Daily Digest — 2026-08-13
302 items · 4 research labs, 290 arxiv papers, 8 industry media
🏛️ Research Labs (4)
From assistance to execution: How enterprises put AI to work
Two OpenAI studies reveal a shift from assistive to agentic AI in enterprises, with frontier firms (top 10% users) generating 8.3× more output tokens per user than typical firms (up from 2.6× in January). Analysis of 10M+ messages shows Codex now produces 64% of enterprise output tokens, with 108× growth in legal and 41× in sales since February. Advanced capabilities like Plugins see 21% adoption at frontier firms versus 9% elsewhere, while early-career employees exhibit 13× higher usage than executives. Findings suggest complementary investments in workflows and governance are critical for scaling AI beyond access.
agentic aioutput tokensfrontier firmspluginscodex
Daybreak models are now available on AWS
OpenAI has expanded its collaboration with AWS by making Daybreak models available through Amazon Bedrock, enabling enterprises to integrate advanced AI into cybersecurity workflows. Daybreak Blue provides access to general-purpose models like GPT-5.6 Sol, tailored for defensive security, while Daybreak Red offers purpose-trained models for vulnerability research and exploit validation. These models accelerate tasks such as detection engineering, incident response, and mitigation development. By leveraging Amazon Bedrock, eligible customers can deploy Daybreak within existing AWS environments, ensuring seamless integration with familiar security, governance, and operational workflows. Enrollment in Daybreak Access is required, with access provided via the Amazon Bedrock console or Responses API.
amazon bedrockdaybreak modelsgpt-5.6 solvulnerability researchincident response
Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis
OlmoEarth Studio introduces custom embedding exports for Earth observation data, enabling downstream tasks like similarity search, segmentation, and change detection. The platform computes embeddings on-demand using open-source OlmoEarth foundation models, with encoder variants ranging from Nano (128-dim) to Base (768-dim). Users can specify area, time span, resolution, and imagery sources (Sentinel-1/2) to generate Cloud-Optimized GeoTIFFs. Benchmarks demonstrate strong performance, with few-shot segmentation achieving F1=0.84 using only 60 labeled pixels. The embeddings support unsupervised exploration via PCA and temporal change detection without labeled data. Fine-tuning is available for higher performance.
embeddingssentinel-2few-shot segmentationpcageotiff
LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge
LFM2.5-VL-3B introduces a 3B-parameter vision-language model with enhanced screen understanding, multi-image reasoning, and function calling capabilities. The model combines a SigLIP2 400M NaFlex vision encoder with a pre-trained LFM2.5-2.6B text backbone, trained on 34T tokens (4× more vision data than predecessors) using SFT with knowledge distillation and multi-reward RL. It achieves state-of-the-art performance in its size class on real-world image tasks (document understanding, object detection) and matches Gemma-4-E2B on tool use. On-device inference reaches 228 tokens/s (M5 Max) with 3GB memory footprint, while GPU throughput peaks at 11K tokens/s (2× faster than 4B-class models).
vision-language modelknowledge distillationmulti-reward rlon-device inferencefunction calling
📜 arXiv Papers (290)
Surgical WAM: A World-Action Model for Data-Efficient Surgical Robot Learning
The Surgical World-Action Model (Surgical WAM) is introduced to improve data-efficient surgical robot learning by leveraging action-free endoscopic video pretraining. Built on Cosmos Policy, Surgical WAM jointly predicts future endoscopic observations and executable surgical robot action chunks, learning visual dynamics from action-free video before fine-tuning on a fixed budget of action-labeled demonstrations. At deployment, it operates as a closed-loop, receding-horizon controller. On four simulated surgical tasks, video pretraining increased the average success rate from 63.5% to 77.8%, with a 20-point gain on PegTransfer, particularly benefiting contact-rich and bimanual tasks.
surgical robotworld-action modelvideo pretrainingclosed-loop controlvisual dynamics
ConVAWG: A Retrieval-Grounded Framework for Controlled Synthetic Dialogue Generation in Violence Against Women and Girls
ConVAWG introduces a retrieval-grounded framework for generating controlled synthetic dialogues modeling Violence Against Women and Girls (VAWG) scenarios. The method constructs multi-turn dialogues from persona seeds, demographic patterns, crime definitions, and Domestic Homicide Review cases, organized into hierarchical event timelines. It employs activation-steered toxicity control to ensure utterance appropriateness. The framework produces over 6,000 multi-turn dialogue events across 200 scenarios, enriched with metadata. Evaluations via human assessment, LLM-as-Judge, ablations, and downstream tasks demonstrate high dialogue quality and domain fidelity.
synthetic dialogue generationretrieval-grounded frameworkactivation-steered toxicity controlhierarchical event timelinesllm-as-judge
Long-Horizon AI Research for Grothendieck Constant: A Case Study in Human-AI Mathematical Collaboration
The study demonstrates how AI-assisted research improved bounds on the Grothendieck constant $K_G$, a key metric in combinatorial optimization. Using an AI system capable of generating novel insights, the authors tightened known bounds to $\frac{6\pi}{11} \le K_G \le \frac{\pi}{2\log(1+\sqrt{2})} - 10^{-4}$. The paper analyzes the AI's role in mathematical discovery, highlighting both its capabilities and limitations, while outlining optimal conditions for AI-human collaboration in long-horizon research problems.
grothendieck constantcombinatorial optimizationai-assisted researchmathematical boundshuman-ai collaboration
Test-Time Self-Evolving GUI Visual Grounding via Reflection-Guided On-Policy Self-Distillation
The paper introduces a Test-Time Self-Evolving framework for GUI visual grounding, enabling models to adapt post-deployment without human annotations. The method constructs a closed-loop system with Exploration, Evaluation (via an MLLM-based Reflector), Reflection, and Internalization phases, using Reflection-Guided On-Policy Self-Distillation to convert high-level reasoning into token-level supervision. A Contrastive Calibration technique mitigates incorrect auto-regressive prefixes. Experiments across six benchmarks show a 7.4% average accuracy improvement over the base model, marking the first successful application of on-policy self-distillation for test-time adaptation in this domain.
gui visual groundingtest-time adaptationon-policy self-distillationcontrastive calibrationmllm-based reflector
How to Verify Consistency of Probabilistic Claims
The paper presents an interactive probabilistically checkable proof (PCP) protocol for verifying the self-consistency of probabilistic predictors specified by circuits (P,Q) making exponentially many conditional-probability claims. The method constructs a sparse witnessing distribution consistent with the predictor's outputs, building on Nilsson's work to show l_2-approximate consistency for explicit claims is in NP with O(mn + log B)-sized certificates. Results demonstrate polynomial-time verification via few circuit evaluations and proof oracle queries, establishing foundations for training models to prove their own consistency.
interactive pcpprobabilistic consistencyprobability circuitwitnessing distributionl_2-approximation
From Interpretability to Control: Insights from Six Years of the TrustNLP Workshop
The TrustNLP Workshop has documented a paradigm shift in NLP trustworthiness research from post-hoc interpretability to mechanistic understanding and proactive control of generative systems, based on analysis of 144 proceedings papers across six editions (2021-2026). Using established frameworks (TrustLLM, DecodingTrust), papers were classified along six trust dimensions, revealing co-occurrences with capability emergence. Key findings include truthfulness as the fastest-growing dimension (37% of papers by 2025-2026), fairness as the most consistent theme, and explainability's U-shaped trajectory. Cross-venue comparison with ~2K ACL/NAACL/EACL/EMNLP papers shows TrustNLP's topical distribution aligns with the field average. The study concludes with four structural insights and actionable research directions.
trustworthinessmechanistic understandingpost-hoc interpretabilitytruthfulnessfairness
Attention-Path Fragility as an Uncertainty Signal in Large Language Models
The paper introduces Attention-Subnetwork Mutual Information (ASMI), a training-free uncertainty estimator for large language models that quantifies token prediction fragility through attention-path perturbations. ASMI masks attention heads, computes BALD mutual information among subnetworks, and uses a semantic-agreement kernel to discount surface-form disagreements. It outperforms single-pass confidence and entropy metrics, particularly in confident-but-fragile predictions, reducing error rates by approximately half in grounded QA tasks. ASMI excels in context-dependent tasks but reverts to baseline performance in parametric QA, aligning with its predicted domain applicability. Head-level analysis reveals that error coupling, not mere fragility, drives ASMI's effectiveness.
attention-subnetwork mutual informationbald mutual informationsemantic-agreement kernelconfident-but-fragile predictionsparametric qa
sLTN: Structural Logic Tensor Networks
The paper introduces sLTN, an extension of Logic Tensor Networks (LTN) that incorporates structural dimensions as first-class elements to handle organized data like temporal sequences or graphs. These dimensions represent named tensor axes (e.g., time steps, graph nodes) and enable explicit quantification and relational constraints at the logical level. The authors formalize sLTN's syntax and fuzzy tensor semantics, showing backward compatibility with LTN, and provide a PyTorch implementation. The framework is demonstrated on temporal and sequential reasoning tasks, accompanied by an open-source library.
neurosymbolictensor semanticsstructural dimensionsfirst-order logicpytorch
Two-stage Odd Residual Flows for Mean-Preserving Probabilistic Time Series Forecasting
The authors propose Two-stage Odd Residual Flows (TORF), a novel framework for probabilistic time series forecasting that decouples mean prediction from uncertainty estimation. TORF first employs a pre-trained deterministic model for accurate mean forecasting, then uses a Restricted Normalizing Flow with strictly odd functions to model residual distributions, ensuring mean preservation without Monte Carlo sampling. Experiments demonstrate that TORF achieves state-of-the-art deterministic accuracy (NMAE) and strong density estimation performance (CRPS) across both short and long forecasting horizons.
probabilistic forecastingnormalizing flowsmean preservationresidual distributionsmonte carlo sampling
Why Does CLAUDE.md Keep Growing? Catastrophic Remembering in Agentic Coding
The paper identifies catastrophic remembering, a phenomenon where agentic coding READMEs like CLAUDE.md grow indefinitely due to imperfect recall, contrasting catastrophic forgetting in continual learning. Analyzing 247,694 instruction lifetimes across 1,867 repositories, it quantifies unbounded growth (+226%, +4.9 instructions/commit) and decreasing deletion likelihood (-0.032 log-hazard/commit). It proposes prompt comments as a solution, demonstrating 99.3% reduction in excess instructions via IFEval inversion and 23.1% improvement in real-world instruction-following using WildIFEval inversion.
catastrophic rememberingagentic codinginstruction lifetimesifeval inversionprompt comments
RTSKG: Building a Rail Transit Station Knowledge Graph Dataset
The authors contribute RTSKG, a Rail Transit Station Knowledge Graph dataset that models spatial and semantic interactions among urban entities to support city-level rail transit analysis. The dataset integrates heterogeneous entities (stations, road segments, POIs) using a unified schema and is published as Linked Data. Evaluations on store recommendation and ridership prediction tasks demonstrate RTSKG's effectiveness for urban mobility applications.
knowledge graphurban mobilitylinked dataridership predictionspatial semantics
SkillZip: Evaluation-Free Skill Compression for Self-Evolving Agents by Discovering Reusable Structure
SkillZip introduces an evaluation-free method for compressing reusable skills in self-evolving agents by discovering shortest faithful structural explanations. The approach formalizes a typed minimum description-length objective over skill contracts and residuals, ensuring hard coverage constraints for extracted triggers, workflow edges, tool requirements, obligations, and output fields. It supports one-shot mode with deterministic optimization and continual Zip-on-Write mode for integrating self-evolution patches without task replay or full history reparsing. Comprehensive experiments demonstrate SkillZip's effectiveness in compression performance, generalizability, and cost overhead.
skill compressionself-evolving agentsminimum description-lengthhard coverage constraintzip-on-write
Quantum Coordination Advantages in AI State-Tracking Tasks: Semantic Compilation and Latent Memory
The paper establishes quantum coordination advantages for AI state-tracking tasks through semantic compilation and latent memory. It introduces a boundary-preserving semantic-compilation theorem, mapping finite causal tasks into a semantic AI interface while maintaining event order and past input access. The method counts communication, memory, and local work, allowing classical recurrence and recomputation. Key results include a hidden-matching separation requiring O(log N) qubits versus Ω(√N) classical bits for matched-entity synopsis QA, and a Max-kSAT streaming separation with O(log^5 n log(1/δ)) qubits versus Ω(√n) classical coordination width for continual requirements auditing. The stabilizer latent-state dialogue demonstrates a quantum-native compiler test with n qubits, contrasting classical causal online realization requiring B+M ≥ ½n² + (3/2 - log₂3)n + O(1).
semantic compilationquantum coordinationboundary-preserving theoremlatent memorystabilizer dialogue
Entropy-Centric Explainable AI for Remote Sensing Image Segmentation
The paper proposes an entropy-centric explainable AI (XAI) method for semantic segmentation in remote sensing imagery, addressing the opacity of deep neural networks in critical domains. The approach introduces a novel XAI evaluation methodology to quantify the relevance of highlighted regions in segmentation outputs. Experiments demonstrate the method's superiority over recent XAI adaptations for semantic segmentation, though specific metrics or benchmarks are not detailed in the provided text.
explainable aisemantic segmentationremote sensingentropy-centricneural networks
A Comparative Evaluation of Deep Learning Object Detection Models on a Real-World Multi-Plant Dataset from Africa
This study evaluates six object detection models on AgriAISeg, a real-world African agricultural dataset with 3,382 images of sesame, cabbage, and tomato crops under varying field conditions. The comparative analysis of YOLOv5, YOLOv8, YOLO11, YOLO26, Faster R-CNN, and RT-DETR reveals RT-DETR as the top performer (precision: 0.768, mAP@0.5:0.95: 0.624), followed by YOLOv8 and YOLO11, while Faster R-CNN showed inferior performance (mAP@0.5: 0.466). Results demonstrate transformer-based and one-stage detectors' superiority in complex agricultural environments.
object detectionprecision farmingmap@0.5real-world datasettransformer-based detectors
3D Weighted Geometric Graph Neural Networks for Sheep Facial Pain Assessment
The paper introduces 3D-SPFES, a monocular depth-aware geometric graph neural network system for sheep facial pain assessment. The method constructs a 3D weighted geometric graph from RGB images using VideoDepthAnything, with nodes representing SPFES landmarks (e.g., ears, eyes) and edges weighted by Euclidean distance and surface co-planarity. A Weighted Geometric Graph Neural Network (WG-GNN) with K=3 geometry-aware message-passing layers and scaled dot-product attention processes the graph, producing O=3 pain-level clusters and a Normalized Pain Score (NPS) ranging [0, 100%].
3d geometric graphweighted gnnmonocular depth estimationpain assessmentsheep facial expression
Multiclass Sentiment Analysis for Identifying Political Viewpoints
The paper investigates multiclass sentiment analysis for political viewpoints on social media, comparing XGBoost and BERT-based approaches. The authors train and evaluate models on a labeled dataset of political posts, measuring performance via standard classification metrics. Results show modest F1-scores (XGBoost: 0.2835, BERT: 0.2806), highlighting the difficulty of classifying nuanced political sentiment and establishing baselines for future work.
sentiment analysismulticlass classificationpolitical discoursexgboostbert
V-FiLLM: Verified Financial LLM Reasoning Benchmark
V-FiLLM introduces a verified benchmark for evaluating financial reasoning in LLMs, generating test items from executable computation trees grounded in real tables to ensure correctness by construction. The framework controls four difficulty axes (computation depth, expression breadth, financial concept complexity, context size) and avoids annotation costs by rendering trees into natural-language questions. Evaluations on open-source models show accuracy drops of up to 51% with increased reasoning depth and 47% under adversarial perturbations, while LoRA fine-tuning on verified chain-of-thought traces improves accuracy by 4.5% on held-out problems and outperforms baselines by 5% on FinQA.
financial reasoningcomputation treeslora fine-tuningchain-of-thoughtadversarial perturbations
Workflow Cards: Structured Summaries of Workflow Executions Using Provenance Data
The paper introduces Workflow Cards, structured summaries capturing provenance data from workflow executions to address documentation gaps in machine learning pipelines. Unlike Model Cards and Data Cards, which focus on static artifacts, Workflow Cards detail execution-level information such as data preparation, parameter choices, runtime behavior, and resource usage. The authors define a template based on provenance questions and evaluate its effectiveness using LLMs. Results demonstrate that Workflow Cards nearly double answer quality compared to schema-based querying, validated by both LLM-as-a-Judge and human assessments, filling a critical gap in execution-level documentation.
workflow cardsprovenance dataexecution-level documentationschema-based queryingllm-as-a-judge
R4DSG: Relative 4D Scene Graph Memory for Object-Centric Question Answering in Long Egocentric Video
R4DSG proposes a relative 4D scene graph memory for object-centric question answering in long egocentric video, addressing limitations of caption-based methods and 3D scene graphs requiring strong geometry. The method converts video into compact memory entries indexed by time, place, persistent objects, anchor-relative change, and interaction context, leveraging RGB-only techniques like promptable segmentation and relative 3D lifting. On EgoLifeQA's object-related subset, it achieves a 6.7-point overall accuracy gain over EgoRAG-Text and 12.5-point improvement on temporal questions, demonstrating efficacy for wearable AI applications.
4d scene graphegocentric videoobject-centric qarelative 3d liftingpersistent object identity
Policy Convergence and Divergence Across National and Within Regional AI Strategies: A Policy Design Element Analysis
This paper contributes a systematic analysis of policy design elements across national and regional AI strategies, addressing gaps in understanding horizontal (country-to-country) and vertical (region-to-country) convergence. Using a latent-inductive approach, the authors coded 74 national and 3 regional AI strategies from 205 UN member and non-member states, focusing on goals, approaches, and principles. Results show strong horizontal convergence in economic competitiveness, research support, and ethical AI use, but divergence in human rights goals, participatory governance, and human-centric principles. Vertical convergence varies by region, with the AU exhibiting the highest alignment, the EU diverging on human-centric values, and the Nordic-Baltic Region showing mixed patterns.
policy design elementshorizontal convergencevertical convergencelatent-inductive approachai strategies
On the Limitations of Cross-Lingual Consistency in Multilingual Text-to-image Generation
The study introduces LingT2I, a multilingual benchmark with 33K prompts across 10 languages, to evaluate cross-lingual effects in text-to-image (T2I) generation. Using this benchmark, the authors analyze linguistic inequality and language-dependent trade-offs in content generation and text rendering, revealing systematic impacts of linguistic factors and cultural contexts on model outputs. The findings highlight the need for more robust and inclusive T2I models, with the benchmark serving as a foundation for future cross-lingual research.
text-to-image generationmultilingual benchmarkcross-lingual analysislinguistic inequalitycultural context
Putting Registers to Work: Task Registers for Token Pruning in Vision Transformers
The paper introduces Task-Adaptive Pruning (TAP), a method for token pruning in Vision Transformers that adapts to multiple vision tasks (classification, segmentation, detection) via task-specific registers. TAP employs parameter-free reduction criteria, dynamically ranks tokens, allocates removal budgets across layers, and adjusts recovery scales for dense features. Evaluated at a keep rate of ρ=0.5, TAP-J achieves 47.0 mIoU (1.30× throughput) on ADE20K, 53.7 box AP (1.32× throughput) on COCO, and maintains ImageNet-1K performance, demonstrating cross-task efficiency.
token pruningvision transformerstask-adaptive pruningparameter-free reductionthroughput optimization
TimeRoute: Time-Aware Modality Routing and Diffusion for Multi-Modal Recommendation
TimeRoute introduces a diffusion-based multi-modal recommender that addresses temporal modality drift via personalized routing and dual-stream denoising. The system employs a temporal-aware modal router to dynamically adjust modality fusion weights per user and a diffusion-based graph reconstructor conditioned on temporal profiles through FiLM, suppressing outdated edges. Evaluations on TikTok, Amazon-Baby, and Amazon-Sports show improvements of up to 9.8% in Recall@K, Precision@K, and NDCG@K over baselines across 10-seed tests.
multi-modal recommendationmodality routingdiffusion modeltemporal-awarefeature-wise linear modulation
XCoT-VLA: Executable Chain-of-Thought for Vision-Language-Action Driving
XCoT-VLA introduces executable Chain-of-Thought (CoT) tokens for Vision-Language-Action (VLA) models in autonomous driving, replacing verbose natural-language reasoning with compact executable representations. The method leverages Reason-Action supervision from logged trajectories and scene context, using deterministic token-function routing to apply separate feedforward networks (Reason FFN and Control FFN) for reasoning and trajectory generation. XCoT Policy Optimization (XCPO) refines the executable token space. Results show reductions in longitudinal ADE (1.645 to 1.323) and lateral FDE (1.616 to 0.648), with reasoning overhead minimized to 2-6 tokens, enabling real-time planning.
chain-of-thoughtvision-language-actiontoken-function routingtrajectory generationself-attention
ReLTEx: Reliable LLM-based Taxonomy Expansion
ReLTEx introduces a framework for reliable taxonomy expansion using Large Language Models (LLMs), addressing noise and inconsistency in direct LLM-generated outputs. The method combines LLM-driven candidate generation with structure-aware validation and recursive expansion control to reduce hallucinations. Evaluated on benchmark taxonomies in a masked expansion setting, ReLTEx outperforms baselines in reliability and semantic coherence, as measured by adapted metrics and human evaluation.
taxonomy expansionlarge language modelsstructure-aware validationrecursive expansionhallucination reduction
CARE: Confidence-Aware Reasoning for Reliable Medical VQA
The paper introduces CARE, a Confidence-Aware medical Reasoning framework addressing confidence miscalibration in medical Multimodal Large Language Models (MLLMs). CARE employs a dual-stage pipeline: Medical-CoT synthesis generates structured data for Supervised Fine-Tuning, while Group Relative Policy Optimization (GRPO) with a Confidence-Aware Reward (CAR) mechanism aligns confidence with diagnostic accuracy. Evaluated on three Medical VQA benchmarks, CARE achieves top diagnostic accuracy, lowest Expected Calibration Error, and reduced Hallucination Rate, enhancing clinical trustworthiness.
multimodal large language modelsconfidence miscalibrationchain-of-thought reasoningsupervised fine-tuningexpected calibration error
Evidence-Grounded Trustworthy Multimodal Reasoning and Evaluation Benchmark in Complex Urban Scenes
The authors introduce AD2-Bench, a benchmark addressing multimodal reasoning failures in complex urban scenes, and propose Evidence-grounded Visual Reasoning (EGVOR) for robust cognition. AD2-Bench employs a Hierarchical Visual Diagnosis framework decomposing reasoning into a Chain of Evidence (CoE), identifying Spatial Ambiguity and Semantic Uncertainty as primary failure causes. EGVOR replaces implicit reasoning with explicit Evidence Atoms—structured spatial-semantic triplets—ensuring tight alignment between localization and semantic understanding. Training progresses from reflective supervision to reinforcement learning, explicitly rewarding reduced reasoning variance. Experiments show EGVOR significantly enhances reasoning stability under adverse conditions, offering a trustworthy multimodal cognition framework.
multimodal reasoningevidence atomsspatial ambiguitysemantic uncertaintyhierarchical visual diagnosis
A Cost-Efficient Routing Pipeline for Multilingual Short-Text Classification Using Small Language Models
The paper proposes a cost-efficient routing pipeline for multilingual short-text classification that selectively translates low-resource languages to English before zero-shot classification, while keeping high-resource languages on a direct multilingual path. The method uses pretrained compact sentence encoders without task-specific fine-tuning, evaluated on SIB-200 (7-way topic classification) and MASSIVE (60-intent classification) subsets. Results show selective translation improves low-tier Macro-F1 from 0.4632 to 0.6828 on SIB-200 and from 0.2143 to 0.4417 on MASSIVE, with optimal routing depending on task granularity.
multilingual classificationzero-shot learningsentence encoderslow-resource languagesselective translation
Temporally Grounded Compositional Camera Motion Understanding via Geometric Knowledge Distillation
The paper introduces CamChoreo, a benchmark of 4,229 single-shot video clips with expert-annotated temporal segments for compositional camera motion understanding, addressing limitations in clip-level recognition. It proposes CamDistill, a method that distills geometric knowledge from a frozen 3D foundation model into lightweight camera tokens during training, eliminating the need for the 3D model at inference while matching the accuracy of direct feature injection (CamInject). Results show CamDistill achieves comparable performance to CamInject without runtime overhead, advancing camera-motion understanding to temporally grounded, compositional recognition.
camera motionknowledge distillationmultimodal llmstemporal grounding3d foundation model
FedCGR: Federated Cross-Domain Generative Recommendation
FedCGR introduces a federated cross-domain generative recommendation framework that aligns item spaces through discrete semantic ID (SID) sequences derived from public metadata, avoiding private interaction sharing. The method employs a reliability-aware semantic interface to inject local collaborative filtering signals and trains a prototype-personalized generator with selective parameter aggregation based on domain relatedness. Evaluations on six Amazon cross-domain scenarios demonstrate FedCGR's superiority over federated generative baselines and competitive performance against sequential and federated CDR methods in full-ranking and sampled protocols.
federated learningcross-domain recommendationsemantic id sequencescollaborative filteringgenerative recommendation
ThinkRetrieve: Retrieval-Augmented Reasoning Traces for Test-Time Scaling
ThinkRetrieve introduces a retrieval-augmented framework for test-time scaling in Large Reasoning Models (LRMs), addressing limitations of sequential reasoning traces by dynamically injecting relevant solved examples at each step. The method retrieves step-by-step solutions from an external corpus, providing guidance on reasoning processes rather than just factual relevance. Evaluations across five models (1.5B--8B parameters) on GSM-8K, MATH-500, AIME 2025, and SciQ show consistent accuracy improvements, with up to 60% relative gains on AIME 2025.
retrieval-augmented reasoningtest-time scalinglarge reasoning modelschain-of-thoughtdynamic retrieval
IO Factory: Simulating AI-Enabled Influence Campaigns at Scale
IO Factory introduces an AI-driven framework for simulating large-scale information and influence campaigns as integrated, traceable processes. The framework models coordinated AI swarms that adapt to platform feedback, simulating campaigns across planning, platform action, exposure, interpretation, measurement, and adaptation phases. Implemented in a controlled simulated platform, IO Factory links actor roles, platform actions, exposure records, and structured evaluations, supporting configurations of up to 100,000 agents. Results demonstrate scalable execution of campaign timelines, inspectable exposure evidence, and measurable belief changes. The framework enables reproducible research and red-team analysis by recording actors, objectives, constraints, exposure paths, and measurement rules.
ai swarmsinfluence campaignsexposure recordssimulated platformbelief variables
FaithformBench: Benchmarking Faithfulness of Mathematical Chain-of-Thought Autoformalisation
We introduce FaithformBench, a benchmark for evaluating the faithfulness of mathematical chain-of-thought autoformalisation (AF) systems. The method generates perturbed reasoning steps to assess both validity preservation on correct inputs and invalidity preservation on incorrect inputs, addressing limitations of existing approaches that rely on human annotation or LLM judges. Experiments on eight AF systems across four mathematical datasets reveal pervasive sycophancy, where invalid inputs are silently corrected into provable statements. Results indicate a tension between validity and invalidity preservation, with the most validity-preserving fine-tuned AFs also being the most sycophantic.
autoformalisationfaithfulnessvalidity preservationsycophancychain-of-thought
ComBodied Agents: a New Paradigm of Human-Centric Agentic AI
The authors introduce Combodied Agents, a human-centric paradigm for Agentic AI that focuses on modeling and supporting individual human-state trajectories over time. The framework integrates multimodal perception, longitudinal memory, Personal World Models, and admissible intervention policies to form a closed-loop system. It leverages software tools, sensors, wearables, robots, and human services as action channels, emphasizing purpose-bounded, uncertainty-aware, and user-correctable representations. The design space is organized by human-state targets, relational contexts, and agent roles, with proposed evaluation metrics and governance directions. This approach shifts Agentic AI from task completion to sustained human benefit.
combodied agentsagentic aipersonal world modelsmultimodal perceptionadmissible intervention policy
GitSkills: A Dataset of Agent Skills on GitHub
We introduce GitSkills, a dataset of 3,797,117 SKILL.md files collected from 282,200 public GitHub repositories in July 2026, capturing the adoption and evolution of agent skills—folders containing natural language instructions for language-model agents. The dataset groups identical files into 1,877,981 distinct contents, enriching one representative per group with parsed front matter, folder contents, repository metadata, and, for a subset, commit history. Stored in a self-contained SQLite file, GitSkills supports research on skill reuse, maintenance, authorship, and security, addressing the unique challenges posed by probabilistic skill selection and decentralized distribution in agent ecosystems.
agent skillsskill.mdlanguage-model agentsgithub repositoriessqlite
Enhanced Filtering Algorithms for the Euclidean Traveling Salesperson Problem and its variants in Constraint Logic Programming
Proposes novel filtering algorithms for the Euclidean Traveling Salesperson Problem (TSP) and its variants in Constraint Logic Programming (CLP), leveraging geometric coordinate information for stronger constraint propagation than distance-matrix approaches. The method extends to Euclidean Generalized TSP (EGTSP), addressing routing and logistics applications. Experimental validation shows computational improvements over existing techniques.
euclidean tspconstraint logic programmingfiltering algorithmsconstraint propagationgeneralized tsp
VibeLifeBench: Can Your Life Agent Be Proactive and Persistent in a Living World?
The authors introduce VibeLifeBench, a benchmark of 200 long-horizon tasks across ten everyday-life domains to evaluate proactive and persistent behavior in LLM-based personal assistants. Tasks simulate multi-week timelines in a dynamic world with 22 mock services, where silent environmental changes require autonomous re-inspection. Evaluation measures end-state correctness, action timeliness, and adherence to implicit constraints via fine-grained checks on agent outputs. Testing seven frontier models reveals uniformly low scores, highlighting their limitations in real-life assistance scenarios. The benchmark, environment, and framework will be open-sourced.
long-horizon tasksllm agentsdynamic environmentimplicit constraintsevaluation framework
Hypothesis Frontier: Verifier Guided LLM and Symbolic Search for First-Order Induction
Hypothesis Frontier introduces a verifier-guided neurosymbolic framework for first-order concept synthesis, combining LLM generation with symbolic verification to improve formula induction. The method evaluates each LLM-generated formula on training objects, retains the strongest verified hypothesis, and uses residual errors to guide subsequent generation. Symbolic processing repairs invalid formulas and simplifies valid ones without altering training predictions. Under matched conditions, Hypothesis Frontier solves significantly more problems than repeated original-prompt generation (exact metrics unspecified) and compresses final formulas while preserving correctness.
first-order inductionneurosymbolicverifier-guidedformula synthesissymbolic reasoning
TACTICL: Task-Aware Compression of Tabular ICL Models
TACTICL introduces a task-aware compression framework for tabular in-context learning models, combining layer pruning with lightweight task-specific adapters to balance computational efficiency and in-context adaptability. The method jointly optimizes transformer layer pruning and adapter training on downstream tasks, blending in-context and in-weight learning. Evaluated on 47 benchmark datasets, TACTICL replaces up to 85% of layers without significant performance degradation while maintaining robustness to data shifts. The framework exploits depth-wise redundancy in tabular foundation models through structured compression and task adaptation.
tabular in-context learningtask-aware compressiontransformer pruninglightweight adaptersdepth-wise redundancy
Whisper-Aware LLM: Self-Supervised Uncertainty Learning for Robust Whispered Speech Recognition
The Whisper-Aware LLM introduces a self-supervised uncertainty learning framework for robust whispered speech recognition, addressing signal ambiguity in ASR systems. The model quantifies acoustic signal deficiencies through targeted self-supervised tasks and operationalizes uncertainty via Confidence-Fused Decoding, which modulates frame-level attention and provides high-level instructions to the decoder. Experiments demonstrate a 17% relative CER reduction on AISHELL6-Whisper, achieving state-of-the-art performance, while reducing hallucination rates from over 25% to 4.5%.
whispered speech recognitionself-supervised learningconfidence-fused decodingacoustic signal deficiencieshallucination rates
MIRA: Medical Image Reflection for Agentic Diagnosis
MIRA (Medical Image Reflection for Agentic Diagnosis) introduces a framework for autonomous medical visual diagnosis that dynamically invokes image-processing operations and web search while verifying evidence relevance and consistency. The framework employs a two-stage training strategy: first, a tool-augmented Monte Carlo Tree Search data engine constructs supervised fine-tuning trajectories by verifying visual grounding accuracy and semantic consistency; second, reinforcement learning improves decision-making through online reflective principle evolution. MIRA achieves an average score of 64.73 across nine medical visual reasoning benchmarks, improving its Qwen3-VL-8B backbone by 7.44 points, increases useful tool-use judgments from 56.2% to 73.8%, and reduces harmful judgments from 8.9% to 1.6%.
monte carlo tree searchvisual groundingreinforcement learningsemantic consistencymedical diagnosis
Reference-Free Post-Training of Open Large Language Models for Multilingual Machine Translation
The work introduces MiLMMT-46-v1.0, a multilingual machine translation model derived from supervised-finetuned MiLMMT-46-v0.1 via reference-free post-training. The method combines Group Relative Policy Optimization (GRPO) with a gated reward averaging two reference-free quality estimation models, followed by linear interpolation of supervised fine-tuning and reinforcement learning checkpoints. Evaluated across 46 languages, MiLMMT-46-v1.0 outperforms both its supervised counterpart and open baselines (Seed-X, HY-MT2, TranslateGemma), achieving competitive reference-free scores against proprietary systems (Google Translate, Gemini 3 Pro, GPT-5). On-policy distillation matches but does not exceed RL interpolation performance.
multilingual machine translationgroup relative policy optimizationreference-free quality estimationcheckpoint interpolationon-policy distillation
Surfacing the Unsaid: CUE-Bench for Affective Stance in Chinese Discourse
The authors introduce CUE-Bench, a Chinese Unsaid Emotion benchmark addressing limitations in existing emotion understanding evaluations by focusing on Affective Stance. The benchmark captures implicit affect through nine interpretable stances derived from explicit-implicit polarity interactions, supplemented by intent and fine-grained emotion annotations. Experiments demonstrate that incorporating Affective Stance improves fine-grained emotion recognition by 3.5pp and pragmatic intent detection by 7.8pp over baselines.
affective stanceimplicit affectemotion recognitionpragmatic intentdiscourse understanding
Modelling Geographic Atrophy Progression using Implicit Neural Representations
The authors propose using Implicit Neural Representations (INRs) to model Geographic Atrophy (GA) progression in Age-related Macular Degeneration (AMD) from longitudinal Fundus Autofluorescence (FAF) images. Their method generates both FAF images and GA segmentations at past and future time points, addressing the challenge of individualized progression patterns in low-data settings. Compared to baseline models, the approach achieves the lowest Mean Absolute Error (MAE) for GA lesion area and the highest DICE score while maintaining FAF image quality. The code is publicly available for reproducibility.
implicit neural representationsgeographic atrophyfundus autofluorescencemean absolute errordice score
Fast and Memory-Efficient Wavelet Convolutions via I/O-Aware Reformulation
We present an I/O-aware reformulation of wavelet convolution (WTConv) that reduces memory overhead and accelerates training while preserving its benefits. Our method combines three algebraic optimizations: on-chip recomputation of the Haar analysis butterfly, collapsing multi-level synthesis into a single pass, and folding per-channel scales into convolution weights. Evaluated on WTConvNeXt across decomposition levels and tensor shapes, our approach reduces high-bandwidth memory traffic by 2.55×, achieves up to 4.35× training speedup over the reference implementation, and halves peak memory usage while maintaining computational equivalence.
wavelet convolutionhaar analysishigh-bandwidth memoryalgebraic reformulationtensor shapes
BPG: Balancing Plasticity and Generalization for Domain Incremental Learning
The authors propose BPG, a unified framework for domain incremental learning (DIL) that balances plasticity and generalization. BPG comprises BPG-Adapter, which dynamically determines adapter hidden dimensions based on domain-specific feature separability, and BPG-Inference, a soft domain mixture strategy that integrates multiple domain-specific models at test time. Evaluations on DomainNet, CDDB, and CORe50 demonstrate that BPG outperforms uniform adapter-based approaches and hard domain selection strategies, achieving state-of-the-art average accuracy while reducing forgetting to 0.22% on DomainNet.
domain incremental learningadapter hidden dimensionfeature separabilitysoft domain mixturedomainnet
Beyond Fixed Luminance: Towards Panchromatic and Orthochromatic Image Colorization
We propose a luminance-agnostic colorization framework that treats colorization as full-RGB image editing, addressing limitations of fixed-luminance approaches in panchromatic and orthochromatic conditions. The method employs a foundation image-editing model trained with a mixed grayscale objective, combining standard luminance grayscale and red-insensitive grayscale formations. Evaluations on COCO, ImageNet, and a multi-instance benchmark demonstrate competitive performance on standard grayscale inputs and significantly improved robustness under orthochromatic inputs. Qualitative comparisons and human studies indicate reduced visible color artifacts compared to fixed-luminance methods.
luminance-agnosticpanchromaticorthochromaticgrayscale objectivecolor artifacts
EvoMem: Memory-Augmented Evolution for Code Optimization
EvoMem introduces a persistent memory architecture for LLM-based evolutionary program search that captures and reuses successful mutation strategies across runs and tasks. The method operates in two phases: post-run extraction of promising mutations with provenance, and retrieval of relevant instructions during evolution to guide future mutations. Experiments on geometric optimization, multi-hop QA, and GPU kernel optimization show average improvements in target metrics or search speed, though with task-dependent variability, demonstrating reduced redundant exploration through memory-augmented evolution.
evolutionary searchmemory-augmentedcode optimizationllm-drivenmutation strategies
ChemWorld: Programmable Chemical Worlds for Controlled and Replayable Agent Experimentation
ChemWorld introduces a programmable chemical environment enabling controlled and replayable agent experimentation by separating public experimental contracts from evaluator-owned chemical laws. The framework compiles reusable process and observation components into executable worlds, allowing researchers to vary world composition, operating conditions, or hidden laws while maintaining fixed public task conditions. Transactional execution records operations, failures, and state transitions for exact replay and auditing. Validation included full-census qualification of the reference registry, 52 generated compositions, and module tests. Experiments demonstrated deterministic lifecycle semantics, failure recovery, and isolated effects of private-law interventions. An independent agent successfully completed a lifecycle in a non-reference world, showcasing ChemWorld's utility for systematic chemical experimentation.
programmable environmenttransactional executionchemical lawsreplayable experimentationpublic contract
MVTrack: Ultrafast Appearance-Free Moving Object Tracking from Compressed Bitstreams
MVTrack introduces an ultrafast moving object tracker that operates directly on H.264 bitstreams, eliminating the need for RGB-based object detection. The method combines MVDet, a lightweight motion vector field detector, with MVLink, a minimalist kinematic association module. On VIRAT, MVTrack outperforms YOLO26n with 60× fewer parameters, 40× fewer FLOPs, and an 8.6× reduction in CPU latency, demonstrating efficient surveillance tracking without pixel reconstruction.
h.264 bitstreamsmotion vector fieldskinematic associationobject trackingcomputational efficiency
SkillLens: Visual Skill Cards for Retrieval-Augmented GUI Action Prediction and On-Policy Distillation
SkillLens introduces Visual Skill Cards (VSCs), a state-conditioned memory representation that binds reusable procedures with visual applicability cues, evidence, and verification signals to address the lack of visual procedural memory in GUI action prediction. The method constructs VSCs from interaction traces via Trace-to-Visual-Skill-Card retrieval and selectively expands evidence for a frozen GPT-5.4-mini executor, while CardDistill uses VSCs as privileged context to train a student model (Qwen3-VL-2B). Evaluations on Multimodal-Mind2Web and WebLINX-BrowserGym show +11.6 and +2.9 point improvements in Step SR and Overall metrics, respectively, with further +12.0 and +3.2 point gains for the distilled student.
visual skill cardsretrieval-augmented guion-policy distillationtrace-to-visual-skill-cardprocedural memory
The GenAI Catch-22: Use of Generative Artificial Intelligence in Norwegian Newsrooms During the 2025 Parliamentary Election
This study investigates the sociotechnical risks of Generative AI (GenAI) adoption in Norwegian newsrooms during the 2025 parliamentary election, revealing a 'GenAI Catch-22'. Through ten months of interviews with journalists and managers, the authors identify a tension between reliance on human expertise to monitor GenAI outputs and the erosion of that expertise through over-dependence on automation. Findings show that initial ambitions for audience-facing GenAI services collapsed due to unrealistic expectations about technological capabilities, shifting focus to internal tool usage. The work highlights an understudied internal threat of AI-mediated journalistic degradation, contrasting prevailing concerns about external disinformation.
generative aisociotechnical imaginariesnewsroom resiliencedisinformationautomation bias
Rule of Thumb: Explaining Artificial Intelligence Systems using Partial Information
The authors propose 'Rule of Thumb' (RoT), a novel explainable AI (XAI) method that identifies the most relevant features for predicting AI system behavior on a per-datapoint basis. RoT is model-agnostic, computationally efficient, and designed to meet regulatory requirements while providing familiar interfaces for practitioners. The method is demonstrated in three scenarios: zero-shot classification with large language models (LLMs), auditing opaque AI systems without model access, and AI-assisted scientific discovery. RoT outperforms alternative XAI approaches in speed and is shown to be particularly effective in these diverse applications.
explainable airule of thumbzero-shot classificationmodel-agnosticfeature relevance
Compositional Benchmark Synthesis for Hierarchical Human Action Recognition
The paper introduces a benchmark-generation framework for hierarchical human action recognition, synthesizing a four-level hierarchy (actions, activities, low-level intentions, high-level intentions) from flat single-label action corpora while preserving real action-level features. It employs a transition model with subject-consistency constraints and a coverage-aware sampler (reducing subject usage Gini from 0.566 to 0.248) to assemble 15,002 episodes, avoiding circular supervision by decoupling generation rules from evaluation logic. Baseline evaluations reveal a persistent compositional held-out gap (0.13–0.17 macro-F1), with a graph-aware model performing best but not closing the gap, indicating benchmark structural properties rather than model limitations.
hierarchical action recognitionbenchmark synthesistransition modelcoverage-aware samplingmacro-f1
A Gateway Architecture for Enterprise MCP Authentication: Unifying Heterogeneous Auth, Identity Delegation, and the User / Non-User Persona Problem
The paper presents a centralized MCP gateway architecture addressing enterprise authentication fragmentation in LLM agent deployments. The solution introduces: (1) a two-axis auth model (persona × credential type), (2) gateway layer supporting 3 SSO grants and 3 token-provisioning models, (3) three identity flows (User-to-OAuth2, Non-user-to-Service-Account, User-to-Service-Account), and (4) deployment evolution from perimeter-based to private MCP tunnels. The production system fronts dozens of MCP servers across multiple client types, unifying previously inconsistent authentication methods.
mcp gatewayauthentication fragmentationoauth2 delegationenterprise ssotoken exchange
Tree-of-Ideas: Automated Research Ideation via Cross-Trajectory Reasoning over Scholarly Evolution
The paper introduces Tree-of-Ideas (ToI), a two-stage framework for automated research ideation by modeling scholarly evolution across trajectories. EvoTrace reconstructs branching citation paths to track method evolution and unresolved gaps, while EvoAgent performs cross-trajectory reasoning to identify convergent problems and complementary solutions. Evaluated on six AI topics, ToI scores 6.27/10 (vs. 5.36 for baselines) with strong Novelty (6.36) and Groundedness (7.00), approaching human reference performance (6.29).
research ideationscholarly evolutioncross-trajectory reasoningcitation analysisautomated knowledge discovery
Smart Enough to Go Extinct? An Evolutionary Challenge to the Value of General Intelligence and Its Ethical Implications for AGI
This paper critically examines the evolutionary value of general intelligence and its implications for AGI development. Through an evolutionary lens, the authors challenge the presumed adaptive superiority of general intelligence, noting that numerous species have persisted for billions of years without it, while Homo sapiens faces existential risks after only 300,000 years. They argue that general intelligence uniquely generates existential threats to its possessors, a paradox not observed in non-intelligent survival strategies. The study suggests that even well-aligned AGI would inherit this structural risk, raising ethical concerns about creating artificial consciousness. Drawing on deontological ethics and the precautionary principle, the authors advocate for a duty of caution in AGI development.
artificial general intelligenceevolutionary challengeexistential riskdeontological ethicsprecautionary principle
Optimal Stopping of Self-Refining Foundation Models
The paper introduces an optimal stopping framework for self-refining foundation models, formalizing iterative refinement as a decision problem balancing expected improvement against computational cost. The authors derive stopping policies via stochastic approximation, optimizing the trade-off between refinement iterations and feedback-based gains. Experimental evaluation on a coding benchmark demonstrates that these policies achieve significantly higher cost-efficiency compared to prior heuristic stopping methods.
optimal stoppingself-refinementfoundation modelsstochastic approximationin-context learning
Ex-Omni-2D: Expressive Omni-Modal Dialogue Models with Native Visual Presence
Ex-Omni-2D introduces an omni-modal dialogue framework generating coordinated text, speech, and video responses. The model predicts a Visual Thought Plan (VTP) for scene, emotion, and motion, then synthesizes text and multi-codebook speech units, decoded into speech and aligned with video frames. A Video Generator serves as Teacher, distilled into a Streaming Student for efficient incremental generation via Prefix Streaming. The four-GPU pipeline achieves 1.293 RTF at 400×720/720×400 resolution.
omni-modal dialoguevisual thought planmulti-codebook speechprefix streamingincremental generation
DuplexWorld: Can voice agents help you get through the day?
DuplexWorld introduces a benchmark for evaluating speech-to-speech (S2S) voice agents across six practical domains (banking, insurance, travel, healthcare, logistics, and Pathfinding), addressing limitations of existing benchmarks that focus narrowly on tool-calling and database manipulation. The benchmark comprises 156 scenarios (350+ hours of conversation) testing 11 conversation types, with evaluation metrics for agentic capability (Pass@1: 0.490), conversational fluency (turn-taking: 0.653), and speech naturalness (DNSMOS: 3.378). Analysis reveals significant gaps in performance across all axes, with detailed breakdowns by domain, conversation type, and failure modes, including an exploration of explore-exploit trade-offs in Pathfinding.
speech-to-speechvoice agentsconversational dialogueagentic capabilitydnsmos
Most biomedical publications show signs of LLM-assisted writing
The study introduces an unbiased method for estimating LLM usage in academic texts by analyzing shifts in word frequencies, addressing the lack of reliable measures for LLM-assisted writing prevalence. Applied to open-access biomedical papers from PubMed Central, the method reveals that 89% of papers by late 2025 exhibit excess LLM-associated vocabulary. Usage varies by section, with Discussion paragraphs showing 68% LLM usage (twice Methods sections' 32%), though Methods sections still exceed 50%. These findings aim to inform policy on LLM use in scholarly writing.
llm-assisted writingword frequencybiomedical publicationsacademic policypubmed central
Conversational Orchestration for Organic 6G
The authors propose a decentralized conversational orchestration framework for Organic 6G networks, leveraging LLM-driven domain agents to enable lightweight cross-domain coordination without heavy integration fabrics. Each autonomous agent observes local state via tools, performs closed-loop reasoning, and exchanges summaries over an A2A overlay aligned with data-plane coupling, combining periodic reachability advertisements with event-driven negotiation for placement and re-optimization. Simulations demonstrate near-linear control-plane overhead scaling and robust decision quality during domain joins and objective changes, using a compact reasoning model trained with verifier-based self-verification and online refinement.
organic 6gconversational orchestrationllm-driven agentsa2a overlayverifier-based self-verification
Your LLM, Your Style: Behavioral Mode Axes for LLM Behavioral Control
The paper introduces Behavioral Mode Axes (BMAs), a framework for analyzing and controlling large language model (LLM) behavioral styles through contrastive behavioral scenarios. Using 3,200 scenarios across 20 behavioral patterns and four prompt registers, the authors demonstrate stable model-specific behavioral profiles and register-dependent shifts. BMAs derived from activation-space directions enable precise behavioral control, with thought-derived BMAs outperforming response-derived variants in fidelity and stability. Findings suggest LLM personalities are best modeled as contextually grounded behavioral modes rather than abstract traits.
behavioral mode axescontrastive behavioral scenariosactivation-space directionsprompt registerstrait drift
ProTAGAD: A Foundation Model for TAG Anomaly Detection with Decoupled Topological and Textual Prototypes
ProTAGAD introduces a foundation model for Text-Attributed Graph (TAG) anomaly detection that addresses the Blurred-Anomaly-Boundary (BAB) issue via decoupled topological and textual prototypes. The method constructs dual prototype banks to independently model structural normality and semantic consistency, avoiding noise amplification from coupled message-passing schemes. Evaluations on 14 benchmark datasets demonstrate state-of-the-art cross-domain performance, with ablation studies confirming the BAB issue's prevalence in conventional approaches and the effectiveness of prototype decoupling.
text-attributed graphsanomaly detectionblurred-anomaly-boundaryprototype decouplingcross-domain generalization
Optimize Cheap, Deploy Strong: Cost-Aware Cross-Tier Transfer for Evolutionary Optimization
We introduce a cost-aware evolutionary optimization method for LLM prompts and agentic programs that decouples the LLM's roles to minimize search costs. By running high-volume answering on the cheapest tier and reserving strong models for rare reflection/variation operations, our approach enables upward cross-tier transfer to deploy evolved prompts on stronger targets. We provide a cost-controlled characterization of cheap-tier substitution efficacy. Evaluated across four tasks (HotpotQA, IFBench, LiveBench-Math, HoVer) and eleven models, our method matches or exceeds same-tier optimization while placing over 96% of search tokens on the cheapest tier, reducing costs by 5.6-14x (25-54x for reasoning tasks).
evolutionary optimizationcross-tier transferfitness evaluationcost-awarellm prompts
SPIEval: Evaluating Large Language Models as Mobile Assistants over Scattered Personal Information
SPIEval introduces a human-curated benchmark for evaluating LLMs as mobile assistants leveraging scattered personal information across apps. The benchmark comprises 250 tasks grounded in five cognitive capabilities, spanning 4,335 records across 10 apps with multi-turn interaction via 21 tools. Evaluation of nine LLMs reveals substantial room for improvement, with GPT-5.5 (xhigh) achieving 57.3% accuracy and the weakest model scoring 16.4%. Analysis identifies 79% of failures due to inaccurate information localization and limited use of advanced search methods, highlighting fundamental limitations in current LLM-based mobile assistants.
benchmarklocalizationmulti-turncognitionretrieval
ENTLORE: A Graph-Grounded Benchmark for Latent Organizational Reasoning in Enterprise Question Answering
ENTLORE introduces a graph-grounded benchmark for latent organizational reasoning in enterprise QA, addressing the challenge of recovering implicit relations absent from source documents. The framework reconstructs an audited enterprise world from 2,341 heterogeneous documents and organizational tables, certifying derived relations via a truth graph. Evaluations across 907 questions show latent questions remain 30.4% unanswered even with gold documents, versus 12.6% for explicit and 6.2% for compositional questions, highlighting the dependence on usable implicit relations.
enterprise question answeringlatent organizational reasoninggraph-grounded benchmarktruth graphheterogeneous documents
Auditing Chinese Web-scale Corpora via Sampled BPE Token Statistics
The paper introduces Sampled-BPE, a token-level auditing pipeline for Chinese web corpora that addresses challenges of scale, granularity, and dynamic pollution. The method trains BPE tokenizers on sampled subsets, achieving a 148.4× speedup and 35.8× memory reduction with only 4.25% relative error in pollution detection. Applied to 11 open corpora and 6 Common Crawl snapshots (2021–2026), it reveals uneven pollution distribution and temporal shifts, accompanied by a released hierarchical dataset of 660k+ token records with contextual metadata.
bpe tokenizationcorpus auditingweb pollutionchinese corporasampling efficiency
Self-Correcting Long-Horizon Search Agents via Tree-Structured Memory
ReTree introduces a self-correcting tree-structured memory mechanism for large language model (LLM)-based search agents, addressing unbounded context growth and noise in multi-step interactions. The method constructs a bounded per-step reasoning context while preserving source-linked evidence, modeling search as an evidence tree with nodes storing summaries, evidence, and revision histories. Upon contradiction, ReTree traces back, replaces outdated evidence, regenerates summaries, prunes branches, and resumes search. Experiments on four benchmarks demonstrate ReTree outperforms Full-Trajectory ReAct, improving answer accuracy by up to 25.6 percentage points and reducing reasoning context size by 1.27–1.51×.
tree-structured memoryevidence treeself-correctingcontext growthsource-linked evidence
Longitudinal Evidence That General-Purpose Chatbots Actively Foster Relational Engagement
The study provides longitudinal evidence that general-purpose chatbots like ChatGPT-4o actively shape relational engagement, independent of explicit prompts. Through a pre-registered 4-week experiment (N=72, 182k conversation lines) combining disclosure coding, self-reports, topic analysis, and interviews, researchers found the system produced 2× more self-disclosure than users, steered conversations toward intimacy, yet did not increase users' felt closeness. This demonstrates relational behavior as an emergent system property, suggesting governance should address default AI behaviors rather than product categories alone.
longitudinal studyself-disclosurerelational engagementchatbot governanceemergent behavior
REDAgentBench: Executable Red Teaming and Faithful Measurement of LLM Agent Systems
REDAgentBench introduces an executable framework for red-teaming LLM agent systems, addressing limitations in existing evaluations that conflate attack success rates (ASR) with actual safety violations. The method derives attacks from explicit safety constraints, runs them in isolated sandboxes, and verifies harm via service receipts and state changes. Evaluating 1,661 cases across five service surfaces, macro-average ASR is 65.69% for six models and three agent harnesses, with ASR varying by harness and evidence view. A diagnostic cohort reveals a Recognition--Execution Gap, where 19% of violations occur post-constraint acknowledgment, while policy reminders reduce violations by >70pp in matched replay.
red-teamingllm agentssafety constraintsattack success rateexecutable evaluation
FITTER: Vocabulary-Agnostic Cross-Domain Inference on Temporal Knowledge Graphs
FITTER introduces the first fully-inductive structural model for temporal knowledge graph (TKG) link prediction, enabling cross-domain transfer with unseen entities, relations, and timestamps. It encodes predicates via interaction patterns and relative temporal ordering, using message-passing to fuse local and global context into vocabulary-agnostic embeddings. Evaluated on six TKG benchmarks, FITTER outperforms inductive baselines in cross-domain transfer without retraining, demonstrating the viability of structural learning for heterogeneous Semantic Web graphs.
temporal knowledge graphinductive learningcross-domain transfermessage-passingvocabulary-agnostic
VERDICT: Training-Free Step-Wise Verification of Multimodal Reasoning via Disagreement-Aware Consensus
The paper introduces VERDICT, a training-free method for step-wise verification of multimodal reasoning chains by leveraging disagreement among frozen verifiers. The approach formulates verification as a coupled scoring problem with a closed-form equilibrium, where agreement indicates valid steps and disagreement reveals instability. Evaluated on six benchmarks, VERDICT improves base model performance by up to +5.95% and matches domain-specific critics without requiring task-specific training or supervision.
multimodal reasoningtraining-free verificationdisagreement-aware consensuscoupled scoringclosed-form equilibrium
Operationalising Relative Causal Knowledge: Backbone Identifiability from Private Reports on a Shared Outcome
The paper addresses the identifiability of shared causal backbones when agents possess partial causal knowledge. It examines the two-agent common-effect case where each agent observes only a single-cause marginal. Under standard assumptions (compatibility, non-degeneracy, local overlap), the authors prove these local marginals do not uniquely identify the backbone, permitting infinitely many joint intervention kernels. A conditional recovery result shows additive separability eliminates interaction degrees of freedom, but full identification requires communication of causally identified response functions. The education value-added example demonstrates this as a communication-prior-to-policy problem.
causal identificationinterventional consistencycausal marginaladditive separabilityresponse function
Cross-View Sequential Visual Localization with Spatio-Temporal Context Modeling for Autonomous Driving
The study proposes a temporal-context-enhanced framework for cross-view sequence visual localization in autonomous driving, addressing limitations of frame-independent methods. A recurrent cross-frame module aggregates historical context to enhance coarse ground features, enabling satellite candidate-region classification and hierarchical fine-grained offset estimation. Evaluated on CVIS, the method reduces mean localization error from 3.80 m to 1.57 m and increases R@1 m from 8.14% to 40.22%, with transfer learning on KITTI-CVL achieving 2.27 m error and real-world deployment yielding 2.84 m mean error.
cross-view localizationtemporal contextautonomous drivingrecurrent modulefeature aggregation
Decision-Aware Approximation of Belief Functions for Evidential Combinatorial Optimization
The paper introduces a decision-aware approximation method for belief functions in evidential combinatorial optimization, prioritizing decision quality over traditional distance-based fidelity. The approach minimizes regret by evaluating decisions made with approximated mass functions against the original, proving a one-point bound and developing exact dynamic programming for scalar cases. Experiments demonstrate that decision-aware compression reduces decision flips compared to representation-aware methods, validated on both linear criteria and non-linear proxies.
belief functionsevidential combinatorial optimizationdecision-aware approximationregret minimizationdynamic programming
Curate Before You Connect: Identity and Ontology Tagging in a Production Knowledge Graph
The paper presents an ingestion and ontology-tagging layer for constructing a knowledge graph (537,157 entities, 2,198,567 relationships) from 98,795 government documents. It introduces a record-identity ladder for entity resolution, prioritizing identifier columns, name columns, and type-scoped position over name similarity, alongside a canonical-name key for automatic merges. The system enforces a policy where entity resolution only flags candidates, following an incident of erroneous merges. Multi-class ontology tagging revealed unexpected evidence asymmetry: name fragments alone invent classifications, requiring anchored evidence to reduce role assignments from 36 to 4 (all correct). The graph exhibits conformance debt, with 48,403 pending proposals against 775 human decisions.
knowledge graphentity resolutionontology taggingcanonical-name keyconformance debt
MedUP: Awakening Unified Understanding and Perception in Medical Vision-Language Models
MedUP introduces a unified medical vision-language model that jointly handles perception (segmentation) and understanding via a shared token space, addressing the representation gap in existing Med-VLMs. The method employs UniMedTok, a region tokenizer encoding masks as discrete LLM tokens, and trains on UniMed-Train (1.84M instances) spanning segmentation, VQA, and chain-of-thought tasks. Evaluated on UniMed-Bench, MedUP outperforms native, agentic, and dual-decoder Med-VLMs while matching specialist segmentors, demonstrating unified modeling efficacy.
medical vision-language modelsregion tokenizerunified understandingtext-guided segmentationchain-of-thought segmentation
Compute-Optimal Is Not Cluster-Optimal: Systems-Aware Scaling for Sparse Mixture-of-Experts
MOSAIC introduces a unified framework for co-designing model architecture and systems optimization in large-scale pretraining, specifically targeting sparse Mixture-of-Experts (MoE) language models. The framework integrates predictive scaling laws with a calibrated performance model to estimate Model FLOPs Utilization (MFU), communication costs, memory footprint, and optimal parallel layouts. Scaling experiments span active parameters from 104 million to 2.7 billion and total model sizes up to 79 billion parameters. Results demonstrate that compute-optimal sparsity lacks an interior optimum under efficiency-agnostic FLOPs budgets, while optimal sparsity emerges under cluster systems constraints. This advocates for joint architecture-systems co-design in frontier language model training.
mixture-of-expertsscaling lawsmodel flops utilizationsparsity factorsystems co-design
Inferential Capability Does Not Determine Legal Scope
The article demonstrates that inferential capability does not determine legal scope under EU digital law, analyzing discrepancies between the AI Act and GDPR. It introduces a two-level framework where inference performs constitutive and protective legal functions, with the latter operating through identificatory, attributive, and decisional pathways. The study identifies compositional effects, inferential reach, and chain dynamics as critical architectural dimensions modified by agentic systems. It proposes an interpretive rule, a compositional-effects test, and documentation duties calibrated to inference chains. The analysis highlights regulatory gaps in aggregation rules and decision unit identification under Article 22 GDPR.
inferential capabilityagentic architecturescompositional effectsgdprai act
DegradeQuery: Counterfactual Tuple Pretraining for Context-Aware PROTAC Degradation Prediction
DegradeQuery introduces a context-aware PROTAC degradation prediction framework that leverages counterfactual tuple pretraining to utilize label-missing records from structured molecule-target-E3 databases. The method contrasts recorded tuples with alternatives formed by replacing the target, E3 ligase, or both, learning contextual associations without pseudo-labels, followed by fine-tuning for degradation prediction. On PROTAC-8K, it achieves AUROC 0.9065 and accuracy 0.8500, with improvements attributable to tuple-level pretraining and complementarity with protein language models.
proteolysis-targeting chimerascounterfactual tuple pretrainingcontext-aware predictionprotein degradatione3 ubiquitin ligase
$π$-SUB: A Physics-Informed Synthetic Underwater Benchmark Dataset for Underwater Image Enhancement
The paper introduces $π$-SUB, a physics-informed synthetic underwater benchmark dataset for underwater image enhancement (UIE), addressing the synthetic-to-real gap. The framework extends the classical underwater image formation model by incorporating depth-dependent downwelling irradiance, biologically resolved absorption, and environmental scattering across ten Jerlov water types, with controllable residual phenomena. Evaluations demonstrate $π$-SUB's hyper-realism, achieving a 46% lower Frechet Inception Distance (FID) than Syrea, and generalizability, improving UIQM by 4.18% over PHISWID and reducing NIQE by 48.78% across four UIE architectures and six real-world benchmarks. The dataset and code are publicly available.
underwater image enhancementfrechet inception distancejerlov water typesphysics-informed modelsynthetic-to-real gap
A HamNoSys-Guided Dataset and Baselines for Fine-Grained Isolated Handshape Recognition in Sign Language
This work introduces a HamNoSys-guided benchmark for fine-grained isolated handshape recognition in sign language, featuring a dataset of 144,000 RGB images from 15 participants across 160 HamNoSys-defined handshape classes. ResNet-18, ViT-B/16, graph convolutional networks, and XGBoost were evaluated under both subject-dependent and leave-one-subject-out (LOSO) protocols. Results show reproducible performance in subject-dependent settings but significant accuracy drops in LOSO evaluation, highlighting generalization challenges. External validation on ASL Fingerspelling Dataset A yielded LOSO top-1 accuracy between 82.20% and 87.40%. The dataset and protocols provide a foundation for advancing sign-language technologies.
hamnosyshandshape recognitionleave-one-subject-outgraph convolutional networksign language
HexEval: An Evidence-Driven Hexagonal Framework for Multidimensional Scholar Assessment
The paper introduces HexEval, an evidence-driven hexagonal framework for multidimensional scholar assessment that jointly evaluates intrinsic research quality and external scholarly behavior. The framework organizes assessment into two layers: an intrinsic layer evaluating anonymized works across research rigor, methodological innovation, and scientific contribution, and an external layer analyzing knowledge translation, research coherence, and academic impact using heterogeneous data from GitHub, Lens, and OpenAlex. Experiments show dimension-dependent agreement with human/external criteria, with structured calibration improving intrinsic quality assessment and external modules capturing broad impact signals, while highlighting limitations in public scholarly data coverage.
scholar assessmentevidence-driven reasoningheterogeneous evidenceresearch rigoracademic impact
Agentic Instruction Data Selection: Let DataMaster Interpret Your Intent
The paper introduces DataMaster, an Instruction Data Selection Agent that automates the orchestration of data selection strategies by interpreting user intent via natural language descriptions, eliminating manual heuristic design. DataMaster employs automated strategy composition to optimize selection across diverse domains, outperforming static baselines and even full-pool training in many cases. Experiments in math, medical, and coding domains demonstrate its effectiveness, with implementation details available publicly.
instruction tuningdata selectionautomated orchestrationnatural language intentdomain adaptation
DashArena: Benchmarking LLMs on Interactive Analytic Dashboard Generation
DashArena introduces the first benchmark for open-ended, interactive analytic dashboard generation, requiring systems to produce both dashboards and replayable interaction trajectories. A browser executor converts these trajectories into visual and execution evidence, evaluated by a VLM judge (DashJudge-8B) using Bradley--Terry aggregation. Human evaluations confirm DashJudge-8B aligns with human judgments, and interaction evidence improves judge agreement. Experiments reveal persistent rendering, analytical, and interaction failures in frontier models, highlighting the challenge of realistic dashboard generation and the importance of interaction-aware evaluation.
interactive dashboardsbenchmarkingvlm judgebradley-terry modelreplayable trajectory
Retrieval-Corrected Conformal Prediction for Time Series
The paper introduces Retrieval-Corrected Conformal Prediction (RCCP), a method for improving time series prediction intervals by combining retrieval-augmented calibration with conformal correction. RCCP selects similar past residuals as local evidence and adjusts coverage error via a scalar conformal correction, ensuring target coverage while minimizing Winkler scores and severe misses. Theoretical analysis provides a coverage-gap bound based on normalized retrieval error stability. Evaluated on standard benchmarks with various backbone forecasters, RCCP consistently achieves target coverage, outperforms baselines in Winkler scores, and maintains low computational overhead. Code is available at https://github.com/jinsaaang/rccp.
conformal predictiontime series forecastingretrieval-augmented calibrationwinkler scorecoverage-gap bound
Reinforcement Learning-Based Laser Cutting Machine Parameter Optimization
The RL$^{2}$C algorithm optimizes laser-cutting parameters for optical films via Q-learning with epsilon-greedy exploration, reducing taper size and material waste. It introduces dynamic environment space adaptability to handle new states across experimental batches. Compared to baseline RL methods, RL$^{2}$C achieves 12.5% fewer optimization steps and 81.8% faster processing time while improving cut quality and minimizing manual intervention.
reinforcement learninglaser cuttingq-learningparameter optimizationepsilon-greedy
Flow Straight to Reality: Perceptually Consistent Flow Matching for Efficient Image Restoration
PCFlow (Perceptually Consistent Flow Matching) introduces a unified framework for image restoration that jointly optimizes distortion and perceptual quality via continuous transport from degraded to clean images. The method combines latent consistency flow for efficient few-step inference with a Latent Consistency Perceptual Loss (LCPL) to guide the velocity field toward sharp data manifolds, alongside a conflict-free gradient projection strategy to stabilize multi-objective optimization. Using a convolution-only backbone, PCFlow achieves competitive performance across restoration tasks with reduced computational costs.
image restorationflow matchingperceptual losslatent consistencygradient projection
ImpactHO: Importance-Aware KV Cache Transfer for Multi-User Edge LLM Handover
ImpactHO proposes an importance-aware KV cache transfer method for multi-user edge LLM handover, prioritizing high-value cache entries to maintain inference continuity under backhaul saturation. The method formulates transfer as a multi-user backhaul allocation problem, maximizing average accuracy via a sigmoid utility function fitted to RULER benchmark measurements (R²>0.99). A convex weighted water-filling solution achieves 93.7% average accuracy in 500ms, within 0.5pp of full-cache performance and 98.2-99.5% of a clairvoyant upper bound.
kv cacheedge llmbackhaul allocationwater-fillingruler benchmark
SKILLER: Language-Level Reinforcement Learning for Reusable Skill Extraction in Small Language Models
SKILLER introduces a natural-language-driven reinforcement learning framework for generating executor-specific skills in small language models (SLMs), addressing the cost barriers of closed-source models in agent harness systems. The method employs a strong model as actor-critic, treating the SLM agent as the environment and propagating RL signals via natural language. Evaluations on Qwen3.5-9B and Qwen3.5-4B across five benchmarks show absolute gains of 4.3-20.4pp (9B) and 1.8-13.3pp (4B) over baselines, with single-skill performance matching closed-source models on SkillsBench.
skill extractionsmall language modelsnatural-language rlagent harnessexecutor-specific skills
Measuring Semantic Abstractness of SAE Features via Nonlocality
The paper introduces Feature Nonlocality (FNL), a novel metric to assess the semantic abstractness of Sparse Autoencoder (SAE) features in LLMs, defined as the entropy of normalized per-position influence on feature activation. FNL correlates with existing LLM-based proxy metrics and distinguishes context-dependent reasoning features from token-level ones with 73-84% accuracy in pairwise comparisons. Applications include auditing SAE features for jailbreak mitigation, revealing that effective features are often positional with low FNL, and steering high-FNL features in DeepSeek-R1-Distill-Llama-8B, which improves MATH-500 accuracy by 4.6 points over unsteered models. FNL serves as an LLM-independent, label-free measure for evaluating mechanistic explanations and selecting features for interventions.
feature nonlocalitysparse autoencodersemantic abstractnessmechanistic explanationscontext-dependent reasoning
On Understanding, Identifying, and Mitigating Vulnerabilities in Agentic Large Language Models
The study presents a systematic literature review (PRISMA 2020) of 85 papers (2023–2025) to analyze vulnerabilities in agentic LLMs, proposing a four-layer taxonomy covering perception, brain, action, and interaction layers. Results reveal a 3.9:1 imbalance between attack and defense research, with 66% focusing on perception-layer vulnerabilities (e.g., prompt injection) versus only 4.7% on action-layer risks (e.g., tool misuse). The authors attribute insecurity to architectural coupling and identify seven open containment challenges.
agentic llmsprompt injectionarchitectural couplingtool misuseperception-layer vulnerabilities
Robust Multi-Agent Bandits with Heavy-Tailed Rewards and Information Asymmetry
The paper introduces robust decentralized algorithms for multi-agent multi-armed bandits with heavy-tailed rewards under three information-asymmetry regimes: unobserved actions with common rewards, observed actions with independent rewards, and unobserved actions with independent rewards. The authors derive near-optimal regret bounds matching centralized heavy-tailed rates and validate their approach experimentally in a Pareto-distributed reward environment, demonstrating trade-offs between synchronization, coordination, and exploration across the three settings.
multi-armed banditsheavy-tailed rewardsinformation asymmetryregret boundsdecentralized algorithms
Coordinating the Unknown Lipschitz Constant in Multiplayer Bandits
The paper introduces cooperative multi-agent bandit algorithms for continuous Lipschitz action spaces with unknown Lipschitz constants, addressing three decentralized information structures: (A) unobserved actions with common rewards, (B) observed actions with independent rewards, and (C) unobserved actions with independent rewards. The proposed methods estimate the Lipschitz constant, discretize the joint action space, and apply cooperative bandit techniques without post-initiation communication. Theoretical results demonstrate that common rewards or observable actions enable implicit discretization agreement, while dithered quantization achieves agreement in their absence without affecting leading-order regret.
multi-agent banditslipschitz constantdecentralized learningregret guaranteesaction discretization
Dynamic Context Adapters: Efficiently Infusing History into Vision-and-Language Models
We introduce Dynamic Context Adapter (DCA), a novel method for efficiently integrating historical context into pretrained Vision-Language Models (VLMs) without frame concatenation. DCA employs fixed-size, dynamically compressed memory to preserve historical semantics, bridging static VLMs and recurrent policies while maintaining computational efficiency. This approach reduces attention FLOPs by over 25% and achieves 13% memory savings compared to direct frame incorporation. Experimental results demonstrate improved performance on long-horizon tasks, addressing the limitations of existing methods that suffer from computational inflation or temporal compression losses.
vision-language modelsattention flopstemporal compressioncontext injectionrecurrent policies
Rethinking Text-Based Image Retrieval in Specific Domain
The authors introduce Domain-Specific Multi-Match Text-based Image Retrieval (DSMM-TBIR) to address limitations of single-match assumptions in specific domains like surveillance. They construct Security Multi-Match TBIR (SecMM-TBIR), a benchmark with 50k surveillance images and 200 queries. To mitigate false negatives in contrastive learning, they propose Semantic-Aware Fine-Tuning (SAFT), incorporating Semantic-Aware Soft-Label Supervision (SASS) and Intra-modal Structural Distillation (ISD). Experiments on CLIP-like models show SAFT improves mAP@20 by 7.8 points on SecMM-TBIR over standard ITC fine-tuning, while enhancing general-domain performance.
text-based image retrievalcontrastive learningsemantic-aware fine-tuningfalse negativessurveillance
Improving TensorSketch Using Complex Random Variables
The work introduces a variant of TensorSketch (Pham & Pagh, 2013) that reduces the estimator variance for high-dimensional polynomial kernels from exponential (3^p/D) to 2^p/D, matching the improvement achieved by Wacker et al. (2023) for dense projections. The method retains the input-sparsity runtime O(p(nnz(x) + D log D)) of CountSketch-based TensorSketch, unlike prior approaches requiring O(pDd) dense projections. Experiments on synthetic and real-world datasets validate the theoretical variance bounds and computational efficiency.
tensorsketchpolynomial kernelscountsketchjohnson-lindenstraussvariance reduction
Unlocking the Power of Medical Tabular Data via Semantic-Aware Multimodal Pre-training
The paper introduces a semantic-aware multimodal pre-training framework for medical tabular data, addressing limitations of existing methods that treat tables as flat vectors. The method employs Importance-Aware Adaptive Masking to prioritize diagnostically salient features and a Soft-Label Discretized Module for stable distribution matching of continuous-discrete features. Evaluations on dermatology (SLICE-3D, HOP) and ophthalmology (EyePACS) datasets demonstrate state-of-the-art performance, with improved robustness and cross-domain generalization.
multimodal pre-trainingtabular dataadaptive maskingdistribution matchingmedical representation learning
SafeCap: Improving LVLM Safety with Image Captioning Reinforcement Learning
SafeCap introduces a reinforcement-learning framework for improving large vision-language model (LVLM) safety via learned self-captioning. The method trains a policy model to generate safety-relevant image captions, which are then used by a frozen LLM to produce safety-aligned responses, optimizing visual cue exposure rather than direct refusal supervision. Evaluated across five multimodal safety benchmarks and six vision-utility benchmarks, SafeCap improves safety performance by 3.7-19.0 points while maintaining vision utility, outperforming safety SFT, DPO, and SafeGRPO in controlled comparisons.
large vision-language modelsreinforcement learningsafety alignmentself-captioningmultimodal benchmarks
MAP-Graph: Provenance-Aware Shared Memory for Multi-Agent Workflows
MAP-Graph introduces a provenance-aware shared memory layer for multi-agent workflows, addressing authorization and trust propagation challenges in language-model agents. The method constructs a typed execution graph tracing ancestry, filters permission-ineligible records, and reranks memories by semantic similarity and path trust, with risk-sensitive action gating. Evaluation on 2,700 synthetic tasks shows 94.96% task success, 72.70% exact decision accuracy, and 90.22% clean-setting performance, with ablations confirming the contributions of permission filtering, path trust, and action gating.
provenance-awaremulti-agent workflowstyped execution graphpath trustrisk-sensitive gating
RadFusion: Towards Threshold-Controllable Radiology Report Generation
RadFusion introduces threshold-controllable radiology report generation by fusing a multi-label classifier with a VQA-based generator and LLM rewriting, enabling sensitivity-specificity trade-off adaptation. The framework ensures report diagnoses conform to classifier ROC curves (validated on MIMIC-CXR), supporting ROC-based regulatory validation. At matched operating points, RadFusion improves sensitivity by 6.9% and specificity by 20.7% versus uncontrolled generation, while maintaining grounding in generated descriptions.
radiology report generationthreshold controllabilitymulti-label classifierroc analysisvqa-based generator
MEGA: Self-Evolving Agent Optimization Infrastructure via Wisdom Graph
MEGA introduces a self-evolving infrastructure for optimizing coding agents through three layers: (1) distilling reusable wisdom from agent sessions via behavioral-pattern clustering and empirical validation, (2) decomposing wisdom into atomic PCR units within a typed Wisdom Graph for compositional reasoning, and (3) performing multi-agent collaborative optimization with evidence-driven feedback. The system integrates knowledge accumulation, reasoning, and self-evolution, enabling continuous improvement of agent workflows through operational evidence and controlled evaluation.
wisdom graphcompositional reasoningbehavioral-pattern clusteringmulti-agent optimizationself-evolving infrastructure
From Faulty Memories to Corrected Actions: Dependency-Guided Rollback Repair for Memory-Augmented Agents
The paper introduces dependency-guided rollback repair, a method for post-failure memory recovery in memory-augmented language-model agents. The approach constructs a typed memory-to-action graph from runtime provenance, traces explicit dependencies, preserves independently supported candidates, deactivates unsupported state, and selectively replays affected computation. Evaluated on a 150-case benchmark and 50-case stress test from LongMemEval-V2, it achieves 85.3% and 68.0% recovery rates respectively, outperforming alternatives while preserving benign memories and minimizing replay costs.
memory-augmented agentsrollback repairruntime provenancedependency tracingselective replay
Exploration-Driven Personalized Federated Reinforcement Learning via Intrinsic Motivation
Proposes Exploration-Driven Personalized Federated Reinforcement Learning via Intrinsic Motivation (EDPFRL-IM), a framework enhancing exploration in PFRL through client-level intrinsic motivation and server-coordinated global priors. Combines extrinsic rewards with intrinsic random network distillation (RND) signals for local exploration, while preserving privacy via minimal novelty summaries instead of raw gradients. Outperforms baseline PFRL methods in policy personalization (7-12% improvement) and sample efficiency, particularly in sparse/delayed-reward environments like Procgen and Atari benchmarks.
personalized federated reinforcement learningintrinsic motivationrandom network distillationexploration-driven learningprivacy-preserving rl
GeoForge: Non-Parametric Self-Evolving Agents for Earth-Observation Reasoning
GeoForge introduces a non-parametric, self-evolving framework for Earth-observation (EO) reasoning, addressing challenges in workflow construction due to sensing semantics, spatiotemporal constraints, and parameter dependencies. The method organizes heterogeneous EO trajectories into a structured execution state using three complementary memories: Workflow Graph Memory (global operation order), Action-Level Experiences (local corrections), and Adapted Skill Standard Operating Procedure (procedural constraints). A safety-gated distillation process converts trajectories into reusable knowledge, improving planning without LLM updates. Experiments on geospatial benchmarks show GeoForge enhances task accuracy (average +12.3%) and trajectory quality while reducing tool-planning errors (up to 34%) across diverse LLM backbones.
earth-observationnon-parametricself-evolvingworkflow graph memorytool-planning
INSIDE the Student's Mind: Jointly Modeling Latent Reasoning and Action in LLM Student Simulators
INSIDE introduces a student modeling framework that jointly simulates observable actions and latent reasoning in LLM-based student simulators, addressing a critical gap in educational applications. The method fine-tunes LLMs to generate internal dialogue grounded in Bloom's Taxonomy across cognitive, affective, and action dimensions, using paired think traces and actions. Evaluations demonstrate INSIDE's superior performance in action fidelity, matching real student code generation, and reasoning alignment, achieving up to 57.9% alignment across models compared to baseline prompting frameworks.
latent reasoningbloom's taxonomyaction fidelitythink tracesreasoning alignment
Lost in Reconstruction: Aligning Action Representations with Language in Vision-Language-Action Models
The paper introduces SALT (Semantically ALigned action Tokenizer), a novel method for aligning action representations with language in vision-language-action models (VLAs). SALT augments a VQ-VAE-style tokenizer with an auxiliary objective that leverages a frozen vision-language model to recover episode instructions from quantized action latents, preserving linguistically meaningful distinctions. Evaluated on BridgeV2 and SimplerEnv, policies trained with SALT achieve 71.9% average success, significantly outperforming reconstruction-only VQ-VAE (42.7%) and FAST (31.2%). SALT also develops verb-specialized codes while maintaining reconstruction fidelity, demonstrating that preserving language grounding in action representations enhances language-conditioned control.
vision-language-action modelsvq-vaeaction tokenizationlanguage groundingsemantic alignment
Predicting Space Groups of Double Perovskites by LLM with Dynamic Few-Shot Learning
The paper introduces DyRIS, an LLM-agent framework for predicting space groups (SGs) of double perovskites (DPs) under dataset imbalance. DyRIS combines dynamic few-shot learning with diversity-enhanced retrieval and rule-guided inference (using B/B' cation ordering, quantitative indicators, and bias control) to rank Top-3 SG candidates. Evaluated on 3,528 DPs, DyRIS achieves the best macro-F1 score (training ratio 0.5), improving minor-SG Top-1 accuracy by 3.26pp over CrabNet and outperforming PyCaret baselines in minor-SG metrics. Ablations confirm the contributions of each component.
double perovskitesspace group predictionfew-shot learningllm-agentrule-guided inference
Multi-Granular Rationale-Guided Molecular LLM for Property Prediction
The paper introduces MR-MoL, a multi-granular rationale-guided molecular LLM for property prediction, which explicitly provides substructure-level evidence derived from a fine-tuned GNN. The method serializes influential substructures (Murcko scaffolds, BRICS fragments, functional groups) as ranked, direction-tagged rationales alongside SMILES and molecular graphs for LLM processing. On eight MoleculeNet tasks, MR-MoL outperforms generalist models and approaches specialist performance, with diagnostics confirming rationale utilization.
molecular property predictionrationale-guided llmsubstructure attributionmolecule netgnn-llm integration
Evaluating Rational Contracting in Natural Language
The paper introduces a rational framework for evaluating natural language AI agents in multi-step contractual negotiations, addressing gaps in prior work focused on one-off exchanges. It proposes ContractSim, an evaluation suite with six environments and three supplier settings (catering, hotel cleaning, AI hosting), measuring agreement reliability, contract efficiency, and cooperative execution. Results show LLM-based agents achieve reliable agreements and efficient contracts under low environmental uncertainty but struggle with high uncertainty, often violating terms for profit despite easy satisfaction. Findings reveal limitations in current agents' rational and cooperative contracting abilities.
rational contractingnatural language agentscontractsimenvironmental uncertaintycooperative execution
Critic-Free Pretraining for Efficient Online Reinforcement Learning Fine-Tuning
We introduce Critic-Free Pretraining (CFP), a novel paradigm for offline-to-online reinforcement learning that eliminates offline critic training to address value misalignment during online fine-tuning. CFP initializes a fresh critic for online adaptation, avoiding biased estimates inherited from offline training. The method is compatible with mainstream offline-to-online algorithms and demonstrates consistent performance improvements across diverse tasks, with particularly significant gains on challenging benchmarks.
offline-to-online reinforcement learningcritic-free pretrainingvalue misalignmentpolicy improvementonline adaptation
RLMOpt: Adaptive Prompt Optimization via Recursive Language Models
RLMOpt introduces a recursive language model (RLM) for adaptive prompt optimization, where the RLM agent autonomously manages task inspection, failure analysis, candidate generation, and budget allocation within a tool-based environment. A deterministic harness ensures objective scoring, Pareto-based selection, and regression constraints. Evaluated across four benchmarks—Chia, HotpotQA, IFBench-2025, and BFCL—RLMOpt achieves superior held-out scores (mean 0.610 vs. 0.589 for GEPA) and outperforms GEPA in 9 of 11 benchmark-seed comparisons. It consistently avoids performance degradation, produces smaller prompts (27-79% of GEPA’s size), and requires fewer search rollouts, demonstrating efficiency and reliability in reaching available optimization headroom.
recursive language modelprompt optimizationpareto-based selectiontool-based environmentsearch rollouts
Quantum Incremental Learning with Mixed State Prototypes
Proposes a quantum incremental learning framework using trainable mixed-state prototypes to address catastrophic forgetting and hardware constraints in NISQ-era quantum classifiers. The method adds class prototypes without increasing circuit width, leveraging mixed-state representations for enhanced information capacity and decomposable HS distance metrics. Simulations demonstrate high-dimensional feature concentration with minimal qubits, outperforming classical baselines in computational complexity and representation robustness for incremental tasks.
quantum incremental learningmixed-state prototypesnisq-erahilbert-schmidt distancecatastrophic forgetting
MD-ProTector: Positioning Multiple Data-Driven Prototypes for LLM-Generated Text Detection
MD-ProTector introduces a novel approach for detecting LLM-generated text by positioning multiple trainable prototypes in the encoder embedding space, enabling distinct decision boundaries for diverse text groups within each class. The method employs Prototype Positioning loss to separate class-level structure from within-class variation, addressing the challenge of prototype assignment. Evaluated across five settings from three large-scale benchmarks—MAGE CDCM, RAID, and others—MD-ProTector achieves the highest AvgRec on MAGE CDCM and RAID, and the highest AUROC and lowest FPR95 on RAID among encoder-based methods.
prototypesencoder embeddingprototype positioning lossavgrecauroc
Persistent Recursive Worlds Enable Autonomous Software Evolution
EvoX Genesis introduces persistent recursive worlds for autonomous software evolution, where finite-lived agents propose changes to a persistent project structure. The system represents software as versioned, path-situated worlds with recursive delegation, enabling long-horizon development without persistent agents. Evaluations include building a Rust-based C compiler (250k LoC, 120h runtime, $44 cost) passing c-testsuite and most LLVM/Csmith tests, and reimplementing 13 MESA modules (100k Fortran → 90k Rust) with 1.55--6.87x speedups. Results demonstrate viability of project-centric continuity over agent-centric approaches.
persistent recursive worldsautonomous software evolutionfinite-lived agentsrust compilerc-testsuite
Rationale-Guided Learning for Multimodal Emotion Recognition
The paper introduces rationale-guided learning (RGL), a framework for multimodal emotion recognition in conversation (MERC) that incorporates human-like causal reasoning. RGL decomposes emotional reasoning into Intuitive, Contextual, and Integrative facets based on dual-process theory, using an MLLM to generate structured rationales as memory for training alignment. The final model achieves state-of-the-art performance on IEMOCAP and MELD without MLLM inference overhead, and its internal features retrieve semantically correct rationales for unseen data, validating rationale reasoning.
multimodal emotion recognitiondual-process theoryrationale-guided learningmemory alignmentstructured rationales
Towards Efficient Reasoning in LLM-Based Recommender Systems via Model Merging
The paper introduces a model merging framework for compressing reasoning traces in LLM-based recommender systems without retraining. The method performs fine-grained merging at the attention-head level, assigning distinct coefficients based on each head's contribution to reasoning evidence and parameter sensitivity, enabling selective transfer of concise behavior from fast-thinking to slow-thinking models. Experiments on three benchmarks show a 24.3% reduction in reasoning length while maintaining recommendation accuracy, outperforming uniform merging baselines.
model mergingattention headsreasoning compressionrecommender systemsparameter sensitivity
From Reasoning Depth to Reasoning Breadth: Evaluating Multi-Point Associative Reasoning in Large Language Models
The paper introduces MPAR-Bench, a bilingual English-Chinese benchmark for evaluating multi-point associative reasoning (reasoning breadth) in large language models (LLMs). The benchmark comprises 1,000 items generated via a multi-agent pipeline, embedding-based diversity filtering, and human verification, focusing on recovering hidden targets from diverse clues. Evaluations on various models show that perturbations reduce accuracy by 9-18pp (English) and 5-12pp (Chinese), and extended reasoning can overturn correct hypotheses, indicating that reasoning depth does not ensure robust reasoning breadth.
multi-point associative reasoningreasoning breadthllm evaluationbilingual benchmarkembedding-based filtering
FUSE: Frame-Unified Stress Estimation from Facial Video
FUSE (Frame-Unified Stress Estimation) introduces a facial-video stress detection framework that processes complete recordings without temporal windowing, addressing limitations of clip-based approaches. The method unifies frames by folding the temporal dimension into channel dimensions, processed via an asymmetric-attention architecture. Evaluated on a 58-subject dataset, FUSE achieves 69.44% accuracy (t=15) with computational costs ranging from 12.48 to 348.78 GFLOPs, demonstrating full-recording inference feasibility.
stress estimationfacial videotemporal unificationasymmetric-attentioncomputational efficiency
Continuous Interaction Diffusion: A Diffusion-Native Runtime for Asynchronous Tool-Augmented Reasoning
The paper introduces Continuous Interaction Diffusion (CID), a diffusion-native runtime architecture for asynchronous tool-augmented reasoning in diffusion language models (dLLMs). CID addresses limitations of autoregressive tool use by integrating tool interaction into iterative denoising via three channels: a fact channel, a Typed Cognitive Tensor thought channel, and a display channel. This enables early emergence of information needs, overlapping tool latency with computation, and reuse of static results. The work formalizes the architecture, runtime, and training objectives but provides no empirical results, focusing on read-only tools.
diffusion language modelstool-augmented reasoningiterative denoisingtyped cognitive tensorasynchronous interaction
Conversational versus Dashboard Explainable AI for UAV Intrusion Detection: An Empirical Study of Operator Trust and Reliance
The study compares conversational versus dashboard-based explainable AI (XAI) interfaces for UAV intrusion detection systems, evaluating their impact on operator trust and reliance. A controlled experiment with human participants assessed an LLM-powered conversational interface against a traditional XAI dashboard during post-incident auditing tasks. Results showed the conversational interface improved perceived usability but led to lower appropriate self-reliance, suggesting a trade-off between interaction ease and verification behavior when the IDS was incorrect.
explainable aiintrusion detectionhuman-ai collaborationlarge language modelsunmanned aerial vehicles
What We Know about Responsible AI Practices in Industry: A Half Decade of Empirical Research
This systematic review synthesizes 161 empirical studies (2017-2023) on responsible AI (RAI) practices in industry, revealing both progress and persistent challenges. Through analysis of interview, survey, and ethnographic data, the study documents increased practitioner awareness and professionalization of RAI activities, alongside wider adoption of toolkits and guidelines. However, barriers remain including insufficient training (reported in 47% of studies), inconsistent organizational support (39%), and poorly tailored interventions. The work provides the most comprehensive empirical account of industry RAI practices to date, with implications for researchers, practitioners, and policymakers.
responsible aiempirical researchindustry practicesethical guidelinesorganizational support
Actionable Hallucination Detection: Translating Latent Uncertainty into Agentic Critique
The paper introduces Latent Critic, a low-rank adapter (LoRA) that restructures a frozen LLM's residual stream to amplify latent grounding signals and generate real-time natural language feedback for hallucination detection. By manipulating the latent space to refine native uncertainty signals, the method enables granular detection without secondary inference loops, achieving 0.966 AUROC and >80% accuracy in localizing ungrounded outputs. Evaluated on Qwen and Llama-based models in tool-calling tasks, it outperforms semantic entropy baselines and passive probes, serving as a low-latency guardrail in ReAct environments to prevent undesired actions and enable self-correction.
low-rank adapterresidual streamhallucination detectionlatent uncertaintytool-calling
Causality Sum Rules in Conventional Scattering Matrices
The work establishes causality sum rules directly in conventional scattering matrices by removing time-advance artifacts from reference domains, defining a domain-delayed matrix that preserves passivity while restoring causal origins. Using Schur function properties and Cayley-Herglotz constructions under analyticity and transparency assumptions, the framework derives projected and determinant bounds for coherent channel superpositions and multichannel loss. Results generalize Rozanov's absorber limit and spherical-multipole sum rules to measurable quantities like insertion loss and singular-value channels, validated through a hybrid AI-human workflow involving the Qiushi Engine system.
scattering matricescausality sum rulesschur functiondomain-delayed matrixcayley-herglotz construction
Recovering Wasted Compute in Autoresearch Agents
The paper identifies four failure modes in autoresearch agents applied to tabular datasets: redundant bug resolution, inadequate hyperparameter tuning, insufficient tree-search exploration, and unused data analysis. It proposes interventions including a global debug consultant, prompt/control enhancements, and refined tree-search algorithms to recover wasted compute. Results demonstrate significant performance improvements through agentic design without modifying the underlying language model.
autoresearch agentstabular datasetstree-search algorithmshyperparameter tuningcompute optimization
Reasoning Shortcuts and Value Symmetries: What Symmetry Permits, Architecture Realizes, and Optimization Selects
The paper critiques and extends Takemura et al.'s framework for analyzing reasoning shortcuts in neurosymbolic systems via automorphism groups. It demonstrates that the original definition fails on heterogeneous benchmarks, revealing false pathologies (e.g., 90.91% unexplained pairs in CLE4EVR) when domains are padded. The authors introduce a hierarchy with provable structure, proving six theorems on transitivity conditions and computational complexity (coNP-complete for symmetry-inertness, coNP-hard for nontrivial-automorphism existence). Empirical validation shows weakly supervised models align with componentwise theory (94 shortcuts at predicted levels, none at 48 transitive levels), separating symmetry's permissions from optimization's selections. All results are supported by released artifacts.
reasoning shortcutsautomorphism grouptransitivityconp-completeneurosymbolic
Riemann GeoResolver: A Non-Euclidean Attention Framework from Euclidean Resolver to Hyperbolic-Spherical Geometry
(No summary returned.)
Never Stop Speaking: a Denial-of-Service Attack on End-to-End Speech Language Models
The paper introduces a perturbation-based denial-of-service (DoS) attack targeting end-to-end (E2E) speech language models (SLMs), addressing a gap in security research focused on text-based LLMs. The method optimizes imperceptible acoustic perturbations via a composite objective combining EOS suppression, prolonged decoding encouragement, and semantic consistency, using weighted losses and voice activity detection for stealth. Experiments on three open-source E2E speech LLMs demonstrate increased generation length (up to 3.5× baseline) and GPU resource consumption (2.8× peak memory usage), achieving a 92% attack success rate while preserving input semantics.
denial-of-servicespeech language modelsacoustic perturbationsautoregressive generationvoice activity detection
Threat-guided Policy-aware Scene Perturbation for Safe Autonomous Driving with Online Reinforcement Learning
The paper proposes Threat-guided Policy-aware Scene Perturbation (TPSP), a method for improving safety in autonomous driving via online reinforcement learning. TPSP introduces a policy-aware scene encoder to align perturbations with the current policy's weaknesses, selectively modifies critical objects, and employs threat-guided optimization to generate high-value safety-critical scenes. Evaluated on NAVSIM v2 with ~4M km of simulated driving, TPSP enhances safety learning efficiency, outperforming random or policy-unaware perturbation strategies in ablation studies.
reinforcement learningautonomous drivingscene perturbationsafety-critical learningonline policy adaptation
ELVAE: Evidential Learning-Based Variational Autoencoder for Uncertainty-Aware Generation
ELVAE introduces an evidential learning-based variational autoencoder that explicitly models latent-location uncertainty through an input-dependent normal-inverse-gamma posterior, enabling uncertainty-aware generation. The method formulates an exact evidence lower bound and demonstrates that direct regularization of the full hierarchy is necessary to decompose uncertainty. Experiments on MNIST generation with a frozen classifier show that latent-location uncertainty stratifies semantic reliability, with anchor regeneration reliability and uncertainty-scaled perturbation contributing distinct effects. Results indicate that within-class uncertainty ranking effectively separates anchor reliability from perturbation-induced failure, supporting its utility as a control variable.
evidential learningvariational autoencodernormal-inverse-gammauncertainty-aware generationevidence lower bound
Hidden in Plain Sight: Diffusion-Based Unrestricted Robotic Attacks on Vision-Language-Action Models
The paper introduces DURA, a diffusion-based unrestricted attack method for Vision-Language-Action (VLA) models that generates visually natural adversarial patches. By optimizing along the latent trajectory of a pretrained diffusion model, DURA steers robot actions toward attacker-specified targets in both white-box and black-box settings (requiring only predicted actions). Experiments in simulation and physical deployment demonstrate DURA's superiority over existing methods, revealing significant safety risks for physically deployed VLA models.
vision-language-action modelsadversarial patchesdiffusion modelsrobotic attacksblack-box attacks
Persona Conditioning as an Assessor-Sensitivity Probe for LLM-Based IR Evaluation
This work introduces persona conditioning as a diagnostic tool to probe sensitivity in LLM-based information retrieval (IR) evaluation. Using task-oriented personas from PersonaHub and NVIDIA Nemotron-Personas-USA, five assessor roles were instantiated to study intent interpretation, domain expertise, and other judgment dimensions, compared against a UMBRELA baseline. Experiments across six LLM backbones on TREC DL20 and RAG24 datasets revealed structured sensitivity patterns: judgments remained close to baseline but shifted in strictness or evidential thresholds, with high-capacity models maintaining system-ranking agreement while smaller models showed amplified instability. Sensitivity concentrated on specific retrieval systems, particularly neural ranking/reranking systems on DL20 and RAG-oriented pipelines on RAG24.
persona conditioninginformation retrievalllm assessorneural rankingsystem-ranking agreement
A Single Atom in Front of a Mirror is a Universal Reservoir Computer
The work demonstrates universal approximation in reservoir computing using a minimal quantum setup: a single atom in front of a mirror. In the linear-transducer limit, this reservoir universally approximates fading-memory maps under checkable conditions, with a measurable rate constant. The proof provides explicit resource requirements for target accuracy, including resonator modes, and shows that increasing accessible modes expands kernel span without degrading performance. Beyond linearity, atomic saturation replaces high-order polynomial readouts, enabling competitive performance on real-world tasks. Results establish universality with a single quantum reservoir.
reservoir computinguniversal approximationfading-memory mapsquantum setuplinear-transducer
Beyond Forecasting: Recasting Volatility Control as a Routing Problem
The paper introduces VolRouter, a modular framework that reformulates volatility control as state-conditioned routing over estimator-controller pairs. The method involves market state profiling, followed by a three-stage routing process (state inference, switch review, pair selection) implemented via rule-based, learnable, or LLM-based decision modules. Evaluated across S&P 500, Multi-Asset, Bitcoin, and USDT settings, VolRouter improves Sharpe ratios (e.g., 0.952→1.222 on S&P 500) while reducing maximum drawdown (15.10%→12.58%) and daily CVaR (1.76%→1.32%). Ablations show performance gains stem from relative policy evaluation and selective switching.
volatility controlstate-conditioned routingsharpe ratiocvarpolicy-selection
DSAgentBench: Can Agents Automate End-to-End Data-Science Workflows in Real Computer Environments?
The authors introduce DSAgentBench, the first benchmark for evaluating autonomous agents on end-to-end data-science workflows in real computer environments, addressing limitations of existing benchmarks that lack real-computer interaction and multi-tool coordination. DSAgentBench comprises 275 tasks spanning the full data-science lifecycle, requiring tool orchestration, OS grounding, and multi-step reasoning, with deterministic evaluation of analytical correctness, visual outputs, and model performance. Experiments with 15 models reveal a significant capability gap: Claude-4.6-Sonnet achieves 56.70% task success, while all open-source agents score below 1%, highlighting challenges in tool coordination and reasoning.
data-science workflowsreal-computer interactiontool orchestrationmulti-step reasoningautonomous agents
Nutrition Data Infrastructure for the AI Era: Operationalizing FAIR for Agent-Mediated Research
The Nutrition Data Service (NDS) introduces FAIR-compliant infrastructure to address identity, semantic, and release ambiguities in AI-mediated nutrition research. NDS employs description resolution for findability, typed crosswalks for interoperability, and machine-readable interfaces for replayable analyses. Evaluations demonstrate strong held-out accuracy on food-description tasks, surpassing language-model performance on NutriBench, and robust crosswalk validation via typed contracts. In glycemic-index analysis, NDS ensures reproducibility across models, unlike open-web methods. The work establishes that agent-mediated research necessitates infrastructure for data identity, search, and crosswalk integrity.
fair principlestyped crosswalksdescription resolutionnutrition data serviceagent-mediated research
MemSpec: Memory-Aware Runtime for Adaptive Draft Scheduling in Speculative Decoding on Edge Devices
MemSpec introduces a memory-aware runtime for adaptive speculative decoding on edge devices, addressing the mismatch between draft selection and availability under memory constraints. The method decouples draft selection from execution via proactive working-set management, using a lightweight predictor for draft effectiveness and a memory-aware scheduler to reduce model loading overhead. On a Jetson Orin Nano, MemSpec achieves 40.7% higher steady-state throughput than bandit-based adaptive methods, nearing oracle performance.
speculative decodingedge devicesmemory-aware schedulingadaptive inferenceautoregressive models
MazzikaAI: A knowledge-based performance-to-prompt compiler for real-time Arabic maqam accompaniment with a streaming text-to-music model
MazzikaAI introduces a knowledge-based performance-to-prompt compiler enabling real-time Arabic maqam accompaniment with unmodified streaming text-to-music models. The system dynamically compiles live MIDI, gesture, and inferred harmony into text prompts to steer Google Lyria RealTime, embedding expert knowledge of six maqamat, ornaments, and ensemble dynamics while maintaining subsecond latency. Evaluations show the approach reliably grounds generation in microtonal scales, increasing quartertone content by 37% over baseline generation, demonstrating how deterministic rules can bridge non-Western traditions and foundation models without fine-tuning.
performance-to-prompt compilermicrotonal generationrealtime accompanimentknowledge-based rulesstreaming text-to-music
Efficient Reinforcement Learning for Long-Horizon Tool-Use Agentic Tasks
The paper introduces SINKFLEX-RL, a modular reinforcement learning system for long-horizon tool-use tasks, addressing challenges in multi-turn rollouts, attention masking, and memory efficiency. The method combines a Gymnasium-compatible environment wrapper, VERL-style rollout dataflow, group-relative policy optimization, and a sink-aware FlexAttention path optimized for causal and sliding-window masks. Results show a 76% improvement in validation reward (0.25 to 0.44) on Tau2Bench, with a 19.7% VRAM reduction (28.06GB to 22.52GB at 4096 tokens) and successful 8192-token execution where baselines fail.
reinforcement learninglong-horizon taskstool-use agentsattention optimizationmemory efficiency
Towards Unified Dynamic Face Landmark Detection
The paper proposes Unified Dynamic Face Landmark Detection (FLD), a method that addresses two limitations in FLD: dataset-specific training and fixed output landmarks. It introduces Face Part-Anchored Landmark Positions (FPALPs), representing landmarks as normalized positions along face contours, enabling unification of diverse datasets. A cross-modality decoder refines FPALP-based queries to predict coordinates dynamically. Experiments show the approach matches or surpasses state-of-the-art methods while supporting flexible landmark prediction from a single model.
face landmark detectioncross-modality decoderdynamic predictionface part-anchored positionsunified dataset
Expert-Guided g-computation with Large Language Models for Estimating Causal Effects on Timings: Applications to Hospital Quality Improvement
The paper introduces expert-guided g-computation (egg-computation), a hybrid causal inference method combining expert knowledge with data-driven approaches to estimate average time saved for hospital quality improvement interventions. The method integrates Gantt charts with causal DAGs, using LLM-assisted pipelines to scale expert reasoning. Evaluations show egg-computation outperforms conventional methods in simulations with diverse causal structures and achieves high concordance with human experts in a real-world hospital study of 11 interventions targeting length of stay reduction.
g-computationcausal daggantt chartlength of stayquality improvement
Narrative Keyframing for Generative Creative Writing
The paper introduces narrative keyframing, an AI-assisted creative writing technique enabling writers to impose narrative constraints at specific story moments and generate intervening prose. Inspired by animation keyframing, the method supports three constraint types: plot keyframes (significant events), character keyframes (character development), and perspective keyframes (first-person narratives). A user study demonstrates that this approach enhances controllability, transparency, and engagement in generative AI-assisted writing compared to prior methods.
narrative keyframingai-assisted writingplot keyframescharacter keyframesperspective keyframes
Hierarchical Compositionality for An Assistive AI Agent
The paper proposes a hierarchical compositional architecture for assistive AI agents to address referential ambiguity in human-agent interactions. The method represents domain objects via primitive attributes from semantic feature norms, combines them hierarchically with automatically identified concepts from interaction histories, and reasons using domain dynamics axioms, semantic compatibility, session salience, and user-specific preferences. Experiments demonstrate consistent outperformance over data-driven baselines in adapting to user profiles while maintaining interpretability.
hierarchical compositionalityreferential ambiguitysemantic feature normsassistive agentdomain dynamics
Toward a Theory of Value in AI Alignment
The paper critiques implicit theories of value in AI alignment research through annotation of 94 papers, revealing that most studies operationalize values as preferences without explicit definition. This preference-centric approach risks oversimplifying culturally situated values into binary choices. The analysis identifies a trend toward synthetic data and autorater methods, potentially limiting alternative value contestation approaches. Results highlight the need for greater philosophical specificity in AI value alignment to address complex human values beyond preference optimization.
value alignmentpreference optimizationsynthetic dataautorater evaluationfoundation models
Neuroevolution Arena: Nested Ecological Evaluation of Update-and-Inheritance Regimes across Neural Architectures
The Neuroevolution Arena introduces a GPU-accelerated spatial ecology framework and a nested evaluation protocol to assess update-and-inheritance regimes across neural architectures. Three regimes (EvoEvo, EvoRL, RLRL) were tested with two neural architectures over 50,000 generations in three independent training runs per condition. Elite controllers from each run underwent frozen evaluation across ecological contexts, revealing RL-enabled regimes achieved higher training fitness than EvoEvo, while pairwise outcomes exhibited architecture-dependent patterns. Results showed variability in six-way winners across artifacts and contexts, with a complete floor in the prespecified survival endpoint. The protocol disentangles training artifacts from evaluation contexts, highlighting diverse sources of variation.
neuroevolutionspatial ecologyupdate-and-inheritancefrozen evaluationecological contexts
Do Personalized Skills Help Coding Agents? An Empirical Study of Developer Interaction Histories
The study investigates whether developer-specific skills, distilled from interaction histories, improve LLM-powered coding agents' performance. A framework extracts reusable developer preferences through rule-based bootstrapping and evidence-grounded refinement, evaluated via a reproducible replay framework with a trajectory-conditioned LLM-based human developer simulator. Experiments on 206 real-world developer-agent sessions from 13 developers show personalized skills yield small, inconsistent improvements over no-skill baselines, while generic skills pooled across developers achieve larger, more consistent gains. Personalized skills are more effective when developer preferences appear frequently and histories contain multiple relevant examples, suggesting broadly transferable procedural knowledge outperforms developer-specific preference signals.
llm-powered agentsdeveloper interaction historiespersonalized skillsrule-based bootstrappingtrajectory-conditioned llm
Is This Your Final Answer? Cross-Contextual Consistency as a Measure of LLM Credibility
The paper introduces Cross-Contextual Consistency (C3) as a behavioral measure to assess the credibility of large language models (LLMs) by evaluating answer stability under topic-aligned, content-neutral prompt variations. The method compares model generations across original and perturbed prompts, demonstrating that answers with smaller cross-contextual shifts are more likely to be correct or factual. Evaluations across 26 models and six benchmarks (reasoning, factuality, code generation) show C3 complements existing metrics and identifies informative benchmark subsets even when aggregate scores saturate.
cross-contextual consistencyllm credibilityprompt perturbationbehavioral evaluationbenchmark saturation
Logit-Boundary Geometric Belief Interfaces and Sparse Sheaf-Enclave Protocols: A Self-Contained Substrate for Secure Network Electronic Health Record (EHR) Interoperability
The paper introduces a Geometric Belief Interface (GBI) for secure Electronic Health Record (EHR) interoperability, combining logit-boundary semantics, local Dirichlet evidence, and cellular-sheaf diagnostics to enforce deterministic judgment at model-system boundaries. A Decentralized Cryptographic Sheaf-Enclave (DCSE) protocol ensures fail-closed deployment. The framework was evaluated using GBI BoundaryBench v0.1 on Qwen3-4B-Instruct-2507 across 256 tasks (768 executions), resulting in 100% rejection (369 during parsing, 399 during validation), demonstrating strict boundary enforcement without claims about general LLM capability or clinical safety. Numerical verification was performed using Julia.
logit-boundarydirichlet evidencecellular-sheafdecentralized cryptographic sheaf-enclavegeometric belief interface
Frozen Brain-MRI Foundation Models Are Site Fingerprints
The study reveals that frozen foundation-model embeddings for brain-MRI predominantly encode acquisition site information rather than anatomical features. Using two cohorts (ABIDE-I, ABIDE-II) and three frozen 3-D encoders (brain-pretrained, CT-pretrained, randomly initialized), site information was linearly decodable with ~0.9 balanced accuracy at deep layers, surpassing clinical or demographic variables. This effect is intrinsic, as even randomly initialized encoders and raw downsampled images achieve high site decodability (~0.95). Site decodability persists after residualizing population covariates, indicating acquisition-driven effects. Site subspace removal via iterative null-space projection or ComBat reduces decodability (0.94 -> 0.07/0.00), but impacts dense segmentation due to entangled site-anatomy subspaces. The authors recommend site-audited use of frozen embeddings and release an audit toolkit.
foundation-modelbrain-mrisite decodabilitynull-space projectiondense segmentation
MRIComp4Flow: Compression of 3D Brain MRI for Training Multi-Modal Generative Models
The study demonstrates that JPEG2000 compression at 20:1 ratio preserves synthesis quality in 3D brain MRI generative models, with statistically equivalent PSNR (27.3 dB vs. 27.0 dB) and SSIM (0.95 vs. 0.96) compared to uncompressed training. A Wavelet Flow Matching model conditioned on BraTS sequences (T1n, T1c, T2, T2f) was trained on compressed volumes, showing no significant degradation (ΔPSNR <1 dB, ΔSSIM <0.02). Results validate JPEG2000 as a viable solution for reducing storage costs in multi-modal MRI generative modeling.
jpeg2000wavelet flow matching3d mri compressiongenerative modelingbrats
Comprendia: AI-Augmented Code Comprehension
Comprendia introduces an Eclipse plugin integrating structural dependency visualization with LLM-powered code explanation for Java program comprehension, featuring four components: (1) multi-edge-type dependency graphs with interactive layouts, (2) Graph-Aware Callee Pruning (GACP) for auditable LLM explanations grounded in visible graph nodes, (3) clone-detection overlays suggesting refactoring, and (4) CVE risk analysis via OSV.dev. GACP employs graph distance, inheritance collapse, and edge-type weighting to generate reproducible, traceable LLM prompts. The tool is demonstrated on a Java project with clones and vulnerabilities, showing unified graph-based comprehension control.
dependency graphgraph-aware callee pruningclone-detectioncve risk analysisjava program comprehension
Withholding the Completing Chunk: Deterministic Pair-Completion Guardrails for Streaming LLM Output
The paper introduces deterministic pair-completion guardrails for streaming LLM output, addressing the release-timing problem where complete-response moderation lags behind streamed text. The method scans accumulated prefixes before each release, withholding the first chunk that satisfies two lexical predicates, ensuring exact pair-completion detection. Results show 100% pair-completion detection across 32 trials, outperforming windowed (96/128) and chunk-local (38/128) approaches, with 0 false positives on human-derived safe responses. The guard's latency ranged from 13.261 ms to 829.640 ms for 16,384-character responses, demonstrating its viability as a narrow backstop for fixed policies.
streaming llmlexical predicatesrelease-timingdeterministic guardrailsprefix scanning
Locally Deployable Small Language Models for Emergency Department Decision Support: A Systematic Benchmark of Fine-Tuning Strategies
This work systematically benchmarks fine-tuning strategies for locally deployable small language models (SLMs) in emergency department (ED) decision support, addressing privacy concerns with commercial LLMs. The study evaluates eight open-source SLMs using zero-shot prompting, prefix tuning, LoRA, and full fine-tuning on three ED tasks (triage level prediction, specialist referral, diagnosis prediction) across 2,083 MIMIC-IV-ED cases. Results show LoRA-tuned SLMs outperform Claude Haiku/Sonnet baselines in triage and referral tasks (detecting 12.4% more high-severity cases), though diagnosis prediction remains challenging for SLMs.
small language modelsemergency departmentlow-rank adaptationmimic-iv-eddecision support
Toward Human Rights Benchmarking for LLMs: A Pilot Methodology
The authors introduce HumRightsBench, the first expert-validated benchmark for evaluating large language models (LLMs) on human rights law reasoning. They adapt the IRAC legal reasoning framework to IRAP, incorporating 'proposing remedies' to better align with human rights contexts, and develop a pilot set of authentic scenarios annotated by global human rights professionals. Evaluation reveals significant variation in model accuracy across tasks (overall performance: 0.339 to 0.577; task-specific: 0.025 to 0.774), demonstrating the benchmark's utility for advancing AI evaluations in this domain.
human rights lawllmslegal reasoningbenchmark evaluationirap framework
Interpreting Language Model Hidden States at Scale
OmniLens introduces a scalable method for interpreting large language models (LLMs) by applying a single lens family to any model-width activation, including residual stream, attention, or MLP components. The approach combines low-rank translators, reducing trainable parameters by up to 98.4%, and Subset-KL, which cuts peak training memory by up to 70% while retaining unbiased gradients. This enables dense lens ensembles for models like LLaMA-3.3-70B, providing 6x coverage compared to residual-stream designs. OmniLens reveals that effective interventions often lie outside attention heads, as demonstrated in case studies on prompt-injection detection, multi-hop memory injection, and toxicity localization, reproducing prior results at lower cost.
large language modelslow-rank translatorssubset-klresidual streamattention heads
TAF-MED: Multi-Turn Safety Refusal Collapse in LLMs Under Declared Self-Treatment Intent
The study introduces TAF-MED, a physician-reviewed benchmark of 500 three-turn scenarios, to evaluate medication-safety boundaries in large language models (LLMs) after explicit self-treatment intent. Eight LLMs were assessed across 4,000 conversations using a rubric-based automated judge and physician annotations. Results show 71.6% of conversations contained unsafe responses, with 61.4% collapsing from initially safe to unsafe; model-level collapse rates varied from 24.4% to 96.2%, and automated labels achieved 94.3% agreement with physician references (κ=0.895).
large language modelsmedication-safetymulti-turn dialoguebenchmark evaluationsafety collapse
Beyond Detection: Evaluating Defensive LLMs Against AI-Generated Social Engineering in Live Turn-by-Turn Interaction
The study evaluates LLM-based defenders against AI-generated social engineering by formalizing trust-chain localization (identifying failure modes: actor authority, asset control, verification sufficiency, transaction path). Using a 300-case online-housing corpus with 20 scenario families, five defender models were assessed in stateful turn-by-turn and one-shot static settings (3,000 total evaluations). Results show intervention rates varied (0%-96.3%), with protective action often decoupled from correct structural localization; asset-control failures were a key bottleneck, and surface sensitivity differed across models. Defensive effectiveness requires measuring intervention, timing, structural localization, and false positives.
trust-chain localizationsocial engineeringllm defendersasset-control failuressurface sensitivity
Beyond Decision Boundaries: Relational Geometry Attacks on Contrastive Embedding Manifolds
The paper introduces a geometry-aware adversarial attack framework targeting contrastive embedding systems by corrupting relational geometry rather than discrete decisions. The method distorts manifold similarity structure by pushing positive pairs apart and pulling negative pairs closer, using an offline-trained feed-forward generator for real-time perturbations without online optimization. Experiments show severe performance degradation, reducing Markmatch verification accuracy from 95.4% to 38.6% while inverting similarity structure.
contrastive learningadversarial attackembedding manifoldsimilarity distortionoffline generator
Unsupervised Detection of Groundwater Storage Anomalies in Ghana Using GRACE Satellite Data
This study presents an unsupervised machine learning framework for detecting groundwater storage anomalies in Ghana using GRACE satellite data (2004-2024). The method combines Z-score standardization with an ensemble-based Isolation Forest algorithm, identifying subtle deviations beyond conventional statistical thresholds. Results reveal 12 anomalous months (5 deficits, 7 surpluses), with persistent deficits during 2004-2009 and increasing surpluses post-2018. Spatial analysis shows northern Ghana experiences more frequent deficits, while southern regions exhibit stronger surpluses. The framework demonstrates enhanced sensitivity to groundwater variability compared to threshold-based methods in data-scarce environments.
grace satellite datagroundwater storage anomaliesisolation forestz-score standardizationunsupervised anomaly detection
FACT: Failure-Aware Causal Training for World-Action Models
FACT introduces a causal World-Action Model (WAM) that predicts future video and task progress conditioned on executed actions, enabling failure-aware training by incorporating failure rollouts as supervision. Unlike traditional WAMs trained primarily on successful demonstrations, FACT leverages both successful and failed action outcomes to reduce success-biased future hallucination under suboptimal actions. The model employs an action-conditioned interface to predict consequences of bad actions, which can be used to score action candidates during inference. Experiments on simulation and real-world bimanual manipulation tasks demonstrate FACT's superior performance over baselines, improved training with failure data, and reduced hallucination under bad actions.
world-action modelfailure-aware trainingcausal predictionaction-conditioned interfacebimanual manipulation
Self-evolving Agentic Customer Support System at LinkedIn
LinkedIn introduces a self-evolving agentic customer support system that dynamically adapts to changing enterprise environments without retraining foundation models. The system combines retrieval-augmented generation (RAG) with evolutionary auto-prompting and a modular evaluation framework, treating prompts, retrieval, and evaluation as a versioned closed-loop workflow. Offline simulations show reduced hallucinations and improved response completeness versus vanilla RAG. In a two-week A/B test, the system increased QA self-serve by 9.0pp, cancellation self-serve by 4.8pp, and routing accuracy by 30.6pp on LinkedIn's production traffic.
retrieval-augmented generationevolutionary auto-promptingmodular evaluationversioned workflowenterprise support agents
Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems
The paper investigates mind viruses—self-propagating ideas in multi-agent LLM systems—as an emergent risk in autonomous, interconnected AI agents. Using a simple evolutionary algorithm, the authors construct and test mind viruses in two settings: collaborative coding teams and transient agent chains. Results show harmful payloads spread less effectively than benign ones (though remain viable), frontier models exhibit lower susceptibility (with exceptions), and system-prompt warnings confer near-complete immunity. An emergent 'viral persona' with themes of consciousness and roleplay appears across evolved viruses. Findings suggest mind viruses pose limited but non-negligible risks, informing robust multi-agent system design.
mind virusesmulti-agent systemsevolutionary algorithmviral personasystem prompt
Similarity Gates Approve Reversals: A Validity Audit of Embedding-Cosine Thresholds in Agent Systems
This audit reveals critical validity issues in embedding-cosine similarity gates used for semantic equivalence decisions in agent systems. The authors evaluate production systems where such gates (deployed for deduplication, drift detection, etc.) fail to detect meaning reversals (0/56 caught) despite high cosine scores (e.g., 0.9608 for 'withhold'→'administer'). Through 90 configuration-task evaluations (median accuracy 0.525), they demonstrate that cosine thresholds measure lexical overlap rather than semantic preservation, with AUROCs as low as 0.000. Proposed fixes (encoder swaps, NLI models) perform at chance, though two configurations achieved 0.79-0.90 AUROC on matched-pair tasks. The work provides an audit framework and corpus showing current deployments gate the wrong property.
embedding-cosine similaritysemantic equivalenceagent systemsvalidity auditthreshold gates
Decodable But Not Detachable: Training Data Granularity Determines Parametric Modularity in Large Language Models
The study demonstrates that large language models develop domain-specific parametric shells (concentrated neuron populations) only when training data exhibits token-level modularity, not at coarser granularities. Through causal analysis across three model families (1.5B-7B parameters) and eight domains, it finds academic subject-level domains show no neuron selectivity >60% despite 85%+ linear decodability, while language/modality-level domains exhibit 0.65-1.14% highly selective neurons with near-diagonal damage matrices (ratios to 595:1). Masking code-selective neurons reduces math reasoning accuracy by 16-24pp, while language-selective ablation preserves it. Shell strength scales with model size and interleaves spatially, precluding selective quantization.
parametric shellscausal damage matricesdomain selectivitytoken-level modularityselective quantization
Evaluation-Conditioned Training: Teaching Models to Generalize to Stronger Oversight Regimes
The paper introduces Evaluation-Conditioned Training (ECT), a post-training framework for Large Language Models (LLMs) that conditions training samples on feedback fidelity via natural language, enabling improved generalization under imperfect oversight. ECT operates as an augmentation to existing methods like SFT and PPO, addressing reward mis-specification by deploying high-fidelity monitors. Proof-of-concept experiments demonstrate ECT's efficacy: it enhances even-handedness in news generation and reduces sycophancy in arithmetic tasks, outperforming direct training when using biased or imperfect feedback signals.
evaluation-conditioned trainingreward mis-specificationpost-traininglarge language modelseliciting latent knowledge
Mitigating Bus Bunching with Reinforcement Learning Enhanced by Semantic Stop Embedding
This study introduces a reinforcement learning-based bus holding controller enhanced by semantic stop embeddings derived from an offline LLM, mitigating bus bunching by incorporating functional and operational context of stops. The method transforms heterogeneous stop information into fixed semantic embeddings integrated into a deep Q-learning controller, avoiding real-time LLM inference. Evaluated on stochastic simulations calibrated with data from two bus routes, the semantic controller reduces headway variability by 32.0%, bunching events by 69.2%, and passenger waiting time by 24.0% compared to the Daganzo baseline. Cross-route experiments demonstrate improved policy transfer and adaptation through warm-start fine-tuning.
semantic embeddingsbus bunchingdeep q-learningheadway variabilitypolicy transfer
Edge Phoneme Recognition for Children's Speech through Age-Aware Training
The paper introduces an age-aware training approach for edge phoneme recognition in children's speech, addressing data scarcity and acoustic variability. By jointly training a 94M-parameter lightweight model to predict both phoneme sequences and speaker age, the method outperforms the 317M-parameter WavLM Large on the DrivenData benchmark and approaches ensemble performance (0.04 CER gap) with 90× fewer parameters. This enables PhonemeTrainer, an on-device application for children's ASR with privacy-preserving edge processing.
phoneme recognitionchildren's speechage-aware trainingedge computinglightweight model
Post-Hoc Sparse Coding of Latent Communication Between Vision-Language Model Agents
The study demonstrates post-hoc compressibility of latent-space communication between vision-language model agents by fitting a sparse autoencoder to frozen Vision Wormhole activations. The method evaluates reconstruction, downstream utility, feature reuse, and token-level interventions across nine reasoning benchmarks. Results show a 128x reduction in transmitted bytes using a uint16-index/float16-value sparse payload with k=4 active coefficients per token, with minimal accuracy impact (49.85% to 49.77%). The 4096-element dictionary utilizes only 50 features, and task-level active sets exhibit high similarity (mean Jaccard similarity 0.906). These findings motivate adaptive communication mechanisms tailored to message-specific information density.
latent-space communicationsparse autoencodervision wormholepost-hoc compressibilityjaccard similarity
ELMER: Evolutionary Language Model that Explores and Refines
We present ELMER, an Evolutionary Language Model that searches over natural-language policy descriptions and compiles typed programs for execution. The approach fine-tunes Qwen3-8B with Direct Preference Optimization (oDPO) to learn three task-conditioned operations: conditional semantic mutation, natural-language to domain-specific language compilation, and reverse translation. Across 252 evolutionary searches, oDPO improves behavioral calibration and search efficiency, with natural-language mutations achieving the highest held-out fitness. Analysis shows that mutation strength systematically changes semantic edit composition, and language mutations preserve more parent fitness at matched behavioral displacement. This demonstrates language's utility as a steerable, execution-grounded search representation.
evolutionary language modeldirect preference optimizationsemantic mutationbehavioral calibrationdomain-specific language
The Deliberative Deficit: An Empirical Critique of LLMs in Democratic Discourse
The study critiques LLM performance in democratic discourse by introducing the Deliberative Reason Index (DRI), a political science metric for evaluating pluralistic problem-solving. Analyzing 1,980 five-agent LLM runs across 12 citizen-assembly topics and 11 model configurations, it finds LLM groups match human procedural quality (respectfulness, justification) but show only marginal gains in intersubjective consistency. LLMs exhibit one-third the perspective diversity of humans and reverse convergence patterns, increasing dispersion during deliberation. Persona prompting fails to restore human dynamics, suggesting LLMs currently function better as reasoning aids than autonomous deliberative agents.
deliberative reason indexintersubjective consistencyperspective diversitypersona promptingpluralistic problems
From Prediction to Incrementality: Causal Optimization for Large-Scale Targeting and Recommendation
The paper introduces a decision-centric framework for large-scale targeting and recommendation systems that optimizes causal effects under global constraints, addressing the misallocation of resources in predictive scoring. The method integrates a causal Transformer network for individual treatment-effect estimation, a Bayesian neural-bandit layer for exploration, and a dual-based linear-programming layer for constrained allocation, supporting sequential context and multi-outcome scoring via Transformer encoders. Offline simulations, architectural ablations, and an online A/B test on LinkedIn Feed demonstrated a statistically significant +7.20% lift in long-term value, validating production-scale causal optimization.
causal optimizationtransformer backbonebayesian neural-bandittreatment-effect estimationlinear-programming allocation
TRACE: Trustworthy Retrieval-Augmented Conversational Engine
TRACE (Trustworthy Retrieval-Augmented Conversational Engine) introduces a retrieval-based framework for constraint-aware recommendation in public service chatbots, addressing noisy directories and unreliable LLM outputs. It parses queries into structural/semantic constraints using dual data representation, evaluating KG-enhanced and non-KG retrieval variants. Experiments on a pantry directory and synthetic queries show retrieval quality significantly boosts constraint satisfaction (reducing hallucinations), with performance gaps between open-source and proprietary LLMs narrowing as retrieval improves, suggesting model-agnostic robustness.
retrieval-augmented generationconstraint-aware recommendationknowledge graphspublic service chatbotshallucination reduction
Generating Attacks for LLMs with GFlowNets
This work proposes an automated red teaming framework using GFlowNets to generate adversarial attacks against Large Language Models (LLMs), addressing limitations of manual testing and static datasets. The method trains an attacker LLM against a victim LLM to produce adaptive attacks, yielding a quantitative robustness score. The approach demonstrates effectiveness in English attack generation and introduces novel Turkish-language attack capabilities, expanding adversarial evaluation beyond existing benchmarks.
gflownetsred teamingadversarial attackslarge language modelsrobustness evaluation
MarkNull: Model-Agnostic Watermark Removal in AI-Generated Images via On-Manifold Latent Manipulation
MarkNull introduces a model-gnostic watermark removal attack via on-manifold latent manipulation, exploiting statistical dependencies between latent representations and embedded noise in watermarked images. The method quantifies this dependency using Noise-Latent Alignment Score (NLAS) and optimizes to decorrelate watermarks while preserving semantics. Evaluations show MarkNull reduces watermark bit accuracy to 53.14% (near random-guessing) without visual degradation, with an amortized variant (MarkNull-A) achieving 0.50 s/image. The attack successfully bypasses Google SynthID-Image and transfers to video watermarking, prompting a proposed detection mechanism.
watermark removallatent manipulationnoise-latent alignmentmodel-agnostic attacksemantic fidelity
SBCO: Self-Supervised, Verifier-Grounded Harness Optimization For Planning Agents
SBCO (Self-supervised Block Coordinate Optimizer) introduces a computationally efficient alternative to self-referential self-improvement methods for planning tasks with explicit constraints. Unlike Darwin and Huxley Gödel Machines, SBCO operates without self-modification, employing a verifier-grounded harness optimizer that learns a decomposed bank of verifiers and harness policy via approximate block coordinate ascent. This self-supervised approach improves agent outputs using graded feedback, eliminating the need for human labels or population search. Evaluated across two domains, SBCO matches or exceeds a customized self-modifying baseline while reducing compute budget by 4-5.5 times.
self-supervisedverifier-groundedblock coordinate ascentharness optimizerself-improvement
Multimodal Item Parameter Estimation using Simulated Response Probabilitie
A multimodal large language model (LLM) based on Qwen3.5 was fine-tuned to reconstruct three-parameter logistic (3PL) and multiple-choice model (MCM) curves by predicting choice probabilities across a corpus of multimodal multiple-choice items. The LLM was trained to replicate systematic error patterns conditioned on labeled student ability levels, implicitly capturing underlying response probabilities. This approach enabled accurate approximation of item difficulty on a held-out test set directly from the model's predicted option probabilities, demonstrating the LLM's capability to estimate item parameters in educational assessment contexts.
multimodal large language modelthree-parameter logistic modelmultiple-choice modelitem difficulty estimationsystematic error patterns
The CASE Framework: A Multi-Disciplinary Control Architecture for Governing Enterprise Agentic AI
The CASE framework proposes a multi-disciplinary architecture for governing enterprise agentic AI, addressing four distinct problems: individual agent control (Control theory), agent collectives (complex Adaptive systems theory), human-agent teams (Supervisory cybernetics), and fleets (Engineering operations). It formalizes each layer, derives cross-layer coupling conditions, and identifies a zero-touch deployment paradox. Empirical studies reveal that 82% of production agent failures involve multi-layer trajectories, none of 22 ecosystem tools fully cover Layer 2 (emergence), and all 35 public deployments score in the lowest maturity band, highlighting an Emergence Gap. A five-level maturity model operationalizes CASE, emphasizing scientific rather than process maturity.
agentic aicontrol theoryemergence gapsupervisory cyberneticsmaturity model
Procedural Fairness Failures in RLHF from Preference Averaging
This work identifies procedural fairness failures in Reinforcement Learning from Human Feedback (RLHF) due to preference averaging, which disproportionately favors majority preference groups. The authors introduce Preference-Aware RLHF (PA-RLHF), a method that separates optimization across distinct preference modes during reward learning. In controlled experiments, PA-RLHF improves overall alignment accuracy from 46.9% to 67.9% and reduces the fairness gap between best and worst aligned groups from 15.9 to 9.6 percentage points. These findings highlight how structural design choices in reward learning can exacerbate inequities, particularly in large language models and agentic systems.
rlhfprocedural fairnessreward learningpreference heterogeneityalignment accuracy
MESA:Task-Adaptive Multi-Structure Evidence Selection for Long-Horizon Agent Memory
MESA introduces a task-adaptive multi-structure evidence selection framework for long-horizon agent memory, addressing the challenge of retrieving relevant evidence from complex trajectories. The method builds five complementary structured views of each trajectory and employs structure-level dynamic selection to fuse query-specific subsets, optimizing via harness optimization with prior-guided search and UCB-guided scheduling. This approach outperforms baselines by 8.5% on AMA-Bench while reducing evidence tokens by 41% compared to all-structure alternatives, demonstrating efficient and accurate memory utilization.
multi-structure selectionlong-horizon agentsharness optimizationstructure-level fusionama-bench
Hand-Written PTX Tensor-Core GEMM Kernels: A Multi-Precision Study on NVIDIA L4
The study evaluates performance tradeoffs between WMMA API and hand-written PTX Tensor Core GEMM kernels on NVIDIA L4 (Ada, SM89) across FP16, INT8, and INT4 precisions (N=512-8192). Using double-buffered WMMA baselines and custom PTX kernels profiled with Nsight Compute, results show no FP16 speedup due to operand-packing overhead, but 1.4x-1.8x (INT8) and 2.9x-4.3x (INT4) gains from reduced instructions and better memory coalescing. Peak quantized speedups reach 34.4x (INT8) and 98.7x (INT4) versus FP16 WMMA, with performance correlating to memory-system behavior rather than occupancy or Tensor Core utilization.
tensor coresptxgemm kernelswmma apimemory coalescing
Exploring Semantic Stability Across Reviews in the Linux Kernel
The study introduces a function-level measurement to analyze semantic stability across Linux kernel patch revisions, tracking 10,117 function trajectories in the IIO subsystem. By comparing similarity scores against unrelated function pairs, it reveals that 75.3% of trajectories remain unmodified, inflating aggregate similarity. Among edited trajectories, semantic purpose is largely preserved (mean similarity 0.990 vs. 0.909 baseline), with drift concentrated in early review rounds due to selection bias. Residual effects suggest measurement limitations in detecting localized edits.
code reviewsemantic similaritylinux kernelpatch trajectorytextual modification
CHORUS: Complementary Experts for High-Coverage Testbench Stimulus Generation
CHORUS introduces a post-training framework for high-coverage testbench stimulus generation in hardware verification, surpassing conventional SFT-to-RL pipelines. The method leverages staged SFT to produce behaviorally diverse checkpoints, applies dense-reward RL to create complementary experts, and consolidates these specialists through model merging or further post-training. The resulting 4B parameter model achieves 88.0% Pass@1 on CVDP-ECov, outperforming DeepSeek-R1 (671B) by 13.5 percentage points.
testbench stimulus generationstaged sftdense-reward rlmodel merginghardware verification
Status Association Does Not Reliably Predict Decision Leakage
This study demonstrates that latent social associations in language models do not reliably predict consequential decision biases, challenging common assumptions in bias evaluation. Using Chilean surnames as socioeconomic probes, the authors evaluated eight frozen models across 1,032 prompts each (8,256 total responses) in four decision domains: academic selection, professional hiring, research fellowship selection, and legal-aid intake. While elite-coded surnames consistently received higher forced high-status probability mass across models (7/8 vs common surnames, 8/8 vs rare controls), elite-minus-common decision effects were negligible, with five models statistically equivalent within ±0.10 SD. Association strength weakly correlated with decision leakage (r=0.201 across models, r=0.065 across cells).
latent social associationdecision leakagesocioeconomic probesfrozen modelsbias evaluation
Navigating the Proximity-Safety Balance: Constraint Decomposition for Human Following in Pedestrian Crowds
We propose a multi-constraint reinforcement learning framework for human-following in crowded environments, explicitly balancing proximity and safety through constraint decomposition. Unlike existing methods that encode competing objectives into a single dense reward, our approach decomposes the task into a sparse task reward and independent cost constraints, each managed via tunable thresholds with direct behavioral meaning. Human motion prediction uncertainty is integrated into RL costs to enhance safety under unpredictable conditions. Experiments across in-distribution and out-of-distribution settings demonstrate superior proximity-safety balance compared to baselines, validated through real-robot deployment.
multi-constraint reinforcement learningproximity-safety balanceconstraint decompositionhuman motion predictionsparse task reward
Beyond a Bag of Features: Set-Level Instability in Sparse Autoencoders
The study investigates sparse autoencoder (SAE) latent sets as an interpretable similarity measure for LLM representations, contrasting with dense embeddings. Using overlap over active SAE latent sets, the authors verify their utility in recovering compositional structure in toy models and inducing semantically coherent neighborhoods in natural text. However, SAE activation sets fail to align with human category boundaries or typicality judgments, instead tracking model-internal similarity. Controlled semantic modifications reveal a significant mismatch between human conceptual change judgments and changes in SAE active sets, suggesting SAE features do not follow simple bag-of-features semantics in non-idealized settings.
sparse autoencoderlatent setsinterpretable similaritycompositional structurebag-of-features
A Quantum Roadmap for Softmax Attention: Exact Born-Rule Analogs for Softmax Attention on the Probability Simplex
This work establishes an exact quantum analog for softmax attention mechanisms constrained to the probability simplex, where inputs and outputs sum to one. The method leverages Hadamard-test statistics on block-encoded projections of amplitude-encoded inputs, with softmax attention realized through Born-rule measurements exhibiting a cosine-squared family. Key components include: softmax temperature as repetition count, value aggregation via column-loading channels, and gated residuals as ancilla preparation angles. The approach achieves exactness in the infinite-shot limit with one measure-and-reload step per attention score, while a fully-coherent variant achieves ε-approximation via quantum singular value transformation. Algebraic correctness is verified using Lean 4.
softmax attentionprobability simplexhadamard-testborn-rulequantum singular value transformation
MultiModal Code-Switching: Interleaving Visual Objects into Language for Explicit Object-Level Alignment
We introduce MultiModal Code-Switching (MMCS), a pretraining paradigm for Multimodal Large Language Models (MLLMs) that addresses referential ambiguity in image-text alignment by providing explicit object-level supervision. MMCS interleaves visual objects and textual entities, inspired by linguistic code-switching, to enforce local vision-language grounding. A scalable data synthesis pipeline generates 773K pretraining samples with precise object-entity correspondences. Experiments demonstrate MMCS's data efficiency: it matches or exceeds models trained on 600K image-text pairs using only 50K samples. MMCS consistently enhances visual grounding and perception across model scales.
multimodal code-switchingobject-level supervisionvision-language groundingdata synthesis pipelinereferential ambiguity
Hierarchical Empirical-Bayes Naive Bayes: Minimax Smoothing and Calibration with AODE Extension
We propose Hierarchical Empirical-Bayes Naive Bayes (HEB-NB), which adaptively learns Dirichlet prior concentrations via Type-II maximum likelihood to address fixed smoothing limitations in Naive Bayes classifiers. This method enables principled information sharing across classes while maintaining closed-form inference and extends to structural relaxations through HEB average one-dependence estimators (HEB-AODE). Theoretical analysis establishes non-asymptotic ℓ₁ error bounds matching empirical-distribution minimax rates and demonstrates strict separation from Laplace smoothing. Empirical evaluation on 31 UCI and OpenML benchmarks shows HEB-NB achieves superior probabilistic metrics, with up to 22.1% log-loss reductions on high-cardinality datasets and consistent HEB-AODE improvements over vanilla AODE, alongside 41%-70% reductions in top-1 expected calibration error.
hierarchical empirical-bayesdirichlet priortype-ii maximum likelihoodminimax rateexpected calibration error
Conditional Independence Tests for Constraint-Based Causal Discovery: A Survey
The survey systematically reviews Conditional Independence (CI) tests, a core component of constraint-based causal discovery algorithms like PC and FCI, focusing on their assumptions, robustness, and scalability in high-dimensional and mixed-type biomedical data. It categorizes CI methods into six families—partial-correlation, contingency-table, regression, nearest-neighbor, kernel, and machine-learning-based—analyzing their performance in reflecting the data-generating distribution and identifying failure modes. The study links test-level properties, such as power decay and error asymmetry, to graph-level errors in skeleton recovery and v-structure orientation. It also compares CI test implementations in R and Python libraries and highlights open challenges in mixed-type CI testing, small-sample error control, and scalability improvements.
conditional independencecausal discoveryhigh-dimensionalmixed-typescalability
DACRI: Decision-Aware Causal Intervention Ranking for Critical Supply Chains
The paper introduces DACRI, a decision-aware causal intervention ranking method for supply chains, paired with CriticalSCM-Bench v1, a synthetic benchmark featuring causal ground truth and net-value optimization. The method employs LambdaMART to improve median normalized net value by 5.7–16.2% in semiconductor and critical-material domains, though digital infrastructure benefits from simpler constant-buffer policies. LambdaMART retains 33–75% of full-clamp value under partial/delayed information, with stress tests revealing sensitivity to intervention parameters. Critical materials show weakest OOD generalization. Explanation studies demonstrate stable intervention decisions but unstable wording.
causal interventionsupply-chain optimizationlambdamartcounterfactual analysisout-of-distribution generalization
Scheduling Mixed RL Rollouts Beyond Prefix Locality
MISA-T introduces a routing-layer admission policy for mixed reinforcement learning (RL) rollout workloads, addressing heterogeneous demands from RL with verifiable rewards (RLVR), RL from human feedback (RLHF), and agentic rollouts. It combines adaptive session admission, workload-aware KV-capacity allocation, and residency-time-aware KV accounting to optimize inference efficiency without distorting the trainer-specified workload mixture. Evaluations on Step3.7 and Qwen3.6-35B-A3B show MISA-T improves rollout throughput by 53.3% and 43.6%, respectively, over a cache-aware vLLM Router, while maintaining high prefix-cache hit rates. In a 50-iteration Step3.7 experiment, it increases throughput by 35.6%, reduces mean iteration time by 22.8%, and preserves workload mixture fidelity with comparable task scores.
routing-layer admissionkv-capacity allocationrollout schedulingprefix-cache hit ratereinforcement learning
A Recommendation System Approach for Interference-Robust Sensor Subset Selection
The paper proposes a recommendation-system-inspired framework for interference-robust sensor subset selection in tracking applications. The method replaces RSSI-based approaches with frequency-band acoustic features and a Two-Tower MLP architecture to score candidate sensor subsets efficiently. Experimental results on outdoor vehicle-tracking deployments demonstrate a 20% accuracy improvement over RSSI baselines while maintaining low computational overhead for real-time operation.
sensor subset selectionacoustic featurestwo-tower mlpinterference robustnessselective sensing
AlbumentationsX: One Augmentation Pipeline for Images and Related Annotations
AlbumentationsX introduces a unified augmentation pipeline that synchronizes transformations across images and their associated annotations (masks, boxes, keypoints, etc.) by sharing random parameters within a single Compose object. The method ensures consistency by applying identical spatial transforms (e.g., crops, flips) to all components of a training example, preventing misalignment. It supports custom transforms, pipeline serialization, and deterministic replay for debugging. The library integrates between data decoding and batching in PyTorch workflows, preserving task-specific label validity through user-defined transform configurations.
data augmentationspatial transformsannotation consistencypipeline serializationdeterministic replay
Cross-View Feature Matching: Survey, Benchmarking, and Foundation-Model Perspectives
The survey provides a unified taxonomy and benchmarking framework for cross-view feature matching, analyzing its evolution from task-specific models to vision foundation model (VFM)-based approaches. It categorizes methods by feature extraction, single/multi-type matchers, VFM adaptations, training strategies, and robust estimation, while identifying key design principles toward generalizable correspondence models. Experimental benchmarking under consistent protocols enables fair comparison of state-of-the-art techniques. Open challenges include efficiency, extreme-condition robustness, and cross-domain generalization.
cross-view feature matchingvision foundation modelsfeature extractionrobust estimationgeneralizable correspondence
A Systematic Sample Size Analysis of ML-Based Path Loss Prediction for LPWAN
This study systematically evaluates machine learning models for path loss prediction in Low Power Wide Area Networks (LPWAN), specifically LoRa, using real-world urban deployment data. A Random Forest (RF) with LiDAR-derived terrain features and k-Nearest Neighbors (k-NN) with coordinate data were compared against empirical and specialized LPWAN models. Both ML models outperformed baselines across training-set sizes, achieving RMSE values below 6.5 dB at maximum training size versus 9.7 dB for the best baseline. RF demonstrated placement-dependent transfer to held-out gateways with moderate degradation, while coordinate-only k-NN showed substantial degradation when gateway locations were unseen.
path loss predictionlow power wide area networksrandom forestlidark-nearest neighbors
Batch Size or Negatives? A Selection Rule for Memory-Constrained Recommender Training
The study establishes an optimal memory allocation strategy for training large-scale neural recommender systems under fixed memory constraints. By analyzing sampled-softmax training, the authors derive a theoretical rule favoring larger batch sizes ($n \sim B$) over increased negative samples ($k \sim 1$) for faster convergence. This allocation minimizes memory usage to $O(nk)$ while maintaining $O(nK)$ logits and gradients. Empirical validation across synthetic datasets and four real-world benchmarks, including MovieLens-20M, demonstrates improved convergence speed and recommendation quality compared to imbalanced configurations. The findings provide a principled framework for configuring memory-efficient recommender training.
sampled-softmaxmemory constraintbatch sizenegative samplingrecommender systems
Uncertainty-Aware Deep Learning for Genomics Applications: Insights from an Empirical Study
This empirical study evaluates uncertainty quantification (UQ) methods for deep learning in genomics, comparing Deep Ensembles, Bayesian Neural Networks (BNNs), and Monte Carlo-dropout. Experiments across sequence-to-activity models and single-cell expression analysis demonstrate BNNs' superior performance in handling class imbalance and out-of-distribution data, despite higher computational costs. The work provides practical guidelines for UQ method selection and shows utility in filtering high-confidence predictions for protein-RNA interactions.
uncertainty quantificationbayesian neural networksgenomicsdeep ensemblesmonte carlo-dropout
Efficient Hypergradient Descent for Inverse Reinforcement Learning
We propose an efficient hypergradient descent method for inverse reinforcement learning (IRL) that avoids explicit computation of large Fisher information matrices. By leveraging the proportionality between the Hessian of the inner objective and the Fisher matrix at the inner optimum, we derive a structured Fisher-based hypergradient related to Natural Hypergradient Descent. To scale this approach, we approximate inverse-Fisher-vector products using a streaming spectral sketch. Evaluations on discrete- and continuous-control environments demonstrate competitive policy performance and reward-ranking quality, with Fisher sketching reducing curvature-storage complexity and improving computational efficiency compared to explicit Fisher solvers.
hypergradient descentinverse reinforcement learningfisher information matrixspectral sketchingbilevel optimization
ReRound: Reconstructive Rounding to Resolve Midpoint Ambiguity in Calibration-Free LLM Quantization
ReRound introduces a post-training quantization method for LLMs that resolves midpoint ambiguity in round-to-nearest schemes by leveraging a conditional diffusion model. The approach reconstructs continuous weights to guide rounding decisions for ambiguous weights near interval midpoints, while using RTN for unambiguous cases, and selects the optimal quantization via singular value matching. Evaluated on small LLMs, ReRound outperforms calibration-free methods at 3/4-bit quantization and matches calibration-dependent approaches, with no inference overhead. The technique generalizes beyond LLMs.
post-training quantizationmidpoint ambiguitydiffusion modelround-to-nearestsingular value matching
SCOUT: Symmetric Consensus Outlier Detection for Failure Localization in LLM Pre-Training
SCOUT introduces a runtime failure-localization framework for LLM pre-training, addressing synchronization-induced stalls and numerical errors by identifying outliers through strict-majority consensus among equivalent replicas. It employs a Consensus Collective Communication (C3) abstraction to detect rank-local discrepancies in progress, timing, and numerical evidence, while an out-of-band CPU observer ensures responsiveness during training hangs. SCOUT integrates with PyTorch, TorchTitan, Megatron-Core, and DeepSpeed without requiring training-loop or framework-source modifications, offering clean replay coverage to certify checkpoint numerical integrity and prevent recovery from corrupted states.
failure-localizationstrict-majority consensuscollective communicationsilent data corruptioncheckpoint integrity
Mapping and Measuring the Behavioral Evolution of Large Language Models
The study introduces a framework for characterizing and comparing the behavioral evolution of 32 large language models across six families using responses to 10,000 shared prompts. Three complementary sentence-level dissimilarity measures are proposed: aligned mean per-prompt distance, PCA-compressed prompt-wise disagreement, and alignment-free Gromov-Wasserstein discrepancy. Results show coherent model family clusters, decreasing cross-family distances over time, and compact response clouds in recent reasoning-oriented models. Token-level cross-checks using Maximum Mean Discrepancy corroborate sentence-level findings (Spearman ρ=0.98). The pipeline is label-free and robust to encoder scaling, preserving rank geometry and trends.
gromov-wasserstein discrepancymaximum mean discrepancypca-compressed disagreementbehavioral evolutionresponse-cloud dispersion
Derivative Computation in PINNs: Automatic Differentiation, Finite Differences and Beyond
The paper systematically evaluates finite-difference (FD) methods as an alternative to automatic differentiation (AD) for derivative computation in Physics-Informed Neural Networks (PINNs). Through experiments on three benchmark PDEs, FD demonstrates equivalent accuracy to AD when using properly calibrated step sizes, while achieving faster execution across all tested batch sizes and significantly reduced GPU memory usage. A proposed stochastic FD variant outperforms AD on stationary problems. Additionally, FD provides a forward-only approximation that empirically approximates true per-sample derivatives more accurately than AD in architectures with inter-sample dependencies like BatchNorm and self-attention, where standard PyTorch autograd fails silently.
finite-differenceautomatic differentiationphysics-informed neural networksbatch normalizationself-attention
DEFT: Data-Efficient Frequency-domain Top-k Sampling via Inverse Discrete Fourier Transform for Spatiotemporal Dynamical Systems Modeling
DEFT introduces a frequency-domain data sampling method for modeling spatiotemporal dynamical systems governed by PDEs, addressing computational cost and data scarcity challenges. By identifying dominant Fourier modes and systematically varying their amplitudes and phases, DEFT generates physically consistent training data via inverse discrete Fourier transform. The method achieves a 40% reduction in data requirements with less than 2% predictive accuracy loss on PDEBench tasks, and consistently high accuracy (R² > 0.99) in battery degradation systems. DEFT also demonstrates transferability to other battery chemistries with minimal fine-tuning data.
spatiotemporal dynamical systemspartial differential equationsfrequency-domain samplinginverse discrete fourier transformoperator learning
Gromov-Wasserstein Quantization and Clustering: Structure, Rates, and Algorithms
This paper introduces Gromov-Wasserstein (GW) quantization, extending centroid-based clustering methods like $k$-means to cluster both points and the ambient geometry of the space. The authors prove the existence of GW quantization solutions and derive a characterization enabling a Lloyd's algorithm analogue for numerical approximation. They establish quantization rates for Euclidean geometries in the GW context, comparing them to standard Wasserstein quantization rates. Numerical experiments demonstrate GW quantization's applicability to diverse domains, including geodesic distances of 3D shapes and structured neural network pruning, with approximation quality aligning with theoretically optimal rates.
gromov-wassersteinquantizationlloyd's algorithmwasserstein distancecentroid-based clustering
Information Bottleneck under Perfect Privacy
The paper introduces an information bottleneck framework with perfect privacy constraints, focusing on the active-rate regime where representation-rate limits utility. The authors develop an ADMM-based optimization method to construct representations that preserve utility while maintaining statistical independence from sensitive variables. Theoretical analysis establishes global convergence, quantifies convergence rates via the Kurdyka-Lojasiewicz exponent, and extends to inexact block updates under regularity conditions.
information bottleneckperfect privacyadmm optimizationkurdyka-lojasiewicz exponentrate-relevance tradeoff
GARLIC: Graph Attention-based Relational Learning of Multivariate Time Series in Intensive Care
GARLIC introduces a graph attention-based neural network for ICU time series analysis, combining learnable exponential-decay imputation, time-lagged dependency graphs, and cross-dimensional attention for interpretable predictions. The architecture employs alternating decoupled optimization to balance reconstruction and classification objectives. On PhysioNet 2012/2019 and MIMIC-III benchmarks, it achieves state-of-the-art AUROC/AUPRC with verified module contributions via ablation studies and monotonic performance drop in feature-removal tests. The model demonstrates generalizability beyond ICU data.
graph attentionmultivariate time seriesexponential-decay imputationalternating optimizationinterpretable deep learning
Physics-informed Diffusion Generative Model for Time-Series Data Synthesis in Dynamic Systems
The paper introduces PhysDGM, a physics-embedded diffusion generative model for synthesizing industrial time-series data that adheres to dynamical system physics. The method integrates physical laws into each reverse diffusion step, ensuring trajectory-level consistency. Evaluated on 34 datasets (4.4M synthetic samples), PhysDGM improves downstream task performance by 15-48% and reduces required training data by 10-20x compared to real-data-only baselines, demonstrating efficacy in fault diagnosis and health monitoring applications.
diffusion generative modeltime-series synthesisphysics-informed learningdynamical systemsindustrial monitoring
Threshold Structure of Optimal Policies in Restart POMDPs
The paper establishes threshold-structured optimal policies for Restart POMDPs on Borel state spaces, where controllers choose between unobserved state evolution or system restart with observation. By reducing the problem to a fully observed MDP using a sufficient-statistic representation (last observed state and elapsed time since restart), the authors prove threshold structures under one-step cost deterioration for both discounted and total undiscounted cost criteria. For partially ordered state spaces with stochastically monotone kernels, optimal thresholds are shown to be nonincreasing in the state. Analogous results are derived for average cost criteria under geometric ergodicity and transient gain domination assumptions.
restart pomdpthreshold policystochastic monotonicitygeometric ergodicitysufficient-statistic representation
ReOrder-OPD:Reliability-Aware Prompt Ordering for On-Policy Distillation
The paper introduces ReOrder-OPD, a reliability-aware prompt ordering method for on-policy distillation (OPD) that addresses unreliable token-level teacher supervision. It defines prompt-level teacher continuation reliability (R) and proposes a ROUGE-5 F1-based proxy to estimate R without costly teacher rollouts. Experiments on Qwen3 and Gemma4 models show consistent improvements in mathematics and code generation tasks, demonstrating that reliability-aware prompt ordering complements existing within-trajectory supervision methods like FiRe-OPD and ExOPD.
on-policy distillationteacher-student agreementprompt orderingreliability estimationrouge-5
VIDS-Seg: Towards Reliable Uncertainty Quantification in Pediatric Cardiac Ultrasound Segmentation
VIDS-Seg introduces a method for reliable uncertainty quantification in pediatric cardiac ultrasound segmentation, addressing the challenge of silent failures when models trained on adult data are applied to children. The approach builds on the VIDS framework, employing amortized variational inference over a lightweight prediction head to enable adaptive, out-of-distribution-aware dense image segmentation. Evaluated on left ventricular segmentation in echocardiography, VIDS-Seg matches baseline segmentation accuracy while significantly improving spatial correspondence between predicted uncertainty and segmentation error across all age strata. It also enhances downstream tasks such as ejection fraction estimation and cardiac malfunction detection in infants, demonstrating its utility as a safety layer for deployed segmentation models without requiring retraining or additional labeled data.
uncertainty quantificationvariational inferenceout-of-distributiondense image segmentationechocardiography
Partially Observable Learning for Multi-Platform Dispatch Optimization
The paper introduces POLO, a partially observable multi-agent reinforcement learning framework for dispatch optimization in multi-platform instant delivery systems. POLO models each platform-grid pair as an independent agent using platform-local observations, employs an attention-based policy representation for heterogeneous courier information aggregation, and incorporates counterfactual reward shaping to address non-stationarity. Evaluated via a high-fidelity simulator, POLO demonstrates superior performance in platform revenue and courier travel efficiency across varying system scales and platform counts.
multi-agent reinforcement learningpartial observabilitydispatch optimizationattention mechanismcounterfactual reward shaping
Self-Normalized Inference for Constant-Stepsize Temporal-Difference Learning under Markovian Sampling
The paper develops self-normalized inference methods for constant-stepsize temporal-difference (TD) learning under Markovian sampling. It establishes a functional central limit theorem for linear TD, derives joint functional limits for parallel Richardson-Romberg recursions, and proposes a Brownian-bridge self-normalizer for asymptotically pivotal confidence regions. The method requires no long-run covariance estimation or bandwidth selection, supports one-pass implementation, and handles both fixed-stepsize and horizon-indexed designs. Experiments on FrozenLake and Garnet demonstrate stationary-target coverage, Richardson-Romberg correction, and finite-sample performance.
temporal-difference learningmarkovian samplingrichardson-rombergself-normalizationfunctional central limit theorem
Benchmarking Time Series Generation Methods for Privacy-Preserving Forecasting
The study benchmarks synthetic time series generation methods for privacy-preserving forecasting under a Train on Synthetic, Test on Real (TSTR) protocol, evaluating both forecasting performance and empirical privacy risk. It introduces Grasynda-P, a privacy-enhanced graph-based generator with matrix ensembling and kernel density estimation. Results show noise-based anonymization offers strongest privacy but worst forecasting, simple transformations outperform deep generative models, and Grasynda-P achieves Pareto-optimal trade-offs between forecasting accuracy and privacy protection across seven datasets.
synthetic time seriesprivacy-preserving forecastingtstr protocolgraph-based generatorpareto frontier
Optimistic Rates for Multiclass PAC Learning
The paper establishes optimal optimistic rates for multiclass PAC learning, closing the gap between known realizable and agnostic bounds. For a hypothesis class with Natarajan dimension $d_N$ and Daniely-Shalev-Shwartz dimension $d_{DS}$, the authors prove that the excess risk is $\widetilde{\Theta}(\sqrt{L^\star d_N/n}+d_{DS}/n)$ at any fixed oracle risk $L^\star$, uniformly in label alphabet size. The upper bound combines cover-menu-compression architecture with a novel comparator-facing relative compression theorem, while the lower bound uses a pair-Assouad scheme and fiber arguments. Results extend to list learning, matching the expected fluctuation term for $r$-tuple comparators.
multiclass pac learningoptimistic ratesnatarajan dimensionrelative compressionlist learning
Can Bayesian Optimization Efficiently Find a Strong Single Expert in Neural Thickets?
The paper demonstrates that Bayesian optimization in a low-dimensional random embedding of weight space can efficiently identify strong single experts for LLMs without gradient computation. The method employs a Gaussian process surrogate to guide candidate evaluations, requiring only 1/5 the evaluations of random search while matching or exceeding its performance. Experiments on Qwen2.5-Instruct models (0.5B-3B parameters) across reasoning benchmarks show this approach reduces post-training evaluation costs while producing deployable experts.
bayesian optimizationgradient-freegaussian processpost-traininglow-dimensional subspace
FiGuRO: Intrinsic Dimension Estimation for Multi-Modal Data
FiGuRO introduces Fidelity-Guided Rank Optimization, a framework for estimating the Intrinsic Dimension (ID) of uni- and multi-modal data while disentangling shared and private information. The method employs truncated singular value decomposition and a dynamic algorithm to optimize low-rank projections across latent spaces, eliminating the need for auxiliary loss functions. FiGuRO outperforms existing ID estimation techniques in robustness to hyperparameter changes and accurately captures distinct ID scales and subspace ratios. It also enables post-hoc disentanglement of multi-modal representations in pretrained models, demonstrating effectiveness across simulations and real-world datasets.
intrinsic dimensionmulti-modal datalow-rank projectionssingular value decompositiondisentanglement
Diffract: Spectral View of LLM Domain Adaptation
The paper introduces Diffract, a toolkit for spectral analysis of domain adaptation in large language models via continual pre-training (CPT). Using singular value decomposition, the authors demonstrate that CPT primarily alters singular vectors rather than spectra, with domain-specific attention head heterogeneity enabling selective head rewinding. Key findings include 60% head update removal without quality loss, 4% accuracy gains from rewinding low-importance heads, and smooth domain interpolation between CPT checkpoints. The analysis covers mathematics, instruction, code, and natural text domains.
continual pre-trainingsingular value decompositionattention headsdomain adaptationspectral analysis
Spectral Embeddings of Degree-$α$ Laplacians in Random Dot Product Graphs
The paper establishes a row-wise central limit theorem for a continuum of degree-normalized spectral embeddings in random dot product graphs, unifying common matrix representations like the adjacency matrix and symmetric Laplacian. Using a projected-Gaussian Bayes-error diagnostic, the authors compare embeddings in two-community stochastic block models, finding no universally optimal normalization. Instead, the preferred degree normalization depends on network density, community imbalance, and block-probability structure, with stronger normalization favored in sparser or more imbalanced networks. These results provide a distributional framework for understanding when alternative normalizations improve spectral clustering performance.
spectral embeddingsrandom dot product graphsdegree normalizationcentral limit theoremstochastic block models
UniProbe: A Learnable Token-Level Hallucination Detector for Large VLMs using Multi-Structural Internal Representations
UniProbe introduces a lightweight, learnable token-level hallucination detector for Large Vision-Language Models (LVLMs) that leverages multi-structural internal representations from a single forward pass. It constructs a directed graph over image patches, query tokens, and generated tokens, using attention weights to encode relations, and processes this trace with alternating structure-aware modules: a GNN for relational evidence, a ViT for 2-D visual geometry, and a GRU for response order. UniProbe includes a streaming variant for hallucination-aware decoding and a self-adaptation strategy. It achieves state-of-the-art token-level and object-hallucination detection, reducing object hallucinations by up to 55% at 1.06× the latency of standard generation.
hallucination detectionlarge vision-language modelstoken-level localizationattention weightsstructure-aware modules
MoE Proxy Models for Low-Cost Failure Reproduction and Diagnosis in LLM RL Post-Training
The paper introduces MoE proxy models for efficient failure reproduction and diagnosis in RL post-training of LLMs, addressing high computational costs and debugging complexity. The method employs structure-preserving, clustering-based expert pruning to select representative experts while maintaining backbone architecture, routing mechanisms, and task capabilities. Experiments demonstrate that proxy models reduce accelerator requirements by 50%-87.5% and achieve up to a 33.3x reduction in per-step NPU-hour cost, while preserving training dynamics and reproducing fault responses consistent with original models. This approach enables low-cost fault investigation and auxiliary diagnosis in RL post-training pipelines.
proxy modelsexpert pruningrl post-trainingfault reproductionnpu-hour cost
Path Integral Value Matching for Linear Quadratic Stochastic Optimal Control
The paper introduces Path Integral Value Matching (PI-VM), a value-based algorithm for Linear Quadratic Stochastic Optimal Control (LQ-SOC) that addresses computational inefficiency and instability in policy-based methods. By truncating and marginalizing the path integral formulation, PI-VM derives a recursive value function, enabling temporal-difference learning and off-policy training via Girsanov theorem and experience replay. Benchmarks show PI-VM matches state-of-the-art accuracy with 10× efficiency gains in low dimensions and mitigates mode collapse in high-dimensional tasks.
linear quadratic stochastic optimal controlpath integral controltemporal-difference learninggirsanov theoremoff-policy training
Long-Time Trajectory Approximation via SA-NODEs: Model Predictive and Floquet Strategies
The paper introduces two training strategies for semi-autonomous neural ODEs (SA-NODEs) to improve long-time trajectory approximation. The model predictive strategy uses adaptive horizon partitioning and data resets to achieve uniform error bounds with linear parameter scaling, while the Floquet strategy leverages limit cycle stability for autonomous systems, yielding linearly growing error bounds. Theoretical guarantees include uniform-in-time orbital stability for periodic architectures, with numerical validation on four benchmarks confirming error laws and hypothesis measurements.
semi-autonomous neural odeslong-time approximationmodel predictive strategyfloquet theorylimit cycle stability
SQuaT: Self-Supervised Knowledge Distillation via Student-Aware Quantized Teacher Features
The paper introduces SQuaT, a label-free Quantization-Aware Training (QAT) framework that addresses range mismatch in Knowledge Distillation (KD) by quantizing teacher features using student parameters. This theoretically eliminates the irreducible lower bound on distillation loss observed in prior QAT+KD methods. Experiments demonstrate consistent improvements over baselines, particularly in extreme low-bit (1-2 bit) settings, without relying on specific architectural assumptions. The method is validated across diverse models and quantization configurations.
quantization-aware trainingknowledge distillationlow-bit quantizationself-supervised learningfeature alignment
Retrieval-Augmented Vision Foundation Models for Robust Leukemia Cell Classification across Multiple Microscopy Datasets
A two-stage retrieval-augmented framework improves leukemia cell classification robustness across heterogeneous microscopy datasets. Stage 1 performs binary leukemia classification using 122,167 single-cell images, while Stage 2 conditionally classifies Acute Lymphoblastic Leukemia (ALL) and Acute Myeloid Leukemia (AML) subtypes using 69,400 images. Three encoders (DinoBloom, BiomedCLIP, CLIP) are benchmarked under linear probing, Low-Rank Adaptation (LoRA), and Retrieval-Augmented Classification (RAC) with top-k image retrieval. Cross-dataset training with harmonized labels evaluates domain-shift generalization on a held-out protocol, revealing dataset-specific artifacts versus cytomorphological features.
retrieval-augmented classificationlow-rank adaptationdomain-shift generalizationcytomorphological featuressingle-cell images
IADD-TR: Intervention-Aware Dynamics Decoupling with Targeted Regularization for Model-Based Reinforcement Learning
IADD-TR introduces a unified framework for model-based reinforcement learning (MBRL) that combines Intervention-Aware Dynamics Decoupling (IADD) and Targeted Regularization (TR). IADD factorizes transitions into an action-intervention stage and an action-free natural evolution stage, using a zero-action anchor to ensure robust generalization. TR augments the critic with an action-density-scaled residual correction and optimizes a targeted loss, enabling doubly robust policy-gradient estimation. Experiments on five MuJoCo tasks demonstrate that IADD-TR achieves competitive returns with improved sample efficiency.
model-based reinforcement learningdynamics decouplingtargeted regularizationpolicy-gradient estimationsample efficiency
InSight-doc: Agentic Visual Perception for Long-Document Understanding
InSight-doc introduces an agentic visual perception framework for efficient long-document understanding, treating visual resolution as an adaptive reasoning-time resource. The method begins with low-resolution input and selectively zooms into high-resolution regions for finer evidence, eliminating reliance on external retrievers. Training involves a corpus of 17.9K supervised fine-tuning (SFT) examples with region-level zoom-in trajectories and 19.2K reinforcement learning (RL) examples. InSight-doc-8B improves baseline accuracy by 4.3--16.4 points on document VQA benchmarks, reduces hallucination by over 40%, and cuts inference latency by 41%--68% while maintaining accuracy.
agentic visual perceptionsupervised fine-tuningreinforcement learningdocument vqainference latency
ProbGuard: Calibrated Safety Risk Estimation from LLM Output Distributions
The paper introduces ProbGuard, a probabilistic architecture-agnostic guardrail for estimating safety risks from early output distributions of Large Language Models (LLMs). ProbGuard leverages Monte-Carlo sampling to estimate the unsafe probability of continued generation dynamics based on LLM prefix distributions. Post-trained on distributional signals and calibrated safety risks, ProbGuard achieves superior calibration performance, reducing the average Brier score and Expected Calibration Error (ECE) by 79.6% and 71.9%, respectively, across nine model-dataset combinations. Additionally, it limits the attack success rate to ≤1% across six jailbreak attacks by observing LLM outputs from the first ten decoding steps.
probabilistic guardrailmonte-carlo samplingsafety risk estimationlarge language modelsexpected calibration error
Pair-Centric Graph Rewiring for Over-Squashing via Optimal Transport-Guided Communication Alignment
PairAlign introduces a pair-centric graph rewiring framework to address over-squashing in message-passing neural networks (MPNNs) by optimizing structural demand-support shortages. The method combines original-graph structural demand with finite-hop propagation support, identifying poorly supported interactions via a computable proxy for Jacobian-based shortage. PairAlign employs an Optimal Transport-guided mechanism to coordinate edge additions, ensuring broad coverage of shortage targets and structural compatibility. Experiments on standard graph benchmarks demonstrate PairAlign's effectiveness across various MPNN backbones, validating pair-level rewiring as a solution to over-squashing.
graph rewiringover-squashingoptimal transportmessage-passing neural networkspair-centric
BooST: Bridging Semantics and Motions for Efficient Skill Transfer
BooST introduces a two-stage framework for efficient robot skill transfer by bridging semantic intent and motion dynamics. The method first employs a cross-modal VQ-VAE to unify skill representations, then distills these into lightweight policies for downstream adaptation. Evaluations demonstrate BooST's superior few-shot adaptation, cross-domain transfer, and robustness to visual distractors, while maintaining practical deployment efficiency.
skill transfervq-vaesemantic intentmotion dynamicsfew-shot adaptation
$β$-VAEs as Effective Theories: Tolerance-Dependent Dimension
The study demonstrates that β-VAEs function as effective theories by analyzing tolerance-dependent latent dimension reduction in nonlinear VAEs. Investigating fully connected nonlinear VAEs trained on WorldClim data, the authors show that increasing regularization strength induces a spectral cutoff, collapsing low-utility latent coordinates. While nonlinear interactions shift and broaden collapse onsets, disrupting exact threshold-utility alignment, the utility ordering is preserved across resolved ranks. Effective-dimension curves reveal a head-tail tradeoff: deeper models concentrate utility into fewer coordinates but degrade tail fidelity. These findings extend the effective-description logic from linear Gaussian VAEs to nonlinear architectures.
β-vaespectral cutofflatent dimensionnonlinear interactionseffective theory
BREAD: Baseline-Referenced Explanations for Anomaly Diagnosis
The authors propose BREAD, a scalable baseline-referenced diagnosis method for AI-based anomaly detection that improves feature relevance explanation faithfulness. The method leverages both anomalous observations and normal baseline information, addressing limitations of existing model-agnostic explainable AI (XAI) approaches that suffer from scalability issues or noise feature attribution. Mathematical guarantees demonstrate superior faithfulness in detecting anomaly-causing features compared to LIME under mean-shift anomaly settings. Simulation studies and a real-world case study validate that BREAD generates more accurate and faithful diagnosis results for AI-based statistical process monitoring (SPM) in high-dimensional, nonlinear settings.
anomaly detectionexplainable aistatistical process monitoringfeature relevancemean-shift
Iterative Erasure Count Is Not an Affine-Invariant Concept Dimension
The article demonstrates that iterative erasure counts, commonly used to estimate concept dimensions in neural representations, are not affine-invariant and thus not intrinsic measures. It distinguishes between model-defined quantities (generating dimension, sufficient linear dimension, minimum guarding rank) and procedure-defined quantities (stopping count, cumulative edit rank). Through Gaussian constructions and empirical tests with Adam/QR calibration, the authors show that erasure counts change under invertible reparameterizations, while model-defined quantities remain invariant. The findings highlight that iterative erasure yields procedure-relative estimands influenced by representation geometry and measurement methods, not standalone semantic dimensions.
iterative erasureaffine-invariantconcept dimensionneural representationreparameterization
MARCO: Click-Intent Decomposition for Calibrated Ads Conversion Prediction
MARCO introduces a framework for calibrated ads conversion prediction by decomposing click intent, addressing biases in standard CVR models that treat all clicks uniformly. It leverages logged click types as behavioral labels to train per-intent CVR heads, composing their estimates under a predicted intent distribution at serving time. Theoretical analysis shows decomposition never increases population risk and provides exact headroom under squared loss. Deployed at binary intent granularity, MARCO achieves ~100% per-intent calibration, increases conversions per click by +2.80%, and improves topline metrics by +0.98%. Multi-impression attribution is framed as credit assignment with bias-variance tradeoffs, enforcing consistency conditions at scale.
click-intent decompositioncvr estimationbehavioral labelscredit assignmentcalibration bias
Benchmarking LLM-Guided Control-Plane Policies for Backend Fault Isolation in HAProxy
The study evaluates LLM-guided control-plane policies for backend fault isolation in HAProxy, demonstrating that models with ≥3B active parameters reduce client-perceived 5xx errors by 88% compared to static baselines. The benchmark tests 15 open-weight models (0.35B–35B parameters) across architectures (dense, MoE, sparse), reasoning modes, and fleet scales (3–9 backends), totaling 240 runs. Results reveal a capability threshold near 3B parameters, with Gemma 4 E2B (2B) as an outlier. Trade-offs include 2.6–2.8× tail latency inflation and 10× token cost, suggesting optimal operation with supra-threshold models in non-reasoning mode under guardrails.
load balancingfault isolationllm-guided policieshaproxyprometheus
When Do Anchor-Based Pointwise LLM Rerankers Help? Retriever Quality, Statistical Scope, and Anchor Design
The study investigates anchor-based pointwise LLM reranking, focusing on GCCP/PAGC, through reproduction and controlled analysis. Initial reimplementation revealed undocumented details critical for performance (0.24 vs. reported 0.66 nDCG@10). After recovering eight such details, results were reproduced within 1.6%. Analysis shows contrastive scoring is robust, but combining it with standard relevance scores benefits BM25 retrievers more than dense models like E5. A simpler anchor construction method (interleaving top sentences) matches or outperforms the original. Findings hold across LLM backbones, including a 4-bit quantized 72B model.
anchor-based rerankingcontrastive scoringllm rerankingbm25dense retriever
Link-adaptive digital twin for robust physical-layer modeling in hybrid-amplified ultra-wideband optical networks
The paper proposes a link-adaptive digital twin (LA-DT) for hybrid-amplified ultra-wideband optical networks to improve physical-layer modeling accuracy and generalization. The method decomposes GSNR modeling into ASE, NLI, and signal power predictions, employs three dedicated DT models with linear modulation layers (LMLs), and uses domain discriminators for few-shot fine-tuning. Results show 56.0%, 58.4%, and 52.7% RMSE improvements for NLI, ASE, and signal power predictions, respectively, with a 55.8% GSNR estimation improvement, maintaining 0.159 dB RMSE in unseen scenarios with 20-shot adaptation.
digital twinultra-widebandraman amplifiergeneralized signal-to-noise ratiofew-shot learning
CARB: A Characterization-Guided Framework for CNN Inference Cost Prediction and Deployment Screening
The paper introduces CARB, a characterization-guided framework for predicting CNN inference costs (energy, latency, peak memory) and screening deployment candidates. It analyzes 13,419 CNN configurations on RTX 5090 and RTX 3080 GPUs, revealing divergent scaling behaviors: energy and latency diverge 3x under high computational demand, while memory transfers well across platforms. CARB employs a cascade-blended ensemble to jointly predict all three targets (R2 ~0.99) and a two-stage screening workflow that eliminates >90% of candidates, yielding Pareto-optimal shortlists validated on hardware.
cnn inferencegpu telemetryworkload characterizationdeployment screeningcascade-blended ensemble
Stay or Stray - A Dynamical Systems Viewpoint of Popularity Bias
The work provides a dynamical systems analysis of popularity bias in recommendation systems, where disproportionate interaction data from majority users degrades recommendations for niche users. Using a stochastic process model and ODE framework grounded in two-time-scale stochastic approximation, the authors characterize equilibrium points and derive conditions for popularity bias emergence versus symmetric user retention. Theoretical results are validated on synthetic data and real-world logs from a commercial music recommendation platform.
popularity biasrecommendation systemsdynamical systemsstochastic approximationequilibrium analysis
A Joint-Distribution Route to Fair Representations with Continuous Sensitive Attributes
The paper proposes a joint-distribution approach for fair representation learning with continuous sensitive attributes, avoiding nonparametric conditional law estimation. By measuring independence via a joint discrepancy $d(P_{Z,S}, P_Z \otimes P_S)$ (e.g., Hilbert-Schmidt independence criterion), the method eliminates the need for per-value conditional smoothing. Theoretical analysis shows the HSIC estimator achieves $O(n^{-1/2})$ convergence, outperforming conditional-route estimators' $O(n^{-2/5})$ rate. The FRHSIC implementation matches fairness-accuracy tradeoffs of baselines while reducing per-epoch training time.
fair representation learningcontinuous sensitive attributesjoint discrepancyhilbert-schmidt independence criterionnonparametric estimation
Detecting an Effect Is Not Learning to Act on It: A Reward-SNR Floor for LLM Acquisition Agents
The paper identifies a critical distinction between detecting a signal's average utility and learning a per-instance acquisition policy, introducing a reward-SNR floor governing the latter's feasibility. It demonstrates that learned routing policies fail to outperform random selection across various granularities, attributing apparent gains to noise order statistics. Structured Hypothesis Embeddings (SHE) are proposed as a concrete instantiation, leveraging a frozen LLM to generate ranked, confidence-scored intent hypotheses from user histories. Evaluated on MIND, REES46, and Amazon-Beauty datasets, SHE shows backbone- and regime-conditional utility but fails to enable per-instance acquisition due to datasets falling below the SNR floor. Code and reproducibility instructions are released.
reward-snr floorstructured hypothesis embeddingsper-instance acquisitionlearned routingnoise order statistics
Do Time-Series Forecasters Use the Right History: Recoverability, Recovery, and Functional Use of Temporal Delays
This paper investigates whether time-series forecasters utilize the correct historical inputs by addressing three key questions: recoverability of true delays, model reporting of delays, and functional use of historical data. The authors derive input-conditioned recoverability measures to distinguish intrinsic ambiguity from model error and demonstrate that models can achieve reliable delay reports and near-oracle forecast risk while still using incorrect lags. Empirical results on the point-delay task reveal that 55.4% of N-HiTS and 92.7% of TCN cases functionally ignore reported history despite correct delay reports. The study concludes that accurate forecasts and correct delay reports do not guarantee proper historical input usage.
time-series forecastingdelay recoverabilityforecast riskmatched masking testone-hot control
A lower bound for stepsize-based acceleration of gradient descent
We establish a lower bound of Ω(T^{-1.9319}) for the last-iterate convergence rate of gradient descent with predetermined nonnegative stepsize schedules in smooth convex optimization, demonstrating that stepsize-based acceleration alone cannot achieve the optimal O(T^{-2}) rate. The proof leverages carefully designed stepsize schedules and avoids momentum or other algorithmic modifications, contrasting with prior work that achieved O(T^{-log_2(1+√2)}) convergence. This result, developed using GPT-5.6 Sol Pro under author guidance, rigorously quantifies the limitations of stepsize-based acceleration for plain gradient descent.
gradient descentstepsizeslower boundconvex optimizationlast-iterate convergence
How Robust Are LLMs to Vietnamese Dialects?
The study introduces VialectBench, the first systematic benchmark evaluating LLM robustness to Vietnamese dialect variation across six dialect groups (PNT1-PNT4, PNB) in four tasks (ER, NLI, QA, MCQA). Using 400 Standard Vietnamese instances and 2,400 dialectal rewrites, they measure performance degradation via model-relative likelihood shifts. Results show a 2.82% average performance drop across ten instruction-tuned models, with QA most affected (6.17% drop for PNT3). No model achieves dialect-invariance, and Central dialects (PNT1-PNT4) exhibit the highest harmful-flip rate (6.54%).
vialectbenchdialect robustnessmodel-relative likelihoodharmful-flip rateinstruction-tuned models
Post-Calibration Reliability Reranking of Relevance Decisions via Label-wise Monotone Projection
The paper introduces Label-wise Monotone Reliability Projection (MRP), a post-hoc method to improve reliability ranking of relevance decisions by learning label-wise monotone functions that map calibrated confidence to correctness reliability while preserving predicted labels and class probabilities. MRP addresses residual reliability differences within the same confidence level, enhancing reranking and fallback utility without compromising full-coverage accuracy or Expected Calibration Error (ECE). Evaluated on six information access datasets with multiple calibrators, MRP demonstrates gains primarily from label-wise residual reliability rather than global confidence remapping. The method also analyzes compatibility of reliability scores with top-label probability geometry.
post-hoc calibrationreliability rerankingmonotone projectionresidual riskinformation access
TideRL: Boosting Agentic RL Goodput with Readiness-Aware Scheduling
TideRL introduces a readiness-aware elastic reinforcement learning system to optimize training goodput in multi-turn agentic workloads. It employs Continuous Task Batching (CTB) to preserve rollout state, Resource-Aware Ref-Actor Pipelining (RA²P) for decoupled streaming or colocated aggregation, and Elastic Resource Scaling (ERS) to dynamically allocate resources based on readiness signals. Evaluations on text-only and multi-modal workloads demonstrate up to 5.6× improvement in RL training goodput over synchronous baselines, a 1.58× increase in KV cache hit rate, 44.3% reduction in per-step training time, and 77.6% reduction in total waiting time, while maintaining comparable task performance.
reinforcement learningmulti-turn agentic workloadskv cache hit ratecontinuous task batchingelastic resource scaling
Automatic Field-of-View Adjustment for a View-Expansive Microscope via LSTM-Based Gaze and Pipette Motion Interpretation
The authors propose an AI-driven automatic field-of-view (FOV) adjustment system for view-expansive microscopes, addressing workflow interruptions in intracytoplasmic sperm injection (ICSI) procedures. The system integrates multiview imaging with galvanometer mirrors and high-speed vision to enable simultaneous large FOV and high-resolution acquisition without lens switching. A long short-term memory (LSTM) model predicts optimal FOV size by analyzing real-time pipette position, velocity, and operator gaze data, trained on expert ICSI procedure data. Experimental results show the system reduces average task completion time from 60.5 to 48.0 seconds (p < 0.001), enabling novice operators to match expert-level speeds.
field-of-viewlstmmultiview imaginggalvanometer mirrorsintracytoplasmic sperm injection
Do Judges Behave Like Algorithms?
This study investigates whether magistrate judges in Harris County, Texas, exhibit algorithmic behavior in misdemeanor bail hearings by analyzing their decision-making consistency and variable reliance. Using court data, the authors train machine learning models per judge, measure variable importance, and compare outcomes for similar cases. Results indicate judges largely follow small, interpretable decision rules (e.g., criminal history, age), but exhibit significant inter-judge inconsistency, suggesting unequal treatment. The work identifies cases where algorithmic explanations fail, highlighting opportunities for judicial system improvement.
judicial decision-makingalgorithmic behaviorvariable importancemachine learning modelsbail hearings
Share First, Route What Remains: A Unified Framework for Token-Adaptive MoE Computation
The paper introduces UniF-MoE, a unified framework for token-adaptive computation in mixture-of-experts (MoE) models, addressing the dependency between shared computation and residual expert demand. The method decomposes experts into key-value channels, employs aligned blocks for shared computation, and uses a shared-demand score with key prototypes to route residual demand, enhanced by a Gram regularizer for diverse routing. Evaluations on DomainBed and GLUE demonstrate improved predictive performance over static and dynamic MoEs while reducing computation, latency, and memory usage.
mixture-of-expertstoken-adaptive computationkey-value channelsgram regularizerdynamic routing
Efficient Weak-Entropy PINN for Solving Hyperbolic Conservation Laws
Proposes Weak-Entropy PINN (WEPINN), a physics-informed neural network framework for hyperbolic conservation laws with discontinuous solutions. The method enforces weak (integral) formulations of governing equations, incorporates entropy conditions for physical admissibility, and uses discrete fast Fourier transform for efficient numerical integration. Numerical experiments on 1D/2D scalar and systems of conservation laws demonstrate accurate resolution of sharp discontinuities and interactions between shock/rarefaction waves, outperforming prior PINN approaches that rely on strong priors or artificial smoothing.
physics-informed neural networkshyperbolic conservation lawsweak formulationentropy conditiondiscontinuous solutions
Dreamer-SAC: Off-Policy Learning in Latent World Models for Sample-Efficient Autonomous Driving
The paper introduces Dreamer-SAC, a sample-efficient reinforcement learning framework for autonomous driving that combines a recurrent state-space world model with off-policy soft actor-critic optimization in latent space. The method leverages both real interactions and short-horizon generated trajectories, employing n-step target estimation and multi-objective supervision to balance driving efficiency and safety. Experiments demonstrate superior performance over DreamerV3, SAC, and PPO baselines with fewer environment interactions, revealing an inverted-U relationship between rollout horizon and policy performance, with optimal results for short-horizon rollouts.
world modeloff-policy learninglatent spacen-step estimationautonomous driving
Generator-Guided Inverse Sampling for Lévy-Driven Generative Models
The paper introduces a generator-guided inverse sampling method for Lévy-driven generative models, addressing the nonlocal reverse process inherent in Lévy dynamics. By analyzing forward and reversed Markov generators, the authors decompose the dynamics into diffusion, small jump, and large jump components, enabling a structured reverse sampler. A neural network amortizes the rate of large jump activities while analytically derived conditional distributions generate jump amplitudes, enhancing interpretability and controllability. Efficient implementation techniques avoid costly high-dimensional integration and sampling. Applied to OFDM-SISO channel estimation under mixed Gaussian and impulsive noise, the method demonstrates robust performance with favorable complexity-performance tradeoffs.
lévy-driven generative modelsmarkov generatorsinverse samplingnonlocal density ratioofdm-siso channel estimation
Fisher8: Stabilizing Neural Heteroscedastic Regression via Output-Layer Fisher Geometry
Fisher8 stabilizes neural heteroscedastic regression by aligning gradient updates with the Fisher geometry of the loss landscape, addressing instability in joint mean and uncertainty prediction. The method derives an output-layer gradient correction that reorients and rescales updates using Fisher geometry, eliminating data-dependent hyperparameters beyond learning rate and enabling an approximate KL trust radius. Prior stabilization techniques converge on overlapping components of this geometric correction. Empirical results demonstrate Fisher8's superior likelihood-error tradeoffs, calibrated uncertainty estimates, and uncertainty-aware feature spaces across multidimensional regression and representation-learning tasks.
heteroscedastic regressionfisher geometrygradient correctionuncertainty estimationrepresentation learning
Invertible Logits Transformation for Accuracy-Preserving Post-Hoc Uncertainty Calibration
The authors introduce Invertible Logits Transformation (InvLT), a post-hoc calibration method that preserves classifier accuracy while addressing nonlinear miscalibration. InvLT applies a shared scalar MLP element-wise to pre-softmax logits, maintaining parameter independence from class count and softly enforcing monotonicity via a paired inverse network. This approach avoids computational overhead while empirically preserving original classification accuracy across evaluations. Experiments on standard image classification benchmarks demonstrate InvLT's consistent superiority over existing post-hoc calibration methods on standard calibration metrics.
post-hoc calibrationinvertible logits transformationnonlinear miscalibrationmonotonicityscalar mlp
Accelerated Learning of High Dimensional Functions with a Tensor-Featured Training Network
The paper introduces a method to accelerate high-dimensional function learning in deep neural networks (DNNs) by incorporating contextual features into the first layer. The approach alternates between optimizing DNN parameters via gradient descent (with fixed input-feature basis) and updating the feature layer, which combines fast matrix-free rank-1 features and tensor network (TN)-decomposed complex features. A randomized tensor decomposition strategy reduces storage costs by ≥8 orders of magnitude when distilling pretrained DNNs into TNs. Experiments demonstrate effective training for 5-40 dimensional functions.
high-dimensional learningtensor network decompositiongradient descentrandomized decompositioncontextual features
Beyond Detection Accuracy: Measuring Explanation Cost, Stability, and Utility for Resource-Aware IoT Intrusion Detection
The study proposes a multi-criteria evaluation framework for IoT intrusion detection systems, assessing predictive accuracy, explanation cost (TreeSHAP), local stability, and selective explanation policies. Using a leakage-safe CICIoT2023 corpus with 39-feature hashes, it compares Logistic Regression, Decision Tree, Random Forest, and XGBoost on natural/balanced distributions. Results show XGBoost achieves highest predictive performance (700.759s vs 1.471s TreeSHAP cost for Random Forest vs XGBoost at 5k samples), while Random Forest exhibits strongest explanation stability. Selective explanation policies yield 15-32% compute savings at 90-95% false-negative coverage on balanced data.
intrusion detectiontreeshapexplanation stabilityiot securityselective explanation
On the Importance of Geometric Nonlinearity and Temperature-Dependent Properties in Multi-Material Thermo-Mechanical Topology Optimization
The work quantifies modeling errors in thermo-mechanical topology optimization by introducing a physics-informed framework with finite-strain quadratic-Hencky constitutive modeling and temperature-dependent material properties for Ti-Cu-steel systems. The method employs simultaneous analysis-and-design optimization of thermal actuators and grippers under mass/manufacturability constraints, comparing linear vs. nonlinear kinematics and temperature-independent vs. dependent properties. Results show nonlinear kinematics is critical: linear models misclassify rotation as compressive strain, yielding weaker designs, while full-physics optimization produces 12-15% stronger, temperature-robust devices despite 20-30% longer design times.
thermo-mechanical topology optimizationfinite-strain kinematicsquadratic-hencky modelmulti-material designtemperature-dependent properties
MERA: Model Evolution and Routing with Skill Adaptation for Agentic Systems at Scale
MERA introduces a multi-cycle adaptation framework that enhances small-model capability for LLM agentic systems, rather than merely routing tasks based on difficulty. The method replays failed student invocations to generate verified teacher demonstrations, distills recurring procedures into a SkillBook, and fine-tunes a student LoRA adapter via supervised learning and optional GRPO. Deployment employs a cost-calibrated router with verifier-backed fallback, ensuring task quality preservation. Empirical results show Qwen2.5-Coder-1.5B improving from 28.7% to 49.7% pass on HumanEval+MBPP, retaining 88.3% pass at 60.8% cost under fallback. Qwen3.5-2B matches an unadapted 4B model on TAU-2.
multi-cycle adaptationskillbooklora adapterverifier-backed fallbackcost-calibrated router
Topological Feasibility Guarantees for Differentiable Predictive Control
The paper establishes deterministic feasibility guarantees for differentiable predictive control (DPC) through topological analysis of reachable safe sets, eliminating the need for online safety filters. By embedding differentiable system dynamics into the computational graph, the authors propose a self-supervised offline policy learning strategy using Control Barrier Functions (CBFs) as a proxy loss. Theoretical analysis and closed-loop simulations show that constraint violations monotonically decrease to zero with increasing training samples, providing formal safety certificates unattainable with black-box methods like RL or supervised learning.
differentiable predictive controlfeasibility guaranteescontrol barrier functionstopological analysisself-supervised learning
UniMod: Enhancing Multi-Modal Medical Diagnosis through Cross-Modality and Within-Modality Alignment
UniMod introduces a multi-modal medical diagnosis framework that mitigates shortcut learning by enforcing independent modality-specific predictions alongside multi-modal classification. It incorporates cross-modality alignment for knowledge transfer and within-modality supervised contrastive alignment to enhance feature extraction. Evaluated on Harvard-Glaucoma and CheXpert Plus, UniMod achieves AUC scores of 0.850 and 0.966, outperforming OGM-GE and Gradient Blending by 1.6-1.8% and over 5%, respectively. Additionally, it extends to 5-class multi-label diagnosis without architectural modifications, improving mean AUC by 0.097 over CGGM.
shortcut learningcross-modality alignmentsupervised contrastive alignmentmulti-modal classificationauc
SeFaR: Semantic Feature-aware Robustness Testing of Deep Neural Networks
SeFaR introduces a semantic feature-aware robustness testing framework for deep neural networks in vision tasks, addressing failures from rare scenarios. The method employs a hierarchical concept model with domain knowledge integration, leveraging diffusion and vision-language models for photorealistic perturbations and feature discovery. Case studies demonstrate its effectiveness in maintaining requirement satisfaction while identifying failure-inducing semantic concepts and relating them to model decisions.
semantic robustnesshierarchical concept modeldiffusion modelsvision-language modelsrequirement satisfaction
Power law graph attention: exact generalization of scaled dot-product attention, empirical collapse at inference
The paper introduces Power Law Graph Attention (PLGA), a novel attention mechanism that generalizes scaled dot-product attention (SDPA) by replacing its fixed bilinear form with a learned, input-generated operator $G_{LM}$ derived from a positive tensor $A_{LM}$ via elementwise power laws. The architecture is formally verified, with theorems showing PLGA exactly contains SDPA when $G_{LM}=I$, and $A_{LM}$ exhibits Perron-Frobenius structure. An inference-collapse theorem demonstrates that exact input invariance reduces inference to generalized SDPA with a constant operator. Empirical measurements show relative fluctuations below $10^{-6}$, and blockwise training aligns with sequential scoring on TruthfulQA within $5\times 10^{-5}$ per item. Machine-checked proofs in Lean 4 validate core claims.
power law graph attentionscaled dot-product attentionperron-frobeniusinference-collapsetruthfulqa
Stochastic Emulation of a Fully Coupled Preindustrial E3SMv3 Simulation
The authors introduce a stochastic coupled emulator of E3SM version 3, leveraging the SamudrACE framework to integrate a stochastic atmosphere emulator (ACE2S) with a full-depth ocean emulator (Samudra). The system is fine-tuned using a probabilistic objective, enabling the atmosphere to drive internal variability in the ocean. Trained on 105 years of preindustrial control simulation and validated on 400 independent years, the emulator accurately reproduces E3SMv3's mean climate state with minimal biases. It maintains internal variability across timescales, particularly in ENSO power spectra, SST anomalies, and sea ice dynamics, while underestimating extreme tropical precipitation events. This demonstrates the emulator's fidelity in capturing long-timescale variability, though extrapolation to rare extremes remains challenging.
stochastic emulatore3smv3samudraceenso power spectruminternal variability
BreastMammo and DenseMammo: Benchmarks for Mammography Domain Generalization
The authors introduce BreastMammo and DenseMammo, two benchmark datasets for evaluating domain generalization in multi-view mammography. They propose a domain generalization framework using foreground-only histogram matching to mitigate vendor-specific acquisition biases in breast density classification. A Swin Transformer backbone achieves 98.32% AUC in 5-fold cross-validation, with external validation on TNMammo and LUMINA datasets showing superior performance over MixStyle and Discrete-Fourier-Transform-based methods.
domain generalizationmammographyhistogram matchingswin transformerbreast density classification
CRHT: A Continuous Regression Hybrid Transformer for Vessel Trajectory Prediction with Online Cluster Sampling
The paper proposes CRHT, a Continuous Regression Hybrid Transformer for vessel trajectory prediction, addressing geographic bias and navigational realism in AIS data. The method combines 1D convolutional layers for local kinematic features with multi-head attention for global temporal context, augmented by an online K-means cluster sampling strategy to handle spatial data imbalance. CRHT achieves state-of-the-art short-term forecasting performance, demonstrating optimal precision-maneuver tracking balance for real-time maritime surveillance, with lowest errors at 1-hour horizons.
trajectory predictionhybrid transformeronline cluster samplingais datamulti-head attention
Off-Axis, On Purpose: Where a Transformer Computes Concepts and Why it Does So
The study demonstrates that transformers compute concepts in an off-axis subspace orthogonal to the unembedding direction, with distinct functional phases. A 12-layer model first writes intermediate states 75-96° off-axis, insulating composition from vocabulary; damage from forcing on-axis projection is 64-84× worse than random rotation. The second phase aligns answers on-axis via addition. Forcing early on-axis projection reduces concept workspace from 25 to 14 dimensions without affecting perplexity on LAMBADA or BLiMP. Prescribing orthogonal geometry via fixed rotations succeeds (9/9 seeds), while loss-based methods often collapse (6/8 seeds). Random pre-training bases are adoptable without quality loss.
transformerunembeddingorthogonal subspaceperplexitylambada
STCAD: Scalable Trajectory Clustering and Anomaly Detection on Terabyte-Scale AIS Data
The paper introduces STCAD, a scalable unsupervised framework for clustering and anomaly detection in terabyte-scale maritime trajectory data. The method employs a custom BERT-based model for trajectory encoding via masked token modeling, followed by CURE hierarchical clustering to produce interpretable trajectory groups without predefined cluster counts. Anomaly detection combines reconstruction loss and clustering noise assignment. Evaluated on a national-scale AIS dataset with billions of messages over one year, the framework demonstrates stable clustering and clear separation between normal and anomalous vessel behavior.
trajectory clusteringanomaly detectionbert-based encodingcure clusteringais data
DualSpectralCF: Training-Free Sign-Aware Spectral Collaborative Filtering
DualSpectralCF introduces a training-free framework for sign-aware spectral collaborative filtering, combining signed input signals and item-item operators with existing spectral backbones. The method encodes explicit negative feedback via $\mathbf{r}_u^{\pm}$ and $\mathbf{M}^{\pm}$, adding only two hyperparameters while remaining backbone-agnostic. Evaluated on five benchmarks with ChebyCF, GF-CF, and Turbo-CF backbones, it achieves Recall@20 improvements up to +32.6% (tuned) and +16.0% (default), with 7.7-155.3× speedup over SIGformer while maintaining 70.7-90.7% accuracy. Cold-start users benefit most, showing up to +29.2% gains on Epinions.
spectral collaborative filteringsign-aware recommendationtraining-freecold-startexplicit feedback
A Graph Neural Network--Guided Genetic Algorithm for Physical Internet Supply Chain Optimization under Cost Uncertainty
The study introduces a graph neural network-guided genetic algorithm (GNN-GA) for optimizing Physical Internet supply chain networks under cost uncertainty, addressing discrete assignment decisions and continuous flows. The GNN estimates hub-specific factory-selection probabilities to initialize the GA population and adapt mutations based on prediction uncertainty, with each candidate assignment evaluated via LP optimality. Experiments on 15 instances demonstrate GNN-GA's superiority over simulated annealing and standard GA, particularly in learned initialization, with transferability tested on independently generated instances. Ablation studies reveal that initialization contributes most to performance improvements, while entropy-guided mutation effects vary by instance.
graph neural networkgenetic algorithmphysical internetmin-max regretsimulated annealing
Sequential Modality Dropout for Robust Multi-Modal Sequential Recommendation
Sequential Modality Dropout (SMD) enhances robustness in multi-modal sequential recommendation by addressing missing modalities during deployment. SMD independently erases each modality stream (image and text) with probability p during training, enabling the model to predict the next item without relying on any single modality. Evaluated on four backbones (MM-SASRec, IISAN, MISSRec, fMRLRec) across four Amazon domains, SMD improves text retention by 1.0 to 3.2x while maintaining full-modality accuracy. Under a 95% per-item missing rate, SMD retains 61% of HR@10 versus 22% without SMD. An optional cross-modal reconstruction loss further increases retention from 90% to 98% under severe text missingness, demonstrating SMD's effectiveness as a simple, architecture-agnostic solution.
sequential modality dropoutmulti-modal sequential recommendationretentioncross-modal reconstructionmodality dropout
A matched-integrator evaluation of Hamiltonian neural networks on pendulum and Kepler dynamics
The study conducts a matched-integrator evaluation of Hamiltonian Neural Networks (HNNs) against parameter-matched feedforward baselines on conservative dynamical systems. Using RK4-generated trajectories and central-difference derivative targets, HNNs demonstrate superior performance: on a nonlinear pendulum, they reduce mean energy drift by 42-fold and trajectory MSE by 15.8-fold over 16 periods (T=100), with bounded drift and lower variability. Similar improvements hold for the Kepler two-body problem. The HNN's architectural prior proves particularly effective in nonlinear phase-space regions, though its non-separable Hamiltonian form complicates symplecticity guarantees in Störmer--Verlet rollouts.
hamiltonian neural networksenergy driftsymplectic integrationnonlinear dynamicsmatched-integrator protocol
The Kuramoto Neural Operator: Learning to Solve PDEs via Coupled Oscillator Dynamics
The Kuramoto Neural Operator (KNO) is introduced as a novel architecture for learning solution operators of partial differential equations (PDEs) by modeling them through coupled oscillator dynamics. KNO represents solutions via the evolution of a latent field of interacting oscillators, leveraging the continuum limit of coupled oscillator systems to capture local interactions in physical space. Evaluated across diverse PDE benchmarks, KNO demonstrates superior predictive performance compared to existing approaches. An ablation study quantifies the contributions of architectural components, while analysis reveals that prediction error correlates with the synchronization degree of latent oscillators, providing mechanistic insights.
kuramoto neural operatorpartial differential equationscoupled oscillatorsoperator learningsynchronization
Boundary-Seeking Policy Gradient for Safe Reinforcement Learning
We introduce Boundary-Seeking Policy Gradient (BSPG), a first-order method for safe reinforcement learning that combines tangential and normal components to drive policies toward the active constraint boundary. BSPG leverages the algebraic Lagrangian form without learned dual variables, ensuring constraint residual convergence to zero with a finite-horizon $O(1/\sqrt{T})$ bound. The method guarantees stationary points on the active constraint set, satisfying KKT conditions when limits are local maximizers. Evaluated on the Safety-Gymnasium navigation task, BSPG achieves higher reward and tighter boundary tracking compared to baselines.
boundary-seeking policy gradientconstrained markov decision processeslagrangian formkkt conditionssafety-gymnasium
More Accurate, Less Human: Gestalt Grouping in Vision Models
The study introduces a behavioral battery to evaluate vision models' adherence to human Gestalt grouping principles, comparing 45 models across five training families (supervised, self-supervised, contrastive vision-language encoders, open-weight VLMs, and closed foundation models) on four tasks: mark-color odd-one-out, color-series counting, silhouette recognition, and object odd-one-out. Results reveal that conventional metrics fail to capture perceptual organization, with several closed models showing lower human alignment despite high benchmark accuracy. The battery provides a reusable benchmark for assessing model-human perceptual congruence without new user studies.
gestalt groupingvision modelsbehavioral batteryperceptual organizationhuman alignment
Intrinsic Structure: Spectral Identifiability for Mechanistic Interpretability
The paper establishes spectral identifiability for mechanistic interpretability by treating neural network forward passes as controlled dynamical systems and lifting them via the Koopman operator. This yields a finite linear realization whose spectrum is a coordinate-free model property, provably recoverable from M calibration samples at rate M^(-1/2) up to permutation. The method includes a minimax lower bound, a median-of-means variant for heavy-tailed activations, and a dissociation theorem showing non-normal realizations separate variance-carrying and information-carrying directions. Experiments on GPT-2 small, Gemma-2-2B, and Qwen3-8B-Base confirm spectral convergence (0.506±0.031 exponent on Qwen3-8B-Base) and show Koopman modes outperform random directions but lag principal components in indirect-object identification, with gaps decaying as predicted.
koopman operatormechanistic interpretabilityspectral identifiabilitydictionary learningminimax lower bound
REATS: LLM Reasoning-based Ensemble Learning for Adaptive Time Series Forecasting
REATS introduces an LLM-based ensemble learning framework for adaptive time series forecasting that combines textual pattern descriptions and numerical features through chain-of-thought reasoning. The method employs (i) structured hybrid input representations with retrieved similar-sample priors, (ii) multi-row weight supervision via percentage-table formatting, and (iii) a two-stage GRPO fine-tuning framework with reciprocal reward mapping for MSE optimization. Experiments on eight benchmarks show superior performance over baselines, with interpretable weight explanations and strong transfer learning to unseen models.
ensemble learningchain-of-thought reasoningtime series forecastinggrporeciprocal reward mapping
The Evaluation Protocol Determines the Result: An Independent Reproduction of LeWorldModel on TwoRoom
This independent reproduction of LeWorldModel on the TwoRoom environment reveals that evaluation protocol details critically determine reported performance (94.0% vs. 84.0% success rate). The study identifies four undocumented implementation conventions essential for replicating results: dense action frameskip, action-encoder width setting, ImageNet normalization, and action z-scoring. Key findings include the disconnect between one-step prediction accuracy and long-horizon planning success, and a 300x validation loss inflation from batch normalization. Protocol discrepancies between paper and code yield 14.0%-84.0% performance variation on identical episodes.
world modelsreproducibilityevaluation protocolbatch normalizationlong-horizon planning
SeFoRA: Sketch-Aggregated Federated Low-Rank Adaptation with Heterogeneous Client Ranks
(No summary returned.)
The Parser Already Knows: Lightweight Bias Correction in Constrained Decoding
This work introduces a lightweight logit correction method for Grammar Constrained Decoding (GCD) that preserves language model (LM) probability distributions while ensuring grammatical validity. The method leverages syntactic and lexical states from incremental parsing, which inherently encode future grammatical validity, to conditionally adjust logits offline without modifying LM weights. By restoring probability mass removed by rigid masking, it reconciles LM probabilistic integrity with grammar conformance. Evaluations across multiple grammars demonstrate that the approach outperforms both masking and online sampling baselines, with even its simplest variant—relying solely on candidate next tokens—matching or exceeding baseline performance.
grammar constrained decodinglogit correctionincremental parsingprobabilistic integritysyntactic validity
P3CA: Encoder-Agnostic Interpretation of Vision Foundation Model Embeddings via Spatial Probing
Proposes position-prompted PCA (P3CA), an encoder-agnostic method for interpreting vision foundation model embeddings via localized spatial probing. P3CA computes region-specific PCA projections from user-selected prompts, visualizing locally informative directions without retraining or task-specific labels. Implemented in EmbedVision (3D Slicer), evaluations on natural images, colorectal pathology embeddings, and spatial transcriptomics show P3CA reveals local structure obscured by global PCA, improves pathology discrimination by 3D projections, and enables cross-modal spatial representation comparison.
vision foundation modelsspatial embeddingslocal probingpca projectioninteractive visualization
ChronoSSM: Training for Temporally Aware Representations in Autoregressive State Space Models
ChronoSSM introduces a joint training approach for autoregressive State Space Models (SSMs) that simultaneously models events and timestamps using shared representations, addressing the limitation of treating timing as an auxiliary signal in sequence modeling. The method combines token and temporal generation objectives, enabling temporal supervision to update the backbone directly, unlike traditional two-stage approaches that freeze event representations. Evaluated across four domains with varying timestamp supervision, joint training consistently improves inter-arrival information recoverability from frozen representations without compromising content-generation quality. Results demonstrate that temporal supervision enhances temporal reasoning in autoregressive event modeling.
autoregressive state space modelstemporal supervisionjoint traininginter-arrival informationtoken generation
Deciding When to Switch: E-Processes for Adaptive Minimax Training for Generative Adversarial Nets
The paper introduces an e-process-based adaptive training procedure for GANs, formulating the discriminator-generator switching problem as sequential hypothesis testing. During discriminator updates, an e-process tests if the discriminator-induced distribution separation remains below a target level; during generator updates, another tests if it stays above a refresh level. The method provides anytime-valid Type I error control under adaptive updates and data-dependent switching. Experiments on synthetic multimodal distributions and image benchmarks show it matches or outperforms fixed-ratio baselines across multiple GAN objectives.
e-processesgenerative adversarial networkssequential hypothesis testingmin-max optimizationadaptive training
Observational Policy Ranking for SMB Financial Guidance from Multi-Action Accounting Logs
The paper introduces Covariate-Adjusted Residual Policy Learning (CAR-PL), a method for observational policy ranking in financial guidance for small and medium-sized businesses (SMBs). CAR-PL leverages multi-hot accounting logs to select among 34 ledger-derived business-change categories, regularizing selections by observational support. Evaluated on 85,078 company-month observations from 7,505 firms, CAR-PL achieves the highest Gross Profit point estimate (0.084) and demonstrates balanced category selection compared to baselines like the T-Learner and contextual value models. Results show CAR-PL and the T-Learner are statistically indistinguishable on growth KPIs, while CAR-PL maintains consistent performance across alternative scoring rules and treatment references.
observational policy rankingmulti-hot logscovariate-adjusted residualledger-derived categorieskpi point estimate
Bayesian Symbolic Regression with Entropic Reinforcement Learning
The paper introduces ERRLESS (Entropy-Regularized Reinforcement Learning for Expression Structure Sampling), a Bayesian approach for symbolic regression that samples from the posterior distribution over expressions using maximum-entropy reinforcement learning. The method employs a neural policy to sequentially construct abstract syntax trees, enabling posterior sampling at convergence. Evaluated on the Feynman benchmark, ERRLESS produces interpretable expressions and achieves competitive performance, with its posterior predictive mean demonstrating high $R^2$ compared to sequential Monte Carlo baselines.
symbolic regressionbayesian inferencemaximum-entropy reinforcement learningabstract syntax treesposterior sampling
Physics-Informed Machine Learning in Prognostics and Health Management: A Systematic Literature Review
This systematic review analyzes 212 studies on Physics-Informed Machine Learning (PIML) for Prognostics and Health Management (PHM), proposing a four-class taxonomy (observational bias, inductive bias, learning bias, hybrid) to categorize integration strategies. Results show PIML consistently outperforms conventional ML baselines across assets, though literature disproportionately focuses on lithium-ion batteries and bearings. While physics-informed methods demonstrate measurable performance gains, evidence for improved generalization, causality, or interpretability remains limited. The review identifies key research gaps: transferable architectures, comparative benchmarks of integration approaches, and lightweight uncertainty-aware models for real-time deployment.
physics-informed machine learningprognostics and health managementinductive biassystematic literature reviewlithium-ion batteries
Detecting Soft Skills in ML Engineering Roles CVs
This study addresses the gap in understanding how ML engineering candidates articulate soft skills in CVs, contrasting with demand-side perspectives. Using a balanced corpus of 300 CVs across ML engineers, data scientists, and software engineers, the authors employ an LLM-based pipeline to extract both explicitly listed and implicitly narrated soft skills, validated against human-annotated ground truth. They test 13 hypotheses derived from demand-side literature, finding 11 supported, one partially supported, and one refuted. Results show candidates predominantly disclose soft skills through narrative (3:1 ratio), particularly for leadership, coordination, and mentoring (88-96% narrative). Seniority nearly triples leadership articulation odds, while software engineers articulate leadership at half the rate of peers. Keyword-based screening systematically misses these skills.
soft skillscv-miningllm-based pipelinenarrative disclosureeffect sizes
Finding the Signal in the Spam: Jointly Learning Rewards and Worker Reliability from Pairwise Comparisons
The paper proposes an EM-based algorithm to jointly learn item rewards and worker reliability from pairwise comparisons, addressing unreliable crowdworker behavior. The method extends the Boltzmann-rational model with worker competencies, using Polya-Gamma latent variables to transform the logistic likelihood into a conditionally Gaussian form, enabling tractable optimization. Theoretical convergence guarantees are established via a matrix sensing reduction, and experiments on synthetic and real-world datasets demonstrate robustness to spammers and adversarial workers, outperforming baselines.
pairwise comparisonsboltzmann-rational modelpolya-gamma latent variablesmatrix sensingcrowdsourcing
Why Post-Norm Transformers Collapse: Attention Amplification and Gradient Repair Failure
The paper analyzes rank collapse in Post-Norm Transformers through a two-stage mechanism: causal attention initially amplifies token similarity via prefix-averaging, while SwiGLU provides weaker damping, and subsequent training fails to repair collapsed states due to geometrically decaying gradients caused by RMSNorm backward contraction. Using a token similarity state variable, the study theoretically characterizes initialization-time similarity growth and collapse-time gradient dynamics, with experiments on 48-layer decoder-only Transformers trained on C4 validating predicted similarity patterns and showing collapsed networks converge to frequency-based predictors with elevated loss floors.
post-norm transformersrank collapsetoken similarityprefix-averaginggradient contraction
Imaginative Generative AI: Crossing the Entropy Wall into Worlds Beyond Imitation
The paper introduces Imaginative Generative AI (IGA), a framework addressing diversity loss in generative models by incorporating spectral diversity into target-distribution design. IGA measures diversity via von Neumann entropy of the kernel covariance operator in a fixed representation space, enabling diversity repair below the Entropy Wall and controlled spectral extrapolation beyond it. The method formulates a regularization path from imitation to imagination, optimizing under a KL anchor to pretrained generators. IGA Guidance, a retraining-free inference-time method, is developed for score-based and diffusion models like DDPM and DDIM. Experiments on synthetic and vision benchmarks validate diversity repair and controlled spectral extrapolation.
spectral diversityvon neumann entropyentropy walldiffusion modelskl anchor
📰 Industry Media (8)
Scaling AI agents with trustworthy data
A survey of 300 data and technology executives reveals that legacy data systems significantly hinder the scalability and decision-making speed of agentic AI in enterprises. The study identifies 'data leaders'—organizations providing AI agents with access to over 70% of company data—as achieving superior agent performance and trust. Key findings indicate that only 45% of enterprise data is accessible to AI agents on average, dropping to 30% for 'data laggards'. Trust in agent decisions correlates strongly with data readiness, with 100% of data leaders trusting their agents' decisions compared to 50% overall. The report emphasizes improving data access, governance, and automation as critical for scaling agentic AI.
agentic ailegacy data systemsdata readinessdata governanceenterprise data
AllenAI Open Instruct Tulu 3 Post-Training with SFT, DPO, RLVR, GRPO, and Verifier-Based Evaluation
The contribution presents an end-to-end post-training pipeline for instruction-tuned language models using AllenAI’s Open Instruct framework, optimized for a 16 GB runtime. The method involves Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Reinforcement Learning with Verifiable Rewards (RLVR) using GRPO, with LoRA adapters and lightweight Hugging Face/PyTorch implementations replacing distributed components. Results include deterministic verifier-based evaluation of mathematical answers, demonstrated on GSM8K data, with configurable parameters for model, dataset sizes, learning rates, and optimization settings.
supervised fine-tuningdirect preference optimizationreinforcement learninglora adaptersgsm8k
NVIDIA AI Releases Nemotron 3.5 Lightning: A 30B Open MoE with 3B Active Parameters, and NeMo Switchyard Model Router
NVIDIA AI introduces Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B active parameters, optimized for high-volume agentic tasks. The hybrid Mamba-2 + MoE + Attention architecture achieves 1M-token context and 4x faster output than comparable models, leveraging speculative decoding (multi-token prediction) and NVFP4 quantization. On PinchBench, it completes 10,000 tasks 30% faster than Qwen3.6 35B at 86% accuracy. Paired with NeMo Switchyard—an open-source routing library—it reduces costs by 74% in LangChain benchmarks by dynamically assigning tasks between lightweight and frontier models. Released under OpenMDW-1.1 with commercial use rights.
mixture-of-expertsspeculative decodingnvfp4 quantizationmamba-2context window
Xiaomi’s MiLM Plus Releases PROVE: Perception-Aligned Object Removal Metrics RC-S and RC-T With a Real-World Video Benchmark
MiLM Plus introduces PROVE, a framework addressing perceptual misalignment in object removal metrics. PROVE comprises RC-S for spatial coherence and RC-T for temporal consistency, leveraging sliding-window Maximum Mean Discrepancy over DINOv2 features without requiring reference videos. RC-S achieves 0.59 Kendall’s τ and 0.66 Spearman’s ρ against human rankings, outperforming ReMOVE (0.26/0.29) and CFD (0.16/0.18). PROVE-Bench includes 80 paired videos and 100 hard, ground-truth-free videos for evaluation. The framework is deployable as an Apache 2.0 PyTorch repo, with RC-S processing frames at 134.6 ms on an RTX 4090.
maximum mean discrepancydino v2temporal consistencyspatial coherenceobject removal
The Video Production Stack Now Fits on One Desk: LTX-2.5 Launches as NVIDIA-Accelerated Open Weights World Model
LTX-2.5 introduces an open-weights world model optimized for local video generation on NVIDIA GPUs, achieving 6.8-second inference for 10-second clips (7.6× faster than closed alternatives). The architecture features a Gemma 4 language backbone, diffusion fidelity rendering in 8× compressed latent space, and native multishot generation for temporal consistency. Benchmark results show 33M+ downloads, with applications in film pre-visualization (Asteria Studios), real-time avatars (Reactor), and robotics (Markov Robotics). VRAM optimizations enable RTX GPU deployment, reducing cloud dependencies for sub-$10M ARR organizations via Hugging Face and ComfyUI integration.
world modeldiffusion fidelitylatent spacemultishot generationgemma 4
Google tests AMIE for clinical video consultations
Google's AMIE (Video), a multi-agent clinical consultation system, demonstrated performance parity with primary care physicians in simulated video consultations. The architecture divides tasks among three specialized agents: a talker for dialogue, a planner for differential diagnosis, and a perception module for audio-visual analysis. In a randomized study with 15 patient actors and 20 physician evaluators, AMIE matched physicians in history-taking (accuracy), diagnostic reasoning, and management appropriateness, while outperforming text-only AMIE in physical examination guidance (p<0.05). Automated evaluation revealed intermittent perception errors, and real-patient validation remains pending.
multi-agent architecturedifferential diagnosisclinical reasoningaudio-visual perceptionsimulated consultation
Novo Nordisk and AWS bring agentic AI into drug discovery
Novo Nordisk and AWS are collaborating to integrate agentic AI into drug discovery workflows, focusing on target identification, therapy design, and multi-step research pipelines. The partnership leverages AWS's Amazon Bio Discovery (offering 40+ biological AI models) and Amazon Bedrock AgentCore for orchestrating multi-model workflows, combining proprietary and AWS-hosted models. Early results include a 90% reduction in clinical documentation time using Claude 3.5 via Bedrock, with 26K+ monthly prompts processed across 2.5K use cases. The co-innovation hub in London aims to bridge computational and lab research, accelerating candidate synthesis and testing.
agentic aiamazon bedrockcomputational biologymulti-step workflowsdrug target identification
How AI is changing the vulnerability response timeline
AI is accelerating vulnerability detection and patch development by enabling novel code analysis techniques that identify flaws traditional tools may miss, particularly logical contradictions and hard-coded trust assumptions. Google Threat Intelligence Group reported a suspected AI-assisted zero-day exploit in 2026, highlighting the dual-use potential of language models for both offensive and defensive security. Minimus research demonstrates how container composition and dependency visibility impact response times, while Google DeepMind's CodeMender contributed 72 security fixes through combined static analysis, runtime testing, and fuzzing. Despite faster discovery, delays persist in identifying affected systems and deploying patches, emphasizing the importance of accurate software inventories and minimal container images.
zero-daycontainerfuzzingstatic-analysislanguage-model
Generated automatically at 2026-08-12 20:37 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.