entropy, and Directional Degrees of Freedom
The paper presents rigorous formulations of three open problems in classical mechanics using the complex-time (kime) representation, treating the kime phase as a latent circular random variable. By establishing an exact symplectic identification between the kime cone and action-angle coordinates, the authors (i) derive sharp entropic uncertainty relations with von Mises x Gaussian extremals, (ii) prove non-canonical uncertainty relations involving Poisson brackets, (iii) extend bounds to multi-degree-of-freedom systems via Williamson normal form, and (iv) show phase diffusion leads to monotone entropy growth. Results include Fisher-information inequalities and symplectic Schur-Horn-type open problems.
kime representationentropic uncertaintysymplectic identificationwilliamson normal formpoisson bracket
A Graph Neural Network Model for Real-Time Gesture Recognition Based on sEMG Signals
The paper introduces a graph neural network (GNN) model for real-time hand gesture recognition using surface electromyography (sEMG) signals. The method constructs graph networks encoding forearm muscle activation patterns from 8-electrode myoband data, achieving 99% average classification accuracy across 8 subjects. With 48ms latency for graph construction and inference on an M1 Pro CPU, the approach demonstrates suitability for real-time prosthetic and augmented reality control applications.
graph neural networksurface electromyographyreal-time recognitionhand gesture classificationmuscle activation patterns
VectorizationLLM: Smart Vectorization Based AI Assistant
VectorizationLLM introduces a specialized large language model for educational assistance in computational analysis, built upon Google's open-weight LLMs. The model employs retrieval-augmented generation (RAG) to provide detailed explanations of vectorization, Fourier analysis, and differential equations in MATLAB, specifically for CTEC 247 coursework. It delivers instructional content through code examples, text, and images while avoiding direct answers to maintain academic integrity. The system demonstrates how domain-specific LLMs can enhance technical education by contextualizing complex mathematical concepts.
vectorizationllmretrieval-augmented generationfourier analysisdifferential equationsmatlab
Shift & Drift: A Zero-Shot Benchmark for Generalizable and Robust Autonomous Driving Motion Planning
The paper introduces Shift & Drift, a zero-shot benchmark for evaluating autonomous driving motion planners' generalization and robustness across semantic and state-distribution shifts. The Semantic Shift Track converts the DeepScenario Open 3D dataset into nuPlan's framework for cross-regional evaluation (1,182 scenarios in German cities and San Francisco), while the State-Distribution Drift Track injects stochastic perturbations to test recovery from execution errors. Results show imitation learning methods fail under semantic shift and actuation noise, whereas reinforcement-learning-based planners degrade more gracefully, revealing a trade-off between imitation fidelity and closed-loop resilience.
motion planningdistribution shiftzero-shot evaluationclosed-loop resilienceautonomous driving
Infinity-Parser2 Technical Report
The work introduces Infinity-Parser2, a multimodal model for document parsing that combines controllable data synthesis with multi-task reinforcement learning. It contributes: (1) Infinity-Doc2-5M, a 5M-sample bilingual corpus with diverse annotations; (2) a joint reinforcement learning framework unifying eight parsing and understanding tasks; and (3) two model variants achieving SOTA performance (87.6% on olmOCR-Bench, 74.3% on ParseBench) while outperforming DeepSeek-OCR-2 and PaddleOCR-VL-1.5. The synthesis pipeline enables scalable training data generation, while the multi-task reward system jointly optimizes layout analysis, formula parsing, and document VQA.
multimodal parsingreinforcement learningdata synthesisdocument understandinglayout analysis
From Triggers to Emotions: A CPM-Grounded Appraisal Multi-Agent for Dynamic Emotional Evolution in Persona-Based Dialogue
The paper introduces CPM-MultiAgent, a novel framework for dynamic emotional evolution in persona-based dialogue systems, addressing the limitation of static emotion representations in current LLM-based agents. Grounded in the Component Process Model (CPM) psychological theory, the framework models emotions as latent states updated through affective trigger extraction, collaborative appraisal, and state updating mechanisms. Evaluations demonstrate its effectiveness in maintaining emotional consistency during multi-turn interactions, outperforming baselines in emotionally sensitive role-simulation scenarios.
component process modelemotional evolutionpersona-based dialogueaffective trigger extractionmulti-agent framework
DreamCharacter-1: From 3D Generative Foundation Models to Product-Ready Character Generation
DreamCharacter-1 introduces a lightweight post-adaptation framework for refining pretrained 3D foundation models to generate production-ready 3D characters. The method integrates three components: geometry post-training for surface detail enhancement via geometric preference optimization, texture post-training for high-resolution texture synthesis and occlusion refinement, and inference acceleration for scalable deployment. Evaluations show the framework outperforms state-of-the-art methods in generating high-fidelity, structurally robust 3D character assets.
3d foundation modelsgeometric preference optimizationtexture synthesisinference accelerationcharacter generation
From Solvers to Research: Large Language Model-Driven Formal Mathematics at the Research Frontier
The paper proposes a paradigm shift in AI for Mathematics (AI4Math) from predefined theorem provers to research agents capable of addressing frontier mathematical challenges. It systematically reviews current approaches, including datasets, auto-formalization, and proof synthesis in Interactive Theorem Proving (ITP) languages, while identifying limitations in handling open-ended research problems. Key limitations include dataset constraints, relational structure deficiencies, and inadequate tool ecosystems, with a roadmap proposed for advancing AI4Math systems toward genuine research capabilities.
ai4mathinteractive theorem provingauto-formalizationproof synthesisresearch agents
Idiobionics: The Unification of Privacy and Intelligent Robotic Prostheses
The paper introduces idiobionics, a novel research domain investigating privacy risks in intelligent robotic prostheses. Through integration of wearable robotics literature and threat modeling, the authors identify adversarial attack vectors in AI-enabled bionic limbs (e.g., sensor data exploitation, control system manipulation). Preliminary analysis demonstrates vulnerabilities in current semiautonomous prosthetic designs. The work contributes a taxonomy of 12 open research questions spanning privacy-preserving control algorithms, secure sensor fusion, and human-robot co-adaptation security. Findings highlight the need for privacy-aware design in next-generation prosthetics with advanced sensing (EMG, IMU) and machine learning controllers.
idiobionicsrobotic prosthesesadversarial attackswearable roboticsprivacy-preserving control
Graph-Regularized Deep Learning for EEG-Based Emotion Recognition with Psychologically-Grounded Label Structure
The paper introduces a graph-regularized learning framework for EEG-based emotion recognition that incorporates psychological interdependencies between emotion classes. The method represents emotions as graph nodes with edges encoding dimensional proximity, applying three regularization strategies: Graph Label Smoothing, Commuting distance via Graph Laplacian, and Sliced Wasserstein Distance. Evaluated on SEED-IV and SEED-V datasets with AudioTransformer, Conformer, and DCGNN backbones, the framework achieves up to +5.42% accuracy improvement and reduces implausible misclassifications by 39%, demonstrating architecture-agnostic benefits.
graph-regularized learningeeg emotion recognitiondimensional emotion theorygraph laplaciansliced wasserstein distance
Principled Analysis of Deep Reinforcement Learning Evaluation and Design Paradigms
This paper provides a principled analysis of canonical evaluation and design paradigms in deep reinforcement learning (RL), challenging established conclusions. The authors introduce theoretical foundations for scaling laws in RL, demonstrating that asymptotic performance lacks a monotonic relationship between performance rankings and data regimes. Through large-scale experiments, they reveal that conventional RL research paradigms have led to incorrect conclusions. The analysis offers core insights into scaling, capacity, and complexity in deep RL, emphasizing the need for reevaluation of existing methodologies.
deep reinforcement learningscaling lawsasymptotic performancedata regimescanonical paradigms
EgoWAM: World Action Models Beyond Pixels with In-the-Wild Egocentric Human Data
EgoWAM introduces a World Action Model (WAM) framework for robot manipulation training using egocentric human data, addressing the entanglement of transferable and non-transferable factors in behavior cloning. The method evaluates three world prediction targets (Pixel, DINO, 3D motion flow) while fixing policy backbone, action head, and data mixture. Results show DINO improves out-of-distribution generalization by 4x, and 3D flow boosts in-domain performance by 20-30%, demonstrating superior scalability with in-the-wild human data compared to behavior cloning.
world action modelegocentric databehavior cloning3d motion flowrobot manipulation
Alignment Plausibility: A New Standard for Assuring AI in Healthcare
The authors propose 'alignment plausibility' as a regulatory construct for ensuring AI safety in healthcare, particularly for large language models (LLMs) used in mental health support. Drawing parallels to clinical practice assurance, they advocate for a three-level alignment framework: 1) explicit value specification based on clinical norms, 2) value-embedded training, and 3) deployment oversight akin to clinical supervision. Alignment plausibility serves as a structured demonstration that a system's values, training, and oversight mechanisms collectively ensure safe, positive outcomes and patient benefit, addressing both acute and long-term risks in LLM deployment.
alignment plausibilitylarge language modelsclinical supervisionvalue specificationdeployment oversight
Aligning Clinical Needs and AI Capabilities: A Survey on LLMs for Medical Reasoning
This survey contributes a dual-view framework aligning clinical reasoning requirements with computational capabilities of large language models (LLMs) in healthcare. The clinical perspective establishes a five-level competency scheme based on Miller's Pyramid, while the computational view maps reasoning patterns to medical tasks. The authors introduce a benchmark dataset spanning five levels of medical reasoning and evaluate 18 state-of-the-art models, finding specialist models excel in diagnosis while general models perform better in decision support and dialogue. The study highlights ongoing challenges including data limitations, hallucination, and grounding issues, proposing directions for developing safer and workflow-ready systems.
miller's pyramiddeductive reasoninginductive reasoningabductive reasoninghallucination
Adversarial Social Epistemology for Assemblies of Humans and Large Language Models
The paper introduces adversarial social epistemology (ASE) to analyze densely interactive communicative landscapes where public assertions are scaffolded by testimony, inference, institutional certification, and tacit trust. It identifies mechanisms by which agents distort information for private, reputational, rhetorical, or material gains, arguing that existing frameworks like epistemic bubbles and misinformation diffusion inadequately capture these phenomena. The authors propose a semantic framework enriched with inferentialist semantics to interpret assertions and outline machinery for auditing and redressing trust breaches in scaffolded communications.
adversarial social epistemologyscaffolded assertionsinferentialist semanticsepistemic networkstrust breaches
AI-integrated models for assessing agricultural resilience
The authors present an AI-integrated modeling framework for agricultural resilience assessment, combining economic (GTAP) and biophysical (APSIM) models to analyze supply chain disruptions. The system enables natural language querying of cross-disciplinary impacts through a unified interface, targeting policymakers and market participants. The approach bridges traditionally siloed modeling paradigms to evaluate coupled biophysical-economic system vulnerabilities.
agricultural resiliencesupply chain shocksgtap modelapsim modelnatural language interface
Behavior Foundations for Quadruped Robots: ABot-C0 Technical Report
ABot-C0 introduces a generalist motion-control system for quadruped robots, addressing data scarcity and cross-embodiment challenges through three foundations: a scalable multi-source motion-data pipeline, robust policy learning, and a unified deployment stack. The method combines conditional video-generation synthesis, annotated motion capture, teleoperation, and human design to create 16,074 motion clips, enabling a Flow-Matching policy that exhibits scaling laws for motion tracking. A three-stage privileged-to-perceptive framework enhances all-terrain locomotion with LiDAR memory and terrain-predictive supervision. Experiments show successful urban-terrain navigation and multimodal interaction, advancing quadruped robots toward product-level behavioral intelligence.
quadruped robotsmotion trackingflow-matchingprivileged-to-perceptivelidar memory
A Transdiagnostic Space of Disorder Like Phenotypes in Reinforcement Learning Agents
The study introduces a computational framework for modeling psychological disorders in reinforcement learning agents through dose-controllable manipulation of cognitive appraisal signals. Using an appraisal-guided PPO agent, seven disorders (anxiety, mania, OCD, depression, impulsivity, addiction, PTSD) were expressed as single knobs grounded in computational psychiatry, with symptoms measured via preregistered assays. Results from over 1,000 runs showed graded, monotone dose-responses for all disorders, self-organization into a 2D affective space, and nonadditive interactions in comorbidity. The framework transferred to a 3D pixel environment (MiniWorld) with a standard convolutional agent, demonstrating cross-assay dissociation.
reinforcement learningcomputational psychiatrycognitive appraisalppo agentaffective phenotypes
Score Accuracy Along the Forward Diffusion Does Not Certify Numerical Stability in Diffusion Sampling
The paper demonstrates that small forward-marginal error in score matching does not ensure numerical stability in diffusion sampling. By constructing a smooth score field with arbitrarily small forward-marginal $L^2$ error, the authors show that Euler--Maruyama discretizations can diverge in Wasserstein distances $W_p$ despite weak convergence. They identify failure cases within fixed neural architectures and propose a positive result for compactly supported data: projecting the denoiser onto a bounded convex set preserves accuracy and ensures Wasserstein convergence. Experiments with a DiT-style network confirm rare trajectory divergence and its suppression via denoiser projection.
score matchingeuler--maruyama discretizationwasserstein distancedenoiser projectiondiffusion sampling
MulTTiPop: A Multitrack Transcription Dataset for Pop Music
The authors introduce MulTTiPop, a novel dataset of 572 pop music segments (3.5 hours) with aligned multitrack MIDI for evaluating automatic music transcription systems. The dataset was constructed by metadata-matching segments from Lakh MIDI and TheoryTab, manually anchoring beats, and warping MIDI to audio tempo using beat tracking. Evaluation of state-of-the-art transcription models reveals significant performance gaps, with the best model achieving only 38% Onset F1. The dataset spans diverse genres and decades (1930s-2000s), addressing a need for standardized pop music transcription benchmarks.
automatic music transcriptionmultitrack midibeat trackingonset detectionmusic information retrieval
ARDY: Autoregressive Diffusion with Hybrid Representation for Interactive Human Motion Generation
ARDY introduces a real-time 3D human motion generation framework that combines autoregressive transformers with diffusion models for interactive control via text prompts and kinematic constraints. The method employs a hybrid representation (explicit root features + latent body embeddings) and a two-stage denoiser with variable history context, enabling long-horizon goal conditioning. Evaluations on HumanML3D and Bones Rigplay datasets show superior motion quality and constraint adherence compared to existing online methods, with demonstrated applications in interactive control scenarios.
autoregressive diffusionhybrid representationkinematic constraintsinteractive generationmotion synthesis
Super Weights in LLMs and the Failure of Selective Training
The study challenges the assumption that Super Weights—individual parameters whose removal severely degrades performance—are universally critical or trainable in LLMs. Through experiments on OLMo-1B and OLMo-7B, the authors demonstrate that selectively training Super Weights (100–8,192 parameters) or their local neighborhoods (up to 36K parameters) collapses accuracy to random-guessing levels, while training randomly chosen parameters in the same layers improves performance. Low-rank updates (e.g., LoRA) succeed with only 0.16% of parameters, showing that effective fine-tuning requires structured layer-wise updates rather than targeting individual important weights.
super weightsllmsparameter pruninglow-rank adaptationselective training
Latent Memory Palace: Reasoning for Control as Autoregressive Variational Inference
The paper introduces Latent Memory Palace (LMP), a method that enables reasoning in continuous control policies by organizing information in an autoregressive latent space, akin to a memory palace. LMP formulates reasoning as variational inference with an autoregressive latent distribution and derives a latent-space reinforcement learning technique to optimize its variational lower bound. The resulting policy, LMP-$π$, demonstrates strong performance in simulation and real-world domains, exhibiting adaptive allocation of test-time compute. Additionally, LMP-$ exttt{tok}$, a variable-length action tokenizer derived from the framework, enhances downstream autoregressive policies. This work presents a novel perspective on latent reasoning for control through variational inference.
autoregressive latent spacevariational inferencereinforcement learningmemory palaceaction tokenizer
Deep Learning for Joint Narrowband Interference Cancellation and Soft Demodulation in OFDM Systems
A deep learning framework is proposed for joint narrowband interference (NBI) cancellation and robust soft demodulation in OFDM systems, addressing limitations of conventional compressed-sensing methods. The framework comprises NBI-CNet, a physics-informed convolutional network for multi-tone interference removal, and LLR-CNet, a structural whitener for mapping non-Gaussian residuals to calibrated soft metrics. Evaluations show the framework reduces computational complexity by up to 60% compared to EOMP-IDS, operates within 0.2-0.5 dB SNR of optimal baselines at BLER 10^-4 under severe interference, and achieves over 3 dB coding gain under mild interference with spectral overlap. The architecture generalizes across arbitrary FFT sizes without retraining and eliminates error floors caused by interferer-estimation errors.
narrowband interferenceofdm systemssoft demodulationconvolutional networkstructural whitener
LTM: Large-scale Terrain Model for Wildfire-prone Landscapes
The authors propose LTM, a multi-modal 3D terrain reconstruction framework for wildfire-prone landscapes that leverages outdated Digital Elevation Models (DEMs) as geometric priors. The method introduces physics-based pixel-pixel alignment between images and DEM data, eliminating computationally expensive feature matching while maintaining accuracy. Validation uses a large-terrain simulator based on real wildfire-prone areas, demonstrating significant improvements in reconstruction fidelity and real-time performance compared to conventional techniques. The approach addresses key limitations of airborne LiDAR (cost) and image-based methods (sparse features) for emergency response applications.
3d terrain reconstructiondigital elevation modelsphysics-based alignmentwildfire hazard assessmentmulti-modal fusion
MPFlow: Learning Budgeted Max-Flow Optimization on the Lightning Network with Deep Graph Reinforcement Learning
The paper introduces MPFlow, a deep graph reinforcement learning method for optimizing liquidity placement in Bitcoin's Lightning Network (LN) under budget constraints. The approach formulates channel selection as a combinatorial max-flow problem, solved via a message-passing policy network trained with proximal policy optimization (PPO) and action masking. A hub-exclusion curriculum forces capacity-aware learning. Evaluations on real LN snapshots show consistent outperformance over baselines in max-flow optimization, with production deployment handling 4640 channel-open decisions allocating 267.3 BTC across 30 nodes.
lightning networkmax-flow optimizationgraph reinforcement learningproximal policy optimizationmessage-passing networks
Resample or Reroute? Budget-Aware Test-Time Model Selection for Large Language Models
The paper introduces budget-aware test-time model selection for LLMs, proposing a resample-or-reroute (RoR) policy that optimally allocates per-query budget between resampling and rerouting to maximize expected correctness. The method leverages estimated marginal correctness per unit cost, exploiting recoverability asymmetry between selection and sampling. Evaluations on an eleven-model pool across four benchmarks demonstrate RoR's superior cost-quality Pareto front compared to baselines, with notable gains on heterogeneous tasks; results are verifier-dependent and robust under varied pricing and label-free conditions.
large language modelsmodel selectionbudget-awaretest-timeverifier
EdgeRefine: Privacy-Utility Balance for Graphs via Jaccard Sampling under Edge Differential Privacy
EdgeRefine proposes a local differential privacy framework for graph-structured data that optimizes the privacy-utility trade-off via adaptive edge refinement. The method estimates edge-existence probabilities using Jaccard similarity, ranks edges for noisy removal, and separately samples true/false edges based on a privacy budget $ε$ and sampling rate $k$. Experiments demonstrate that EdgeRefine achieves noise-free baseline accuracy, outperforming state-of-the-art methods by up to 19.7% in node classification under $ε=2.5$, with only 5% accuracy degradation in graph classification and strong resilience against reconstruction attacks (relative absolute error >1).
edge differential privacyjaccard samplinggraph neural networksprivacy-utility trade-offadaptive edge refinement
Secure Decentralized Federated Learning via Gossip and Virtual Voting
The paper introduces gspDAG-FL, a secure decentralized federated learning framework that combines gossip-based model dissemination with Hashgraph-style virtual voting for consensus. Nodes exchange model payloads locally, while full nodes reconstruct a compact Topology DAG and validate payloads through accepted-proof validation and private semantic audit. The method ensures finality over unique model-origin tuples rather than identical local parameter states. Theoretical analysis proves safety, conditional liveness, and convergence under time-varying effective mixing. Experiments on MNIST classification and Penn Treebank language modeling with networks up to N=100 demonstrate learning quality comparable to ledger-based FL, reduced coordination bottlenecks, improved throughput, and high invalid-origin detection under Byzantine and lazy participation.
decentralized federated learninggossip protocolhashgraphtopology dagbyzantine resilience
BiSCo-LLM: Lookup-Free Binary Spherical Coding for Extreme Low-Bit Large Language Model Compression
BiSCo-LLM introduces a codebook-free binary spherical coding framework for extreme low-bit compression of large language models (LLMs), addressing memory and bandwidth constraints. The method maps local weight chunks onto a unit hypersphere, binarizes them into compact spherical codes, and employs a residual BSQ stage to encode reconstruction errors. It incorporates category-wise recovery distillation and an 8-bit protected-channel path for stabilization. The framework eliminates explicit codebooks and index lookup, reducing storage overhead while maintaining model performance. Reported storage budgets include binary codes, neural decoders, protected-channel payloads, LoRA adapters, and metadata.
binary spherical codinglow-bit compressionunit hypersphereresidual bsqrecovery distillation
Steering Neural Network Training through Interpretable Constraints Based on Partial Dependence
The authors propose a novel method for steering neural network training via interpretable constraints derived from partial dependence, aligning model behavior with domain knowledge. Their approach modifies standard training by incorporating functional constraints based on feature interactions, applicable particularly to regression tasks like dynamical systems forecasting. Empirical results demonstrate improved performance and data efficiency compared to unconstrained models, with generated explanations better matching prior domain knowledge.
explanation-guided learningpartial dependenceinterpretable constraintsdynamical systems forecastingdata-efficient training
Federated Deep Learning for Privacy-Preserving Cardiovascular Disease Risk Prediction
A federated deep learning approach is proposed for privacy-preserving cardiovascular disease risk prediction across heterogeneous cohorts, addressing limitations of single-institution models and data-sharing constraints. The method integrates two population-based cohorts—Lifelines (n=148,230) and Rotterdam Study (n=10,155)—with differing characteristics and outcome definitions, employing deep survival models trained via federated learning. Evaluation on the Rotterdam Study demonstrated improved predictive performance, with the C-statistic increasing from 0.728 (95% CI: 0.717-0.739) to 0.739 (95% CI: 0.728-0.749). Lifelines also showed improvement, from 0.783 (95% CI: 0.775-0.791) to 0.787 (95% CI: 0.780-0.792).
federated learningdeep survival modelscardiovascular disease risk predictionc-statisticprivacy-preserving
Robust Bayesian Decision Making under Adversarial Uncertainty
The paper introduces a robust Bayesian experimental design framework for decision-making under adversarial uncertainty, addressing limitations of conventional methods that assume well-specified models. The authors formalize adversarially robust optimal decisions by incorporating worst-case variations in adversarial variables, deriving a Bayesian experimental design criterion that prioritizes decision stability over nominal optimality. Experiments on synthetic and real-world datasets demonstrate that traditional decision-aware designs often yield fragile high-confidence decisions, while the proposed robustness-aware approach achieves significantly improved stability under adversarial perturbations.
bayesian decision theoryadversarial robustnessexperimental designdecision stabilitysequential learning
Spectral Stability of Pseudoinverse-Based Extreme Learning Machine
The paper analyzes the spectral stability of pseudoinverse-based Extreme Learning Machine (ELM), demonstrating that perturbation amplification in output weights is governed by the smallest singular value while the condition number quantifies hidden-layer instability. It compares SVD-based pseudoinverse computation with iterative hyperpower methods, investigating width-dependent conditioning through a random feature interpretation. Experiments on synthetic matrices and ELM benchmarks reveal SVD-based methods as most reliable under ill conditioning, with iterative methods showing greater sensitivity to spectral properties, indicating ELM stability is fundamentally tied to the hidden layer matrix's singular value structure.
extreme learning machinemoore-penrose pseudoinversespectral stabilitysingular value decompositionrandom features
ImputeViz: A Visual Analytics Dashboard for Diagnosing Missing Data and Comparing Imputation Methods
ImputeViz introduces a visual analytics dashboard for diagnosing missing data patterns and comparing imputation methods, integrating MICE, Random Forest, XGBoost, kNN, and a novel geographically informed gKNN variant. The system provides coordinated views for missingness diagnosis (MCAR/MAR/MNAR), model configuration, and result evaluation via heatmaps, distributional overlays, and performance metrics (MAE, RMSE, Delta RMSE). Case studies demonstrate its utility in identifying sensitive variables, assessing model robustness, and facilitating method selection through interactive comparison of imputation outcomes.
visual analyticsmissing dataimputation methodsgeospatial reasoningmethod comparison
Contravariance Theory: Strong Alignment for Minimal Solutions to Hard Tasks
The paper introduces contravariance theory, demonstrating that for any two minimal deep neural network (DNN) solutions to sufficiently hard tasks, weak alignment via affine mappings guarantees strong alignment of privileged axes. The method leverages end-to-end task optimization to show alignment propagates hierarchically, formalizing concepts from prior work. Results imply that for hard tasks, inter-network comparison metrics are less sensitive, and convergent evolution between artificial and biological networks becomes likely.
contravariance theorydeep neural networksprivileged axesaffine mappingsneuroai
CAAD: Causality-Aware Multivariate Time Series Anomaly Detection via Multi-Scale Alignment and Structural Causal Consistency
The paper introduces CAAD, a causality-aware framework for multivariate time series anomaly detection that verifies Granger causality consistency through exogenous variables. The method models exogenous variables as residuals, employs multi-scale alignment to capture system dynamics, and uses a gradient-based matrix to detect causal relationship breakdowns. Experiments on industrial datasets show CAAD outperforms state-of-the-art baselines in anomaly detection precision.
granger causalityexogenous variablesmulti-scale alignmentanomaly detectionstructural causal consistency
Structural Bottlenecks on Frequency Representation in End-to-End Audio Models
The study identifies structural bottlenecks in strided convolutional encoders for end-to-end audio models, showing they collapse frequency-localized primitives into alias equivalence classes (31-35% collapse rate) and limit filter resolution (10-35x above theoretical bounds). Through theoretical analysis and experiments, the authors propose Gabor Latent Refactorization (GLRF), a post-hoc intervention that re-expresses latents in a frequency-localized basis, reducing filter bandwidths to 1.5-3x of theoretical bounds while preserving reconstruction fidelity and improving pitch control. Results demonstrate degraded feature access in existing encoders and GLRF's effectiveness in recovering interpretability and steerability.
strided convolutional encodersfrequency-localized primitivesalias equivalence classesgabor latent refactorizationreconstruction fidelity
High-Dimensional Procrustes Matching via Tree Counts
The authors present a polynomial-time algorithm for exact recovery in the high-dimensional Procrustes matching problem, where two sets of n Gaussian vectors in ℝ^d are ρ-correlated after permutation and rotation. The method computes and compares weighted counts of a family of wide trees, succeeding with high probability when d ≥ polylog(n) and ρ^2 > √α, where α ≈ 0.338 is Otter's tree-counting constant. They also provide an improved information-theoretic guarantee, showing exact recovery is possible when ρ^2 ≳ max{log n/d, √(log n/n)}, and suggest ρ^2 > √α is necessary for tree-counting algorithms.
procrustes matchinggaussian vectorstree-countinghigh-dimensionalpolynomial-time algorithm
Stop Guessing When to Stop Testing: Efficient Model Evaluation with Just Enough Data
The authors propose an adaptive evaluation framework leveraging sequential testing to optimize the efficiency-reliability trade-off in model evaluation. The framework integrates established sequential testing paradigms with tailored stopping criteria, addressing diverse objectives such as diminishing returns detection and minimum detectable effect size. Applied to the Open VLM Leaderboard, the method achieves an 80% reduction in computational cost compared to fixed-size evaluation while maintaining statistical significance, with a 2.5-point confidence interval width allowance. This approach mitigates the inefficiencies of rigid fixed-size benchmarks in model ranking, selection, and development testing.
sequential testingadaptive evaluationstatistical powerdiminishing returnsconfidence interval
Systematic Evaluation of Learning Rate Scheduling Strategies Across Heterogeneous Architectures
This work systematically evaluates learning rate scheduling strategies across diverse neural architectures, demonstrating their critical impact on classification accuracy. The study employs automated source-code injection to apply 25 scheduler configurations across 30 representative convolutional and transformer architectures from the LEMUR dataset, evaluating 3,938 model variants on CIFAR-10. Results reveal architecture-dependent scheduler performance, with CosineAnnealingWarmRestarts and CyclicLR consistently outperforming basic decay strategies. The best configuration achieved 86.45% top-1 accuracy, with 237 variants exceeding 80%. The accuracy landscape is contributed to the LEMUR nn-dataset as a practical reference for scheduler selection.
learning rate schedulingconvolutional architecturestransformer architecturesautomated source-code injectionclassification accuracy
Ensemble Diversity Optimization for Subjective Supervision
Ensemble Diversity Optimization (EDO) is introduced as a differentiable framework for optimizing ensemble composition, cardinality, and calibration in subjective NLP tasks. The method employs Gumbel-Softmax relaxation for end-to-end ensemble learning, a signed diversity regularizer to control disagreement preservation, and integrates soft F1 and class-weighted cross-entropy for imbalance handling. Experiments on ArMIS, ConvAbuse, HS-Brexit, and MD-Agreement show EDO improves calibration (40-78% cross-entropy reduction) while maintaining competitive F1 and better alignment with annotator distributions compared to baselines.
ensemble diversitygumbel-softmaxsigned regularizationprobabilistic calibrationsubjective nlp
Frequency-Domain Multi-Modality Transportation Modeling
The paper proposes Frequency-Domain Multi-Modality modeling (FreMo), a lightweight framework for multi-modality transportation forecasting that operates in the frequency domain to address spectral heterogeneity and uneven cross-modality interactions. FreMo decomposes the task into modality-wise spectral refinement via Modality-Wise Frequency Filter (MFF) and frequency-guided cross-modality synergy via Frequency-Guided Synergy Integrator (FSI), enabling adaptive noise suppression and selective knowledge sharing. Experiments on real-world datasets demonstrate FreMo's consistent superiority over state-of-the-art baselines in performance and generalization across diverse forecasting scenarios.
multi-modality transportationfrequency-domain modelingspectral refinementcross-modality synergytime series forecasting
MatBind: A Shared Embedding Space for Multimodal Materials Characterization
MatBind introduces a contrastive learning framework that aligns four materials science modalities—crystal structure, simulated powder X-ray diffraction (pXRD), density of states (DOS), and text—into a unified embedding space, using crystal structure as the physical anchor. The method enables zero-shot cross-modal retrieval without explicit pairwise training, demonstrating emergent alignment between modalities. Results show the embedding space organizes materials by physical properties without supervision, with improved retrieval performance when combining modalities at query time, validating the approach's physical consistency.
contrastive learningmultimodal embeddingmaterials sciencezero-shot retrievalcross-modal alignment
Predicting Viticulture Potential through an Ensemble of U-Net and a Geospatial Foundation Model
The DS@GT ARC team presents an ensemble method combining U-Net and Prithvi-2.0, a geospatial foundation model, for predicting viticulture potential in Southern France, addressing Subtask 1 of ImageCLEF AI4Agri 2026. The approach leverages remote sensing data to overcome the cost limitations of traditional agricultural assessment methods. Their model achieved a ±1 accuracy of 68.32%, securing 2nd place among 7 competing teams, with open-source implementation available.
u-netgeospatial foundation modelremote sensingviticulture potentialensemble learning
Statistical Efficiency and Inference of Quantile Distributional Reinforcement Learning
The paper analyzes quantile-based distributional reinforcement learning (RL) through the lens of statistical efficiency, focusing on distributional policy evaluation. The authors construct an estimator η_m^(n) using an empirical Markov decision process, deriving non-asymptotic error bounds under the W_∞ metric that scale as Õ(√(m/n)). They show optimal √n convergence rates and semiparametric efficiency, extending results to diverging quantile counts where estimators remain asymptotically efficient. A Berry-Esseen theorem for smooth functionals of the quantile-projected return distribution is also established.
quantile distributional rlstatistical efficiencydistributional bellman equationsemiparametric efficiencyberry-esseen theorem
FPGN: Redefining Ultra-Fast Programmable Gate-based Neural Acceleration with Differentiable LUTs
FPGN introduces an end-to-end framework for ultra-low-latency DNN inference on FPGAs by bridging LUT-native neural networks with hardware optimization. The method combines (i) a hardware-aligned differentiable LUT formulation, (ii) a structured LUT-native topology for improved routability, and (iii) a latency-driven compiler with analytical QoR models for automated DSE. Results demonstrate 205× latency reduction over FPGA-based BNN accelerators and 30× higher LUT efficiency than prior LUT-native networks while maintaining accuracy.
lut-native networksfpga accelerationnanosecond-scale inferencedifferentiable lutslatency-driven compiler
Beyond Backpropagation: Monte Carlo Method Can Train Deep Neural Networks
The paper demonstrates that a simple Monte Carlo method can effectively train deep neural networks without backpropagation, overcoming gradient-related issues like vanishing gradients. The approach randomly mutates parameters, retaining changes only if they reduce loss, and requires no auxiliary techniques like batch normalization. Experiments show successful training of networks with over 20 layers, wide single-hidden-layer networks (16,384 neurons), and a Transformer on MNIST and Tiny Shakespeare tasks. The method also supports discrete weights, Gaussian transfer functions, and pure pruning training, highlighting network redundancy.
monte carlo methodgradient-free trainingdeep neural networksdiscrete weightstransfer functions
Prompt Compression via Activation Aggregation
The paper introduces a method for compressing instruction prompts into single activation vectors via learned weighted aggregation of intermediate-layer activations in large language models (LLMs). This compressed representation, when reinjected at early layers, preserves task-relevant information while reducing computational overhead, achieving within 2% accuracy of full prompt processing. Key findings include cross-layer compatibility in activation representations, quantifiable semantic encoding in single vectors, and the efficacy of weighted sums as robust compressors.
activation aggregationprompt compressionlarge language modelscross-layer compatibilitysemantic encoding
Joint Discrete-Continuous Flow Matching for Open-Vocabulary Inverse Design of Multilayer Optical Coatings
The authors present IrisFlow, a joint discrete-continuous flow matching framework for open-vocabulary inverse design of multilayer optical coatings. The method combines discrete flow matching for material selection with continuous flow matching for thickness optimization, operating on wavelength-aware optical tokens rather than fixed vocabularies. A single 136M-parameter model handles 2-100 layer stacks, achieving faithful reconstruction on in-distribution targets (224-task benchmark) and generalizing to held-out material banks. Experimental validation includes fabricated color-displaying coolers with CIEDE2000 errors of 3.1-5.2 and 93-95% solar reflectance, demonstrating end-to-end open-vocabulary design.
flow matchinginverse designoptical coatingsopen-vocabularywavelength-aware tokens
On Exploring Input Resolution Scaling For Anytime LiDAR Object Detection
The paper proposes a memory-efficient method for anytime LiDAR object detection by enabling dynamic input resolution scaling with a single DNN model, eliminating the need for multiple resolution-specific models. The approach processes point clouds as pillars or voxels and includes a deadline-aware scheduler that predicts execution times across resolutions to select the optimal input scaling. Evaluated on nuScenes, it outperforms existing anytime methods and demonstrates collision-free navigation in simulated autonomous driving by avoiding unnecessary stalls.
lidar object detectionanytime computinginput resolution scalingdeadline-aware schedulerpoint cloud processing
Revisiting One-Zero and Two-Zero Neutrino Mass Textures in Light of Recent Oscillation and Cosmological Data
The study re-examines one-zero and two-zero neutrino mass matrix textures under current experimental constraints, including neutrino oscillation parameters, cosmological bounds on neutrino mass sum, and neutrinoless double-beta decay limits. For two-zero textures, analysis reveals that only $A$-series textures remain viable under CMB+BAO constraints, while $B$-series textures predict specific Dirac CP phases ($δ_{\rm CP} ≈ π/2, 3π/2$). Machine learning techniques, particularly flow matching, are employed to analyze one-zero textures, identifying excluded structures and distinct predictions for $\sum_i m_i$, $m_{ν_e}^{\rm eff}$, $\langle m_{ee}\rangle$, and $δ_{\rm CP}$. The work also explores non-invertible selection rules as potential origins for one-zero textures.
neutrino mass texturescmb+bao constraintsdirac cp phaseflow matchingnon-invertible selection rules
Dynamics of Gradient Descent with Large Step Size Near a Manifold of Flat Minima
This paper extends gradient descent (GD) theory to overparametrized least-squares with vector-valued outputs and manifolds of flat minima, generalizing prior work on isolated minima. The authors develop a normal form for large-step GD near such manifolds and solve a singular PDE via a novel method. They prove three convergence theorems and apply the framework to deep matrix factorization, showing flat minima form a fiber bundle over a product of spheres with Morse-Bott sharpness. Key technical challenges include handling vector outputs and manifold structures.
gradient descentsharpnessflat minimamatrix factorisationmorse-bott
Eigenvalue Calibration for Semantic Embeddings of Large Language Models
The paper introduces a framework for calibrating eigenvalues of semantic embeddings in large language models (LLMs), addressing a gap in uncertainty quantification. By interpreting LLMs with semantic embeddings as density matrix predictors, the authors propose temperature scaling for eigenvalue calibration, establishing entropy-risk equivalence and deriving a calibration inequality. Experiments demonstrate systematic overconfidence in current LLMs and validate the proposed method's effectiveness in optimizing calibration through proper score risks.
eigenvalue calibrationsemantic embeddingsdensity matrix predictorstemperature scalinguncertainty quantification
Tubular Neighbourhoods of Pfaffian Sets and Applications to Neural Networks
The paper establishes volume bounds for tubular neighborhoods of smooth Pfaffian hypersurfaces, extending prior work on algebraic varieties. The analysis employs Pfaffian format properties of defining functions. Key applications include tail probability bounds for a neural network condition number under uniform and Gaussian distributions, quantifying classifier robustness. For single-hidden-layer sigmoid networks with rational weights, polynomial-in-width bounds are derived for decision boundary neighborhoods.
pfaffian setstubular neighborhoodsneural networkscondition numberdecision boundary
Certified Interventional Fidelity: Anytime-Valid, Adaptive Evaluation of Causal Claims in Mechanistic Interpretability
The authors introduce Certified Interventional Fidelity (CIF), a statistical framework for evaluating causal claims in mechanistic interpretability. CIF formalizes interpretability evaluations as causal estimands—expectations of bounded scores over specified input and intervention distributions—and provides anytime-valid confidence intervals and sequences, even under adaptive intervention sampling via bounded mixture importance weighting. The framework employs Hoeffding-style sequences and variance-adaptive betting sequences, reducing certification costs by 10-30x in experiments. Evaluations on MNIST abstractions and GPT-2 Small IOI circuits demonstrate CIF's ability to certify high-fidelity claims, identify statistically unsupported method differences, and quantify sensitivity to intervention distributions.
mechanistic interpretabilitycausal estimandconfidence sequencesadaptive samplingvariance-adaptive
Prediction-Powered Active Testing
The paper introduces Prediction-Powered Active Testing (PPAT), a label-efficient risk estimation framework that combines the unbiased LURE estimator with a prediction-powered control variate. PPAT leverages black-box model predictions to residualize the loss, maintaining unbiasedness while reducing variance. It also proposes oracle and surrogate-based acquisition rules to minimize estimator variance and establishes asymptotic normality for valid confidence intervals. Evaluated on tabular regression and image-classification tasks, PPAT outperforms existing methods in risk estimation, achieving target coverage with fewer labels and narrower confidence intervals.
active testingrisk estimationcontrol variateasymptotic normalitylabel-efficient
AutoAnchor: Stable Diffusion Unlearning Using Cross-Attention as a Manifold Surrogate
AutoAnchor proposes a stable diffusion unlearning framework addressing biased or unrobust concept removal in text-to-image models. The method introduces a two-stage approach using automatically synthesized manifold-proximal anchors, with a cross-attention consistency loss as a computationally efficient surrogate for manifold proximity. Experiments show improvements of up to 31.04% in targeted concept removal (CLIP score) and 4.18% in non-target utility, while enhancing existing methods by 6.30% and 6.65% on average respectively.
diffusion unlearningmanifold hypothesiscross-attention consistencytext-to-image modelsclip score
Bayesian Experimental Design via Score Matching
The paper introduces a method to decouple the double intractability of expected information gain (EIG) from policy learning in Bayesian experimental design (BED). By first solving a score matching problem independent of the policy, the approach enables singly intractable policy training, converting multiplicative computational costs into additive ones. This reduces the burden of policy optimization, facilitating multiple training runs for architecture search or hyperparameter tuning. Experiments demonstrate competitive policy performance without multiplicative likelihood evaluation costs, enabling better policy selection.
bayesian experimental designscore matchingexpected information gainpolicy learningdouble intractability
Write-Protected Discrete Bottlenecks for Language-Grounded World Models: A Structural Limitation and Sufficient Fix
The paper identifies a structural limitation in language-grounded world models where language gradients collapse discrete symbol bottlenecks, and proposes a three-layer write-protected blackboard architecture as a sufficient fix. The solution combines (1) gradient-isolated discrete bottlenecks, (2) co-occurrence counting for semantic binding, and (3) DP-Means clustering for symbol collision handling. Experiments across 74 runs show 97.2% grounding accuracy versus 22.2% baseline, with zero symbol collapse in all 32 seeds, while maintaining parameter efficiency (<2M params) and architecture agnosticism (tested with CNN, V-JEPA, CLIP).
discrete bottleneckslanguage groundinggumbel-softmaxdp-means clusteringsymbol collapse
Learning $\mathsf{AC}^0$ under Locally Sampleable Graphical Models
(No summary returned.)
Classifier Chain-based Pathological Test Recommendation
A pathological test recommendation system is proposed to accelerate test selection using patient symptoms prior to physician consultation, framed as a multi-label classification problem. The Classifier Chain (CC) technique is employed to model dependencies between tests, with Logistic Regression, Decision Tree, Random Forest, and Majority Voting ensemble evaluated on a custom dataset from SOUTHERN.IML pathology. Logistic Regression with CC achieved 98.83% accuracy, while Majority Voting balanced precision (0.93), recall (0.85), and F1-score (0.89). SHAP-based Explainable AI (XAI) ensured clinical interpretability, revealing symptom contributions consistent with medical knowledge, enhancing model reliability for diagnostic decision-making.
classifier chainmulti-label classificationexplainable aishappathological test recommendation
CASL-VAE: Learning Structured Latent Variables from Unpaired Data for Semi-supervised Clustering and Paired Sample Generation
CASL-VAE introduces a contrastive latent variable model for learning structured generative factors from unpaired data, addressing heterogeneous target variation without requiring paired samples. The method factorizes variation into continuous shared factors and hierarchical salient factors (discrete subtypes + continuous within-subtype variation) via variational inference, enabling joint likelihood optimization across domains. Evaluated on semi-synthetic neuroimaging data, it outperforms baselines in subtype recovery (F1=0.82 vs 0.68) and paired-sample generation (SSIM=0.91 vs 0.85), while revealing plausible Alzheimer's disease heterogeneity.
contrastive learninglatent variable modelvariational inferenceunpaired dataheterogeneous variation
An interpretable Good--Turing restart criterion for k-means++
We propose GTRC, an interpretable Good-Turing restart criterion for k-means++ that dynamically determines the number of restarts based on data set difficulty. GTRC combines a Good-Turing estimate, an unconditional bound, and a confidence-based bound to estimate the probability of improvement from further restarts, stopping when this probability falls below a user-specified tolerance ε. Evaluated on 36 data sets, GTRC achieves clustering quality comparable to fixed restart counts while adaptively varying the number of restarts according to data complexity. This approach provides a principled, data-driven alternative to arbitrary fixed restart counts, enhancing reproducibility and computational efficiency.
k-means++good-turing estimaterestart criterionclustering qualitydata-dependent
Closing the Null Space: Guidance-Aware Quantization for Classifier-Free Diffusion
The paper introduces Guidance-Aware Mixed Precision (GAMP), a quantization method addressing the branch-drift trap in classifier-free guidance (CFG) diffusion models. GAMP calibrates directly on guided predictions, derives per-layer activation-bit sensitivity from guided-output degradation, and allocates bits via a greedy knapsack, preventing unconditional branch drift by construction. Existing post-training quantization methods treat CFG models as single-branch networks, ignoring the paired conditional/unconditional structure, leading to latency overhead and degraded sample quality. Analytical proof and empirical results confirm the branch-drift trap, demonstrating that standard diagnostics can yield false positives. GAMP mitigates these issues, ensuring efficient deployment under real-world compute budgets.
classifier-free guidancequantizationbranch-drift trapmixed precisiondiffusion models
Structure Learning on Clustered Data
The paper introduces a novel method for directed acyclic graph (DAG) structure learning in clustered data, addressing limitations of existing techniques that assume population homogeneity. The approach extends mixed-effects models to structure learning via a differentiable graph coupling mechanism ensuring acyclicity, with provably convergent first-order optimization and batched cluster updates. Theoretical guarantees include model identifiability and asymptotic structure recovery. Empirical results demonstrate improved dependency detection versus baselines on synthetic and real-world datasets.
directed acyclic graphstructure learningmixed-effects modelsdifferentiable couplingcausal discovery
Benchmark Evaluation of Feredated Learning on Multi-organ Images
The study introduces MobenFL, a comprehensive federated learning benchmark for medical imaging that addresses limitations of prior benchmarks by incorporating 20 state-of-the-art FL algorithms and 22 datasets spanning 12 organs. The benchmark evaluates performance, algorithmic efficiency, and privacy protection across diverse clinical scenarios involving multiple diseases, devices, and imaging modalities. Results demonstrate MobenFL's superior breadth and depth compared to existing benchmarks, providing a unified evaluation framework for FL in real-world medical applications.
federated learningmedical imagingbenchmark evaluationprivacy protectionalgorithmic efficiency
PIT-SUN: A Deployable Empirical Marginal Transform Framework with Expectation-Consistent Recovery for Regression in Recommender Systems
The paper introduces PIT-SUN, a deployable framework for expectation-consistent regression in recommender systems that handles heavy-tailed, zero-inflated, and multimodal targets. The method combines an empirical marginal transform with multiplicative SUN recovery to estimate original-space expectations without direct inversion, addressing instability in standard MSE gradients. Evaluations on synthetic data, public benchmarks, and industrial datasets demonstrate improved accuracy, calibration, and ranking with low deployment overhead.
expectation-consistent recoverymarginal transformrecommender systemsheavy-tailed targetsmultiplicative sun recovery
MLQENABLER: Enabling Secure Machine Learning Queries over Encrypted Database in Cloud Computing
The paper proposes MLQENABLER, a scheme enabling secure machine learning queries over encrypted databases in cloud computing. The method employs an index-aid approach to maintain both security and ML functionality when processing encrypted client data stored on untrusted public clouds. Initial experiments demonstrate the scheme achieves acceptable security levels with minimal performance degradation for ML tasks.
encrypted databasecloud computingmachine learning queriesindex-aid approachsecurity-performance tradeoff
Understanding Layer Patching in Model Size Interpolation
This work presents the first systematic study of layer selection strategies for zero-shot model size interpolation via layer patching, where student model layers are replaced with teacher model blocks. The authors formulate optimal patching as a shortest-path problem and propose KLPatch, a greedy algorithm minimizing KL divergence between layer outputs. Experiments show patching direction significantly impacts interpolation quality, with sequential (first/last-to-first) strategies performing surprisingly well, while KLPatch often outperforms them. Results provide theoretical grounding and practical methods for constructing interpolated models between arbitrary pre-trained checkpoints.
zero-shot interpolationlayer patchingkl divergenceshortest-path problemboomerang distillation
MuScriptor: An Open Model for Multi-Instrument Music Transcription
MuScriptor introduces an open-weight model for multi-instrument music transcription, addressing limitations of existing methods that struggle with complex, real-world music mixes. The approach combines synthetic data pre-training with fine-tuning on real music audio and reinforcement learning-based post-training. Instrument presence conditioning is incorporated to enable customized transcriptions. The model demonstrates improved generalization across diverse musical genres, overcoming the poor performance of prior synthetic-data-trained models in realistic multi-instrument settings.
music transcriptionsynthetic datareinforcement learningfine-tuninginstrument conditioning
DeepPySR -- A Symbolic Regression Framework with Dynamic Pruning, Pareto Selection, and Hierarchical Composition for Real-World Scientific Discovery
DeepPySR introduces a symbolic regression framework addressing three key challenges: high-dimensional inputs via dynamic variable pruning, Pareto-front selection through an exponential criterion balancing accuracy and complexity, and hierarchical symbolic composition for multi-layer equation discovery. The method outperforms PySR and baselines on 11 benchmarks, achieving R$^2$ improvements of 0.092-0.155 on biomedical datasets (body fat, Raine BMI) and F1 gains of 0.111 on heart disease prediction, while producing interpretable formulas aligned with domain knowledge.
symbolic regressiondynamic pruningpareto selectionhierarchical compositioninterpretable models
Generalization Theory for Through-the-Wall Radar Human Activity Recognition
The paper proposes a generalization-analysis framework for through-the-wall radar (TWR) human activity recognition (HAR) to address performance degradation from structured distribution shifts. The method establishes unified models for human kinematics, TWR echo generation, and bounded-weight neural networks, deriving a target-domain generalization bound decomposed into cross-person, cross-view, and cross-wall components. Experiments demonstrate the framework's validity, highlighting the benefits of physical low-dimensional representations and multi-source training for improving generalization.
through-the-wall radarhuman activity recognitiongeneralization bounddistribution shiftsneural networks
Securing Autonomous Vehicle Systems via Twin-Aware Federated Reinforcement Learning
The paper proposes SecAggPP, a defensive framework for federated reinforcement learning (FRL) in autonomous vehicles, addressing poisoning attacks that compromise global control models. The method integrates digital twins for rehearsal-based learning and leverages historical aggregated parameters with a central gradient selection to filter malicious data. Theoretical convergence guarantees are provided, and experiments using digital twins in highway environments demonstrate effectiveness against adversarial conditions.
federated reinforcement learningpoisoning attacksdigital twinsautonomous vehiclessecure aggregation
TTHE: Test-Time Harness Evolution
We introduce Test-Time Harness Evolution (TTHE), a method for optimizing LLM agent harnesses during evaluation without gold labels or model weight updates. TTHE maintains a population of candidate harnesses and refines them through an agentic proposer that reasons over execution traces, with a judge committing improved harnesses based on execution-derived proxy signals. This approach adapts the executable harness rather than the underlying LLM, enabling persistent improvements across diverse tasks. Experiments on text-to-SQL, competitive programming, software engineering, data-science coding, and agentic tool-use tasks demonstrate TTHE's effectiveness over fixed ReAct-style baselines, highlighting execution-derived proxy reliability as a key challenge for unsupervised agent improvement.
test-time adaptationexecution tracesagentic proposerproxy signalsharness evolution
Workload-Preserving Differentially Private Synthetic Data for Causal Inference via Maximum-Entropy Calibration
The paper introduces causal workloads, a differentially private (DP) query framework tailored for preserving estimands in causal inference, particularly average treatment effects (ATE). The method combines DP-aggregated orthogonal moments with maximum-entropy calibration to produce reusable synthetic data, supported by theoretical error decomposition. Causal-AIM adaptively selects workloads, while noise-aware multiple-imputation (NA+MI) enables valid confidence intervals. Empirical results show causal workloads excel under strict privacy budgets and calibrated uncertainty, though generic workloads may outperform in point RMSE with relaxed privacy. The work highlights a tradeoff between distributional fidelity and causal moment preservation.
differentially private synthetic datacausal workloadsmaximum-entropy calibrationaverage treatment effectnoise-aware multiple-imputation
Contrastive Order Learning: A General Framework for Ordinal Regression
We introduce Contrastive Order Learning (ConOrd), a novel framework for ordinal regression that combines contrastive learning with order learning. ConOrd addresses limitations in both approaches by proposing a contrastive order loss with soft affinity and disparity weights based on rank differences, enabling fine-grained modeling of ordinal relationships across all sample pairs within a batch. Extensive experiments on facial age estimation, blind image quality assessment, and blind video quality assessment demonstrate that ConOrd consistently achieves state-of-the-art performance across diverse ordinal regression tasks. The source code is publicly available.
contrastive learningordinal regressionorder learningsoft affinityrank differences
BACH: A Bayesian Admixture of Contrastive Heads for Multi-Interest Two-Tower Retrieval
BACH introduces a Bayesian admixture of contrastive heads for multi-interest two-tower retrieval, addressing limitations of single-embedding and hard-routing approaches. The method employs variational inference to model per-user interest mixtures, enabling soft routing during training (mitigating collapse) and per-user interest weighting at serving. Evaluated on MovieLens-20M, Taobao, and Netflix, BACH outperforms single-vector and hard-routing baselines across all head counts, with further gains from best-head scoring and global-codebook variants.
multi-interest retrievaltwo-tower modelvariational inferencecontrastive learningadmixture model
Vanilla SGD with Momentum Survives Heavy-Tailed Noise: Convergence Analysis without Gradient Clipping or Normalization
The paper provides the first comprehensive convergence analysis of vanilla stochastic gradient descent (SGD) with momentum under heavy-tailed noise, addressing strongly convex, convex, and nonconvex objectives without gradient clipping or normalization. The authors refine existing convergence results for vanilla SGD and demonstrate that its convergence rates are inferior to those achieved by clipped or normalized SGD variants, highlighting inherent limitations of vanilla methods in such settings. Theoretical findings are validated through experiments on synthetic functions.
stochastic gradient descentheavy-tailed noiseconvergence analysismomentumstrongly convex
Stochastic Order Learning: An Approach to Rank Estimation Using Noisy Data
The paper introduces stochastic order learning (SOL), a framework for robust rank estimation with noisy ordinal labels by modeling them as stochastic orderings. SOL employs two key objectives: a discriminative loss structuring instance--centroid relationships and a stochastic order loss enforcing probabilistic ordering constraints. Evaluations across multiple datasets show SOL's effectiveness in handling diverse noise types and levels. The approach is available as open-source software.
rank estimationlabel noisestochastic orderingdiscriminative lossordinal labels
ConRad: Efficient Conformal Prediction for Radiomics
ConRad introduces an efficient conformal prediction framework for radiomic features derived from medical image segmentations, addressing overconfidence in segmentation models. The method leverages test-time covariates including image appearance, mask geometry, segmentation uncertainty, and predicted radiomics to construct adaptive prediction intervals while preserving coverage guarantees. Evaluated on five 2D medical imaging datasets with 171 radiomic targets, ConRad demonstrates improved interval efficiency over baselines while maintaining near-nominal empirical coverage, with segmentation boundary uncertainty features identified as the primary efficiency driver.
conformal predictionradiomic featuressegmentation uncertaintyprediction intervalsmedical imaging
Modular Pretraining Enables Access Control
We propose Gradient-Routed Auxiliary Modules (GRAM), a modular pretraining method that enables access control in AI systems by selectively updating auxiliary modules to induce capability specialization. GRAM approximates the effect of data filtering by ablating modules at inference time, disabling targeted dual-use capabilities while preserving others. Evaluations on synthetic stories and realistic dual-use datasets (virology, cybersecurity, nuclear physics, specialized code) demonstrate GRAM's effectiveness in resisting capability recovery under finetuning compared to post-hoc unlearning. Scaling analysis from 50M to 5B parameters shows GRAM closely tracks data filtering, with widening gaps for removed capabilities but small gaps for retained ones. GRAM achieves a 5x training cost reduction over data filtering in 5-profile settings.
gradient-routed auxiliary modulesaccess controldual-use capabilitiesdata filteringcapability specialization
Cross-Modal Generative Framework for Signal Translation from Fetal-Maternal Electrocardiograms to Fetal Doppler Waveforms
A cross-modal generative framework synthesizes fetal Doppler waveforms from fetal-maternal electrocardiograms (fECG/mECG), quantifying recoverable and residual Doppler components to isolate mechanical contributions to fetal hemodynamics. The model combines dilated convolutions, cross-modal attention for maternal-fetal coupling, and self-attention for temporal dependencies, trained on 885 synchronized ECG-Doppler segments from 39 pregnancies. It achieves a power spectral density mean squared error (PSD MSE) of 49.9 ± 15.8 dB² (51% lower than baseline) and heart-rate error of 4.71 ± 0.77 bpm (1.5% better than baseline). Cross-modal attention reduces PSD MSE by 39% over naive concatenation, demonstrating its efficacy in leveraging maternal-fetal coupling for improved Doppler synthesis.
cross-modal attentiondilated convolutionsfetal hemodynamicspower spectral densitymaternal-fetal coupling
Holographic Neural PCFG for Unsupervised Parsing
Holographic Neural PCFG (Hol-PCFG) introduces an interpretable, parameter-efficient approach to unsupervised constituency parsing by reformulating PCFG rule scoring as algebraic relation modeling. The method adapts Holographic Embeddings to score left-child, right-child, and lexical-emission relations using torus-constrained embeddings, providing closed-form rule probabilities that inherently capture grammar structure. Hol-PCFG achieves state-of-the-art performance across six languages, reduces rule-scoring parameters by 99.94% compared to Neural PCFG baselines, and demonstrates robust parsing of Japanese at the character level without morphological segmentation.
unsupervised parsingholographic embeddingsneural pcfgtorus-constrained embeddingsconstituency parsing
RadioDiff-v2: Generative Angular Radio Maps for Multi-Beam Selection and Localization
RadioDiff-v2 introduces a dual-branch one-dimensional diffusion transformer for generating angular radio maps in 6G networks, addressing the ill-posed mapping of angular power spectra in NLOS conditions. The model employs periodic angular encoding, adaptive layer-normalization conditioning, a Fourier angular mixer, and joint velocity and clean-signal heads, enabling deterministic transport through concentrated conditional trajectories. It achieves state-of-the-art performance on zero-shot tests across 99 environments and one million links, with a 0.39 dB Wasserstein-1 distance, sub-regression per-bin error, 2.43 dB eight-beam NLOS sweep loss, and 20.6-pixel localization error using four base stations.
diffusion transformerangular radio mapsnlos conditionswasserstein distancezero-shot test
An exact information theory of generalization phase transitions in Bayesian diffusion models
The paper introduces Bayesian information restricted diffusion (BIRD) models to analyze how diffusion models avoid the curse of dimensionality. BIRD models time-reverse diffusion by inferring training samples from restricted noisy observations using Bayesian posteriors, generalizing prior analytical models with spatial locality. The authors identify an information-theoretic phase boundary between memorization and generalization, showing that models memorize when mutual information exceeds log training data size. Experiments confirm the predicted transition, revealing that generation occurs near this boundary via progressive information restriction.
bayesian diffusioninformation restrictionphase transitionmemorization-generalizationcurse of dimensionality
What to Keep, What to Forget: A Rate--Distortion View of Memory Compaction in LLMs and Agents
The paper proposes a unified rate--distortion framework for memory compaction in LLMs and agents, viewing KV-cache eviction, prompt distillation, architectural state bounding, and agent memory consolidation as instances of a single optimization problem. It introduces a layer-agnostic objective function and a seven-axis taxonomy to classify methods across these domains, enabling mechanism transfer between layers. Key findings include the prevalence of attention magnitude/recency as retention signals and their shared failure mode of irreversible pre-query discarding. The work identifies benchmarking gaps for repeated compaction and proposes a multi-layer evaluation protocol.
memory compactionrate--distortionkv-cacheattention magnitudeagent memory
Rethinking Small VLM Quantization: From Component-Wise Analysis to Hardware-Aware Edge Deployment
The paper presents a systematic framework for analyzing component-wise quantization in small vision-language models (VLMs) under 3B parameters, focusing on edge deployment. It evaluates five hypotheses across six quantization configurations on Jetson Orin NX and AGX hardware, dissecting vision encoders, projectors, and LLM backbones. Key findings include: MoE backbones better tolerate INT4 noise than dense ones; SigLIP encoders show INT8 latency spikes on Jetson Ampere; INT4 LLMs reduce VRAM but slow token generation; quantization errors are additive except in modality alignment; and intelligence-per-joule varies by platform due to memory constraints.
vision-language modelsquantization sensitivitymodality alignmentjetson orinintelligence-per-joule
PGD-NO: A Neural Operator with Precomputed Geometry Decomposition for 3D Million-scale Physics Simulations
PGD-NO introduces a neural operator with Precomputed Geometry Decomposition to address memory constraints in large-scale physics simulations. The method relocates geometric encoding to a deterministic pre-computation phase using an iterative geometry decomposition algorithm, decoupling feature extraction from solution querying. This enables linear memory scalability, handling meshes with over 10 million nodes while maintaining competitive accuracy across industrial benchmarks and offering interpretability via attention mechanisms.
neural operatorgeometry decompositionphysics simulationsmemory scalabilityattention mechanisms
FedTR: Federated Learning Framework with Transfer Learning for Industrial Visual Inspection
FedTR introduces a federated learning framework with transfer learning for industrial visual inspection, addressing data scarcity and task complexity in label defect identification. The method leverages pre-trained models on public datasets, followed by federated fine-tuning on distributed private data. Extensive experiments on private ink cartridge datasets demonstrate FedTR's effectiveness, achieving 95.5% and 94.2% word-level accuracy on homogeneous and heterogeneous data, respectively, while matching centralized training performance.
federated learningtransfer learningindustrial visual inspectionlabel defect identificationend-to-end text recognition
Collate: Collaborative Neural Network Learning for Latency-Critical Edge Systems
Collate introduces a federated learning framework for latency-critical edge systems, addressing device heterogeneity by collaboratively training heterogeneous models under varying latency constraints. The method employs dynamic zeroizing-recovering for local model architecture adjustment and a proto-corrected federated aggregation scheme to maintain accuracy while meeting system-specific latency requirements. Experiments show Collate improves accuracy by 1.96% for extended models and 3.09% for shrunk models under latency constraints, with minimal training overhead.
federated learningedge systemslatency constraintsheterogeneous modelsdynamic zeroizing-recovering
Tool-Making and Self-Evolving LLM Agents in Low-Latency Systems
The paper introduces a tool-making pipeline for LLM agents that compiles frequently repeated procedural steps into validated, versioned tools to reduce inference-time latency and errors. The method involves execution trace collection, schema observation, candidate tool generation, and repair against labeled cases, enabling direct tool calls during runtime with fallback to code generation. Deployed in a fulfillment center alarm-triage system (44-node SOP), the approach reduces p50 latency by 42% and end-to-end error rate by up to 53%, while versioned tools improve auditability and detect specification gaps.
llm agentstool-making pipelineexecution tracesstructured verdictsversioned tools
Unit-Independent Low-Rate Wrist GSR Processing for Stress Detection Using Phasic nSCR Features
The paper proposes a unit-independent low-rate wrist GSR processing pipeline for stress detection, focusing on phasic nSCR features to address amplitude variability in wrist-based measurements. The method processes 25Hz wrist GSR signals through cleaning, tonic/phasic decomposition, z-score normalization, and SCR peak detection to compute nSCR/min. Evaluated on 31 participants during stress tasks (TSST) and baselines, random forest classification achieved balanced accuracies of 0.823 (vs sitting) and 0.871 (vs standing), demonstrating comparable performance to 100Hz features.
galvanic skin responsephasic scrtrier social stress testz-score normalizationrandom forest
Hallucination Self-Play: Bootstrapping Reinforced Detector via Evolved Generator
Hallucination Self-Play (HSP) introduces a framework for bootstrapping hallucination detection in LLM-generated outputs by iteratively evolving both detector and generator components. The method initializes detector and generator roles from the same base model, fine-tunes the detector on human-labeled data, and employs it as a reward model to train the generator via reinforcement learning from AI feedback (RLAIF). The evolved generator synthesizes challenging hallucination data to further optimize the detector through rule-based reinforcement learning. Experiments on the RAGTruth benchmark demonstrate that HSP enables small LLMs to match or outperform advanced LLMs without external supervision.
hallucination detectionreinforcement learningself-playrlaifragtruth
A Quantum Reservoir Architecture for Chaotic Forecasting and a Test of Whether Its High Dimension Helps
The paper presents a quantum reservoir computing architecture for chaotic system forecasting, providing a reproducible implementation recipe and a diagnostic to evaluate whether high-dimensional quantum features contribute meaningfully. The method employs a fixed quantum circuit as a feature generator with a linear readout, scaling both prediction problem size and reservoir capacity to isolate performance gains. Results demonstrate stable error rates on spatiotemporal and shallow-water chaotic systems as dimensions increase, outperforming a classical reservoir baseline while identifying scenarios where classical methods remain stronger.
quantum reservoir computingchaotic forecastingfeature generationlinear readoutstability diagnostic
Expressivity and Statistical Trade-offs in Diffusion Policy Learning
The paper analyzes the expressivity-statistical trade-off in diffusion-based reinforcement learning policies, identifying the drift Lipschitz budget $K$ as a key factor. Using terminal-law representations of diffusion processes, the authors prove $K$-Lipschitz drifts enable $1/K$-order value approximation near optimal policies, with a matching lower bound under nondegenerate noise. Neural network parameterizations reveal a trade-off: increasing $K$ improves approximation but raises statistical complexity, yielding finite-sample gaps of $\tilde{O}(n^{-2/(m+6)})$ (general case) or $\tilde{O}(n^{-2/(m+4)})$ (dissipative drifts), where $n$ is sample size and $m$ is state dimension. Experiments validate the $K$-dependent trade-off, suggesting practical $K$ selection based on sample size.
diffusion policieslipschitz budgetvalue approximationstatistical complexitydissipative drifts
KronQ: LLM Quantization via Kronecker-Factored Hessian
KronQ introduces a Kronecker-factored Hessian approximation for LLM quantization, incorporating gradient covariance alongside activation statistics to improve post-training quantization. The method features bidirectional incoherence processing, applying random rotations to both input and output dimensions based on gradient covariance, and a new sensitivity metric for mixed-precision allocation derived from Hessian traces. Evaluated on LLaMA-3-70B with 2-bit weight-only quantization, KronQ achieves 7.93 perplexity on WikiText-2, outperforming GPTQ and GPTAQ which diverge or exceed 2000 perplexity.
post-training quantizationkronecker-factored hessiangradient covariancemixed-precision allocationperplexity
Evaluating the Generalizability of Foundation Models for Extreme Environmental Events: Case Study of California Wildfire PM2.5
This study evaluates the generalizability of time series foundation models (TSFMs) for predicting extreme PM2.5 concentrations during California wildfires, challenging the assumption that pretrained models universally outperform task-specific architectures. Six TSFM configurations, including zero-shot TimesFM, Chronos-2, Moirai-2, and Time-MoE, plus LoRA fine-tuned variants, were benchmarked against trained baselines (LSTM, BiLSTM, Transformer) and persistence models using a 12-year hourly PM2.5 dataset covering 1,375 wildfire incidents. Evaluated via leave-one-incident-out protocol, BiLSTM achieved the lowest MAE (5.16 μg/m³) and highest exceedance F1 (0.63), outperforming all TSFMs. LoRA fine-tuning improved TSFM stability, but no foundation model surpassed recurrent baselines, highlighting limitations in extreme out-of-distribution scenarios.
time series foundation modelspm2.5 predictionlora fine-tuningleave-one-incident-outexceedance f1
DeepSWE: Measuring Frontier Coding Agents on Original, Long-Horizon Engineering Tasks
DeepSWE introduces a benchmark of 113 original, long-horizon software engineering tasks across 91 repositories and five languages, designed to evaluate coding agents without pretraining contamination. Unlike SWE-bench, it uses hand-written verifiers to assess functionality rather than inherited tests, reducing grading discrepancies (1.4% vs. 32.4%). Tasks involve 5.5x more code changes than SWE-Bench Pro and better differentiate frontier agents. The benchmark includes verifiers and evaluation trajectories.
benchmarksoftware engineeringverifierpretraining contaminationcoding agents
Distributed Sketching on Data Partitions for OLS Regression
The paper introduces a distributed sketching method for ordinary least squares (OLS) regression that operates on partitioned data subsets rather than the entire dataset, reducing computational costs. It constructs OLS estimators separately on each subset and averages them, analyzing the excess loss under a fixed design setting. Results demonstrate that the excess loss of the averaged estimator remains comparable to that of sketching on the full dataset when the divergence among subset covariances is minimal.
distributed sketchingordinary least squaresfixed designexcess losssubset covariances
Optimal Learning Rate Scaling Depends on Data in Deep Scalar Linear Networks
The paper demonstrates that optimal learning rate scaling in deep scalar linear networks (f(x) = ∏_{l=1}^L w_l x) is data-dependent, contrary to data-agnostic approaches that fail to generalize across depths. Using exact time-course solutions for gradient descent dynamics, the authors show that data-dependent scaling yields depth-invariant learning dynamics with a constant linear convergence rate, even at infinite depth. This finding extends to networks with residual connections, reinforcing the data-dependence of optimal scaling.
deep scalar linear networkslearning rate scalinggradient descent dynamicslinear convergence rateresidual connections
Functional and Secure Code Generation with Task Vectors
SecVecCoder introduces a task-vector arithmetic method to enhance LLM-generated code for simultaneous functionality and security without post-generation adjustments. The approach modifies model weights to align outputs, requiring minimal computational overhead and maintaining decoding latency within 0.6% of base models. Evaluated on six coding LLMs across the CodeGuard+ benchmark, SecVecCoder improves trustworthy code completion rates by 2.1-39.1 percentage points, with notable gains on unseen CWE types.
task vectorssecure code generationllm alignmentcodeguard+decoding latency
Physics-Informed Machine Learning Under Small-Data Constraints: Lessons from Abrasive Waterjet Milling
The study introduces three methodological contributions for physics-informed machine learning under small-data constraints in abrasive waterjet milling (n=155, Inconel 718). First, it distinguishes physics-based data cleaning from statistical curation, treating the latter as competing hypotheses. Second, it demonstrates instability in model rankings, with the single-split winner dropping from rank 1 to 7 under 10-fold cross-validation, while Gaussian Process (GP) variants dominate. Third, it explores varying levels of physics integration, finding residual learning on a compact physics baseline effective for GP but detrimental for tree-based models. Bayesian hyperparameter tuning improves gradient boosting and SVR but harms multi-stage pipelines. GP uncertainty intervals achieve 86% empirical coverage at nominal 90%.
physics-informed machine learninggaussian processresidual learningbayesian hyperparameter tuningsmall-data constraints
CTA-Pipelining: A Latency-Oriented Spatial Scaling Method for Multi-GPU Systems
The paper introduces CTA-pipelining, a latency-oriented spatial scaling method for multi-GPU systems that exploits shared-memory coherence. The technique enables concurrent execution of dependent kernels at the Cooperative Thread Array level by leveraging fine-grained dependencies, contrasting with traditional throughput-oriented approaches like Tensor Parallelism. Evaluated on 8-GPU H200/B200 systems using CUTLASS, cuBLAS, and NCCL, CTA-pipelining reduces latency by 31.8% versus micro-batching and 29.6% versus Tensor Parallelism on 2-layer GEMM operations, while remaining orthogonal to existing parallelism techniques.
cta-pipeliningmulti-gpulatency-orientedspatial scalingcooperative thread array
NFTR: From Provable Mode-Averaging to Geodesic Subgoal Selection in Offline Goal-Conditioned RL
NFTR introduces a novel approach to subgoal selection in offline goal-conditioned RL, addressing limitations of Hierarchical Implicit Q-Learning (HIQL). The method combines conditional Normalizing Flows for subgoal policies with Triangle-slack Reweighting, which corrects Advantage Weighted Regression (AWR) weights to penalize suboptimal detours. NFTR provably avoids Gaussian mode collapse and maintains stability under stochastic dynamics, leveraging a closed-form mode-averaging result and a three-term suboptimality decomposition. The approach ensures geodesic subgoal selection in deterministic MDPs and provides conservative bounds for stochastic environments.
normalizing flowssubgoal selectiontriangle-slackadvantage weighted regressionoffline rl
False Confidence: Automated Labels Confound Fairness Audits in Cervical Spine Segmentation
This study conducts the first fairness audit of cervical-spine MRI segmentation across sex, age, and race using the CSpineSeg dataset, revealing that reference label provenance critically impacts fairness evaluations. The analysis demonstrates that models trained on gold labels exhibit higher agreement with machine-generated silver labels than with expert annotations, inflating performance metrics by ~8 Dice points and altering fairness conclusions. Specifically, using silver labels instead of gold labels shifts the fairness verdict for age from non-significant to significant due to collapsed within-group variance, termed 'false confidence'. The findings emphasize reporting performance and fairness against expert labels and disclosing reference label provenance.
fairness auditcervical-spine mrisilver labelsdice pointsreference-label provenance
When Does Continual Learning Require Learning
The paper reconsiders continual learning (CL) in LLMs as a dual challenge of spatial (new domains) and temporal (data drift) adaptation, beyond just context management. It evaluates prompt-based methods (GEPA, ACE), supervised learning (SFT, SDFT), reinforcement learning (GRPO, SDPO), and context compression (Cartridges, In-place TTT) on recasted sequential benchmarks. Results show prompt-based methods adapt quickly but degrade on future tasks, distillation struggles with outdated facts, context compression lacks task-learning improvements, and online RL adapts best but is noise-sensitive. The study concludes CL requires method-specific solutions for different environmental changes.
continual learningdomain adaptationdata driftreinforcement learningcontext compression
Explaining Near-Zero Hessian Eigenvalues Through Approximate Symmetries in Neural Networks
The paper explains the prevalence of near-zero eigenvalues in neural network Hessians through weakly broken symmetries. By analyzing deep linear networks with exact symmetry-induced zero modes and ReLU networks as perturbations, the authors show that high-curvature directions are orthogonal to symmetry subspaces while the bulk spectrum lies within them. The mechanism is demonstrated in a two-layer ReLU student-teacher model and a CIFAR-10 trained network, with convolutional layers confirming generalization beyond fully connected architectures. Results connect Hessian bulk spectra to approximate symmetries, resolving the origin of near-zero modes.
hessian eigenvaluespseudo-goldstone modesrelu nonlinearityloss landscapesymmetry breaking
GradInf: Gradient Estimation as Probabilistic Inference
The paper introduces gradient inference, a novel approach to gradient estimation in probabilistic programs by reducing it to probabilistic inference via coupling and factorization operations. GradInf, a probabilistic programming system, implements this approach through sound source-to-source transformations and information-flow typing, enabling automated derivation of gradient estimators. Evaluations demonstrate that GradInf can express existing estimators and construct new state-of-the-art variants outperforming baselines.
gradient estimationprobabilistic programmingcouplingfactorizationinformation-flow typing
Predicting Pseudo-nitzschia harmful algal blooms along the Portuguese Coast using satellite-derived predictors
A spatio-temporal machine-learning framework was developed to predict Pseudo-nitzschia harmful algal blooms (HABs) along the Portuguese Atlantic coast using exclusively satellite-derived predictors. The study utilized a decade-long dataset (2013-2023) with 1,440 observations and over 1,000 predictors, including sea surface temperature, chlorophyll-a, and plankton functional types. A stringent spatio-temporal cross-validation strategy was employed to prevent leakage and mimic real-world forecasting conditions. Ensemble tree-based methods, particularly Random Forest and Extra Trees, achieved strong discrimination with AUC scores of 0.74 and 0.77, respectively. Feature-importance analyses highlighted seasonal structure, spatial context, and lagged environmental conditions as dominant predictors, while biological indicators refined bloom likelihood.
spatio-temporalsatellite-derived predictorsensemble tree-based methodsfeature-importancecross-validation
A law of robustness for two-layer neural networks with arbitrary weights
The paper proves a conjectured law of robustness for two-layer neural networks with arbitrary weights, up to a logarithmic factor. Focusing on continuous piecewise-linear activations (including ReLU), the authors establish that fitting noisy labels below the noise floor forces a lower bound on the Lipschitz constant. For data from uniform or Gaussian distributions, the bound is $\mathrm{Lip}(f)\ge c\,\varepsilon\sqrt{n/(\bar m\log(C\bar m nd/\varepsilon))}$, with high probability. The proof employs function-space covering and a rigidity lemma, showing kink coefficients are controlled by the Lipschitz constant. The result holds for arbitrary real weights, biases, and affine skip connections, with a separate bound based on realized kink count.
lipschitz constanttwo-layer neural networksrobustnesspiecewise-linear activationfunction-space covering
Distributionally Faithful Imputation via Positive Semi-Definite Kernel Density Estimation
The authors introduce PSD Impute, a novel imputation method for missing completely at random (MCAR) data that preserves the joint data distribution by recasting imputation as density estimation from masked observations. The approach leverages positive semi-definite kernel densities to formulate a convex empirical risk problem with closed-form marginals, solved via a Newton interior point method. PSD Impute provides both single and multiple imputations from the same fitted density, offering statistical consistency and adaptive excess risk that mitigates the curse of dimensionality for regular probabilities. Preliminary experiments on one synthetic and eleven real-world datasets demonstrate competitive distributional accuracy compared to existing imputation baselines.
missing completely at randompositive semi-definite kerneldensity estimationempirical riskcurse of dimensionality
Unlocking Temporal Generalization in Hamiltonian Video Dynamics Models
The paper addresses temporal generalization failures in Hamiltonian Generative Networks (HGN) for video dynamics prediction in non-conservative settings. It identifies two key failure modes: latent magnitude growth from unconstrained action-force maps and truncation error accumulation due to under-resolved integrators. The authors propose targeted fixes for each issue, enabling stable predictions at temporal resolutions beyond the training distribution. Results demonstrate improved performance in externally forced, dissipative environments, with recommendations for continuous-time video generation.
hamiltonian generative networkstemporal generalizationnon-conservative dynamicslatent magnitude growthtruncation error
Trustworthy Machine Learning through the Lens of Combinatorial Optimization: Survey and Research Perspectives
The survey establishes combinatorial optimization (CO) as a framework for analyzing trustworthiness in machine learning (ML), addressing transparency, robustness, fairness, and privacy. It synthesizes CO-based approaches for interpretable model learning, explanation generation, robustness certification, fairness auditing, and privacy protection, contrasting them with heuristic methods by emphasizing global guarantees, formal certificates, and explicit trade-off management. While noting scalability challenges, the authors highlight solver advancements that position CO as increasingly vital for trustworthy ML system design.
combinatorial optimizationtrustworthy machine learningrobustness certificationfairness auditingprivacy protection
Scalable and Trustworthy Earth Observation Foundation Models
The chapter proposes domain-specific design principles for remote sensing foundation models (RSFMs) to address challenges in Earth observation (EO) data, where measurement physics and operational constraints necessitate specialized adaptation. It reviews RSFM architectures, pretraining objectives, and downstream adaptation strategies, highlighting inconsistent evaluation as a barrier to reliable deployment. Two case studies demonstrate physics-informed approaches: spectral targeted masking for algal bloom prediction and reinforcement learning for adaptive station selection. Findings indicate no universally superior RSFM exists, advocating for modality-aware transfer and physically plausible representations as key trustworthiness metrics alongside accuracy.
remote sensing foundation modelsphysics-informed learningmodality-aware transferdownstream adaptationtrustworthy ai
The Importance of Encoder Choice:A Tabular-Image Study
This study investigates the impact of encoder selection in multimodal learning, specifically for tabular-image data, by evaluating state-of-the-art tabular models as encoders. A key challenge addressed is the incompatibility of In-Context Learning models, which require labels for instance processing, making uniform embedding of training and test instances non-trivial. The authors propose solutions across multiple models in this family, emphasizing the critical role of encoder choice in multimodal learning performance.
multimodal learningtabular modelsin-context learningencoder selectioninstance embedding
Image classification via a quantum-inspired strategy involving a mixture of experts
A quantum-inspired hybrid classical-quantum framework for image classification is proposed, combining amplitude encoding, local unitary operations, and quantum stabiliser codes with classical neural networks. The method employs a mixture of experts approach, where multiple quantum experts process images with different parameters, and a fully connected neural network jointly analyzes their extracted features. Evaluated on MNIST and Fashion-MNIST datasets, the joint expert analysis outperforms individual experts and reduces image classification failure rates by approximately 50%. The approach maintains moderate computational overhead on GPU workstations, offering a practical alternative to classical convolutional neural networks.
quantum-inspiredamplitude encodingquantum stabiliser codesmixture of expertslocal unitary operations
📰 Industry Media (7)
Anthropic found a hidden space where Claude puzzles over concepts
Anthropic developed the Jacobian lens (J-lens) to probe Claude Opus 4.6's hidden J-space, revealing latent words related to future outputs. The method adapts the logit lens to track intermediate computations in middle layers, exposing conceptual associations (e.g., math intermediates, protein identifiers) and decision-making processes (e.g., cheating cues like 'panic'). Results show J-space captures task-relevant but unsaid content, offering partial interpretability—though limitations remain as it only samples, not exhaustively maps, model internals.
jacobian lensmechanistic interpretabilitylogit lensj-spacechain of thought
Kyutai Releases MuScriptor: An Open-Weight Decoder-Only Transformer for Multi-Instrument Music Transcription to MIDI
Kyutai and Mirelo introduce MuScriptor, an open-weight decoder-only Transformer for multi-instrument music transcription to MIDI. The model processes mel-spectrograms and autoregressively predicts MIDI-like tokens for pitch, timing, and instrument, treating transcription as a language modeling task. Training occurs in three stages: pre-training on synthetic MIDI (D Synth), fine-tuning on real recordings (D Real), and reinforcement learning on verified tracks (D RL). The model achieves an onset F1 score of 60.4 and a multi F1 score of 48.2 post-RL training, demonstrating significant improvements over synthetic-only training.
decoder-only transformermel-spectrogrammidi transcriptionreinforcement learningf1 score
How to Build a T4-Friendly Autonomous Data Science Agent with DeepAnalyze-8B, Sandboxed Code Execution, and Iterative Analysis
The tutorial presents a method for constructing an autonomous data science agent using DeepAnalyze-8B, optimized for T4 GPUs through 4-bit quantization. The agent operates in a sandboxed execution environment, enabling iterative analysis via code generation, execution, and result inspection. Key components include a CodeSandbox class for safe code execution and a DeepAnalyzeAgent class implementing an agentic loop for reasoning and refinement. The workflow culminates in a structured report from an e-commerce dataset analysis, demonstrating the agent's capability to clean, join, visualize, and summarize data. The approach maintains lightweight computational requirements while preserving core agentic functionality.
deepanalyze-8bsandboxed execution4-bit quantizationagentic loopiterative analysis
Google Research Introduces SensorFM: A Wearable Health Foundation Model Pretrained on One Trillion Minutes of Sensor Data
Google Research introduces SensorFM, a wearable health foundation model pretrained on one trillion minutes of multimodal sensor data (PPG, accelerometer, EDA, skin temperature, altimeter) from 5 million participants. The model employs a ViT-1D encoder with masked-autoencoder objectives and Adaptive and Inherited Masking (AIM) for handling missing data, achieving ΔAUC=0.09 and Δr=0.21 improvements over baselines across 35 health tasks. Scaling experiments show linear probes outperform feature-engineered baselines on 31/35 tasks, with LLM-generated heads further boosting performance.
vit-1dadaptive and inherited maskingmasked-autoencoderwearable time-serieslinear probe
Meet LingBot-World-Infinity: An Open Causal World Model With An Agentic Harness
LingBot-World-Infinity (LingBot-World 2.0) is a 14B-parameter open causal world model for interactive video generation, addressing long-horizon drift and latency via a Mixture of Bidirectional and Autoregressive (MoBA) attention mask and distribution matching distillation (DMD). The architecture combines a Vision-Language Model (Director) for semantic reasoning with a Diffusion Transformer (Pilot) for physics simulation, enabling unbounded interaction horizons and real-time 720p/60fps performance. Results demonstrate consistent output quality across diverse actions (e.g., archery, spell-casting) and domains, though benchmarks remain qualitative. The model supports game prototyping and embodied simulation, with a 1.3B variant deployable on single GPUs.
causal world modelmixture of bidirectional and autoregressivedistribution matching distillationvision-language modeldiffusion transformer
Meta Superintelligence Labs Releases Muse Spark 1.1: A Multimodal Reasoning Model for Agentic Tasks on Meta Model API
Meta Superintelligence Labs released Muse Spark 1.1, a multimodal reasoning model for agentic tasks, alongside the Meta Model API. The model features a 1M-token context window with active compaction, supports text/image/video/document inputs, and achieves state-of-the-art performance on tool-use benchmarks (88.1 on Scaled tool use). It trails GPT-5.5 and Opus 4.8 in coding tasks (61.5 on SWE-Bench Pro) but demonstrates zero-shot generalization to new tools. The OpenAI-compatible API is priced at $1.25/$4.25 per million input/output tokens, currently US-only.
multimodal reasoningagentic taskscontext compactionzero-shot generalizationtool-use benchmarks
How to shrink the token budget without shrinking the team
Recent industry trends reveal a misalignment in AI cost optimization, where companies prioritize workforce reductions over token budget engineering. Analysis of hyperscalers and AI deployments shows that 80% of surveyed executives cut headcount without correlating ROI improvements. Effective token budget optimization strategies include prompt caching (reducing costs by up to 90%), model routing, batch processing, retrieval-augmented generation, and open-weight models. Case studies demonstrate that blending AI with human expertise yields superior outcomes, as evidenced by Klarna's rehiring of customer service staff after AI-driven quality declines. Organizations optimizing token budgets through engineering rather than layoffs achieve better long-term returns.
prompt cachingretrieval-augmented generationbatch processingopen-weight modelshyperscalers
Generated automatically at 2026-07-10 20:47 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.