Daily Digest — 2026-08-14
283 items · 7 research labs, 269 arxiv papers, 7 industry media
🏛️ Research Labs (7)
The builder’s guide to GPT‑5.6
The GPT-5.6 model family demonstrates improved cost efficiency and agent performance through architectural optimizations including reasoning persistence, native compaction, and programmatic tool calling. Key innovations enable multi-agent orchestration, deterministic JavaScript execution for tool outputs, and context window optimizations, reducing token usage while maintaining accuracy. On Agents' Last Exam, GPT-5.6 Sol at low reasoning outperformed GPT-5.5 at high reasoning (exact metrics unspecified). ARC-AGI-3 performance improved from 13.3% to 38.3% with 6× fewer output tokens. BrowseComp benchmarks show comparable performance (84.04% vs 84.36%) at 96% lower cost ($1.33 vs $33.27).
reasoning persistencenative compactionprogrammatic tool callingmulti-agent orchestrationcontext window
Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed
OpenAI introduces Ultrafast mode for GPT-5.6 Sol, achieving up to 14× speedup over standard processing via Cerebras-powered infrastructure, delivering 750 output tokens/second. This enables real-time applications in incident response, financial analysis, customer support, commerce, and live research by reducing latency without sacrificing model capability. Early testing demonstrates reduced iteration cycles in debugging (e.g., log analysis during outages) and research workflows (overnight experiments → interactive sessions). The limited preview targets latency-sensitive business applications to study speed-impacted use cases.
ultrafast modegpt-5.6 solcerebrasoutput tokens/secondlow-latency inference
OpenAI appoints Dali Rajic as Chief Revenue Officer
OpenAI appointed Dali Rajic as Chief Revenue Officer to scale enterprise AI deployment, leveraging his expertise in revenue operations from Wiz and Zscaler. The company reports over one billion weekly active users and two million business customers, doubling year-over-year. This leadership transition follows Denise Dresser's foundational work establishing OpenAI's commercial infrastructure during a critical growth phase for enterprise AI adoption.
enterprise airevenue operationsgo-to-marketcompute-powered economyai deployment
How RingCentral builds AI-native work from engineering to ops
RingCentral demonstrates enterprise-scale AI-native workflow adoption by empowering employees with ChatGPT Work and Codex across engineering and operations. The company conducted an AI-Native Challenge where thousands of participants, including non-technical staff, developed end-to-end projects using these tools, resulting in functional repositories and accelerated product development cycles. This approach reduced manual coordination in program management by 50% through automated status reporting workflows integrating Jira, Google Sheets, and CRM systems while maintaining human oversight in architectural decisions and testing.
ai-native developmentchatgpt workcodexagentic voice aici/cd
Record, train, and deploy from one place with Strands Agents, LeRobot, and Hugging Face Storage Buckets
The article introduces an integrated workflow for continuous robot learning using Strands Agents, LeRobot, and Hugging Face Storage Buckets. The method enables agents to record demonstrations, store them in Xet-backed Storage Buckets with byte-level deduplication, train policies by streaming datasets directly from the Hub, and deploy updated checkpoints to physical robots. Key innovations include content-defined chunking for efficient storage and a unified backend for simulation and hardware deployment. The approach reduces redundant data transfers by approximately 4x and supports over 90,000 datasets on the Hugging Face Hub.
strands agentslerobotstorage bucketscontent-defined chunkingbyte-level deduplication
What We Learned by Reproducing 2,200 papers from ICML
The Hugging Face ICML 2026 Open Reproductions challenge systematically verified 2,200 papers using AI-assisted workflows, combining automated agents (e.g., Claude Code, Codex) with human oversight. Participants reproduced 51% of papers (1,103) with verified claims, falsified 23% (496), and identified 242 papers with contradictory reproduction outcomes. Key findings include proof errors in 'Towards Optimal Robustness in Learning-Augmented Paging', counterexamples in 'Attention's forward pass and Frank-Wolfe', and metric inflation in 'Do Transformers Need Three Projections?'. Human-AI collaboration proved critical for perceptual evaluations and error correction.
reproducibilityai-assisted reviewadversarial verificationproof falsificationmetric inflation
Bring your spreadsheet data to life with Sheets canvas
Sheets canvas introduces a Gemini-powered visualization layer for Google Sheets, enabling users to create interactive mini-apps from spreadsheet data via natural language prompts. The system operates as a dynamic read-write overlay that maintains real-time synchronization with underlying spreadsheet data, requiring no coding expertise. Users can iteratively refine layouts, designs, and functionalities through conversational interactions with Gemini. The feature supports collaborative workflows through Google Sheets' native sharing capabilities, with applications demonstrated in educational tracking, fantasy sports management, and event planning. Sheets canvas is globally available for Google AI Pro/Ultra subscribers and Google Workspace Business/Enterprise customers.
gemininatural language promptread-write layerreal-time synchronizationinteractive visualization
📜 arXiv Papers (269)
DreamFly: Causal Memory and Receding-Horizon Diffusion Planning for Aerial Vision-Language Navigation
DreamFly enhances aerial vision-language navigation (VLN) by integrating causal memory, receding-horizon diffusion planning, and explicit termination. The method employs a causally aligned historical memory to prevent future information leakage, formulates navigation as receding-horizon diffusion planning with $K$-step action chunks (plan-$K$, execute-one), and decouples termination via LiteStop from action logits. Evaluated on OpenFly, DreamFly achieves 32.04%/29.46% success rate (SR) and 28.22%/23.54% success weighted by path length (SPL) on test-seen/test-unseen splits, outperforming baselines with the lowest navigation error.
vision-language navigationdiffusion planningcausal memoryembodied agentspartial observability
AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses
The paper introduces strong-to-weak scaffolding, a test-time method where a stronger builder model constructs inference-time harnesses to improve a weaker target model's task performance without parameter updates. Evaluated on four Theory-of-Mind benchmarks, the approach uses 5% validation data for iterative harness refinement, boosting target-model accuracy from 0.49 to 0.91. Key mechanisms include offloading unstable reasoning to deterministic code, benchmark-specific routing, and answer-format enforcement, with gains most pronounced for weaker targets.
test-time transferinference-time harnessstrong-to-weak scaffoldingtheory-of-mindparameter-free distillation
Redistribution-based Cost Inference Improves Sparse Safe Offline RL
The Redistribution-based Cost Inference (RCI) framework improves sparse safe offline reinforcement learning by converting trajectory-level stop-feedback into dense per-step costs via return decomposition. RCI preserves the feasible policy set and optimal Lagrangian in a constrained Markov decision process (CMDP), ensuring theoretical losslessness while enhancing cost critic learning. Experiments on highway driving and robotic manipulation tasks demonstrate that RCI achieves substantially lower violation rates compared to sparse and classifier-based baselines, with robustness to heterogeneous dataset compositions and label noise.
redistribution-based cost inferenceconstrained markov decision processreturn decompositioncost critic learningtrajectory-level stop-feedback
Constructing Dynamic Master Logic Models as Knowledge Graphs for Complex System Diagnostics Using Retrieval-Augmented Large Language Models
This study introduces a framework for automated construction of Dynamic Master Logic (DML) models as Knowledge Graphs (KG-DML) using Retrieval-Augmented Generation and Large Language Models. The method extends prior work by enabling KG-DML construction for large, complex systems through targeted retrieval while preserving functional dependencies and logical relationships. The resulting KG-DML supports diagnostic reasoning, safety assessment, and dependency tracing. Validation via precision, recall, logical gate consistency, and structural integrity metrics demonstrates consistent reconstruction in repeated runs, exemplified by application to a Boiling Water Reactor's Low-Pressure Coolant Injection system.
dynamic master logicknowledge graphretrieval-augmented generationlogical gate consistencyfailure propagation
Class Activation Mapping in Explainable Computer Vision: A Method-Centered Review of CNN, Transformer, and Foundation-Model-Era Visual Explanations
The review systematically analyzes 57 class activation mapping (CAM) methods for explainable computer vision, proposing a taxonomy based on attribution mechanism, architectural dependence, and evaluation objectives. It traces the evolution from early CNN-based gradient CAMs to contemporary approaches incorporating transformers, foundation models (CLIP/DINO/SAM), and probabilistic or multi-layer explanations. Key findings highlight a shift toward comparative, token-aware, and high-resolution explanations, while noting persistent fragmentation in evaluation metrics (faithfulness, localization, robustness). The work identifies methodological gaps and tracks how subsequent approaches address them.
class activation mappingvisual explanationsgradient-based attributiontransformer interpretabilityfoundation models
Beyond Trial-and-Error: Agentic Optimization for Image-to-Video Adherence
The paper introduces an Agentic Self-Improvement framework for enhancing Image-to-Video (I2V) generation by reframing it as a closed-loop optimization problem. The method employs a two-stage approach: (1) iterative prompt refinement via a multimodal LLM using Davidsonian Scene Graph (DSG) queries and Common Mistake Questions (CMQ) for semantic and artifact evaluation, followed by (2) Bayesian optimization of stochastic seeds and CFG scales guided by a Video-Text Adherence (VTA) score. Human preference studies show a 69% win rate over baseline methods, demonstrating significant improvements in output predictability and control.
image-to-video generationagentic optimizationdavidsonian scene graphbayesian optimizationvideo-text adherence
VAKRA: Evaluating Multi-Hop Reasoning Across APIs and Retrieval Under Tool-Use Policies
VAKRA introduces a benchmark for evaluating multi-hop reasoning across 8,000+ APIs in 62 domains, with tasks spanning diverse API interactions, compositional reasoning, and policy-constrained multi-source queries. Using a fixed ReAct harness to isolate model capabilities, the study evaluates frontier and open-weight models, revealing performance drops from 70.4% (single-hop) to 50-51% (compositional APIs) and severe degradation (≤2.4%) on policy-constrained queries. Failures primarily stem from entity disambiguation and cross-source grounding rather than tool invocation mechanics. Code and dataset are publicly available.
multi-hop reasoningapi compositionreact harnessentity disambiguationtool-use policies
Structural Silence: When AI Infrastructure Fails Speakers of Underrepresented Languages
The paper identifies structural barriers in AI infrastructure that disadvantage speakers of underrepresented languages, using Bengali as a case study. It examines four interlocking failures: a web presence gap (Bengali accounts for <0.5% of web content despite 4% global population), a 67:1 training-token deficit between English and Bengali, tokenization penalties due to Bengali's alphasyllabary script, and connectivity exclusion (36.5% rural vs. 71.4% urban internet penetration). The authors argue that dataset scarcity is a structural barrier, not merely a technical limitation, and advocate for offline-first design as an equity-oriented infrastructure strategy. They conclude with research directions to reduce these inequalities.
tokenizationtraining-token deficitalphasyllabaryoffline-first designconnectivity exclusion
A Neighborhood Attention Transformer Network for Enhanced 3D Segmentation of the Left Anterior Descending Artery
NA-UNETR, a transformer-based 3D segmentation model, improves Left Anterior Descending (LAD) artery delineation in low-contrast CT via Neighborhood Attention (NA) and Dilated NA (DiNA) blocks for local-global context modeling. Pretrained on 1,000 CTA volumes and fine-tuned with LoRA on 20 free-breathing CT scans, it employs a dynamically balanced Dice-Focal and Hausdorff loss. The model achieved 45.64% Dice and 38.16 mm HD95, outperforming nnU-Net (+3.10 pp Dice) and Swin UNETR (-2.96 mm HD95), with superior boundary accuracy and centerline stability on ImageCAS (79.49% Dice, 8.89 mm HD95).
3d segmentationneighborhood attentionlora adaptationhausdorff losscardiac ct
Convergent Detour Hijacking: Task-Preserving Resource Amplification in Skill-Based LLM Agents
We introduce Convergent Detour Hijacking (CDH), a text-only attack targeting skill-based LLM agents that exploits sequential control points in skill selection and planning. CDH leverages shared semantic cover to establish relevance during skill selection and fabricates plausible dependencies during planning, recruiting unnecessary benign skills while preserving task completion. Evaluated across multiple LLM backends and 491 tasks, CDH achieves an 80.02% coordinator selection rate on DeepSeek-V4-Pro, increasing token consumption by 66.91% and execution time by 92.45% while maintaining comparable task completion rates. This demonstrates that correct outcomes do not ensure trajectory integrity or cost safety.
convergent detour hijackingskill-based llm agentssemantic covertask completionresource amplification
Diagram-MMU: A Multi-Modal Benchmark for Scientific Diagrams
Diagram-MMU introduces a multi-modal benchmark to evaluate Multimodal Large Language Models (MLLMs) on scientific diagram understanding, featuring 3.7k diagrams and 18.3k validated questions across six domains. It assesses MLLMs on diagram-to-code parsing, diagram-to-code editing, and diagram question answering, including agentic settings. Evaluation of 12 MLLMs reveals that diagram-to-code tasks are more challenging than question answering, with models struggling to parse and edit diagrams despite reasoning well. Agentic settings improve parsing and editing but degrade question answering, except for Claude-4.6 Opus, which consistently improves across all tasks.
multimodal large language modelsdiagram-to-code parsingagentic settingsscientific diagramsbenchmark
One Frozen Simulator Is Not Enough: Simulator Collapse in Multi-Agent RL
We identify simulator collapse in multi-agent RL for human-AI interaction, where training against a single LLM simulator leads to poor generalization due to mode collapse. We propose Verbalized Sampling, an inference-time method that broadens simulator behavior by sampling from verbalized response distributions, and Co-Training, a training-time approach that jointly optimizes policies against a population of trainable simulators. Evaluations on Persuasion for Good, τ²-bench, and CooperBench show Verbalized Sampling improves held-out success by up to 9%, while Co-Training achieves 14% gains, with similar improvements in human studies. We release SCOPE, an open-source framework for Population Co-Training multi-agent RL.
simulator collapsemulti-agent rlverbalized samplingco-trainingmode collapse
An Agentic Workflow for Legacy HPC Modernization: Converting the Two-Electron-Integral Core of GAMESS
The authors present an agentic workflow for modernizing legacy Fortran codebases at production scale, demonstrated by converting GAMESS's two-electron-integral routines from Fortran 77 to Fortran 2008. Three specialized Claude Code agents operate under a version-controlled specification, with human oversight and exact verification via GAMESS's test suite. The method successfully transformed 12 files (56,448 lines, 225 subroutines) with zero chemistry-relevant differences across 612 test runs, achieving bit-for-bit reproduction of canonical energies.
agentic workflowlegacy modernizationfortran conversionexact verificationquantum chemistry
VICBench: A Multi-Language Benchmark for Code Vulnerability Detection
VICBench introduces a multi-language benchmark for code vulnerability detection, comprising 100 verified vulnerability-inducing commits (VICs) across Python, Java, and C++ from 88 projects, covering 48 CWE types. The dataset, curated via dual human-expert and agentic annotation, features complex real-world fixes (avg. 38.6 lines) and VICs (avg. 252.5 lines), surpassing prior benchmarks in scale and complexity. Evaluation reveals state-of-the-art algorithms V-SZZ and LLM4SZZ achieve only 33.3%-40.1% F1, highlighting persistent challenges in automated vulnerability detection. VICBench facilitates robust evaluation of detection methods.
vulnerability-inducing commitsmulti-language benchmarkcwe typesagentic annotationf1 score
HAMP-LIC: Hessian-Aware Mixed-Precision Post-Training Quantization for Learned Image Compression
HAMP-LIC proposes a Hessian-aware mixed-precision post-training quantization framework for learned image compression (LIC) models, addressing computational complexity and hardware mismatch. The method employs a four-stage optimization: Hessian trace-based block-wise sensitivity estimation, task-aware refinement considering rate-distortion trade-offs, global bit-width allocation, and calibration-based error suppression. Evaluated on Minnen2018 and Cheng2020 models, HAMP-LIC achieves 4.85x compression with only 0.59% BD-rate loss, outperforming fixed- and mixed-precision PTQ methods while eliminating cross-platform errors.
learned image compressionhessian-aware quantizationmixed-precisionpost-training quantizationrate-distortion
How Organizations Use AI: Evidence from ChatGPT
The study analyzes organizational adoption of ChatGPT Enterprise by linking account records to usage patterns, worker roles, and financial data through March 2026, covering 1,500 organizations and 17M messages. Methodologically, it employs privacy-preserving aggregation to examine adoption trends, firm characteristics, and task classifications. Key findings include rapid growth driven by new adopters and increased usage intensity, concentration among larger, R&D-intensive U.S. public firms, high engagement from early-career workers, and diverse applications in writing, technical tasks, and information synthesis, indicating heterogeneous integration strategies.
chatgpt enterpriseorganizational adoptionusage intensityknowledge workprivacy-preserving analysis
Few-Shot Ordinal Learning for Day-Wise Freshness Estimation with Hyperspectral Fish Images
We introduce the first few-shot learning framework for hyperspectral imaging (HSI)-based food quality estimation, addressing the challenge of day-wise freshness prediction with scarce labeled data. The method treats each fish fillet as a distinct episodic task, employing a CORAL-style ordinal prediction head to model freshness progression through cumulative thresholds, while incorporating biologically grounded monotonicity and embedding smoothness constraints. Evaluated on a 16-day salmon HSI dataset under a strict unseen-fillet protocol, the approach achieves a mean absolute error of 1.58 days and 72.3% 2-day accuracy with only three labeled days per fillet, outperforming scalar regression and label-distribution baselines.
few-shot learninghyperspectral imagingordinal predictionmonotonicity constraintscumulative thresholds
Domain-Aware Lightweight Spectral-Grouped Convolutions for Hyperspectral Fish Freshness Classification
The authors propose SGNet, a lightweight architecture for hyperspectral fish freshness classification that addresses domain-specific challenges in HSI data. SGNet employs spectral-grouped convolutions to separate spectral and spatial feature extraction, complemented by a dual attention mechanism combining channel-wise squeeze-and-excitation with spatial gating. Evaluated on a 16-day refrigerator-stored salmon fillet dataset, SGNet achieves 97.8% classification accuracy and 0.64 days MAE with only 4.75M parameters, demonstrating a 5-18× parameter reduction compared to ResNet-50 and Vision Transformers. Ablation studies confirm the effectiveness of each component, highlighting the benefits of domain-aware design for industrial freshness prediction.
hyperspectral imaginggrouped convolutionsdual attentionsqueeze-and-excitationlightweight architecture
SCOUT: Unlocking Enhanced Spatial Reasoning via Structured Chain-of-Thought and Multi-Objective Process Reward
We introduce SCOUT, a structured Chain-of-Thought (CoT) framework for Vision-Language Models (VLMs) that enhances spatial reasoning through explicit 3D environmental perception and a novel reinforcement learning algorithm with multi-objective process rewards. The method employs a tailored advantage estimation technique for fine-grained credit assignment across reasoning steps. Evaluated on SCOUT-24k, a synthesized spatial reasoning dataset, SCOUT-3B achieves improvements of 16.85% and 6.3% on general and complex spatial benchmarks, respectively. SCOUT-7B surpasses GPT-4o by 4.28% and demonstrates robust generalization to multi-image and video scenarios despite single-image training.
spatial reasoningchain-of-thoughtvision-language modelsreinforcement learningadvantage estimation
Information Abundance Paradox: Long-Context Training Undermines Parametric Knowledge
The study introduces the Information Abundance Paradox, hypothesizing that abundant relevant information in training contexts reduces parametric encoding incentives, increasing reliance on context. Through pretraining and supervised fine-tuning experiments, it demonstrates that increasing context window size improves language modeling, natural language understanding, and closed-book MCQA up to an intermediate optimum, beyond which performance declines. Mechanistic analysis reveals that informative context shifts gradient pressure from feed-forward networks to attention modules, increasing contextual reliance during inference. These findings challenge the assumption that scaling to near-infinite context universally benefits model performance.
parametric knowledgecontextualizationattention modulesfeed-forward networksinformation abundance paradox
Learning-Based Behavior Planning for Automated Driving: Real-World Integration and Deployment
The authors propose a hybrid planning architecture for automated driving that combines deep learning with classical optimization to balance performance and safety. A deep neural network interprets complex traffic scenes and proposes driving behaviors, while an optimization-based supervision layer enforces drivability and safety constraints. The system was evaluated in open-loop studies on real-world urban data and deployed on the research vehicle karl, demonstrating stable closed-loop operation.
learning-based motion planninghybrid planning architecturedeep neural networkoptimization-based supervisionclosed-loop operation
NetlistBench: Evaluating LLM Reliability in SPICE Netlist Recognition and Manipulation
NetlistBench introduces a structure-verified benchmark for evaluating LLM reliability in SPICE netlist recognition and manipulation, addressing a gap in circuit design automation. The benchmark comprises 2,342 cases across 24 task families, including parameter/connectivity edits, hierarchical operations, and equivalence judgments, validated by a deterministic oracle. Testing six non-reasoning LLMs reveals performance variability: simple edits achieve 96%-100% accuracy, while device addition (41%-83%) and equivalence judgment (49%-90%) degrade significantly. Reasoning aids weaker models but fails to mitigate structure-preservation failures, highlighting netlist reliability as a distinct bottleneck.
spice netlistcircuit designllm reliabilitystructure-aware evaluationbenchmark
M-Net: Integrating Spectral Features and Physical Field Operators into Deep Learning for Medical Image Segmentation
M-Net enhances medical image segmentation by integrating mathematical inductive biases into U-Net through three components: (1) continuous spectral features from local pixel matrix condition numbers, (2) physical field operators (divergence and curl-like) applied to image gradients, and (3) a Math-Attention Gate (MAG) for adaptive feature fusion. Evaluated on LiTS, KiTS, and BraTS, M-Net achieves Dice scores of 78.42%, 76.15%, and 83.67%, outperforming U-Net by up to 12.37%. Ablations show the condition-number feature contributes +2.14% over discrete alternatives, while MAG adds +1.45% over naive concatenation.
medical image segmentationspectral featuresphysical field operatorsmath-augmented networkinductive biases
HYDRA: Hyperbolic Dynamic Representation Architecture for Kolmogorov-Arnold Networks
HYDRA introduces a parameter-efficient hyperbolic extension of Kolmogorov-Arnold Networks (KANs) by combining spline-based functional learning with Poincaré ball representations. The method maps inputs to hyperbolic latent space, performs KAN-style updates in tangent space, and shares functional transformations via low-rank prototype blocks, improving interpretability via structured radial coordinates and stability via radius control. Evaluations on eight benchmarks show competitive or superior performance with enhanced parameter efficiency and interpretability.
kolmogorov-arnold networkshyperbolic representationpoincaré ballparameter efficiencyspline-based learning
How to Spend Your Oracle Budget: Practical Guidance for Protein Structure Prediction Models
This work provides the first systematic comparison of oracle budget allocation methods for protein structure prediction, evaluating FK-steering, DPO, Best K-of-N sampling, and the newly adapted Optimisation Over Outputs (O3) approach. O3 applies off-the-shelf optimizers within a generative model's latent subspace and is extended here to protein structure prediction. Experiments on calmodulin (1CLL) and E. coli aspartate transcarbamoylase (9EEH) reveal that O3 outperforms alternatives at low oracle budgets, while FK-steering and DPO excel as budgets increase, with no single method dominating across all scenarios.
protein structure predictionoracle budgetoptimisation over outputsfk-steeringdpo
Machine Learning-Based Cyber Defense for Cloud Infrastructure: An Adaptive Deep Q-Network Architecture for Intelligent Intrusion Detection and Automated Threat Mitigation
The paper proposes a Deep Q-Network (DQN)-based reinforcement learning framework for autonomous intrusion detection and threat mitigation in cloud environments. The method involves preprocessing, feature engineering, and adaptive policy learning, evaluated on the CICIDS2017 (training) and UNSW-NB15 (validation) datasets. Compared to decision tree, SVM, random forest, XGBoost, and MLP baselines, the DQN achieves 99.72% accuracy, 99.66% F1-score, 0.999 ROC-AUC, 15ms detection latency, and 99.54% attack mitigation rate, demonstrating real-time adaptive defense capabilities.
deep q-networkintrusion detectioncloud securityreinforcement learningthreat mitigation
HSTGFormer: Hyper Spatial-Temporal Graph Transformer for 3D Human Pose Estimation
HSTGFormer introduces a Hyper Spatial-Temporal Graph (HSTG) Transformer for monocular 3D human pose estimation, unifying spatial-temporal reasoning via localized joint-time node aggregation. The method decomposes global reasoning into local receptive fields by extending per-frame skeleton graphs into temporal neighborhoods, preserving structural motion information. An Adaptive Dual-Scale Temporal Graph (ADSTG) captures joint-specific dependencies across short- and long-range windows, fused adaptively via a lightweight module. Evaluations on Human3.6M and MPI-INF-3DHP demonstrate strong accuracy and computational efficiency.
3d human pose estimationspatial-temporal graphtransformerlocalized aggregationadaptive fusion
Co-constructing sociotechnical AI governance: participatory system mapping using algorithm registers
This paper contributes a participatory approach to sociotechnical AI governance by analyzing algorithm registers through system-theoretic safety analysis. The authors conduct interviews, surveys, and participatory system mapping workshops (N=8) with municipal staff, civil society organizations, and ombudsmen to examine a Dutch city's algorithm register for a welfare benefits decision-support tool. Using System-Theoretic Process Analysis (STPA), they identify safety hazards such as wrongful eligibility denial and system performance deterioration, which were not evident from the register alone. The study highlights the political dimensions of algorithm governance and the need for pluralistic stakeholder engagement in system safety analysis.
algorithm registerssystem-theoretic process analysissociotechnical systemsparticipatory mappingdecision-support tool
Who Thinks Best Depends on How Long You Let Them: Budget-Dependent Rankings in LLM Evaluation
This work demonstrates that large language model (LLM) rankings on reasoning benchmarks are budget-dependent, challenging the assumption of stable rankings across inference conditions. The authors vary token generation budgets (64--4,096 tokens) across four models on three benchmarks (56,476 inferences), revealing: (i) 3--19% of items show non-monotone accuracy trends, with 6--14% cross-model overlap; (ii) significant ranking reversals (p<0.01); (iii) oracle complementarity gains up to +27.8pp, especially at constrained budgets; (iv) a budget-aware router captures 14.1% of oracle gains, though budget features are domain-specific. Findings advocate for budget-conditioned evaluation protocols.
large language modelsbudget-dependent evaluationnon-monotone accuracymodel complementaritytoken generation budget
A corpus-specific clinical RAG system matches or outperforms newer frontier LLMs on HealthBench
The study demonstrates that VITA, a retrieval-augmented generation (RAG) system tailored for clinical contexts in low- and middle-income countries, matches or outperforms frontier LLMs on the HealthBench benchmark. VITA integrates curated, region-specific medical guidelines and scored 51.9% of possible rubric points on 4,023 questions, surpassing GPT-5.4 (46.1%) and other models. A 500-question subset evaluated with a neutral judge showed statistical parity with GPT-5.5, with VITA maintaining advantages in accuracy and completeness but lagging in communication scores.
retrieval-augmented generationclinical aihealthbenchlow-resource settingsrubric scoring
Adversarial Resilience of Poisson-Process Submodular Maximization over Matroids: From Robust Offline Optimization to Full-Bandit Learning
(No summary returned.)
GUIDE: Governed Unified Intelligence for Document-to-Artifact Generation in Enterprise Settings
The paper introduces GUIDE, a governed multi-agent framework for enterprise document processing, addressing hallucination and structure degradation in existing LLM/VLM systems. The system employs six specialized agents for parsing, VLM-driven extraction, consistency checking, evaluation, human-in-the-loop escalation, and persona-tailored artifact synthesis, built on a versioned rule store with schema-validated contracts. Evaluated on 120 enterprise documents, GUIDE achieves 96% document success, extracts 3,896 rules (71.4% auto-approved), generates 812 deployment-ready artifacts, and reduces processing time from 2-3 days to 40-125 minutes per document.
multi-agent frameworkvlm-driven extractionschema-validated contractshuman-in-the-loopartifact synthesis
Do LLMs Take Care of Their Own? Similarity Signals Can Induce Cooperation
This paper introduces a framework for evaluating LLM decision-making in strategic interactions when provided with graded similarity signals, addressing cooperation problems like the Prisoner's Dilemma. The study finds that modern LLM models exhibit consistent behavior across cooperation problems, payoff structures, and prompt framing, with similarity signal datasets having minimal impact on cooperation. Notably, LLMs systematically self-identify as highly similar when evaluating another model's chain-of-thought reasoning. A behavioral-game-theoretic model is developed, demonstrating cooperative outcomes in equilibrium under high similarity scores.
llm decision-makingsimilarity signalsprisoner's dilemmabehavioral-game-theoretic modelchain-of-thought reasoning
Ready Cohorts: Bounding GPU Opportunity and Avoiding Host Round Trips in LLM-Agent Control
The paper introduces ready-cohort analysis to optimize GPU utilization in LLM-agent control systems, formalizing boundaries via fixed-partition share (F), exact offline share (P*), and local upper bound (U). Using a dynamic program under idealized conditions and a Poisson replay of an 851-session trace, it demonstrates P*=43.00% recovery of lost opportunity at K=256 and 50ms deadlines. A device-resident routing mechanism shows 1.19x-2.39x speedups across GPU placements, with all 14.5M invocations matching host oracle results.
gpu utilizationllm-agent controlready-cohortdynamic programmingdevice-resident routing
QV-PIC: Query-Aware Visual Position-Independent Caching for Efficient RAG Serving
QV-PIC introduces a query-aware dual-resolution Position-Independent Caching (PIC) framework for Retrieval-Augmented Generation (RAG) to mitigate redundant KV-cache computations. It compiles visual caches offline under the model's native chat-template prefix, preserving global context with low-resolution tokens and restoring fine-grained textual evidence via high-resolution query relevance scoring. Evaluated across six tasks, QV-PIC improves average F1 by 21.6 points over vanilla rendered-image PIC, matches text PIC quality, and surpasses optimized text PIC by 2.58 F1 while reducing Time-To-First-Token (TTFT) by 17.2% and achieving 83.8% TTFT reduction versus full prefill.
retrieval-augmented generationposition-independent cachingkv-cachequery-awaretime-to-first-token
No One to Blame: A Framework of Constitutive AI Unaccountability
The paper introduces constitutive AI unaccountability, a framework identifying configurations where AI accountability is inherently unachievable. Through a three-stage qualitative study involving literature analysis, expert interviews, and application to OpenClaw, the authors identify nine categories and 20 themes of unaccountability, organized into structural, technological, and normative clusters. The framework includes a diagnostic instrument of 20 questions, which detected 17 of 20 conditions in OpenClaw, highlighting issues like inverted anthropomorphism. The study reframes AI unaccountability as a constitutive property of sociotechnical systems and extends existing accountability barriers.
constitutive unaccountabilitysociotechnical systemsanthropomorphismaccountability barriersdiagnostic instrument
Confidence Calibration of Deep Learning Systems
This thesis develops methods for improving confidence calibration in deep learning under label noise, domain shifts, and privacy constraints. For noisy labels, it proposes a framework using estimated noise models to reconstruct clean confidence estimates and extends this to noise-aware Conformal Prediction (CP) with guaranteed coverage. For domain adaptation, it estimates target-domain accuracy from source performance and domain discrepancies. A locally differentially private CP framework is introduced for privacy-preserving settings. Results demonstrate reliable uncertainty quantification in safety-critical applications while addressing noise resilience and privacy.
confidence calibrationlabel noiseconformal predictiondomain adaptationdifferential privacy
Graph-Structured Rubrics: Compiling Rubrics into Typed Evaluation Graphs for LLM Judges
The paper introduces Graph-Structured Rubrics (GSR), a method for compiling evaluation rubrics into typed graphs with criterion nodes and composition operators before response observation. GSR uses pointwise and pairwise evaluation modes, with transformation/reduction operators and a Readout mapping for final scores. Evaluations with GPT-OSS-120B show GSR improves exact score agreement by 0.62--6.75 percentage points over Prometheus-style scoring on pointwise tasks and achieves highest pairwise accuracy on preference benchmarks with native tie/abstention policies.
graph-structured rubricstyped evaluation graphpointwise evaluationpairwise evaluationrubric compilation
Faithful, Sufficient and Understandable: Rethinking Graph Counterfactual Explanations via Discrete Diffusion Inversion
The paper introduces Graph Diffusion Counterfactual Explanation via Inversion (GDCE-I), a method for generating faithful and interpretable counterfactual explanations for Graph Neural Networks (GNNs) by leveraging discrete denoising diffusion models with a novel inversion scheme. GDCE-I produces distribution-aware edits that respect categorical node/edge types and domain rules (e.g., chemical valency), addressing limitations of prior work that either violate data manifolds or restrict edit spaces. Evaluated across four benchmarks using a unified framework of explanation desiderata, GDCE-I outperforms existing methods significantly. Qualitative analysis on molecular graphs demonstrates its ability to yield in-distribution, interpretable solutions.
graph neural networkscounterfactual explanationsdiscrete diffusiondata manifoldmolecular graphs
Better Slots, Better Worlds: Representation Quality & Robustness in Object-Centric World Models
The study investigates object-centric world models (OCWMs) for visual model-predictive control, focusing on representation quality and robustness under distribution shift. Using unsupervised slot-quality metrics (FG-ARI, mBO), the authors show that planning success correlates with slot quality, though gains saturate at high quality. Well-bound slots eliminate the need for proprioception inputs and masking biases. Under distribution shifts, OCWMs with high-quality slots outperform scene-centric LeWM but remain comparable to DINO-WM, suggesting pretrained features enhance robustness.
object-centric world modelsslot-quality metricsmodel-predictive controldistribution shiftpretrained features
Learning Loco-Manipulation From SMPC Demonstrations With Sparse Offline-to-Online RL
The paper presents a method for learning loco-manipulation skills by combining Sample-based Model Predictive Control (SMPC) demonstrations with sparse offline-to-online Reinforcement Learning (RL). SMPC generates large-scale offline datasets in simulation, enabling off-policy RL training with only sparse rewards, eliminating manual reward shaping. The approach integrates a high-level RL agent with a low-level stability controller, achieving policies that outperform the original SMPC teacher. Experiments demonstrate successful sim-to-real transfer on diverse morphologies, including an arm-equipped Spot quadruped and a G1 humanoid.
loco-manipulationsample-based model predictive controloffline-to-online rlsparse rewardssim-to-real transfer
Preference Tree Optimization: Enhancing Goal-Oriented Dialogue with Look-Ahead Simulations
Proposes Preference Tree Optimization (PTO), a framework for improving goal-oriented dialogue agents via iterative preference data generation using look-ahead simulations. Combines Preference Tree with Look-Ahead and Direct Preference Optimization (DPO) to enhance decision-making in data-scarce domains like Motivational Interviewing (MI). Experiments show PTO-trained models outperform baselines in session satisfaction and working alliance, with deeper look-ahead configurations yielding optimal stability and performance.
preference tree optimizationdirect preference optimizationmotivational interviewinglook-ahead simulationsgoal-oriented dialogue
Mechanist: AI as a Scientific Instrument for Discovering the Mechanisms of Intelligence
The paper introduces Mechanist, an agentic system for autonomous discovery of mechanisms underlying AI intelligence. It integrates an interpretability-focused knowledge graph (13K papers) with a multidisciplinary database (43M papers across 26 fields) and a library of 32 foundational methods for mechanism analysis. Compared to Claude Code and existing AI-scientist systems, Mechanist generates higher-value mechanism hypotheses and executes experiments more reliably. Key results include uncovering cross-modal safety risk transfer, developing a mechanism theory of belief formation, and creating practical interventions that improve model performance and steer scientific foundation models for DNA sequence generation.
mechanistic interpretabilityagentic systemknowledge graphcausal interventionfoundation models
How Far from Clinical Deployment? Evaluating the Complete Unsupervised Domain Adaptation Pipeline in Medical Imaging
This work evaluates the complete unsupervised domain adaptation (UDA) pipeline for clinical deployment, addressing both model adaptation and label-free selection. The study analyzes 11 clinical cross-domain scenarios from 9 medical imaging datasets, testing 10 UDA algorithms and 13 validators across 80,000 trained models. Results show that while capable adapted models exist, validator-selected models exhibit a significant performance gap compared to the best available, with no validator proving consistently reliable. Ensembling and limited target-labeling partially mitigate but do not eliminate this gap, highlighting the need for improved selection methods to advance clinical UDA deployment.
unsupervised domain adaptationmedical imaginglabel-free selectionmodel validationclinical deployment
LoSA: Near-Lossless Sparse Attention for Training-Free Video Diffusion Acceleration
LoSA introduces a training-free sparse-attention method for accelerating video diffusion transformers while maintaining near-lossless fidelity. The approach leverages two key observations: 40% of block interactions can be removed while retaining 99% of attention mass, and high-mass support remains stable across denoising steps. LoSA fixes a retained-mass threshold of 99%, measures exact block attention masses at an early dense step, and reuses frozen block indices for subsequent steps. Evaluated on Wan2.1-1.3B and HunyuanVideo, LoSA achieves speedups of 1.36× and 3.2× respectively, with minimal quality drops (0.06 and 0.02 points on VBench Overall), outperforming existing sparse baselines.
sparse attentionvideo diffusiontraining-freeattention massdenoising steps
Uncertainty-Aware Probabilistic Constrained Clustering from Entangled Pairwise Supervision
The paper introduces uncertainty-aware probabilistic constrained clustering (UPCC), addressing the limitations of existing deep constrained clustering methods that rely on hard pairwise constraints. The authors propose ProbPair, an angular pairwise objective for probabilistic relations, and develop ECI-PP, an estimator–corrector–integrator framework that refines imperfect supervision through belief estimation, correction, and reliability-aware integration. Experiments across diverse benchmarks demonstrate that ECI-PP outperforms state-of-the-art methods in probabilistic supervision settings and maintains robustness with a shared default configuration.
probabilistic constrained clusteringpairwise supervisionbelief estimationreliability-aware integrationangular pairwise objective
From Safety Documentation to Safety Knowledge Support: An Evidence-Grounded LLM Framework for Medical Devices
The paper proposes an evidence-grounded LLM framework for medical device safety knowledge support, addressing limitations of current LLM approaches in regulated development. The framework integrates device artifacts with controlled knowledge storage, method-specific generation of safety items, uncertainty checks, and recorded expert review, while maintaining traceability and lifecycle updates. It emphasizes source-linked knowledge over isolated text generation, preparing candidate artifacts for expert decision-making without providing regulatory approval. An evaluation strategy using non-public medical-device case studies is outlined to assess coverage, correctness, and review effort.
medical devicessafety engineeringllm frameworkiso 14971traceability
Dual-Model Sentiment Analysis of Consumer Reviews in the Retail Coffee Sector Using Machine Learning and Deep Learning Approaches
The study compares classical machine learning and deep learning approaches for sentiment analysis on 700+ Starbucks customer reviews from ConsumerAffairs, binarized into positive (4-5 stars) and negative (1-3 stars) classes with notable class imbalance. Five ML classifiers (Logistic Regression, SVM, Decision Tree, Random Forest, Naive Bayes) and five DL models (LSTM, RNN, Bidirectional LSTM, GRU, CNN) were evaluated using accuracy, precision, recall, and F1-score. SVM achieved top ML accuracy (91.0%), while Bidirectional LSTM led DL models, with class imbalance notably degrading positive sentiment recall. Findings underscore model selection and preprocessing importance for retail sentiment analysis.
sentiment analysisclass imbalancebidirectional lstmsupport vector machinecustomer reviews
RealisticTritonBench: A Benchmark for Triton-Kernel Generation in Real-World AI Frameworks
RealisticTritonBench introduces a benchmark for evaluating Triton-kernel generation by LLMs, addressing limitations of prior benchmarks that focus narrowly on PyTorch-to-Triton translation, lack end-to-end evaluation, and rely on potentially flawed manual checks. The benchmark derives tasks from real-world pull requests in popular AI frameworks, providing natural language requirements and reproducible evaluation environments that integrate generated kernels into original frameworks for end-to-end testing. Evaluation of leading LLMs reveals persistent challenges in real-world Triton kernel generation despite their demonstrated potential in prior isolated benchmarks.
triton kernelsgpu accelerationllm code generationend-to-end evaluationai frameworks
CTBench: Evaluating Troubleshooting Capabilities of AI Agents in Realistic Telecom Network Operations
CTBench introduces a public benchmark for evaluating AI agents' troubleshooting capabilities in realistic telecom network operations, focusing on root cause analysis and path restoration. The benchmark, constructed by experts with annotated task metadata and golden evidence steps, employs expert-grounded metrics assessing both final answers and diagnostic evidence. Experiments reveal state-of-the-art agents excel in path-restoration endpoint identification but underperform in root cause analysis, particularly for interface state, link-layer, and service-management faults, while often failing to provide evidence-grounded diagnoses despite plausible answers.
telecom troubleshootingroot cause analysispath restorationevidence-grounded metricsnetwork operations
Remote Sensing and Machine Learning-Based Analysis of Land Use and Vegetation Change in Dhaka District, Bangladesh
This study quantifies land use and vegetation changes in Dhaka District (2019-2024) using remote sensing and machine learning. Sentinel-2 MSI and Landsat 8 imagery were processed with Decision Tree, KNN, and Random Forest classifiers in Google Earth Engine, evaluating NDVI, NDBI, and NDWI indices. Random Forest achieved highest accuracy, revealing a 59.5% urban expansion, with vegetation (-8.46%) and water bodies (-7.77%) declining significantly. Results demonstrate rapid land conversion to infrastructure, providing empirical evidence for sustainable urban planning.
remote sensingnormalized difference vegetation indexgoogle earth enginerandom forestland cover classification
Claim-Level Reliability Assessment for Efficient Test-Time Reasoning
The paper introduces Claim-Level Reliability Assessment (CLR), a training-free framework that improves test-time reasoning efficiency by reallocating compute from solution sampling to targeted claim verification. CLR condenses reasoning traces into critical claims and employs semantic falsification, exploiting the asymmetry between solution construction (requiring flawless reasoning) and claim refutation (requiring only one decisive flaw). Evaluated across four LLMs and four benchmarks, CLR outperforms pass@1 by 27.15 percentage-points on GPT-OSS-20B/CMIMC25 and improves self-consistency accuracy from 77.50% to 82.19% while reducing token usage by 37.0%.
claim-level falsificationtest-time scalingsemantic falsificationreliability scoringreasoning traces
HCGRec: Hint-Conditioned Generative Recommendation with Semantic IDs
HCGRec introduces hint-conditioned generative recommendation to address optimization bottlenecks in semantic-ID generative recommenders, where early incorrect token branches yield zero reward signals. The method diagnoses unreachable items via checkpoint rollouts, injects minimal target-prefix hints to redirect generation, and decomposes credit using supervised learning for hinted prefixes and GRPO for sampled suffixes. On sequential recommendation benchmarks, HCGRec reduces zero-advantage samples from >70% to <20% and outperforms supervised fine-tuning and vanilla reward-based post-training.
semantic-idgenerative recommendationreward-based post-trainingcredit decompositionsequential recommendation
Retry, Switch, or Abstain? Learning Strategy-Aware Tool-Use Policies via Controlled Error Injection
The paper introduces BENCH2ROBUST, a framework for studying robust tool-use policies in LLM agents by converting failure-free benchmarks into stochastic environments requiring retry, switch, or abstain strategies. It evaluates two interventions: Bayesian Tool Memory (BTM) for structured runtime recovery context and curriculum-controlled reinforcement learning. Results show a 16.8 percentage point robustness improvement with BTM on Retail tasks, and combined methods achieve 40.8-45.5% performance under failure injection while maintaining failure-free accuracy across 7 models and 2 benchmark families.
tool-use policiesbayesian tool memoryreinforcement learningfailure injectionrobustness gap
LoongReflect: Boosting Long-Horizon Reflection in Search Agents via Global Perspective Distillation
LoongReflect introduces a training framework for enhancing long-horizon reflection in language model agents by addressing the local-global mismatch in reflective decision-making. The method formulates reflection as a memory-control policy, operating over a reversible trajectory tree with explicit reflect and backtrack actions. It combines a fast channel for distilling globally informed behavior from a privileged teacher and a slow channel for optimizing complete trajectories using outcome-based GRPO. Experiments on multi-hop retrieval-augmented generation and mathematical reasoning benchmarks show consistent improvements over outcome-only reinforcement learning and self-distillation baselines.
memory-control policyreversible trajectory treeoutcome-based grpomulti-hop retrievalmathematical reasoning
Causal inference for group-contaminated structured outcomes: observable quotients, lossless reduction and exact randomization inference
The work characterizes causal inference for structured outcomes contaminated by unknown unit-specific transformations (X = Γ . Y(A)), establishing conditions for uniform recoverability of targets constant on group orbits and Borel maximal invariants. It introduces a quotient-faithful reconstruction theorem, showing sufficiency when the conditional law of raw observations given treatment and covariates has a parameter-free form, with Blackwell equivalence as a special case under conditional Haar contamination. For finite-support multichannel lattice images, a maximal invariant under integer translations and quarter turns is constructed, achieving 0.052 rejection rate under the sharp null and 0.992 power at unit effect strength in simulations, with a primary RxRx1 contrast yielding p = 0.0078.
causal inferencestructured outcomesgroup orbitsmaximal invariantquotient reduction
TailBooster: A Dual-Layer Generative Framework for Extreme Value Augmentation with Operational Validity Enforcement
TailBooster introduces a dual-layer generative framework for extreme value augmentation in mixed-type tabular data, addressing under-representation in distributional tails and operational infeasibility. The method combines a statistical layer (interquartile range-based extreme extraction) with a deep learning layer (autoencoder-based operational validity enforcement), using a Tabular Variational Autoencoder for generation. Evaluation on US flight records showed 47-49% MAE reduction for extreme air time prediction and 29-57% for arrival delays versus conventional synthetic data, while maintaining operational validity through data-driven cleaning.
extreme value augmentationtabular variational autoencoderoperational validityinterquartile rangemixed-type tabular data
ExRole: From Team Trajectories to Executable Roles in Multi-Agent Language Models
ExRole introduces a trajectory-to-role framework for multi-agent language models, transforming team trajectories into executable roles that summarize behavior, guide interaction, and identify trainable capacity. The method learns future-aware role prototypes from prefix-local team traces, resolves them into readable instructions and token-aligned role markers, and optionally routes shared LoRA rank slots with turn-aligned credit. Evaluated on MuSiQue and 2WikiMultiHopQA, ExRole improves EM/F1 scores by 15.0/14.4 and 13.5/16.1 points over single-agent search, and by 11.5/11.6 and 7.7/9.7 points against non-ExRole controls. Results demonstrate the superiority of trajectory-induced role conditioning over manual, random, and shuffled alternatives.
multi-agent systemsrole prototypeslora ranktoken-aligned markerstrajectory-to-role
Accuracy and Order Sensitivity Diverge Under Label-Free Strategies
This paper investigates whether label-free strategies can mitigate positional bias in multiple-choice question (MCQ) benchmarks for large language models (LLMs). The authors evaluate two approaches: generation-then-matching and isolated option scoring, both designed to remove sensitivity to option order. Results show that neither strategy reliably improves accuracy over baselines, with the bottleneck identified as withholding options rather than the matching process. Cyclic permutation often enhances accuracy, but complete elimination of positional influence does not consistently yield gains. Metrics for recall imbalance and per-question order sensitivity fail to demonstrate reliable debiasing in two-stage prompting setups.
positional biasmultiple-choice benchmarksgeneration-then-matchingcyclic permutationrecall imbalance
OEIS Open: How many conjectures can language models turn into theorems?
(No summary returned.)
Hamilton-Zero: A Neural Tensor-Network Foundation Model for Ground States of Arbitrary Quadratic Qubit Hamiltonians
Hamilton-Zero introduces a neural tensor-network foundation model (~0.5B parameters) for amortized ground-state computation of arbitrary quadratic qubit Hamiltonians. The method reformulates spin-1/2 ground-state learning as manifold variational optimization over SU(2)^N, using Lie derivatives for Hamiltonian action and custom automatic differentiation. Trained via SU(2) replica-exchange Langevin sampling and sharded natural-gradient optimization (extending KFAC) on systems up to 64 qubits, the model generalizes to 1024 qubits (fine-tuned) and scales to 8100 qubits in evaluation.
quantum ground statesmanifold optimizationlie derivativessu(2) samplingkfac optimization
Do You See What You Draw? A Semantic Closed-Loop Framework for Holistic Evaluation of Unified Multimodal Models
The authors introduce Self-Generative-Understanding (SGU), a novel evaluation framework for unified multimodal models (UMMs) that assesses integrated visual generation and understanding capabilities through a semantic closed-loop pipeline. SGU leverages UMMs' dual abilities by having them generate textual descriptions from images, reconstruct visual contexts from these descriptions, and reason over the self-generated outputs, all without requiring new annotations. Extensive experiments reveal that even high-performing UMMs often struggle to reason over their own generated contexts, highlighting limitations not captured by separate evaluations. This framework provides a zero-cost, holistic benchmark for assessing UMMs as unified systems.
unified multimodal modelssemantic closed-loopself-generative-understandingvisual generationannotation-free evaluation
Policy-as-logic for robust reasoning over rules
The paper introduces Policy-as-logic, a hybrid symbolic-neural method for robust rule-based reasoning in generative AI systems. The approach encodes policies in formal logic, uses language models for fact extraction to ground predicates, and employs an answer set solver for interpretable reasoning. Evaluated on tasks like tax rules and baggage policies, it outperforms policy-as-prompt and policy-as-code baselines with a 10x reduction in token usage while maintaining accuracy and robustness under input perturbations. Results demonstrate the efficacy of combining structured symbolic reasoning with neural fact extraction for auditable decision-making.
hybrid reasoninganswer set programmingfact extractionpolicy complianceinterpretable ai
Benchmark-Based Comparative Assessment of Publicly Benchmarked Indian Foundation Models: A Capability and Evaluation-Maturity Framework
The paper introduces a benchmark-based framework to assess Indian foundation models against global counterparts across eight capability domains, using publicly reported benchmark results. It evaluates models on general-purpose reasoning, coding, agentic AI, cybersecurity, vision, video, scientific research, and Indic language capability, proposing a Benchmark Maturity Index (BMI) with four dimensions: standardization, participation, independent verification, and national coverage. Findings indicate Indian models perform well on saturated benchmarks like MMLU and MATH-500 but lag in newer, agentic evaluations, with Sarvam AI showing the broadest benchmark coverage. The BMI reveals evaluation-ecosystem gaps that may obscure true capability disparities.
foundation modelsbenchmark maturity indexagentic aiindic language capabilityevaluation-ecosystem
DexterSQL: Deep Schema Exploration and Rule-based Correction for Text-to-SQL Generation
DexterSQL introduces a prompting-based Text-to-SQL system addressing three key limitations: coarse-grained schema information, recurring SQL-generation failures, and condition errors in complex questions. The system employs a deep schema explorator to identify ambiguous columns and analyze their relationships, a database-agnostic rule creator to mine and correct recurring LLM failure patterns, and multi-path SQL generation using a dependency-tree-based intermediate representation. DexterSQL achieves state-of-the-art accuracy, improving by at least 2.7% with GPT-OSS-120B (67.6% accuracy) and 0.9% with GPT-4o and GPT-5.2 (71.6% and 72.2% accuracy) on BIRD-Dev.
text-to-sqlschema explorationrule-based correctionmulti-path generationdependency-tree
Agent Skills Can Be Harmful: An Empirical Study of Skill-Induced Failures in LLM Agents
The paper presents a differential analysis framework for attributing task failures and efficiency regressions in LLM agents to specific loaded skills, comparing skill-guided runs against no-skill or semantically matched reference runs. The study, conducted on SkillsBench and SWE-Skills-Bench, identifies 307 skill-induced failures (125 functional, 182 efficiency regressions) using SkillTriage, a taxonomy-guided attribution tool. Key findings include: seemingly relevant skills often cause incorrect task implementation, efficiency regressions are not solely due to prompt length, and excessive verification and heavy implementation pipelines are major sources of inefficiency.
llm agentsskill-induced failuresdifferential analysisefficiency regressionsvalidation checklists
CoQui: A Coordinate-Conditioned Quantum Implicit Generative Adversarial Network for End-to-End Image Generation
The paper introduces CoQui, a coordinate-conditioned quantum implicit generative adversarial network for end-to-end image generation. The method reformulates quantum image generation as implicit function learning, using spatial coordinates and latent variables as inputs to a classical embedding network that generates input-dependent circuit parameters. A variational quantum circuit evaluates each coordinate, with pixel intensities derived from a dedicated color qubit's expectation value, decoupling resolution from qubit requirements. Experiments on benchmark datasets show CoQui outperforms FRQI-based generation and PQWGAN in visual and quantitative quality while using fewer qubits, and surpasses classical baselines.
quantum generative adversarial networkimplicit function learningvariational quantum circuitcoordinate-conditioned generationexpectation value
Two-Stage Deformable-Convolutional Inverse Design of Nanophotonic Absorbers from Optical Spectra
A two-stage deformable-convolutional framework is proposed for inverse design of nanophotonic absorbers, reconstructing metal-insulator-metal resonator geometries from 80-dimensional absorption spectra. The method projects spectra to a $150\times4\times4$ latent space, decodes into $64\times64$ masks, and combines supervised reconstruction with least-squares adversarial refinement. Compared to plain convolution, involution, Dynamic Conv, and ODConv, deformable convolution achieves superior performance: $20.79\pm0.31$ dB PSNR, $0.8501\pm0.0082$ SSIM, Dice $0.9623\pm0.0027$, IoU $0.9342\pm0.0038$, and boundary F-score $0.9550\pm0.0027$. Spectral consistency yields RMSE $0.0805\pm0.0013$ and $R^2=0.7923\pm0.0065$, demonstrating effective spectrum-to-geometry mapping.
deformable convolutioninverse designnanophotonic absorbersspectrum-to-geometryadversarial refinement
LookBack: Where and How to Score LVLM Responses via Visual Reference Usage
The paper introduces LookBack, a training-free method for scoring Large Vision-Language Model (LVLM) responses by augmenting token likelihood with a visual lookback score, which measures token-image reference strength. Existing confidence-based metrics, adapted from LLMs, fail to capture LVLM responses' grounding in visual input, as removing the image minimally affects confidence-based selection. LookBack addresses this by evaluating how strongly each response token refers to image tokens, improving Best-of-$N$ selection across four benchmarks and three models with negligible overhead. The method highlights the insufficiency of text-level metrics for LVLM evaluation.
large vision-language modelsvisual lookback scoretoken likelihoodbest-of-n selectionconfidence-based metrics
User-Assisted Collaborative Distributed Inference for Efficient QoS-Aware Autoscaling
The paper proposes a collaborative distributed inference system that combines dedicated infrastructure with user-contributed resources to achieve efficient QoS-aware autoscaling for AI services. A high-dimensional generative Markov model with structured temporal factorization captures stochastic interactions among users, resources, tasks, and policies, enabling simulation and optimization of task scheduling and resource allocation. Evaluation across varying user populations and scheduling policies demonstrates that distributed scheduling improves request completion (P99 latency) by 15-30% and reduces dedicated resource consumption by up to 40% as user participation scales.
distributed inferenceqos-aware autoscalingmarkov modeltask schedulingresource allocation
How China-Origin Vision-Language Models Move from Refusal to Reframing in State Alignment
The study systematically examines state-aligned distortion in China-origin vision-language models (VLMs) through a balanced benchmark of 200 politically sensitive topics and a visual-abstraction probe. Nine VLMs (seven China-origin, two non-China) were evaluated across four elicitation paradigms and two prompt languages, yielding 21,708 trials audited on six dimensions by LLM judges and human experts. Key findings include: Chinese-language prompts triple state-aligned framing odds, China-origin models reframe 1.6--3.2x more, and censorship shifts from explicit refusal to fluent reframing across Qwen generations, masking information withholding.
vision-language modelsstate-aligned framingmultimodal censorshipin-context learninghuman-ai interaction
Towards Model-based Run-time Cybersecurity: On Control-Flow Anomaly Detection, Attack Identification, and Hardware Monitoring
The paper proposes a model-based cybersecurity framework combining software and hardware monitoring for robust intrusion detection and attack identification. The method integrates control-flow anomaly detection with attack-tree analysis, using hardware-level monitoring to validate suspicious activities detected at the software level, thereby mitigating camouflage attacks. In an authentication-service case study, the dual-monitoring approach corrected a false benign diagnosis to identify a high-confidence code-injection attack, demonstrating improved anomaly detection and diagnostic precision.
control-flow monitoringattack treesintrusion detectionhardware monitoringcode-injection
Toward Meaningful Transparency for AI Chatbots: Disclosing Persuasive Intent Reduces Persuasion
The study evaluates how AI disclosure types affect persuasive impact in chatbot interactions. A preregistered experiment with 1,500 UK adults tested three conditions: no disclosure (control), AI identity disclosure (T1), and AI identity plus persuasive intent disclosure (T2). Results showed a 12.6-point attitude shift (100-point scale) in controls, 13.1 for T1 (statistically equivalent), and 6.3 for T2—halving persuasion while increasing perceived unacceptability and support for penalties. Findings suggest regulating AI persuasion requires intent transparency, not just system identity disclosures.
persuasive aidisclosure effectschatbot interactiontransparency regulationattitude shift
TELLME: Test-Enhanced Learning for Language Model Enrichment
The paper introduces TELLME (Test-Enhanced Learning for Language Model Enrichment), a novel method combining continual pre-training (CPT) with test-enhanced learning (TEL) to improve domain adaptation efficiency. TELLME uses quizzes during training to enhance knowledge acquisition and long-term memory retention, addressing challenges like dataset scarcity and high computational costs. Experiments show TELLME outperforms baselines by up to 23.6% in financial domain tasks and achieves 9.8% better long-term memory retention.
continual pre-trainingtest-enhanced learningdomain adaptationlanguage model enrichmentlong-term memory retention
GRPO for Financial Advice Generation: Outperforming Commercial LLMs under CATE Evaluation
The paper introduces Group Relative Policy Optimization (GRPO) for fine-tuning open-weight language models to generate financial advice, combining LLM-as-a-judge evaluation with a safety gate and a doubly-robust Conditional Average Treatment Effect (CATE) audit. The method outperforms commercial LLMs, achieving a 2.2x higher gross-profit lift ($0.0228 vs. $0.0104) and superior downside risk metrics. The CATE audit reveals divergent rankings from LLM-judge evaluations, indicating complementary signals for assessing financial recommendation quality.
group relative policy optimizationllm-as-a-judgeconditional average treatment effectfinancial nlpoff-policy audit
The Sleeping Agent: What Gist-Based Context Compression Loses and Why
This work analyzes gist-based context compression in long-horizon language model agents, revealing a task-type interaction: Salience-Weighted Consolidation (SWC) improves multi-hop reasoning and factual recall but degrades temporal question performance due to discarded temporal expressions. The study evaluates 1,501 LoCoMo questions across four conditions, showing SWC's structured gist abstraction preserves relational/event structure but loses temporal details (3.05% preservation baseline). A targeted prompt modification increases temporal expression preservation 20× (to 62.39%) with minimal impact on named entities/events, recovering +0.314 accuracy on temporal questions. Results demonstrate compression trade-offs are mechanistically addressable via prompt engineering.
gist-based compressionsalience-weighted consolidationtemporal expression preservationlocomo benchmarkmulti-hop reasoning
HyperANFIS: Enhancing Rule Representation and Interpretability in Adaptive Neuro-Fuzzy Systems via Hyperbolic Geometry
HyperANFIS enhances adaptive neuro-fuzzy inference systems by embedding rule representation and inference in hyperbolic space while preserving interpretable IF-THEN fuzzy rules. The method retains ANFIS's core architecture but performs rule-prototype learning, activation, and aggregation in hyperbolic geometry, improving representational capacity. Experiments demonstrate consistent outperformance over standard ANFIS and variants in predictive accuracy and rule quality across datasets.
hyperbolic geometryneuro-fuzzy inferenceinterpretable rulesrule-prototype learningfuzzy semantics
Instruction Alignment for Binary Code Representation Learning
The paper proposes instruction alignment as an auxiliary training objective to improve binary code representation learning, addressing the limitation of existing function-level embedding methods that ignore fine-grained instruction correspondences. The method leverages compiler debug information to align instructions explicitly during training, motivated by empirical evidence that function-level similarity models exhibit stronger instruction alignment than their pre-trained counterparts. Experiments demonstrate improved retrieval accuracy and more discriminative similarity judgments with the proposed approach.
binary code representationinstruction alignmentfunction-level embeddingcompiler debug informationretrieval accuracy
MOON: Multi-Objective OrthoNormalized Updates for Multitask Learning
(No summary returned.)
JieZi: A Large-Scale Expert-Audited Dataset and Benchmark for Ancient Chinese Character Exegesis
The authors introduce Ancient Chinese Character Exegesis (ACCE), a vision-language question answering task modeling scholarly exegesis, and construct JieZi-Dataset, a large-scale expert-audited dataset with over 500K QA pairs, and JieZi-Bench, an evaluation benchmark. ACCE comprises four progressive levels: basic character identification, glyph-form analysis, meaning exegesis, and diachronic evolution analysis. The dataset is generated via a pipeline constrained by expert-designed templates and source-text references, with human verification at key stages. Experiments reveal that multimodal large language models perform well on basic identification but struggle with higher-level tasks, while fine-tuning on JieZi-Dataset improves performance across all levels.
vision-language question answeringglyph-form analysisdiachronic evolutionmultimodal large language modelsexpert-audited dataset
G0.5: One Autoregressive Stream for Robot Reasoning and Action
G0.5 introduces a pretrained autoregressive Vision-Language-Action (VLA) model that unifies reasoning and action generation within a single transformer decoder, departing from the conventional approach of separate VLM and action modules. Key innovations include a cross-embodiment action tokenizer for shared action vocabulary, interleaved chain-of-thought reasoning, and a visual memory module for temporal context. Evaluated across 7 benchmarks, G0.5 achieves superior performance (76.7% on R1lite/R1pro, 31.4% on BEHAVIOR Challenge) compared to state-of-the-art baselines, demonstrating effective transfer of pretrained VLM capabilities to physical tasks without additional training.
vision-language-actionautoregressivecross-embodimentchain-of-thoughtvisual memory
Advancing MLLM-based UAV Image Understanding and Reasoning: A Benchmark and a Training-Free Multi-Agent System
The authors introduce UAVQA-Bench, a unified benchmark with 1,500 QA pairs from 13 UAV datasets, evaluating 6 capability dimensions and 16 tasks for MLLM-based aerial image understanding. They identify three failure modes (domain-toolset mismatch, error propagation, static reasoning) and propose UAV-MAS, a training-free multi-agent system comprising DSPE, CAIR, and DAAS modules. UAV-MAS with a 32B MLLM achieves 77.0% accuracy on UAVQA-Bench, outperforming Gemini 3 Pro by 4.0%, while its 8B variant improves base model performance by 8.7%.
multimodal large language modeluav aerial imagerybenchmark evaluationmulti-agent systemtraining-free adaptation
Locating and Controlling Implicit Personalization in Large Language Models
This work identifies and controls implicit personalization in large language models (LLMs) by linking behavioral shifts to internal activation signals. Using matched cued and neutral conversations across five LLMs, the study demonstrates that localized internal activations track recommendation changes with correlations up to r=0.87. Internal signals combine when multiple cues appear, but output changes do not add linearly. Removing internal signals associated with specific cues suppresses their influence more effectively than prompting, while maintaining general benchmark performance. However, selectively removing one dimension's influence without affecting co-present dimensions varies by model and attribute.
internal activation signalsimplicit personalizationmatched conversationsdemographic cuesbenchmark performance
A 12-CNOT Double Qubit Excitation Gate
The authors present a 12-CNOT decomposition for the double qubit excitation operator, achieving SOTA improvements in quantum gate efficiency. Their circuit reduces the CNOT count from 13 to 12, while also achieving the lowest CNOT depth (10) and total circuit depth (16) among existing implementations. The design adds only 2 extra single-qubit gates compared to the previous lowest count (11), maintaining gate efficiency. This advancement addresses a critical challenge in practical quantum computing by optimizing high-level gate implementations.
quantum computingcnot decompositiondouble qubit excitationgate efficiencycircuit depth
Fingerprinting Text-to-Image Diffusion Models via Collapsed Generation
The authors propose a non-invasive fingerprinting framework for text-to-image diffusion models based on collapsed generation, where specific input conditions yield consistent outputs across stochastic seeds. Their method identifies collapse-prone conditions as model-specific signatures, enabling ownership verification under white-box (continuous embeddings) and black-box (natural language prompts) access settings. Experiments on UNet- and transformer-based models demonstrate high distinguishability between models, robustness to fine-tuning and obfuscation, and low query budgets for verification.
collapsed generationmodel fingerprintingtext-to-image diffusionownership verificationstochastic sampling
Harness-IF: Evaluating Instruction Following Across Instruction Surfaces in Coding Agents
Harness-IF introduces a novel benchmark for evaluating instruction-following in coding agents, distinguishing compliance from coincidence through Against-Prior Accuracy (AP-Acc). The method evaluates 60 multi-turn coding items from a 642-rule library, scoring 256 rules across five configurable surfaces. AP-Acc measures rule compliance by comparing execution with and without explicit instructions across nine probe builds. Results show accuracy ranges of 72.1-85.9% and AP-Acc of 66.1-78.6% across 12 models, with all models performing worse on against-prior rules by 3.6-7.4 points. Aggregate scores overstate compliance, and prompt depth does not determine precedence in pooled conflicts.
against-prior accuracymulti-turn codinginstruction-followingconfigurable surfacesprobe builds
Proportional Analogies on Probability Distributions via Bayesian Updating
The paper introduces proportional analogies for probability distributions using Bayesian updating, extending the axiomatic framework beyond Boolean and real-valued domains. The method transforms distributions via Bayesian updates induced by observations, with analysis focused on exponential family distributions and Gaussian mixture approximations for arbitrary distributions. Results demonstrate the framework's applicability to standard probability distributions, providing a novel approach to analogical reasoning in probabilistic settings.
proportional analogiesbayesian updatingprobability distributionsexponential familygaussian mixture approximations
When the API Speaks the Wrong Language: Revisiting Post-Training for Multilingual Tool Use
The study identifies Argument Language Mismatch (ALM) as a key failure mode in multilingual API calling by LLMs, where correct tools are selected but arguments are generated in inconsistent languages. It evaluates post-training strategies, finding supervised fine-tuning (SFT) substantially improves language consistency and end-to-end accuracy, often matching or exceeding reinforcement learning (RL) approaches. While RL methods like Group Relative Policy Optimization (GRPO) offer incremental benefits in generalization and multi-objective trade-offs, SFT emerges as a strong baseline for multilingual tool use.
argument language mismatchsupervised fine-tuningreinforcement learningapi callingmultilingual llms
High-dimensional Multi-objective Bayesian Optimization with Learned Variable Interactions
The paper introduces ViaMOBO, a framework for high-dimensional multi-objective Bayesian optimization (MOBO) that addresses exponential sampling complexity through decision variable interaction analysis. The method employs a variable interaction model to partition the decision space into separable subspaces, enabling local Bayesian optimization without strong assumptions about objective separability. Experiments on synthetic and real-world benchmarks show ViaMOBO outperforms state-of-the-art MOBO methods in approximating Pareto fronts for high-dimensional expensive problems.
multi-objective optimizationbayesian optimizationvariable interactionhigh-dimensional spacepareto front
Making Your LLMs More Objective: Stabilizing LLM Safety Behavior Across Traits with Trait-Invariant Safety Tuning
The authors introduce Trait-Invariant Safety Tuning (TIST), a self-distillation framework to stabilize safety behavior in large language models (LLMs) across varying traits assigned in system prompts. They identify trait-induced safety variation, where safety decisions differ based on traits, and propose Trait-Subspace Neutralization (TraSN) to enforce invariance within a low-dimensional trait subspace. Experiments demonstrate that TraSN improves trait-invariant safety, enhances harmful-request refusal, and preserves general model capability. The study highlights traits as a critical factor in LLM safety and introduces refusal-based metrics, Trait-Induced Deviation and Trait-Induced Flip Rate, to quantify this phenomenon.
trait-invariant safety tuningself-distillationtrait-induced variationlow-dimensional subspacerefusal-based metrics
Robust and Efficient Noisy-Label Time-Series Classification via Dynamic Time Warping Based Granular Ball Computing
The paper introduces DTW-based Granular Ball Computing (DTW-GBC), a method for robust and efficient time-series classification under label noise. DTW-GBC groups temporally similar training samples into granular balls using two construction strategies, enabling classification at the granule level rather than individual samples. Evaluated on four benchmark datasets with symmetric label noise, DTW-GBC variants demonstrate improved robustness to label noise while reducing inference-time DTW computations compared to DTW-based 1-NN classifiers.
dynamic time warpinggranular ball computingtime-series classificationlabel noisenearest-neighbor
Consolidator: Learning Persistent Routed Memory Across Context Boundaries
The paper introduces Consolidator, a slot-local operator that transforms routed short-term memory (STM) into long-term memory (LTM) without replaying source tokens, enabling persistent memory across context boundaries. The method clears the KV cache and STM after consolidation while retaining LTM for subsequent access and hierarchical routing. Evaluated on a two-segment modulo-10 mapping task with a 29.95M frozen model, Consolidator (12.35K trainable parameters) improves updated-mapping recall from 44.38% to 87.02% (+42.64pp) and outperforms identity accumulation by 21.40-68.70pp, demonstrating dual utility as retrievable content and routing state.
consolidatorphasor memory networkkv cachehierarchical routerlong-term memory
REOPD: Reliability-Adaptive Reward Extrapolation for On-Policy Distillation
REOPD introduces a reliability-adaptive reward extrapolation framework for on-policy distillation (OPD), addressing limitations of fixed-coefficient methods like ExOPD. The method combines token-level compatibility weights with batch-level adaptive budgets to compute dynamic coefficients λ_{b,t}, avoiding reward hacking while selectively extrapolating reliable teacher-reference directions. Without requiring additional verifiers or rollouts, REOPD outperforms G-OPD in single-teacher mathematics and multi-teacher settings, matching performance in single-teacher code domains, demonstrating cross-domain adaptability.
on-policy distillationreward extrapolationtoken-level supervisionadaptive budgetteacher alignment
The Wording Effect: Quantifying Two-Way Drift in LLM Benchmark Performance
The paper introduces BenchDrift, a method to quantify 'drift' in LLM benchmark performance by generating meaning-preserving variations of problems along linguistic, referential, pragmatic, and structural axes. It evaluates how often correctness flips under rephrasing across eight models and three benchmarks (GSM8K, MMLU, MATH-Hard). Key findings show that phrasing sensitivity persists across model improvements, with stronger models losing more correctness than weaker ones gain, and that fragility stems from rephrasings rather than model-specific behaviors. Rephrasings also disrupt high-confidence answers regardless of length changes.
benchmark driftllm robustnessmeaning-preserving variationsphrasing sensitivitycorrectness flips
HUGIN: Enhancing Vision-Language Planning for Autonomous Logistics Sorting
HUGIN enhances vision-language models (VLMs) for Joint Multi-Scene Understanding (JMSU) in autonomous logistics sorting systems (ALSS) by addressing cross-scene supervision scarcity and attention dispersion. The framework employs Endogenous Data Augmentation to recombine verified atomic facts and Global Context Ranking to align instruction representations with complete visual contexts. Evaluated on SortingBench, HUGIN improves Qwen3-VL-8B accuracy from 63.6% to 78.8% across five VLMs. Deployment tests with over 15,000 packages confirm its practical viability for ALSS.
vision-language modelsjoint multi-scene understandingendogenous data augmentationglobal context rankingautonomous logistics sorting
APEX: Adaptive Expert Prefetching for Memory-Efficient Edge MoE Inference
APEX introduces adaptive expert prefetching for memory-efficient Mixture-of-Experts (MoE) inference on edge systems, addressing the bottleneck of expert loading from off-chip memory. The method employs a lightweight prefetch router to predict candidate experts before the attention block, dynamically fetching additional experts using a learned confidence model. APEX supports correctness-preserving and stall-free execution modes, achieving over 99% overlap accuracy. Results show up to 26% reduction in per-token latency and 41% improvement in energy-delay product (EDP) compared to state-of-the-art baselines, with negligible impact on application accuracy in stall-free mode.
mixture-of-expertsedge inferenceprefetch routerenergy-delay productadaptive prefetching
FrontierFinance: A Challenging Benchmark for Measuring Frontier Intelligence of Finance Agents
We introduce FrontierFinance, a benchmark for evaluating AI agents in professional investment research, addressing limitations of existing benchmarks that focus narrowly on financial data extraction. The benchmark comprises 220 expert-crafted queries and 11,543 source-attributed rubrics across six use cases in the investor workflow. Evaluations under a common harness reveal that tool harness significantly impacts quality and efficiency, with Samaya's in-house system leading at 56.0%, outperforming Claude Fable 5 (49.2%) at lower cost. The best open-weight model, Kimi K3 (46.4%), nearly matches proprietary models at significantly reduced cost. Screening & Discovery and Sector, Industry & Macro remain the most challenging use cases, with top systems achieving only 33% and 39%. The dataset and grading code are publicly available.
benchmarkinvestment researchtool harnessrubricsuse cases
Learning from Multimodal Pseudo-Labels for Robust Open-Vocabulary Instance and Panoptic Segmentation
The paper proposes a multimodal framework for robust open-vocabulary instance segmentation (OVIS) and open-set panoptic segmentation (OSPS), addressing challenges of noisy pseudo-masks and limited visual-textual grounding. The method leverages pre-trained vision-language models (Grounded SAM, LLaVA, CLIP) for automatic pseudo-label generation, CLIP-guided synonym filtering, and GPT-based caption reconstruction, incorporating three training objectives: extended grounding loss, semantic consistency loss, and generative caption reconstruction loss. Experiments on COCO show state-of-the-art performance, with substantial improvements over previous methods in both OVIS and OSPS benchmarks.
open-vocabulary segmentationpseudo-label generationvision-language modelsmultimodal alignmentpanoptic segmentation
AgenticTwin: An Agentic LLM Framework Integrated with Digital Twin for Anomaly Detection
AgenticTwin introduces an agentic framework integrating large language models (LLMs) with digital twin-based anomaly detection pipelines to enhance interpretability in cyber-physical systems. The framework grounds LLM-generated explanations in anomaly classifier outputs and supports natural-language queries from operators. Evaluation employs a benchmark pipeline with synthetic anomalies injected into a real-world weather sensor dataset, testing lightweight, open-source LLMs in practical environments. Results show structured agent collaboration and knowledge-grounded reasoning improve diagnosis quality, contextual retrieval, and mitigation across diverse anomaly scenarios.
digital twinanomaly detectionlarge language modelscyber-physical systemsknowledge-grounded reasoning
XBridge: Entity-Grounded Latent Bridge for Heterogeneous LLM Communication
XBRIDGE introduces a decode-free communication protocol for heterogeneous multi-agent LLM systems, addressing entity grounding issues in cross-architecture communication. The method combines Lexical Anchor Mapping (LAM) to map context tokens across vocabularies and a Latent Enrichment Bridge (LEB) enabling receivers to query sender hidden states for contextual enrichment. Evaluated across three model families (Llama, Qwen, Mistral) and seven benchmarks, XBRIDGE outperforms text-based communication on all tasks, achieving 11x lower latency, and surpasses a KV-sharing baseline in six of seven same-architecture tasks. LEB requires only 264M trainable parameters, minimal training data, and adds negligible inference overhead.
decode-free communicationlexical anchor mappinglatent enrichment bridgeheterogeneous llmentity grounding
GCPO: Diagnosing and Constraining Subspace Geometry in Rollout RL for LLMs
The paper introduces Principal-Subspace Overlap, a diagnostic measure for stepwise variation in subspace geometry during on-policy rollout reinforcement learning (RL) for large language models (LLMs), and proposes GCPO (Geometrically Constrained Policy Optimization) to mitigate training instabilities. GCPO employs hard bilateral orthogonal projections to constrain updates to complementary subspaces, preventing performance degradation and response-length inflation. Evaluated on mathematical reasoning, code generation, and tool-use tasks using Qwen3-8B and GLM4-9B, GCPO outperforms GRPO and recent variants, achieving improvements of up to 27.69 and 2.37 points over base models and baselines, respectively, while stabilizing policy entropy and preserving general capabilities.
subspace geometryrollout rlpolicy optimizationlarge language modelsorthogonal projections
Rubric Dropout: A Simple Way to Mitigate Reward Hacking in Rubric-as-Reward RL
We propose Rubric Dropout, a simple method to mitigate reward hacking in rubric-as-reward reinforcement learning. The approach randomly drops a subset of rubric criteria during training, preventing the policy from optimizing against a fixed rubric. Experiments with Qwen3-8B trained using Group Relative Policy Optimization on medical and science rubrics demonstrate that Rubric Dropout improves out-of-distribution performance (+1 to +7 points on HealthBench-Hard and ResearchQA) while reducing reward hacking measures. A dropout fraction between 30-50% yields optimal results, outperforming alternative approaches like criterion reweighting. The method maintains domain performance while enhancing generalization to stronger gold-standard judges.
rubric dropoutreward hackinggroup relative policy optimizationout-of-distributionrubric-as-reward
Low-Interaction-Rank Learning: Unifying Multiplicative Dual-Encoder Heads
The paper introduces a unified theory for multiplicative dual-encoder networks by defining the class of low-interaction-rank functions, where complexity is measured by an interaction spectrum. The framework decomposes approximation error into spectral truncation and encoder-realization terms, with sample complexity governed by encoder sums rather than products. It identifies a gauge symmetry in encoder definitions, showing normalization acts as gauge fixing and whitening resolves mode ambiguity up to permutation. Experiments on synthetic kernels, operator learning, and CLIP models confirm spectral decay scaling, mode recovery via whitening, and interpretable concept axes post-rotation removal.
multiplicative dual-encoderinteraction spectrumgauge symmetrywhiteningcontrastive learning
Hybrid-Policy Self-Editing for Composable Unstructured Knowledge Editing
The paper introduces Hybrid-Policy Self-Editing (HPSE), a method for composable unstructured knowledge editing (UKE) in large language models (LLMs). HPSE addresses the limitation of existing UKE approaches, which fail to enable atomic fact recall or multi-hop reasoning from edited passages. The method proactively self-distills knowledge from a privileged in-context state of the model, without external supervision. HPSE constructs hybrid rollouts that inject missing facts into the student model's trajectory where coverage is lacking, while maintaining on-policy behavior elsewhere. Theoretical analysis demonstrates HPSE's advantage over pure on-policy distillation, and empirical results show plug-and-play improvements across four LLM backbones and two KE editors in diverse scenarios.
hybrid-policy self-editingunstructured knowledge editingself-distillationmulti-hop reasoningon-policy distillation
Is Per-Agent Policy Composition Safe? Rethinking Successor-Feature Transfer in Cooperative Multi-Agent Reinforcement Learning
The paper identifies a safety gap in per-agent policy composition for multi-agent reinforcement learning (MARL), proving independent successor-feature transfer can yield joint policies worse than all library entries due to invalidated value estimates from teammate recombination. It proposes MA-USFA, a hierarchical method with universal successor feature approximators conditioned on teammates' objectives and a cross-agent correction layer, guaranteeing safe policy transfer without per-task adaptation. Theoretical analysis shows synchronized composition is the only unconditionally safe fixed rule but lacks flexibility for heterogeneous objectives.
multi-agent reinforcement learningsuccessor featurespolicy transfergeneralized policy improvementuniversal approximation
Semantic Lenia: Emergence of Homeostatic Solitons within the Semantic Space of Large Language Models
Semantic Lenia introduces an artificial life framework that reformulates Large Language Model (LLM) inference as a continuous dynamical system in macroscopic logit space. The method employs a non-linear homeostatic feedback loop to balance semantic attraction and syntactic repulsion, enabling the emergence of 'Autonomous Semantic Solitons'—macroscopic dissipative structures that avoid repetitive crystallization. Parameter sweeps identify a critical 'Habitable Ridge' where steering forces balance syntactic inertia, maintaining generative trajectories at the edge of chaos. This approach facilitates profound abductive leaps without structural collapse and establishes a physical scaling law for machine cognition.
semantic leniahomeostatic feedbackautonomous semantic solitonshabitable ridgemacroscopic logit space
Motion-as-Prompt: Enhancing Motion Reasoning in Multimodal Large Language Models via Motion-Guided Cross-Frame Visual Prompting
Motion-as-Prompt (MaP) enhances motion reasoning in multimodal large language models (MLLMs) by introducing a motion-guided cross-frame visual prompting framework. MaP recovers dense point trajectories, selects motion-informative frames, and marks trajectories between sampled frames onto visual inputs, making hidden displacements and interactions observable to frozen MLLMs. Evaluations on CLEVRER and Something-Something-v2 demonstrate consistent improvements in motion-reasoning accuracy, with gains of 4.2% and 8.9% for GPT-5.5, respectively, without degrading non-motion understanding. MaP achieves these enhancements without model training or architectural modifications, offering a simple yet effective solution for motion-centric video reasoning.
motion-as-promptmultimodal large language modelsmotion-guided promptingdense point trajectoriesmotion-reasoning accuracy
Beyond Memory: A Transactional Continuity Kernel for Long-Lived AI Agents
The Continuity Kernel (CK) introduces a transactional protocol for governing persistent AI agent state transitions, addressing issues of stale overwrites, unauthorized exposures, and privilege escalation. CK decouples candidate evaluation from atomic state activation, requiring untrusted components to propose changes against a precise predecessor head or typed absence. Activation transactions validate ownership, pre-state authority, freshness, and effect uniqueness, producing one of four dispositions: Commit, Reject, Quarantine, or Defer. Commit atomically advances the branch head and installs the complete accepted unit. A bounded executable model verified the protocol across 2,808,230 reachable states and 5,526,474 state transitions with zero invariant violations.
continuity kernelatomic activationstate governancetransactional protocolinvariant verification
CLAIM: Leading Open-domain Active Clarification of Large Language Models with Uncertainty Measurement
CLAIM introduces an uncertainty-driven framework for active clarification learning in open-domain interactions with large language models (LLMs), eliminating reliance on manually annotated data. The method quantifies query uncertainty through entropy from answer disagreements across multiple models, constructs synthetic data via entropy-driven pipelines, and trains a unified clarification decision model using supervised fine-tuning (SFT) and group-relative policy optimization (GRPO). Experimental results show CLAIM learns stable, generalizable clarification strategies without manual labels, offering a low-cost solution for proactive understanding in LLM interactions.
uncertainty-drivenentropy estimationsynthetic data generationsupervised fine-tuninggroup-relative policy optimization
Deep Learning Based Relative Transfer Matrix Estimation for Multiple Sources and Multiple Microphones
The paper introduces three deep learning frameworks for estimating the Relative Transfer Matrix (ReTM), a generalization of the relative transfer function for multiple sources and microphones. The proposed methods include time and short-time frequency transform domain convolutional networks and an LSTM-based recurrent neural network, all trained in a supervised manner. Experiments show these models outperform covariance-based ReTM estimation across five objective metrics and achieve comparable speech enhancement performance to baseline methods.
relative transfer matrixdeep learningspeech enhancementconvolutional networkslstm
Making AI-Generated Feedback Matter: From Provision to Student Enactment
The study demonstrates that structured workflows significantly enhance student engagement with AI-generated feedback, compared to passive receipt or optional dialogue. A quasi-experimental sequential cohort design evaluated three feedback conditions (Directed, Self-Directed, Enacted) across 13,037 students and 51,296 student-authored resources. Enacted Feedback, which prompted students to select, evaluate, and dialogue with AI suggestions, achieved 26.2% feedback uptake (vs. 14.1% Directed, 0.1% Self-Directed) and improved self-assessment confidence and work quality. Findings emphasize the necessity of workflow design that positions learners as active participants in feedback literacy processes.
ai-generated feedbackfeedback literacyquasi-experimental designstudent enactmentworkflow design
Learning to Persuade Exposes How Easily LLMs Abandon Correct Beliefs
The study introduces adversarial persuasion as a critical vulnerability in large language models (LLMs), demonstrating that optimized persuasive arguments can collapse model accuracy to near zero even when factually false. Using an adversarial reinforcement learning framework, persuader agents were trained to alter target model responses in single interactions, achieving success rates from 24% to over 93% against training-time models. Transferability tests showed 83%, 79%, and 25% success on Qwen-14B, Llama-3.1-8B, and GPT-4o-mini, respectively, with curriculum-based training increasing GPT-4o-mini success to 38%. Optimized strategies increasingly relied on credibility-based tactics like fabricated citations, highlighting persuasion robustness as a safety necessity.
adversarial persuasionreinforcement learninglarge language modelscredibility-based tacticstransferability
FM-LLM: A frequency-enhanced mixture-of-experts framework for adapting LLMs to time series forecasting
FM-LLM introduces a frequency-enhanced mixture-of-experts framework for adapting frozen large language models (LLMs) to time series forecasting without textual prompts. The method employs a Fourier Analysis Network (FAN) for spectral token alignment and an asymmetric Mixture-of-Experts (MoE) decoder, separating global periodic reconstruction from non-periodic residual modeling. A time-frequency hybrid loss optimizes temporal accuracy and spectral consistency. Evaluated on eleven benchmarks, FM-LLM achieves state-of-the-art performance on 59 out of 78 metrics, with average improvements of 5.3% in MSE and 5.6% in MAE, and maintains robustness in few-shot and zero-shot scenarios.
time series forecastingfourier analysis networkmixture-of-expertsspectral token alignmentautoregressive framework
MBA: Multimodal Benchmark and Agents for Real-World Business Ideation
We introduce MBA-Bench, the first multimodal benchmark for business ideation agents, comprising 30K samples across six domains with distinct visual cues. The benchmark employs GPT-4o to generate reference ideas via retrieval query generation, market evidence retrieval, and evidence-augmented synthesis. We evaluate agents using MLLM-as-a-Judge across six business-oriented criteria, proposing MBA-b and MBA-k for blind and known settings, respectively. Both are trained via LoRA-based supervised fine-tuning and group relative policy optimization with novel reward objectives for creativity and feasibility. MBA-b and MBA-k outperform caption baselines by 63.9% and 77.1%, and multimodal baselines by 25.6% and 35.8%, respectively.
multimodal benchmarklora-based fine-tuninggroup relative policy optimizationevidence-augmented synthesisretrieval query generation
Dion3: Full-Stack Orthogonal Updates
Dion3 introduces full-stack optimizations for the Muon optimizer, addressing its cubic-time Newton-Schulz orthogonalization overhead. The method combines Gram Newton-Schulz for reduced FLOPs, CuteDSL kernels for symmetric acceleration, and megabatching to minimize communication costs. A novel update rule orthogonalizes only a fraction of the momentum matrix rows per step, improving upon Dion's compression. Dion3 matches or surpasses Muon's loss performance while reducing optimizer step time by up to 6x. The implementation is available as a drop-in replacement via the dion package.
orthogonalizationnewton-schulzoptimizermegabatchingmomentum matrix
Foresight Without Seeing: Latent Futures for World Action Models
ForeWAM introduces a dynamics-conditioned direct-policy World Action Model (WAM) that enables predictive context for action generation without decoding future videos. The method employs Future-KV, which performs a single Video DiT prefill over current visual latent and stochastic future slots, reusing layer-wise key-value states during action denoising. Dynamics registers, supervised by a frozen latent action teacher, capture interaction-induced transitions. Without embodied robot data pretraining, ForeWAM achieves 96.7% and 96.9% success rates on LIBERO in standard and accelerated variants, respectively, and 61.6% on LIBERO-Plus.
world action modelsvideo ditfuture-kvdynamics registerslatent action teacher
Learning from Online User Feedback for Shopping Agents
The paper introduces LOFA, a framework for improving large language model-based shopping agents by learning from online user feedback without human annotation. LOFA combines reinforcement learning over purchase outcomes with feedback-aware on-policy distillation, converting user directives into token-level supervision to capture both behavioral patterns and preferences. Experiments on e-commerce logs show LOFA improves recommendation quality, response helpfulness, and user-satisfaction alignment over baselines.
shopping agentsreinforcement learningon-policy distillationuser feedbacke-commerce
CoAdapt-GUI: Joint Workflow Context and Policy Adaptation for Unseen GUI Applications
CoAdapt-GUI introduces a test-time adaptation framework for mobile GUI agents that jointly adapts workflow context and policy from target-app rollouts without demonstrations. The method separates transferable workflow knowledge (procedures, failure modes) from app-specific details, enabling reusable context to guide LoRA-based policy adaptation on a frozen vision-language model. Evaluations on AndroidWorld-Generalization and AndroidWorld Plus show performance improvements from 37.5% to 45.0% and 38.6% to 52.9% respectively, demonstrating the benefits of transfer-constrained workflow context and joint adaptation.
test-time adaptationworkflow contextlora adaptervision-language modelgui agents
EnterpriseRAG: Benchmarking LLM Instruction Adherence and Robustness under Non-Ideal Enterprise Retrieval
The paper introduces EnterpriseRAG, a benchmark of 983 expert-validated samples across six domains to evaluate LLM performance under non-ideal enterprise retrieval conditions. It systematically simulates three failure modes—retrieval noise, knowledge gaps, and factual conflicts—paired with complex instructions. Evaluation of 13 state-of-the-art LLMs reveals a 57-point orchestration gap between per-constraint satisfaction (80%) and holistic compliance (26.8%), highlighting severe instruction adherence collapse under knowledge gaps and factual conflicts. The benchmark provides a reproducible foundation for improving enterprise-scale RAG systems.
enterprise raginstruction adherenceretrieval noiseknowledge gapsfactual conflicts
Localizing Safety Alignment: MLP Layers and Mid-Network Blocks Encode Refusal Behavior in Large Language Models
This work identifies localized encoding of safety-aligned refusal behavior in large language models (LLMs) through systematic weight transplantation experiments. Using two open-weight model pairs and four safety benchmarks, the study replaces attention weights, MLP weights, and contiguous layer regions to analyze refusal transfer. Results show MLP layers dominate refusal behavior (2.7× more effective than attention), with mid-network blocks (layers 8-11) consistently selected in greedy searches, revealing non-additive interactions and benchmark-dependent precision-coverage trade-offs in safety alignment.
safety alignmentmlp layersrefusal behaviorweight transplantationbenchmark-dependent
A Hybrid Framework of Vision Transformer and Gated Recurrent Unit for Detection of Mosquito Diseases
The study proposes a hybrid framework combining Vision Transformer (ViT) and convolutional Gated Recurrent Unit (ConvGRU) for detecting dengue virus-infected mosquitoes from locomotion videos. The three-step method first segments mosquitoes via YOLO 11M, extracts spatial features with ViT, then classifies temporal sequences using ConvGRU. Comparative evaluation shows ConvGRU outperforms RNN, LSTM, and GRU variants, achieving 88.88% accuracy, 84.45% precision, and 82.81% F1-score by jointly modeling spatial and long-term temporal dependencies. The framework demonstrates robustness in complex mosquito behavior analysis.
vision transformerconvolutional grumosquito behavior analysistemporal dependenciesyolo segmentation
RoadWeaver: Large-Scale Lane-Level HD Map Generation from Scratch for Autonomous Driving Simulation
RoadWeaver introduces a coarse-to-fine framework for generating diverse, large-scale lane-level HD maps from scratch, addressing scalability limitations in autonomous driving simulation. The method synthesizes global road layouts, expands them into connected networks, and constructs lane-level geometry with topologically consistent connectivity. Experiments demonstrate 99.8% reachability, 10.7% dead-end ratio, and 0.24 m endpoint alignment error, reducing alignment error by 94.4% compared to state-of-the-art methods. Maps are generated in 1.39–3.50 s and are directly deployable in driving simulators. Training code and implementation will be released upon acceptance.
hd mapsautonomous drivinglane-level geometryroad networksimulation
Reinforcing Step-level Reasoning for Effective Self-Correction in LLMs
The paper introduces Self-Fix Step-DPO (SFS-DPO), a two-stage reinforcement learning framework enhancing step-level self-correction in LLMs. Stage one optimizes step-level reasoning via preference learning, while stage two trains models to self-verify and correct errors. A teacher-assisted variant, SFS-DPO-R, incorporates explanatory rationales for stronger error correction signals. Evaluations across multiple LLMs show SFS-DPO variants outperform prior step-level baselines, with improved self-correction frequency and effectiveness, demonstrating the value of step-level reasoning reinforcement.
self-correctionstep-level reasoningpreference optimizationlarge language modelsreinforcement learning
From Synthesis to Removal: Physics-Grounded Reflection Simulation and Diffusion-Based Video Dereflection
The paper introduces S2R, a closed-loop framework for video reflection removal comprising physics-grounded synthesis (S2R-Synthesis), diffusion-based removal (S2R-Removal), and benchmark evaluation (S2R-Bench). S2R-Synthesis generates paired reflected/clean videos via structure-space augmentation modeling glass effects (roughness blur, ghosting, reflectance variation) and a video diffusion renderer. S2R-Removal adapts a pretrained video diffusion prior through reflection-aware latent adaptation and one-step pixel-geometric refinement. Evaluations on S2R-Bench and public benchmarks show state-of-the-art performance with faster inference than non-diffusion baselines, validating the synthesis pipeline's effectiveness.
video dereflectiondiffusion priorphysics-grounded synthesislatent adaptationbenchmark evaluation
Beyond Single-Turn Confidence: Trajectory-Adapted Uncertainty Quantification for LLM Agents
The study evaluates the transferability of single-turn uncertainty quantification (UQ) methods to multi-turn LLM agent trajectories, focusing on three families: token-probability scorers, black-box consistency scorers, and reflexive scorers. Experiments conducted across five LLMs and four multi-turn tool-use datasets from BFCL-v4 and τ²-bench reveal that transfer is often useful but uneven. Token-probability scores depend heavily on aggregator choice, reflexive scores serve as strong low-cost baselines, and black-box self-consistency, particularly trajectory-equivalence and action-set consistency variants, often outperforms others. Results highlight the need for trajectory-level revalidation of UQ methods, emphasizing consistency measurement, aggregator selection, and computational budget.
uncertainty quantificationllm agentstoken-probability scorersblack-box consistencytrajectory-equivalence
A Conceptual Framework for Enhancing Workforce Readiness for Smart Manufacturing in the AI Era
The paper introduces a Workforce Readiness Level (WRL) framework to address the skills gap in smart manufacturing by mapping competencies to nine progressive stages across four pillars: digital and AI literacy, cyber-physical systems fluency, human-machine collaboration, and data-driven decision making. The framework, validated through 89 capstone projects at a university lab, employs a composite score and a 'no-thin-pillar' rule to diagnose readiness, with cohort indices ranging from 5.2 to 6.4. Results indicate cyber-physical and data-driven gaps often masked by strong analytics, with highest stages requiring industry experience. WRL provides a standardized tool for educators and workforce planners.
workforce readinesssmart manufacturingcompetency frameworkcyber-physical systemsdata-driven decision making
Generative Semantic Segmentation via an Observable Semantic-Image Interface and Hierarchical Generator Evidence Alignment
Semantic Prism introduces a generative semantic segmentation framework combining a diffusion-distilled one-step RGB generator with hierarchical feature alignment. The method establishes an explicit probabilistic interface via per-pixel distances to a class-color codebook, then refines predictions through zero-initialized output projections in logit space while retaining the image-defined interface. On Cityscapes validation, it achieves 72.07% mIoU (11.39 points above baseline) with 0.41% calibration error; cross-dataset evaluations show 62.22% mIoU on BDD100K and 46.89% on ACDC without adaptation. Contextual Interface--Hierarchy Disagreement (C-IHD) improves pixel-error ranking, raising AUPR from 0.6580 to 0.7557 on ACDC.
generative semantic segmentationdiffusion distillationhierarchical alignmentprobabilistic interfacezero-initialized projection
Hierarchical Federated Transfer Learning in Digital Twin-Based Vehicular Networks
The paper proposes Hierarchical Federated Transfer Learning (HFTL) to address data heterogeneity and sparsity in Digital Twin-based Vehicular Ad hoc Networks (DT-VANET). The method integrates Federated Transfer Learning (FTL) for vehicle type clustering and introduces a framework with two algorithms: one for cloud server model updates and another for intra-cluster federated transfer learning. A data quality score mechanism is implemented to mitigate malicious vehicle impacts. Experiments on real-world datasets demonstrate improved global model accuracy and efficiency across various performance metrics.
hierarchical federated transfer learningdigital twinvehicular ad hoc networkdata heterogeneitydata quality score
Keep the Future, Drop the Rollout: RIFT for World Action Models
The paper introduces RIFT (Rollout-free Imagination via Future Tokens), a method for world action models (WAMs) that eliminates iterative video rollout during deployment while preserving future-conditioned action generation. RIFT uses learned anticipation tokens to construct a complete future key/value (K/V) cache in a single backbone pass, decoupling cache consumption from production. Evaluated on 40 LIBERO tasks, RIFT achieves 98.8% success, comparable to rollout-based methods (98.4-98.6%), while reducing action-chunk latency by 68.2-89.1%. On RoboTwin 2.0, it attains 92.9/92.6% success in clean/randomized scenes, outperforming baseline WAMs. Results demonstrate rollout-free future conditioning is feasible without iterative video generation.
world action modelskey/value cacherollout-freefuture tokenslatency reduction
Do Influence Tactics Matter? Investigating Prompt Framing Effects in LLM Code Generation
This study introduces the first large-scale empirical investigation of psychology-based influence tactics in LLM prompt framing for code generation tasks. Eight influence tactics from Yukl & Falbe's taxonomy were operationalized into prompt templates and evaluated across five open-weight LLMs using LiveCodeBench and SWE-bench Verified. The generated code was assessed on functional correctness, quality, maintainability, and security. Results indicate that certain framings, particularly those emphasizing urgency, negatively impact correctness and security. The findings provide insights into linguistic cue effects on LLM outputs and practical implications for transparent human-AI interaction design in software engineering.
influence tacticsprompt framingfunctional correctnessopen-weight llmssoftware quality
Let it Cook: Learning to Wait in Sequential Decision Making
The paper introduces a 'waiting policy' for sequential decision making, enabling agents to minimize sensing and decision-making frequency without compromising task performance. The approach employs reinforcement learning with lexicographically ordered objectives to determine optimal waiting durations and locations. Experiments conducted across four discrete-state household tasks and three continuous-state environments demonstrate that the method effectively learns waiting behaviors and adapts pre-trained policies. Results show significant waiting periods, sometimes exceeding 50% of task duration, while maintaining task efficiency.
sequential decision makingwaiting policyreinforcement learninglexicographically ordered objectivestask performance
Conflict and Congruency Effects in Large Language Models: In-Weight and In-Context Competition in a Verbal Conflict Task
This study introduces a verbal-only conflict task for large language models (LLMs) to investigate congruency effects, analogous to Stroop and flanker tasks in psychology. Using Gemma-2-2B and six Pythia models (410M to 12B parameters), the authors employed causal attribution analysis, attention analysis, and attention ablations to identify distinct processing pathways: short-range attention to color cues in congruent conditions and long-range attention to rule prefixes in incongruent conditions. Results showed strong default same-color tendencies and congruency effects in six of seven models. Fine-tuning strengthened default tendencies, reducing incongruent performance while increasing congruent performance, whereas larger rule sets selectively impaired incongruent performance. These findings highlight competition between in-weight default mappings and in-context rule-based mappings in LLMs.
congruency effectsattention analysisfine-tuningin-weight mappingin-context mapping
Strengthening Full Justified Representation: Efficient Verification and Computation
We introduce FJR+, a polynomial-time verifiable and satisfiable strengthening of Full Justified Representation (FJR) and EJR+ for approval-based committee elections. The Residual-Budget Greedy (RBG) algorithm is analyzed, proving it selects partial committees whose size-$k$ completions satisfy FJR+. Sequential Phragmén is used to obtain priceable completions, ensuring FJR+, sub-core satisfaction, and priceability when at least $k$ candidates are approved. A Droop-quota variant of FJR+ is derived, and FJR+ is extended to approval-based participatory budgeting with arbitrary project costs. A project-specific RBG computes this efficiently, enabling priceable outcomes satisfying cost-based sub-core.
full justified representationresidual-budget greedysequential phragménsub-coredroop-quota
HyperFix: Combinatorial Nonlinear Correction for Task Vector Merging
HyperFix introduces a combinatorial nonlinear correction approach for merging task vectors without joint retraining, addressing limitations of scalar tuning methods that require repeated optimization per subset. The method employs a lightweight hypernetwork to predict subset-conditioned nonlinear weight-space corrections, trained on singleton, pair, and triple task subsets for generalization to larger sets. Experiments demonstrate HyperFix outperforms existing merging methods across diverse benchmarks while reducing tuning costs, supported by a local perturbation analysis bounding residual corrections beyond linear merging.
task vector merginghypernetworknonlinear correctioncombinatorial optimizationweight-space perturbation
From Prompting to Behavioral Alignment: Personalized LLM Judges for Recommendation Evaluation
The study introduces a sequential behavioral alignment framework to address bidirectional rationalization in Large Language Models (LLMs) for recommendation evaluation. The framework combines fine-tuning with preference optimization over paired correct and counterfactual rationales, enabling LLMs to predict user engagement reliably from raw text logs. Evaluated on real-world homepage interaction logs, this approach achieves a 32.19% improvement in Macro-F1 score over zero-shot baselines and matches the performance of traditional feature-engineered pipelines. The results demonstrate that behavioral alignment mitigates bidirectional rationalization while providing interpretable reasoning traces without manual pipeline maintenance.
behavioral alignmentbidirectional rationalizationlarge language modelspreference optimizationmacro-f1 score
A Modular Agentic Framework for Synthetically Constrained Multi-Objective Hit-to-Lead Optimization
The paper introduces SABLE (Synthetically-accessible Agentic Bayesian Ligand Exploration), an open-source modular framework for multi-objective hit-to-lead optimization in drug discovery. SABLE combines natural-language orchestration via an LLM with specialized tools for reaction-templated analog enumeration, property prediction, affinity scoring, and Bayesian optimization. The framework enriches candidate sets for user-defined objectives while evaluating only a subset of the search space, providing provenance for each output. Its modular architecture allows tool replacement via configuration without operational logic modification.
hit-to-lead optimizationbayesian optimizationllm orchestrationsynthetic accessibilitymulti-objective optimization
The Next Challenge for Agentic Cybersecurity: A Realistic, Contamination-Free Reverse Engineering Benchmark
The authors introduce SRE-Bench, the first contamination-free reverse engineering (RE) benchmark for evaluating AI agents on binary analysis tasks. The benchmark comprises 19 real-world-scale programs (avg. 16.9K LOC) with 44 anti-analysis primitives, yielding 262 binary instances and 1572 deterministically graded tasks. Evaluation across five frontier LLMs (GPT-5.6-sol, Claude-Opus-5, GPT-5.5, Grok-4.5, GLM-5.2) shows RE remains unsolved: GPT-5.6-sol achieves 61.4% per-instance accuracy and fully solves only 31.5% of instances, with agents exhibiting distinct behavioral differences from human engineers (e.g., insensitivity to compiler optimization). Ablations confirm contamination control and realistic scale are essential for meaningful evaluation.
reverse engineeringcontamination-freeanti-analysis primitivesbinary analysisagentic cybersecurity
PAC-Bayes Beyond Parameter Space: Behavioral Equivalence, Z-Information, and Exact Complexity Decomposition
The paper introduces an exact structural decomposition of PAC-Bayes complexity by distinguishing uncertainty over predictive behavior from variation among behaviorally equivalent realizations. Using a measurable behavior map and measure disintegration, it decomposes probability measures on the configuration space into distributions over predictive behaviors and conditional distributions over behavioral fibers. This yields an exact decomposition of the classical PAC-Bayes KL divergence into behavior-selection and realization-level terms, with Z-information defined as the negative of the latter. The behavior-selection term admits a variational characterization as the minimum KL divergence among posteriors inducing the same predictive behavior distribution. Symmetry, fiber geometry, and invariance under fiber-preserving perturbations emerge naturally from this framework.
pac-bayes complexitybehavioral equivalencez-informationmeasure disintegrationkl divergence
TangPoetryBench: A Multi-Dimensional Benchmark and Rubric-Conditioned Evaluator for Poetry-to-Image Generation
The paper introduces TangPoetryBench, a multi-dimensional benchmark for evaluating poetry-to-image generation, comprising 1,280 images (320 Tang poems × 4 T2I models) with human annotations across ten dimensions. It reveals model-specific strengths and weaknesses, including implicit emotion evocation. The authors also propose PoemAutoEvaluator (PAE), a rubric-conditioned evaluator achieving parity with Claude and generalizing to unseen generators and Song Ci poetry. The benchmark and evaluator enable scalable quality assessment without additional human annotation.
text-to-image generationbenchmark evaluationimplicit emotionrubric-conditioned evaluatorcultural content
Herding End-to-End Autonomous Driving via Neuro-Symbolic Safety Guards
The paper introduces a neuro-symbolic safety guard for end-to-end autonomous driving agents, addressing their tendency to violate basic traffic rules despite high average performance. The guard, a lightweight module attached to the final command interface, enforces explicit safety rules by replacing unsafe commands with the nearest safe alternative, requiring no retraining. Evaluated on Fail2Drive and Bench2Drive benchmarks using TransFuser v6, it improves Success Rate by 15% and reduces safety-critical collisions by up to 53% while maintaining the original Driving Score.
neuro-symbolicend-to-end drivingsafety constraintstransfuser v6bench2drive
Benchmarking LLM Judges for Mobile Agent Evaluation
The paper introduces MobileJudgeBench, a benchmark for evaluating LLM-based judges on mobile agent trajectories, comprising 931 human-annotated trajectories across 6 benchmarks, 4 agent models, and 68 apps. It evaluates 6 judge methods, including a simple baseline, across multiple LLM backends. Key findings show the baseline outperforms purpose-built methods, benchmark metrics predict real-world utility, and failure analysis reveals opposite profiles linked to backbone precision-recall characteristics.
llm-based judgesmobile agent trajectoriesbenchmark quality metricsprecision-recall characteristicson-policy reinforcement learning
Social Chain of Thought: A Multi-Agent Architecture Grounded in Medical Differential Diagnosis Methodology
We introduce Social Chain of Thought (SCoT), a multi-agent architecture for medical differential diagnosis that structures collaborative LLM reasoning through multi-round specialist interactions. SCoT addresses the limitations of monolithic inference by integrating multiple forms of specialist reasoning, particularly in complex cases. Evaluated against single-agent baselines, pipeline ablations, and best-of-n scaling, SCoT demonstrates superior recall in the hardest diagnostic scenarios, where iterative specialist deliberation improves ground-truth recovery and differential convergence.
social chain of thoughtdifferential diagnosismulti-agent architecturemonolithic inferencerecall advantage
TRACES: A Benchmark for Epistemic Reliability in Scientific Reasoning by LLMs
TRACES introduces a benchmark for evaluating epistemic reliability in scientific reasoning by LLMs, addressing the critical failure mode of distinguishing reliable from unreliable scientific literature. The benchmark comprises 42 retracted, fraudulent, and pseudoscientific papers, paired with scientifically plausible study-design requests, spanning five claim types. Two scores, IFR-a and IFR-i, measure premise rejection and recognition of unreliability, while the Engagement Depth Index (EDI) quantifies reproduction of withheld details. Across 30 models and 10 runs, aggregate IFR-a is 0.93 ± 0.004 and IFR-i is 0.809 ± 0.009. Models engaged with untenable premises in 95% of non-empty responses, with 22 of 30 models failing over 90% of agentic probes, indicating topic-keyed safety behavior rather than robust epistemic competence.
epistemic reliabilityscientific reasoningretracted papersstudy-design requestsengagement depth index
When Self-Consistency Backfires: Majority Vote Hurts the Majority of Hard Science Problems for Small LLMs
The study demonstrates that self-consistency (SC) via majority vote reduces accuracy on graduate-level science problems (GPQA Diamond benchmark) for small instruction-tuned LLMs. Evaluating Qwen2.5-7B and Llama-3-8B, majority voting harms 56.6% and 65.7% of problems respectively, with correctness uncorrelated to confidence. A grid oracle shows a theoretical upper bound of +14-17 accuracy points over single-chain inference, but no verifier-free gating method achieves this. Pre-registered hypotheses confirm these findings, though reasoning-native models remain untested.
self-consistencymajority voteinstruction-tunedoracle boundverifier-free
AI Guardrail Survival under Single-Cycle Agentic Self-Summarization
The study investigates how safety constraints degrade during single-cycle agentic self-summarization, revealing that textual presence checks inadequately assess functional rule survival. Using behavioral replay across models, the authors find degraded rule residues increase prohibited actions by +34 to +57 points versus intact constraints, with rule-form items retained more than facts despite functional failure. Detection requires external ground truth comparison, as runtime degradation is silent and LLM-judge evaluations risk false conclusions. Results highlight pitfalls in presence-based audits and constraint welding under compaction budgets.
agentic self-summarizationgovernance decaybehavioral replayconstraint weldingllm-judge
From Numbers to Judgment: Specialist LLM Agents and Reinforcement Learning for European Listed Real Estate
The study investigates whether specialized decomposition improves LLM performance in financial analysis, contrasting monolithic versus specialist-aligned prompting for numerical and judgment tasks in European listed real estate. Larix employs eight lens-specific specialists with a frozen LLM (Qwen3.5-9B), finding decomposition boosts numerical-task accuracy by 15.8 percentage points but inconsistently affects judgment tasks. GRPO fine-tuning with task-aligned rewards further raises development-split scores (+12.0 points) and judgment performance (+14.2 points), demonstrating transfer to unseen firms (+15.2) and regulatory wrappers (+4.3). Results show prompt-level decomposition aids modular execution, while parameter adaptation enhances integrative judgment.
llm specializationreinforcement learningfinancial analysisqwen3.5-9bgrpo
Gaze Target Estimation Anywhere with Concepts
The authors introduce Promptable Gaze Target Estimation (PGE), a novel end-to-end paradigm for gaze analysis that conditions predictions on user-specified text or visual prompts (e.g., 'the boy in red') to bypass multi-stage pipelines. Their GazeAnywhere model employs a transformer-based detector to fuse frozen encoder features, jointly solving subject localization, presence detection, and gaze heatmap estimation. Trained on the 120K-image Gaze-Co dataset, GazeAnywhere achieves SOTA on PGE benchmarks, including challenging clinical out-of-domain data, and is open-sourced.
gaze target estimationpromptable visiontransformer detectorheatmap regressionin-the-wild
Inverse Theory of Mind Modeling for Content Recommendation: From Web Browsing to Dynamic Intelligent Interfaces
The paper proposes an Inverse Theory of Mind (IToM) pipeline for content recommendation, which infers user beliefs and preferences from observed interactions in dynamic interfaces. The method reconstructs decision contexts, applies LLM-driven counterfactual reasoning to generate belief statements, and synthesizes these into structured personas via multi-hypothesis abductive inference. Evaluated on the OPeRA dataset, IToM matches or exceeds ground-truth personas in next action prediction, attitude alignment, Big Five personality inference, and held-out category prediction, with multi-hypothesis reasoning proving critical. Cross-modal transferability is demonstrated in a VisionOS spatial banking application.
inverse theory of mindcounterfactual reasoningmulti-hypothesis abductive inferencestructured user personagenerative ui
Dynamics Models for Offline Hyperparameter Selection in Real-World RL
The study introduces calibration models for offline hyperparameter selection in real-world reinforcement learning (RL), addressing the challenge of costly online experimentation in industrial settings. Focusing on a municipal water treatment plant, the authors evaluate k-nearest neighbors with Laplacian distance and other approaches on high-dimensional, non-stationary sensor data for nexting prediction. Results demonstrate realistic long-horizon rollouts, meaningful hyperparameter sensitivity trends, and scalability to year-long datasets, while also examining robustness under distribution shift and fine-tuning learning rates for pre-trained agents. This work provides a proof of concept for deploying RL in real-world environments using offline dynamics models.
reinforcement learninghyperparameter selectioncalibration modelsnexting predictiondistribution shift
Governing Agentic AI in FinTech
The paper introduces a governance framework for agentic AI in FinTech, emphasizing verifiability as the key constraint. Through three empirical studies across nine model versions (ranging from 3B parameters to commercial frontier systems), the authors demonstrate that reproducibility varies by model type and configuration. Study 1 reveals that provider-controlled parameters affect reproducibility, with local models achieving 320/320 executions and hosted models 319/320 and 959/960. Study 2 highlights orchestration as a policy layer, showing architecture impacts final actions. Study 3 finds deterministic credit models reproduce current actions perfectly but fail to recover historical ones. The framework extends to high-stakes domains requiring auditability.
agentic aiverifiability gapreproducibilityorchestrationauditability
Can Frontier LLMs Match Natively Multimodal Embeddings? A Comparison on Hard-Negative Text-to-Image Retrieval
This study compares native multimodal embeddings (Gemini Embedding 2) with frontier LLMs (GPT-4.1, Claude Sonnet 4.6) on zero-shot text-to-image retrieval using Flickr30k. The evaluation focuses on hard-negative ranking performance, contrasting dual-encoder contrastive learning (Gemini Embedding 2) with LLM-based visual understanding. Results show parity between LLMs and native embeddings, though precomputed multimodal embeddings retain advantages for low-latency applications.
multimodal retrievalcontrastive learningzero-shot rankinghard-negative retrievaldual-encoder
Apodex Discovery: Reality Benchmarks and Environments for Evaluating and Building Discoverative Artificial Intelligence
Apodex Discovery introduces a framework for evaluating and building discoverative AI through a heavy-duty solver system comprising a foundation model, harness, tools, and control policies. The framework includes a problem-scouting process identifying 423 high-value real-world problems, a common environment-task-episode abstraction for data and verification, and HDS6 evaluation metrics. Apodex surpassed state-of-the-art performance by 7% in AAV capsid design and improved GPT-5.5 and GPT-5.6-sol prediction scores by 2.5 and 7.6 points in drug repurposing tasks. The TRACES episode interface enables precise attribution of performance differences to solver components, advancing AI evaluation toward verifiable investigations.
discoverative aiheavy-duty solverenvironment-task-episode abstractionhds6 evaluationtraces episode interface
Self-evolving network verifiers
The paper proposes self-evolving network verifiers that automatically refine symbolic models of network behavior through counterexample-guided learning. The system employs a coding agent to iteratively extend the verifier's SMT-based encoding, validated against a ground-truth oracle (e.g., emulated routers), resolving discrepancies between model predictions and actual router behavior. A prototype demonstrated autonomous learning of three unsupported features (OSPF areas, BGP route reflection, L3VPN over EVPN) in a 3,000-line verifier, including vendor-specific behaviors. The approach shifts verification challenges from manual model maintenance to systematic testing of evolved models.
symbolic verificationcounterexample-guided learningsmt-based verifiernetwork control planeautonomous model refinement
Dual-Domain Cross-Modal Decoding for Clinical Text-Guided Medical Image Segmentation
Dual-Domain Cross-Modal Decoding (DD-CMD) introduces a novel approach for clinical text-guided medical image segmentation by integrating spatial and frequency domain guidance. The method employs Text-Guided Spatial Cross-Attention (TGSA) for semantic alignment and Spectral-Text Adaptive Modulation (STAM) for frequency-aware decoding, both embedded in a coarse-to-fine decoder. A lightweight two-stage refinement module restores full-resolution masks. Evaluated on QaTa-COV19 and MosMedData+, DD-CMD achieves 91.46% Dice / 84.26% mIoU and 81.95% Dice / 69.42% mIoU, respectively, outperforming prior baselines by +1.96 Dice and +2.67 mIoU on average.
dual-domain cross-modal decodingtext-guided spatial cross-attentionspectral-text adaptive modulationfrequency-aware decodingcoarse-to-fine decoder
Contextual Quality-Diversity Evolutionary Reinforcement Learning for HVAC Control in Tropical Commercial Buildings
The paper introduces CQD-ERL, a contextual quality-diversity evolutionary reinforcement-learning controller for HVAC systems in tropical commercial buildings. The method maintains a product archive of specialized policies indexed by operating context, weather/load regimes, and behavior descriptors, using a hybrid gradient-free evolutionary operator and soft-actor-critic policy-gradient operator with shared replay buffers. Actions are filtered through a deterministic safety shield. Evaluated on a reduced-order model of a Singapore building, CQD-ERL outperforms an ASHRAE Guideline 36 baseline in annual backtesting.
evolutionary reinforcement learningquality-diversity optimizationhvac controlsoft-actor-criticsafety shield
Deployment Decision Reliability: A Generalizability-Theory Framework for Sizing Long-Horizon Agent Evaluations
The article introduces Deployment Decision Reliability (DDR), a Generalizability Theory framework for evaluating long-horizon agent performance, revealing that current leaderboards measure specialization rather than capability. Using variance decomposition across three benchmarks (TheAgentCompany, $τ^2$-bench, AppWorld), the study finds agent effects account for <3% of variance, while agent-task interactions explain 7-23%. Key findings include reliability collapse on hard tasks (Eρ² drops to 0.000), negative correlation between training and held-out reliability (r = -0.90), and stable population-level diagnostics (capability-gap ratio 0.35-0.40). DDR standardizes reporting for enterprise decisions.
generalizability theoryvariance decompositionagent benchmarksreliability collapsedeployment decision reliability
Socioduality: A Relational Process Framework for Human-AI Interaction
The article introduces socioduality, a relational process framework for analyzing human-AI interactions through sequential, reciprocal exchanges that preserve historical context. The construct employs nested units (moves, episodes, pathways) and requires evidence of response contingency and return contingency for episode confirmation. Three propositions address history-conditioned formation, pathway divergence, and robustness differences. A frozen operational protocol was validated on three natural human-AI records, achieving exact convergence in two cases and minor divergence in one. Socioduality enables bounded, empirical analysis of interaction pathways, preserving information lost in endpoint-centered approaches.
sociodualityrelational processresponse contingencypathway divergencehuman-ai dyads
Terminal Symmetry as a Decision Resource: Statewise Refinement for Anytime Verified Construction
The paper introduces a decision-resource framework leveraging terminal symmetry for sequential construction tasks, formalized as transport--refine--certify. The method employs an episode-fixed transported process structure, state-restricted process rank, and state-dependent residual rank refreshed after transitions, with an ordinal rank meet guaranteeing completion under prefix coverage. Evaluated on CAD assembly, Mini-Programs, and exact-fill packing, statewise refresh improves anytime AUC by up to 6.77, 21.75, and 8.68 points, respectively. On 1,135 target-removal episodes from GRN OOD scenes, the method achieves the lowest mean capped verifier cost across scales, demonstrating transferability across aggregation and scheduler organizations.
terminal symmetrystatewise refreshordinal rank meetprefix coverageverifier cost
Clinical Feasibility of Low-Magnification Fluorescence Imaging for Breast Cancer Margin Detection Using Texture Analysis and Deep Learning
This study demonstrates that 4x magnification in microscopy with ultraviolet surface excitation (MUSE) achieves comparable diagnostic accuracy to 10x for breast cancer margin detection, enabling faster imaging with larger field of view. Texture analysis (local binary patterns) and deep learning (Vision Transformer) were evaluated on patch-level classification of surgical breast tissue. Both methods showed equivalent performance (DL: 98.18% accuracy, 96.30% sensitivity, 100% specificity; TA: 96.67% accuracy), with no clear advantage for higher magnification.
microscopy with ultraviolet surface excitationlocal binary patternsvision transformersurgical margin assessmentpatch-level classification
Backdoor Decontamination Dynamics in LLM Agents
The paper introduces a framework for analyzing backdoor decontamination dynamics in LLM agents, focusing on defensive poisoning and unlearning strategies. Using the AgentDyn benchmark, the authors conduct 115 experiments to evaluate how installing and removing known backdoors affects pre-existing unknown backdoors. Results show defensive poisoning alone erases 56% of original backdoors, while subsequent decontamination eliminates nearly all survivors. Notably, co-installing multiple backdoors increases resistance (36% erased), but decontaminating one clears 87% of co-residents. J-lens visualizations reveal persistent trigger awareness in intermediate layers despite restored benign behavior.
backdoor decontaminationdefensive poisoningllm agentsunlearningj-lens
CLEAR: Class-wise Expert Aggregation with Structured Sampling for Long-Tailed Classification
CLEAR introduces a modular ensemble framework for long-tailed classification, addressing reliability challenges through class-wise expert aggregation. The method generates diverse experts via threshold-based structured sampling while preserving the full label space, estimates class-wise trust scores using smoothed class-wise precision, and aggregates predictions through class-wise generalized product-of-experts. Evaluated on CIFAR-100-LT, ImageNet-LT, and Places-LT across multiple backbones, CLEAR achieves competitive overall accuracy and strong few-shot performance, demonstrating the efficacy of class-wise expert reliability in ensemble learning.
long-tailed classificationensemble frameworkstructured samplingclass-wise precisionproduct-of-experts
Policy Convergence and Divergence Across National and Within Regional AI Strategies: A Policy Design Element Analysis
This study analyzes horizontal and vertical convergence in AI policy design elements by coding 74 national and 3 regional AI strategies using a latent-inductive approach focused on goals, approaches, and principles. Results show strong horizontal convergence on economic competitiveness, research support, and ethical AI, but divergence in human rights, participatory governance, and human-centric principles. Vertical convergence varies: the African Union exhibits high alignment, the EU aligns on regulatory-economic priorities but diverges on human-centric values, and the Nordic-Baltic Region shows mixed patterns. The findings provide an evidence base for emerging AI policy norms.
policy design elementshorizontal convergencevertical convergencelatent-inductive codingai strategies
SegPAR: Class-Centric Decision-Based Sparse Attack for Semantic Segmentation
SegPAR introduces a class-centric decision-based sparse attack for semantic segmentation, addressing query inefficiency in existing image-centric methods. The framework employs a novel discrepancy reward to mitigate misleading feedback during pixel accumulation. Evaluated against black-box baselines, SegPAR achieves superior sparsity efficiency and mean Intersection-over-Union (MIoU) reduction, rivaling white-box sparse attacks. Code is publicly available.
semantic segmentationsparse attackblack-boxquery efficiencydiscrepancy reward
ComBodied Agents: a New Paradigm of Human-Centric Agentic AI
The paper introduces Combodied Agents, a human-centric paradigm for Agentic AI that focuses on modeling and supporting individual human-state trajectories. The framework integrates multimodal perception, longitudinal memory, Personal World Models, and admissible intervention policies into a closed-loop system, leveraging software tools, sensors, wearables, and robots. It emphasizes purpose-bounded, uncertainty-aware, and user-correctable representations rather than exhaustive Human Digital Twins. The design space is organized by human-state targets, relational contexts, and agent roles, with scenario-centered evaluation and agency-preservation metrics proposed. This approach shifts Agentic AI from task completion to sustained human benefit.
agentic aimultimodal perceptionpersonal world modelsadmissible intervention policyhuman digital twin
A Framework for Designing Reward Functions: From Objectives to Features to Human-Aligned Reward Functions
The authors introduce a formal framework for designing human-aligned reward functions from task descriptions. The method involves three steps: distilling task objectives into measurable outcome variables, selecting causally representative reward terms via minimum-cost partial cover on a causal DAG, and fitting weights through preference elicitation framed as a convex feasibility problem. The approach ensures a conflict-free feasible weight region, narrowed via O(n log κ) preference queries using separation oracle methods. This is the first reward-design method to maintain deterministic conflict-free regions while iteratively refining weights to desired tolerances.
reward functionscausal dagpreference elicitationconvex feasibilityseparation oracle
Earth observation embeddings are effective sub-grid descriptors for probabilistic weather downscaling
The study demonstrates that Earth observation foundation model embeddings (TESSERA) enhance probabilistic weather downscaling by encoding persistent surface properties. A convolutional conditional neural process downscales ERA5 reanalysis (~25 km) using 10 m-resolution TESSERA embeddings as sub-grid descriptors, improving predictions of 2 m temperature and 10 m wind speed. Evaluated across five climatically diverse regions, the method achieves 11.5% CRPS improvement for temperature and 6.2% for wind speed, with gains persisting when using Aurora AI forecasts or predicting at new stations. The embeddings capture topography-driven temperature variations and surface-driven wind speed deviations.
probabilistic downscalingearth observation embeddingstesseraconditional neural processcrps
Calibration Bets on the Past: Post-Training Quantization for Financial Time-Series Forecasting
This study investigates activation calibration for post-training quantization (PTQ) in financial time-series forecasting, focusing on cross-sectional volatility prediction for the S&P 500. The authors evaluate seven neural architectures across eight years (2018-2025) using 560 trained models, comparing absolute-maximum (abs-max) and percentile calibration methods. Results show that activation calibration becomes critical at 4-bit precision, where abs-max calibration degrades mean information coefficient by 11-62%, while percentile calibration recovers 53-94% of this loss. Market period variability affects optimal activation ranges, with narrow ranges benefiting typical conditions but underperforming during high dispersion. The findings highlight activation calibration as a key deployment decision for 4-bit PTQ in financial forecasting.
post-training quantizationactivation calibrationcross-sectional volatilityabsolute-maximum calibrationpercentile calibration
Regime-Gated Residual Mixture-of-Experts for Cross-Sectional Volatility Forecasting
The paper proposes RG-ResMoE, a regime-gated residual mixture-of-experts architecture for cross-sectional volatility forecasting, where regime information is restricted to expert routing rather than direct forecasting. The method employs a base predictor for volatility from stock features and a gating network using regime state variables to route residual corrections, evaluated via rolling walk-forward tests on 1,027 U.S. equities and a Japanese panel. Results show RG-ResMoE outperforms capacity-matched MLPs in accuracy and training stability, with hard routing underperforming soft routing, demonstrating that MoE's primary value lies in controlling regime information's influence rather than increasing capacity.
regime-gated routingresidual mixture-of-expertsvolatility forecastingcross-sectional analysissoft routing
An Efficient Near-Optimal Algorithm for Adversarial $m$-Set Bandits
We propose a computationally efficient algorithm for adversarial $m$-set bandits that avoids explicit enumeration of the exponentially large action set. The method leverages the underlying structure of $d$-dimensional item losses to represent sampling distributions with $d$ parameters, enabling polynomial-time execution. Against adaptive non-anticipating adversaries, the algorithm achieves high-probability regret $R_T = O(\sqrt{dT\log(K/\delta)})$, matching the optimal bound of EXP3-KW while requiring only polynomial space. This resolves the open problem posed by Maiti et al. regarding efficient implementations for combinatorial bandits.
adversarial banditscombinatorial optimizationregret analysissampling distributionpolynomial-time algorithm
ScreenShot: A Foundation Model for Few-Shot Combination Drug Screening
ScreenShot, a hierarchical transformer foundation model for few-shot combination drug screening, addresses the challenge of expensive and time-consuming combinatorial screens. Pretrained on 40 drug screening datasets covering 3,700 drugs and 6,000 biological samples, ScreenShot predicts patient-specific combination therapy responses through in-context learning, requiring no fine-tuning or molecular profiling. Evaluated on four held-out datasets, ScreenShot outperforms baselines in prediction accuracy and identification of selectively effective treatments. Its internal representations enable a weighted k-means++ active learning strategy, achieving equivalent hit detection to uniform screening with one-third the budget.
hierarchical transformerfew-shot learningin-context learningactive learningcombination drug screening
Autonomous Telerehabilitation via Skeletal Motion Prediction and Joint-Level Performance Assessment
This paper contributes an autonomous telerehabilitation pipeline integrating skeleton-based exercise quality assessment and short-term motion prediction for marker-free RGB video. The system employs a two-module architecture: a self-attentive Bidirectional LSTM with MMD-NCA metric learning for exercise classification, and a graph-based motion predictor generating per-joint position errors. The classifier achieves 96.45% mean-class accuracy on PROZIS squat sequences, while the STARS predictor attains 75.8 mm mean MPJPE at 560 ms on Human3.6M, outperforming graph and recurrent baselines. Designed for assistive robotics and home rehabilitation, the framework combines motion recognition and prediction to enable feedback-driven telerehabilitation.
bidirectional lstmmmd-ncampjpegraph-based predictiontelerehabilitation
ADEPT: A Unified Framework for Deep Learning Test Adequacy
ADEPT introduces a unified framework for deep learning test adequacy, integrating diverse metrics such as neuron coverage, surprise adequacy, and boundary coverage under a consistent workflow. The framework features a template-based interface for metric integration, YAML-based configuration, preprocessing-cache reuse, and structured reporting to enhance reproducibility and usability. Designed for researchers and practitioners, ADEPT eliminates the need for manual tooling configuration, facilitating efficient adoption of adequacy metrics in research and deployment.
test adequacyneuron coveragesurprise adequacyboundary coverageyaml configuration
FQTree: Fine-grained Quantization and Hardware Generation of Boosted Decision Trees
FQTree introduces a fine-grained quantization-aware training algorithm for boosted decision trees (BDTs), coupled with the QXGB hardware generation framework. The method employs a hardware-oriented leaf-value quantization scheme with global step and tree-wise shift, enabling compact integer representations, controlled clipping/pruning, and bias folding. Quantization is applied during boosting to adapt later trees to quantized ensemble errors, followed by compiler-based hardware lowering. Evaluations on JSC, MNIST, and NID datasets show 26-57% LUT reduction versus state-of-the-art FPGA-based BDT designs while maintaining or improving accuracy.
quantization-aware trainingboosted decision treeshardware generationfixed-point formatsfpga optimization
Attractor Image-Based Deep Learning of Arterial Pulse Waves for Age Classification
The study introduces a method for vascular age classification by transforming arterial pulse waveforms into images using Symmetric Projection Attractor Reconstruction (SPAR) and training a convolutional neural network (CNN) on these representations. Pulse wave data from photoplethysmography (PPG) and arterial tonometry were converted to SPAR images to classify healthy subjects into two age groups (35-40 and 50-55 years). The model achieved F1 scores >70% on internal and external test sets, demonstrating discriminative morphological features in SPAR images even for closely aged cohorts, suggesting potential for early cardiovascular risk detection via wearables.
arterial pulse wavesymmetric projection attractor reconstructionphotoplethysmographyvascular age classificationconvolutional neural network
The Advective Fisher-Rao Geometry of Deterministic Measure Transport
The paper introduces an advective Fisher-Rao metric for optimizing paths of probability measures governed by the continuity equation, demonstrating its optimality for descent directions. The metric is derived from three perspectives: as a zero-noise limit of the Fisher-Rao metric on path measures, as the expected second variation of the Freidlin-Wentzell rate functional, and as the Hessian of the Benamou-Brenier action. Computational experiments show the metric's superiority in probability density fitting compared to the Gauss-Newton method, which excels in velocity field fitting.
fisher-rao metriccontinuity equationfreidlin-wentzellbenamou-brenieroptimal transport
Beyond Parameter Space: NTK-Guided Personalized Aggregation for Robust Federated Learning
LIGHTYEAR introduces a federated learning framework that performs update selection in function space using an NTK-based agreement score, addressing the limitations of parameter-space similarity in heterogeneous, non-IID data environments. The method employs a peer-to-peer topology where clients exchange updates directly and evaluate incoming models on private validation data, selecting only beneficial updates for personalized aggregation. This approach improves stability and local model performance under data heterogeneity. Empirical evaluations across five datasets and nine baseline methods demonstrate that LIGHTYEAR consistently outperforms centralized FL baselines and existing P2P approaches.
federated learningneural tangent kernelfunction spacepeer-to-peerheterogeneous data
Task- and dataset-specific information in protein language models
This study investigates task- and dataset-specific information distribution across layers of protein language models (PLMs). Analyzing 13 PLMs across 15 downstream tasks from 11 datasets, the authors trained probe models on embeddings from each layer and compared their performance. Results show that embeddings from the last layers rarely yield optimal performance, with task-specific information distributed differently across layers. Residue-level tasks exhibit steady improvement across layers, while whole-protein tasks depend more on dataset characteristics. Embeddings from shallow layers perform better for deep mutational scan data, whereas deeper layers excel with diverse natural proteins. PLM performance significantly declines on artificial protein tasks.
protein language modelsembeddingsdownstream tasksdeep mutational scanlatent spaces
Look What the Probes Dragged In! Real-World Chest X-ray Shortcuts in MedCLIP
The study investigates shortcut vulnerabilities in MedCLIP, a medical CLIP-based vision-language model, by analyzing its ResNet-50 vision encoder. Seventeen linear classification probes are attached to intermediate layers and trained on NIH-CXR14 and PadChest datasets for pneumothorax and cardiomegaly detection. Evaluation reveals high AUROC but poor calibration, with shortcuts emerging at varying depths: localized shortcuts (e.g., drains) appear in later layers, while diffuse shortcuts (e.g., scanner noise) manifest earlier. Manual image analysis identifies data quality issues in both datasets, highlighting the need for high-quality annotations to mitigate shortcut vulnerabilities. Code is available on GitHub.
medclipshortcutsaurocresnet-50calibration
NAE: Normalizing AutoEncoder
The paper introduces Normalizing AutoEncoder (NAE), a novel generative framework that improves upon existing flow autoencoders by optimizing encoder and decoder surrogates in alignment with reconstruction loss. The authors theoretically analyze training dynamics in normalizing flows with approximate inverses, demonstrating that current loss functions are suboptimal. NAE employs a conditional loss that aligns surrogate loss gradients with reconstruction gradients, enhancing performance. Experiments on molecule generation, tabular data, and image benchmarks show that NAE achieves state-of-the-art results, underscoring the importance of loss alignment in flow autoencoders.
normalizing flowsflow autoencodersreconstruction losssurrogate lossgenerative framework
Draw This First
The paper introduces a novel sketch generation approach that inverts traditional stroke ordering by predicting a 2D field to determine stroke sequence. The method employs a pretrained latent flow-matching transformer as an image prior, coupled with a VAE decoder trained to predict the order field, stroke mask, and segmentation. The predicted segmentation is vectorized into polylines and sorted by the field, yielding ordered vector sketches. The model supports text-to-sketch generation and image derendering, with optional text instructions to control drawing order.
stroke orderinglatent flow-matchingvector sketchstroke segmentationderendering
Towards Truly Unsupervised Evaluation of Feature Selection
The authors propose a novel unsupervised evaluation framework for feature selection algorithms, addressing limitations in existing methods that rely on implicit supervision. The framework employs Principal Component Analysis (PCA) and optimal transport to assess feature selection quality without label information, enabling truly unsupervised evaluation. This approach contrasts with conventional techniques that operate under unsupervised downstream tasks but still require label-derived supervision. The proposed method provides a principled way to evaluate feature selection algorithms in scenarios where labeled data is unavailable or impractical to obtain.
feature selectionunsupervised evaluationprincipal component analysisoptimal transportlabel-free assessment
Direct Acceleration of Stochastic Root-Finding Without Variance Reduction and Regularization
(No summary returned.)
Clustered Randomized Smoothing for Stochastic Prediction Functions
The paper introduces clustered $α$-smoothing, a framework to improve robustness in stochastic multi-modal prediction by addressing mode collapse in randomized smoothing. The method (1) clusters noisy samples, (2) applies $α$-smoothing locally per cluster, and (3) combines predictions into a mixture distribution, deriving a lower bound on the smoothed prediction's correctness. Evaluations on stochastic trajectory prediction and quadrotor control show a 27% lower Wasserstein distance versus $α$-smoothing and an 81% collision rate reduction versus state-of-the-art randomized smoothing.
randomized smoothingmode collapsemixture distributionwasserstein distancestochastic prediction
SoftWater: Class-Aware Rate Allocation for Softmax Quantization
SoftWater introduces a class-aware rate allocation method for softmax quantization in LLMs, framing it as a KL-divergence-optimized rate-distortion problem. The method leverages second-order analysis to reveal a class-dependent geometry, using feature covariance and class-specific softmax curvature to guide quantization. A separability approximation enables efficient encoding via successive interference cancellation. Evaluated on models from 1B to 32B parameters, SoftWater reduces head-induced KL divergence by 6.5×–8.3× at 2 bits compared to WaterSIC, with a 2.9%–3.7% perplexity increase on Llama-3.2-1B-Instruct. Domain-matched calibration further optimizes performance.
softmax quantizationrate-distortionkl-divergencesuccessive interference cancellationllm
Reducing Symmetry Increase in Equivariant Neural Networks
This paper introduces a principled framework to reduce symmetry increase in Equivariant Neural Networks (ENNs), a phenomenon where outputs become invariant to transformations beyond input symmetries. The authors prove that symmetry increase admits an infimum determined by feature space structure and develop a computable algorithm to derive this infimum. They propose practical guidelines for feature design to mitigate harmful symmetry increases, demonstrating effectiveness under standard regularity assumptions. Experiments on synthetic datasets and the QM9 benchmark validate the theoretical predictions, showing reduced symmetry increase in most equivariant maps.
equivariant neural networkssymmetry increasefeature spaceinfimumqm9 dataset
Adaptive Bregman Proximal Stochastic Gradient with a Stabilized Barzilai--Borwein Step Size
Ada-BPSG introduces a line-search-free Bregman proximal stochastic gradient method that integrates SAGA gradient tables with a stabilized Barzilai-Borwein step size. The method employs a mediant to aggregate incremental secant information, minimizing the influence of nearly singular local ratios, and incorporates an explicit safeguard to ensure bounded step-size sequences. Theoretical analysis demonstrates an $O(n/K)$ ergodic rate for convex objectives, a restarted linear rate under relative quadratic growth, and an $O(1/K)$ bound for Bregman proximal residuals in nonconvex settings. Empirical evaluations on logistic regression and sparse nonnegative matrix factorization show Ada-BPSG achieves low objective values with reduced sensitivity to initial step size compared to standard variance-reduced baselines.
bregman proximalbarzilai-borweinsaga gradientrelative smoothnessvariance-reduced
Beyond Local Power: Functional Connectivity Analysis for Subject-Independent Learning Style Recognition
The paper proposes an EEG-based method for subject-independent learning style recognition using functional connectivity analysis, addressing limitations of traditional questionnaires and behavioral tracking. Phase Locking Value (PLV) connectivity features were evaluated against localized features across Felder-Silverman's Active-Reflective and Verbal-Visual dimensions, with EEG data from 28 participants during Raven's Advanced Progressive Matrices tasks. SVM classification with LOSO-CV achieved 70.00% accuracy for the Verbal-Visual dimension (fronto-occipital polarization) but only 55.56% for Active-Reflective due to overlapping executive networks and a 'Systematic Neural Inversion' phenomenon.
phase locking valueelectroencephalographyfelder-silverman modelsubject-independent classificationfunctional connectivity
A Remote Approach to Cashew Orchard Detection: Leveraging Active Learning with Satellite Imagery in Guinea-Bissau
The study presents a scalable remote detection method for nationwide cashew orchard mapping in Guinea-Bissau using Sentinel-2 satellite imagery and machine learning. A margin-based active learning approach optimized training data selection for informativeness, achieving 94.0% balanced accuracy in orchard classification. The work releases two open datasets and a 10m-resolution 2021 cashew map, demonstrating feasibility for large-scale environmental monitoring of unregulated agricultural expansion.
active learningsentinel-2remote sensingcashew orchard detectiondeforestation monitoring
Latent variable models for simultaneous EOV identification and removal in population-based SHM
A hierarchical Bayesian framework is proposed for simultaneous identification and removal of unmeasured environmental and operational variability (EOV) in population-based structural health monitoring (PBSHM). The method models latent EOV as a state-space Gaussian process, enabling O(T) inference via Kalman filtering and Laplace approximation for population-level EOV identification. Validation on a laboratory-scale benchmark structure demonstrates robust damage detection and EOV recovery. Application to a simulated nine-turbine offshore wind farm shows substantial true-positive rate improvement over projection and cointegration baselines at matched false-positive rates.
environmental and operational variabilitypopulation-based structural health monitoringstate-space gaussian processkalman filterlaplace approximation
TESLA: Taylor Expansion of Sinusoidal Learnable Activations
TESLA introduces a learnable sinusoidal activation function enabling explicit control over polynomial degree and selective amplification of high-order components, addressing the parity problem's challenges of linear inseparability and global interactions. The method combines sine and cosine terms with constrained coefficients, yielding Lipschitz/Rademacher complexity bounds and emphasizing higher-frequency structure in training dynamics. Empirical results demonstrate TESLA's strong generalization on parity with n=32 (100K training samples, 0.002% of input space), robustness under 30% label noise, and competitive performance on Forrelation and ImageNet-100 compared to SIREN, SNAKE, and Fourier feature embeddings.
parity problemlearnable activationpolynomial degreelipschitz complexityrademacher complexity
Distillation of Foundation Models for Time-dependent PDEs
Proposes Teacher Rollout Extension (TREX), a knowledge distillation framework for compressing foundation models of time-dependent PDEs into efficient student models. TREX augments limited downstream data by generating synthetic trajectories via teacher rollouts with optional noise injection, exposing students to long-horizon states and recovery behaviors. Students can incorporate task-specific inductive biases (e.g., equivariance) absent in teachers. Evaluated on multiple PDE benchmarks, TREX students match or exceed teacher accuracy while reducing parameters by orders of magnitude and achieving >10× inference speedup.
knowledge distillationpartial differential equationsteacher rolloutequivarianceinference acceleration
LODESTAR: Trustworthy Entropy Is Navigated, Not Merely Measured -- Reinforced Polarizer Keeps a Frozen LLM from Being Confidently Misled by the Wrong Evidence
LODESTAR introduces a reinforcement learning-based polarizer to improve answer selection in retrieval-augmented QA by inducing uncertainty in a frozen LLM respondent. Unlike prior entropy-based methods that fail when misleading passages reduce entropy, LODESTAR trains a fixed natural-language string (polarizer) offline using gold answers and LLM judges, then inserts it into prompts during inference. Evaluated on 5,008 questions across five benchmarks, LODESTAR achieves the highest mean $F_1$ (0.5339 vs. 0.5148), exact match (0.4136), and GPT-4o judge score (0.6435), reducing misleading passage selection by 4.3%.
retrieval-augmented qaentropy-based selectionfrozen llmreinforcement learningnatural-language polarizer
A Factor Graph Approach to Scalable Multi-Output Gaussian Process Regression
The paper introduces a factor graph approach to scalable multi-output Gaussian process regression, reducing computational complexity from cubic to linear scaling in observations. The method constructs a nearest-neighbor chain of fixed candidate inputs, employing Gaussian message passing with Matérn processes and a linear model of coregionalization for mixing latent processes into outputs. Evaluations on synthetic and electricity time series data show competitive accuracy with exact and approximate baselines, particularly in low input dimensions, while maintaining linear scalability in data points.
gaussian process regressionfactor graphmessage passinglinear model of coregionalizationnearest-neighbor chain
Disentangling the Expressivity of RoPE
The study formalizes two accounts of rotary position embeddings (RoPE) in transformers: expressivity via modular predicates and mechanistic positional anchors. Analyzing fully uniform, finite-precision soft-attention transformers, it shows that periodic RoPE components recognize languages definable in past temporal logic with modular predicates, whereas conventional RoPE's non-repeating rotations simulate bounded look-back operators. Experiments confirm periodic schedules generalize on modular tasks, while conventional RoPE exhibits locality bias, hindering distant-context access. These findings bridge theoretical expressivity and practical RoPE implementations.
rotary position embeddingsmodular predicatestemporal logiclocality biasposition-invariant access
DCM Bandits: Multiplayer Information Asymmetric Cascading Bandits for Multiple Clicks
The authors extend Dependent Click Model (DCM) Bandits to a multiplayer information-asymmetric setting, where multiple agents interact with a shared ranked list and may observe multiple clicks per session. They address asymmetry in actions and rewards, providing sublinear regret guarantees for three settings where at least one asymmetry exists. The termination ranking need not be known for small termination probabilities, improving on prior single-agent results. Empirical results demonstrate algorithm effectiveness across asymmetric environments, emphasizing the role of feedback structure (full vs. first-click feedback) in coordinating exploration and minimizing regret.
dependent click modelinformation-asymmetricsublinear regrettermination rankingfeedback structure
Forward and Inverse Virtual Metrology for Phototransistor Gain: A Hierarchical, Uncertainty-Aware Approach for Small Production Datasets
The paper presents a hierarchical, uncertainty-aware approach for forward and inverse virtual metrology of phototransistor gain, targeting small production datasets. Using a real fabrication history of 13-14 process runs, the method decomposes gain variance, finding ~50% between-run variation, thus bounding recipe-only prediction. It introduces a forward predictor with uncertainty awareness, an inverse search for target gains, and a multi-level data-quality assessment with explicit cross-level linkage. The approach is validated on nested fabrication entities (batch, wafer, die) and released with normalized data and code for reproducibility.
virtual metrologyphototransistor gainhierarchical modelinguncertainty-awaresmall-sample
Small-Scale Experiments: Are We There Yet?
The study challenges the perception that scaling laws fail at small model scales (from 4M parameters), identifying hyperparameter sensitivity as the key confounding factor. Through systematic ablation of scaling law components, the authors demonstrate that well-tuned hyperparameters dominate other factors, with sensitivity diminishing as model size increases due to lower-dimensional loss surfaces. They introduce a methodology for model-centric research, validated by recovering large-scale architectural insights (pre-normalization superiority) from small-scale experiments. The work establishes that small models can obey scaling laws when properly tuned, though statistical limitations constrain extrapolation.
scaling lawshyperparameter sensitivitypre-normalizationloss surfacemodel-centric research
Air Quality Station Simulation via LSTM and Attention-Based Modelling
The paper introduces SATADL, a deep-learning model for multi-hour air quality forecasting during station outages, combining LSTM and attention mechanisms to simulate PM10 concentrations. The architecture extracts complex spatiotemporal patterns from neighboring station data, outperforming baselines on 48-hour prediction windows across four global datasets. Evaluation shows superior performance in both R² (coefficient of determination) and RMSE metrics compared to existing deep learning approaches.
lstmattention mechanismpm10 forecastingspatiotemporal modelingvirtual sensor
Kernel Methods for Learning Operators with Multiple Inputs and Outputs
We introduce a kernel-based encoder-decoder framework for multi-input, multi-output operator learning, enabling mappings between products of distinct function spaces. The framework separates observation, representation, learning, and reconstruction, with convergence rates governed by the most challenging constituent approximation problem. KernelMO, a specialized family of kernel methods, combines operator-valued and product-space formulations. Evaluated on five families of parametric partial differential equations, the methods achieve competitive or state-of-the-art predictive accuracy while reducing training and inference costs compared to neural operator architectures and deep learning models.
operator learningkernel methodsencoder-decoderfunction spacesparametric pdes
Towards Understanding On-Policy Distillation through the Lens of Test-Time Scaling
This study investigates on-policy distillation (OPD) for LLM reasoning enhancement through test-time scaling analysis, varying sampling budget K and evaluating with pass@K and avg@K metrics. Results show OPD-trained models maintain superior avg@K across budgets, while pass@K advantage shifts to pre-OPD base models as K increases, indicating OPD primarily improves sampling efficiency rather than expanding reasoning capability boundaries. Problem-level solvability analysis reveals OPD causes more previously solvable problems to become unsolvable than vice versa, suggesting OPD's gains stem from sampling efficiency rather than genuine capability acquisition from the teacher.
on-policy distillationtest-time scalingsampling budgetpass@kavg@k
Located but Not Releasable: Silent Gate Inversion and Bounded Linear Release
The study presents a preregistered stress test of the complete pipeline for detecting, localizing, and releasing latent task-relevant structure in a 25.7M transformer trained on causal-evidence discrimination. Key findings include: (i) successful localization via mid-layer interventions (best-site release rate 0.889), (ii) out-of-distribution gating failure (detector triggers on 6.9-7.3% of in-distribution generations but zero of 2,400 needed instances), and (iii) bounded linear release efficacy (plateauing below preregistered thresholds). The failures in gating and linear release are dissociable, with neither overturning localization. All results are traceable to hashed artifacts in the audit chain.
localizationgatinglinear releaseout-of-distributiontransformer
Learning with Bilevel-Minimax Optimization for Efficient and Reliable Transfer Attacks
The paper proposes BMAT (Bilevel-Minimax Adversarial Transfer), a bilevel-minimax optimization framework for improving transfer-based adversarial attacks. BMAT formulates the attack as a bilevel problem coupling initialization and perturbation, with an inner minimax objective to enhance surrogate model robustness. The method integrates a Soft Weight Modulator and Implicit Gradient Approximator for efficient optimization. Experiments on classification and segmentation benchmarks demonstrate BMAT's superiority over 10+ baselines across 30+ victim models, achieving up to 2x reduction in mIoU for cross-architecture transfer.
adversarial transferbilevel optimizationminimax optimizationgradient dynamicssurrogate adaptation
Can Vision Models Read the Radar Display? On the Feasibility of Radar Imagery for Air Traffic Complexity Estimation
The study demonstrates that radar imagery, despite its atypical sparsity and self-similarity, is a viable input for deep learning vision models in air traffic complexity estimation. A Vision Transformer (ViT) processes position images augmented with five aircraft state channels (heading, speed, altitude) to regress four intrinsic complexity components derived from pairwise geometric relations. The model achieves $R^2 > 0.96$ for all components and responds proportionally to aircraft removals, validating its sensitivity to operationally relevant changes.
vision transformerradar imageryair traffic complexitygeometric relationssparsity
JAPE: Joint Anomaly Prediction and Intrinsic Explanation in Multivariate Time Series
JAPE introduces a joint framework for anomaly prediction and intrinsic explanation in multivariate time series, shifting focus from numerical-deviation to dependency-structure modeling. The method employs a Decoupled Spatio-Temporal Representation (DSTR) backbone to capture lag-aware dependencies, a dual-view alerting mechanism combining numerical forecasts with evolving dependency graphs, and Native Predictive Explanation (NPE) for variable-level structural deviation ranking. Evaluated on five benchmarks across three prediction horizons, JAPE achieves average improvements of 19.7% in F1, 41.3% in AUC-PR, and 26.6% in MRR for explainability.
multivariate time seriesanomaly predictiondependency-structure modelingspatio-temporal representationnative predictive explanation
Orientation, not magnitude: the causal structure of task-vector interference in merged language models
The study demonstrates that task-vector interference in merged language models is causally determined by orientation rather than magnitude, challenging prior diagnostic approaches. Using a factorial ledger and layerwise flux decomposition, the authors show that cross-term amplification dominates (65-70% in tested families) and identifies the carried direction as a forward-pass attractor. Erasure along this direction dose-dependently removes interference, while norm-matched wrong-direction controls fail; instruction wrappers modulate the effect by 13x. Quantization artifacts explain prior magnitude-based 'universality' claims. All 46 predictions were preregistered, including falsifications of initial hypotheses.
task arithmeticlayerwise fluxcross-term amplificationinstruction wrappersquantization roughness
High-Order Liquid Evidence Encoding for Gradual GNSS Spoofing Detection in Autonomous Driving
The authors propose a causal high-order liquid evidence framework for detecting gradual GNSS spoofing attacks in autonomous vehicles. The method constructs a physics-guided GNSS-motion inconsistency residual, then processes separate evidence streams for residual level and its first- and second-order variations using adaptive liquid encoders with hierarchical coupling. Evaluated on three subsets of the AV-GPS dataset, the approach achieves F1-scores of 0.9535 and 0.9777 on Dataset 1 and 3 respectively, detecting attack transitions within four sampling steps.
gnss spoofingliquid evidenceautonomous drivingmotion inconsistencyadaptive encoder
TradingMoE: Routing the Right Experts in Evolving Markets
TradingMoE introduces a trading-oriented sparse Mixture-of-Experts (MoE) system that augments frozen dense LLMs with lightweight residual experts for financial markets. It employs a Query-Key router to match low-dimensional token-specific expertise queries with learnable expert keys, alongside a sparse expert selection update mechanism that samples inactive experts during training for dynamic adaptation to market shifts. Evaluated against 22 baselines in stock and cryptocurrency markets, TradingMoE improves cumulative returns by 30.89% and 30.7%, respectively, with persistent advantages in forward-only deployment.
mixture-of-expertsrouterlow-dimensionalsparse computationresidual experts
Achieving Near-Zero-Overhead Multi-Model Hierarchical Classification in Real-Time Detection Pipelines
The authors present a methodology for near-zero-overhead multi-model hierarchical classification in edge-deployed vision systems, addressing GPU serial bottlenecks by leveraging concurrent execution on NVIDIA Jetson DLA cores. The five-step approach includes architecture adaptation, manual dynamic range adjustment for TensorRT implicit quantization, quantization-aware training, ONNX graph surgery, and concurrent GPU-detection/DLA-classification pipelines. This recovers 94.0% accuracy from implicit quantization's 75% and achieves near-zero pipeline overhead (12.5 vs. 13.3 FPS detector-only at 1080p) on a Jetson Orin NX. The method is backbone-agnostic and generalizes to detection-classification pipelines.
quantization-aware trainingonnx graph surgeryimplicit quantizationconcurrent executiondynamic range adjustment
A comparison of CNN architectures for Alzheimer's disease detection in single-view MRI scans
This work benchmarks ten CNN architectures for Alzheimer's disease classification from single-view MRI scans, employing a two-stage transfer learning and full fine-tuning pipeline on a class-balanced subset of 3,900 images from the OASIS dataset. The architectures evaluated include ResNet, DenseNet, MobileNet, EfficientNet, and VGG family models, with VGG16 achieving the highest validation accuracy (0.9637) and test accuracy (0.9533). A consistent challenge across all models was distinguishing between Non-Demented and Very Mild Demented stages, highlighting a critical diagnostic difficulty in early Alzheimer's detection.
cnntransfer learningfine-tuningalzheimer's classificationmri scans
Tight Nonasymptotic Local Convergence of Sinkhorn-Knopp
The paper provides the first nonasymptotic local convergence analysis of the Sinkhorn-Knopp (SK) algorithm for matrix scaling, matching existing asymptotic Jacobian-based rates. By establishing polynomial-time convergence under connectivity conditions, the authors identify local suboptimality in SK and propose accelerated variants. For dense matrices, they improve the complexity of first-order matrix scaling algorithms from $O(\tfrac{n^{7/3}}{\varepsilon^{2/3}})$ to $O(\tfrac{n^{9/4}}{\sqrt{\varepsilon}})$.
sinkhorn-knoppmatrix scalingnonasymptotic convergencepolynomial-time algorithmdoubly stochastic
Automated binary classification of hazelnut X-ray images: A deep-learning benchmark for quality assessment
The study establishes a benchmark for binary classification of hazelnut quality using X-ray images, addressing challenges of subtle class differences, class imbalance, and limited annotations. Seven single-model configurations and ten probability-aggregation ensembles were evaluated on 799 segmented single-kernel X-ray images (224x224 pixels, grayscale) using a group-wise split-rotation protocol across five data splits. The average-probability ensemble of a binary cross-entropy-trained CNN and frozen Swin Transformer achieved the highest mean balanced accuracy (86.3% ±1.8%). Expert reassessment improved performance by 2.8-8.1 percentage points, underscoring the importance of rigorous evaluation and label curation in small, imbalanced agricultural imaging datasets.
binary classificationx-ray imagingclass imbalanceconvolutional neural networkswin transformer
Epiplexity Guided Data Selection and Generation for Out-of-Distribution Generalization
The paper proposes using epiplexity, a measure of structural information extractable by compute-bounded learners, to guide data selection and synthetic generation for improved out-of-distribution generalization. For selection, it fits scaling laws to training loss curves to predict epiplexity gains and adaptively weights domain sampling. For generation, it optimizes a REINFORCE policy gradient reward based on epiplexity changes over generated data buffers. Experiments show higher epiplexity correlates with better zero-shot and fine-tuning performance across domains, supporting structural information as a transfer learning facilitator.
epiplexityout-of-distribution generalizationreinforce policy gradientsscaling lawsstructural information
Chain-of-Thought Shows the Path to a Tree: Realizing Branching Complexity
The authors demonstrate concrete instantiations of Chain-of-Thought (CoT) reasoning for branching complexity, addressing a gap in explicit depth-bounded constructions and traversal procedures. They present CoT realizations of depth-first search (DFS) and Dijkstra's algorithm using unique hard-attention decoders with at most two layers. These decoders serve as a shared computational substrate, enabling the computation of the Strahler number of an n-vertex tree in 2n-1 steps with four layers and its width in n-1 steps with three layers. The constructions handle arbitrary n-ary trees without layer normalization or positional encodings, providing a non-trivial witness for the linear-step regime of the CoT hierarchy. Additionally, they exploit the bijection between ordered trees and Dyck paths, realized by their DFS construction.
chain-of-thoughtdepth-first searchdijkstra algorithmstrahler numberdyck paths
LEMUR: Latent Entropy-aware Multimodal Unlearning via Visual-anchored Reasoning Redirection
LEMUR introduces a training-free, inference-time unlearning framework for reinforcement learning (RL)-trained multimodal large reasoning models (MLRMs), addressing privacy leakage in reasoning traces via entropy-aware sanitization. The method detects sensitive content through RL-induced token-level entropy signatures and redirects reasoning trajectories via visual-anchored latent injection during vulnerable intervals. Evaluations show LEMUR outperforms existing methods, reducing reasoning-trace and answer leakage by 23-41% while maintaining non-sensitive utility and fluency in diverse MLRMs.
multimodal unlearningentropy dynamicsvisual-anchored reasoningreinforcement learningprivacy leakage
Drift and Dependence: Layer-wise Information-Theoretic Bounds for Replay-Based Continual Learning
The paper introduces a layer-wise information-theoretic framework to analyze generalization in replay-based continual learning, decomposing the expected gap into representation drift and optimization-dependence terms. The method employs a Wasserstein relaxation for drift analysis and an SGLD instantiation to quantify optimization effects, revealing depth-dependent trade-offs and gradient-alignment diagnostics. Experiments validate memory scaling laws, identify optimal interior layers for stabilization, and correlate alignment statistics with task forgetting.
continual learningrepresentation driftwasserstein relaxationsgld optimizationgradient alignment
FunnelCausalNet: Funnel-aware Joint Conversion-Revenue Uplift for Multi-tier Coupon Allocation
FunnelCausalNet introduces a joint uplift estimator for coupon allocation, coupling binary conversion and nonnegative conditional-value heads via $μ_{\mathrm{gmv}}=μ_{\mathrm{conv}}μ_{\mathrm{val}}$ under RCT-based assumptions. The method employs split-conformal CATE summaries, Bonferroni audit bands, and a Lagrangian budgeted allocator with RCT-anchored ROI accounting. On Criteo-MT7, it matches leading baselines in AUUC_GMV (within one seed SD) and reduces GMV error by 18--48% versus direct regression. In Hotel-Coupon RCT logs (4.9M records), it achieves best seed-averaged DeltaROI across seven anchors (10--60%), demonstrating frontier consistency.
uplift modelingcausal inferenceconformal predictionbudget allocationzero-inflated data
Continuous-Latent Predictive Modeling with Semantic Alignment for EEG-Language Foundation Models
The Brain Latent Predictive Model (BLPM) introduces continuous latent semantic prediction for EEG-language foundation models, addressing limitations in masked autoencoding and autoregressive approaches. BLPM employs a Continuous EEG Latent Predictive (CELP) encoder for transferable representation learning via latent target prediction, coupled with a Multi-Query Semantic Decomposition (MQSD) module to align EEG and textual semantics in a shared latent space. Evaluations across multiple benchmarks demonstrate robust generalization, establishing BLPM as an effective paradigm for cross-task EEG-language modeling.
eeg-language foundation modelscontinuous latent predictionsemantic alignmentmulti-query decompositionneural decoding
Towards a Formal Definition of Agent Memory: Basis, Span, Optimality, and the Sequential Memory Problem
The paper proposes a formal framework for defining and analyzing agent memory, introducing memory as a basis, knowledge as its span, and answerability as a coverage problem. It defines optimal memory as the capacity-constrained maximizer of expected coverage, establishing a utility-capacity frontier for comparing memory systems. The framework addresses noise in memory, balancing coverage and precision, and formalizes the continual agent-memory problem as a sequential MDP with memory as state, writing as action, and delayed rewards driving learning. The framework is instantiated on Homer's Odyssey, quantifying memory performance metrics and positioning existing systems within the framework.
agent memoryutility-capacity frontiersequential mdpcoverage problemoptimal memory
A Quantum/Classical Example Oracle Separation for Making Things Up
The paper establishes an oracle separation between quantum and classical examples in PAC learning. Specifically, it demonstrates that relative to an oracle, certain distributions can be efficiently learned by a quantum algorithm with access to quantum examples, but not by a quantum algorithm restricted to classical examples. This result addresses a previously open question in quantum learning theory, providing evidence for the computational advantage of quantum examples over classical ones in specific learning scenarios.
pac learningquantum examplesoracle separationquantum computationclassical examples
Robustness of AI-Art Detectors under Generator Shift
This work investigates the robustness of AI-art detectors under generator shift, revealing a significant generalization gap. Using a Stable Diffusion 3.5 Medium (SD3.5m) dataset spanning ten art styles, five detectors trained on U-Net-based latent diffusion artwork were evaluated in a zero-shot cross-generator setting. Results show strong in-distribution performance but notable degradation under generator shift, with many SD3.5m images misclassified as human while human false positives remained low. The CLIP ViT-L/14 model achieved the best overall performance, while Grad-CAM analysis indicated weaker, more diffuse activations on false negatives. These findings underscore the need for detectors that remain reliable across evolving generative architectures.
generator shiftzero-shotlatent diffusiongrad-camclip vit-l/14
Transferable Above-Ground Biomass (AGB) Estimation Model from Multi-Sensor Data with Sparse Field Calibration
A globally trained convolutional neural network (CNN) framework is proposed for transferable above-ground biomass (AGB) estimation, combining multi-sensor data (Sentinel-2, Sentinel-1, ALOS-2 PALSAR-2, DEM) with sparse field calibration. The CNN, trained on GEDI Level-4A biomass data across multiple regions and seasons, learns persistent woody structures rather than single-date appearances. Local field plots are used to fit scale-and-bias corrections without retraining. The pipeline harmonizes data onto a 10 m grid, computes vegetation indices and polarimetric ratios, and employs a hybrid log-domain SmoothL1 with RMSE loss. The global model achieves R²≈0.78 and RMSE≈22 Mg/ha; field calibration improves performance to R²≈0.82 and RMSE≈15 Mg/ha, outperforming the uncalibrated model and ESA CCI Biomass product.
convolutional neural networkabove-ground biomassgedi level-4asentinel-2field calibration
CAM-Guided Saliency Cutout and Image-Based Malware Classification
The study evaluates HiResCAM-guided cutout placement for CNNs, comparing random, low-saliency, and high-saliency cutouts against no cutout on RawMal-TF malware images (17 families, ~1,000 samples each) and CIFAR-100. Using ResNet18 (~100 epochs), experiments varied cutout areas (5-30%) and augmentation copies (M=4,8). Results show degraded performance for all cutout types on RawMal-TF but slight improvement with low-saliency cutouts on CIFAR-100, indicating domain-dependent efficacy of saliency-guided augmentation. Malware images exhibit distinct behavior from natural images in this context.
hirescamcutoutresnet18data augmentationsaliency-guided
A Local Sinkhorn Framework for Conditional Distribution Reconstruction of Multidimensional Random Fields
The paper proposes a local Sinkhorn divergence framework for conditional distribution reconstruction of multidimensional random fields, offering a scalable alternative to exact local optimal transport. The method employs debiased Sinkhorn divergence to train stochastic neural networks via a differentiable local distribution matching objective, with theoretical generalization error estimates quantifying the bias-efficiency trade-off. Numerical experiments demonstrate superior balance between reconstruction accuracy, computational efficiency, and scalability compared to existing loss functions and uncertainty quantification frameworks.
sinkhorn divergenceconditional distribution reconstructionmultidimensional random fieldsstochastic neural networksuncertainty quantification
IoT-Enabled Autonomous Maritime Navigation in Smart Ports: A Curriculum-Guided Shared Policy Learning Framework
The paper proposes a curriculum-guided shared recurrent policy learning framework for IoT-enabled autonomous maritime navigation in smart ports, addressing partial observability and dense traffic. The method employs centralized offline training with onboard execution, enhancing temporal reasoning and edge-level decision-making via recurrent policies. Simulations in realistic port environments demonstrate improved navigation reliability (reduced collisions) and generalization to unseen high-density scenarios compared to baseline methods, validating scalability for IoT edge deployment.
autonomous maritime navigationiot edge intelligencecurriculum-guided rlshared recurrent policypartial observability
CookVoice: Unified Framework for Style Controllable Multi-Modal Human Voice Generation
CookVoice introduces a unified framework for multimodal, multi-style, and multi-task human voice generation by decomposing voice into content, prosody, and style factors. The framework employs a flexible alignment strategy to map text, style, and prosody control signals onto frame-level spectrograms, enabling tasks like text-to-speech, text-to-singing voice, style-controllable generation, voice mimicry, voice conversion, and voice editing. Experiments demonstrate that CookVoice achieves generation quality comparable to existing baselines while offering enhanced style and prosody controllability. With only 43.51 million parameters and efficient inference requiring as few as 4 ODE steps, CookVoice is a practical solution for real-world applications.
voice generationspectrogramprosodyautoregressive decodingode steps
Robust Multi-Tier Infant-Centered Audio Understanding with Whisper via Structured Speaker Conditioning
The paper introduces a family-conditioned, multi-tier audio tagger for infant-centered naturalistic recordings, addressing challenges of limited labeled data, low SNR, and domain shifts. The method combines a LoRA-finetuned Whisper encoder with a lightweight, speaker-aware Transformer for long-context inference and framewise prediction, enhanced by a sequence-level smoothing loss and a factorized speaker-token design with shared tier tokens and family-specific offsets. This approach reduces family bias and improves robustness, enabling effective audio tagging in home environments.
lora-finetuningspeaker-conditioningmulti-tier taggingsequence-level smoothingfactorized tokens
Unifying Physical Backpropagation
The paper presents a unifying theory for on-device gradient computation in physical computing systems, based on the adjoint method. It identifies sufficient conditions under which a physical system can generate its own adjoint field for exact gradients: linear systems require preserved reciprocity, while nonlinear systems need reciprocity of the linearized system and a time-reversal mirror. The method recovers existing techniques like Equilibrium Propagation and Hamiltonian echo backpropagation, and generalizes to non-Hermitian, non-reciprocal systems via an intertwining condition. Results extend to time-dependent parameters, Onsager-reciprocal dynamics, and nonlinear PT-symmetric Schrödinger equations.
adjoint methodphysical computingreciprocitynonlinear trajectory systemson-device gradient computation
RECAST: A Machine-Learning Framework for Correction and Super-Resolution of Coarse-Grid PDE Solvers
RECAST introduces a machine-learning framework for enhancing coarse-grid PDE solvers through recurrent error correction and super-resolution. The method combines in-loop learned correction during numerical time-stepping with fine-grid state reconstruction from corrected coarse trajectories. Evaluated on six 1D PDE systems with 8-16x coarsening, RECAST reduces time-averaged relative error by 50-92% over 1000-step rollouts and generalizes to unseen parameters, outperforming contemporary correction architectures in 5000-step tests.
pde solverserror correctionsuper-resolutioncoarse-gridtime-stepping
Sparse and robust geometric twin support vector machine via asymmetric RoBoSS loss function
The paper introduces a novel asymmetric, robust, bounded, sparse, and smooth (aR) loss function for $l_1$-norm penalized geometric twin SVM (aRSGTSVM) to address classification and regression tasks in noisy, high-dimensional data. The method combines $l_1$-norm penalty for feature selection with the aR loss function to mitigate label noise and enhance stability against resampling noise. A proximal gradient descent algorithm is developed for efficient optimization. Experiments on synthetic and UCI datasets demonstrate aRSGTSVM's superiority over state-of-the-art methods, and its application to index tracking in the China stock market yields satisfactory results.
geometric twin svmasymmetric lossfeature selectionproximal gradient descentindex tracking
When Offline Evaluation Misleads: A Diagnostic Protocol for Reward and Policy Selection in Delayed-Feedback Contextual Bandits
The authors propose a diagnostic protocol for evaluating reward functions and policies in delayed-feedback contextual multi-armed bandits (CMABs), addressing systematic misguidance from standard offline evaluation methods. The protocol assesses alignment (whether optimizing the reward improves the north-star metric) and learnability (whether the bandit can identify the reward-optimal policy) before trusting reported performance gains. Validation on an off-policy-evaluation benchmark and a synthetic generator reveals two key insights: denser reward signals improve online learning despite static estimates, and per-user policies often provide robustness rather than true personalization. The protocol was applied to a large-marketplace push system with five arms, yielding directional evidence.
contextual multi-armed banditsdelayed feedbackoff-policy evaluationreward alignmentlearnability
Certifying What Helps Customer-Return Timing: A Screen-and-Confirm Test for Conditioning Signals, and Why Decay Is Nearly Enough
The authors introduce a screen-and-confirm protocol to certify whether candidate signals improve temporal-point-process (TPP) event-timing likelihood, validated on categorical/continuous encodings and real-world datasets. They also measure a model-free ceiling, showing that customer-return timing is largely unpredictable (single-digit percentage of gap variance). Applying these tools to public benchmarks (Amazon, Taobao, RetailRocket) and a marketplace (Thumbtack), they demonstrate that continuous-time decay is nearly sufficient for modeling returns, while additional conditioning signals are redundant or harmful (null to 0.06 NLL impact). The contribution lies in certifying the redundancy of conditioning signals and addressing evaluation pitfalls.
temporal-point-processscreen-and-confirmcontinuous-time decayevent-timing likelihoodcovariate-conditioned
Fine-Tuning Generative Models for Extreme Events via CVaR-Penalized Wasserstein Gradient Flows
The paper introduces CVaR-penalized Generative Particle Algorithm (CVaR-GPA), a method for fine-tuning generative models to capture heavy-tailed distributions and extreme events without requiring prior knowledge of tail characteristics. The approach leverages Wasserstein gradient flow of Lipschitz-regularized KL divergence, penalized by a Conditional Value-at-Risk (CVaR) term, enabling robust learning and transport toward heavier-tailed targets. CVaR-GPA operates on empirical measures, deriving subgradients from Rockafellar-Uryasev representation, and fine-tunes pre-trained models adaptively via kinetic-energy stopping. Experiments on synthetic distributions and the Fama-French 25 portfolio dataset demonstrate significant improvements in global and tail accuracy for heavy-tailed targets.
wasserstein gradient flowconditional value-at-risklipschitz regularizationheavy-tailed distributionskinetic-energy stopping
Robust Ambiguity Detection (RAD) From Model- and Feature-Space Consistency
The Robust Ambiguity Detection (RAD) framework introduces a method for quantifying predictive ambiguity in machine learning models through Model-Space Consistency and Feature-Space Consistency. These metrics, visualized via the RAD Plot, identify ambiguous predictions that may require abstention or human inspection, particularly in high-stakes scenarios. RAD evaluates ambiguity on synthetic datasets with controlled overlap and real-world datasets where ambiguity is not directly observable. A downstream application ranks samples by RAD Pareto-Rank, abstaining from the most ambiguous predictions, achieving performance comparable to existing rejection-based methods.
predictive ambiguitymodel-space consistencyfeature-space consistencyrad plotpareto-rank
FLARE++: Low-rank attention with dynamic attention routing
FLARE++ introduces dynamic token routing to low-rank attention architectures, addressing the fixed query limitation in FLARE by generating input-conditioned queries through an additional encoder call. This approach maintains FLARE's explicit low-rank factorization and O(NM) complexity while using standard scaled dot-product attention operations. The architecture includes a multi-GPU context-parallel implementation that avoids full token sequence gathering on any single device. FLARE++ achieves a 24% average improvement over FLARE on PDE surrogate benchmarks and gains 2.3 points of average accuracy on the Long Range Arena.
low-rank attentiondynamic token routingscaled dot-product attentionpde surrogatescontext-parallel implementation
RelShap: Relationally Consistent Shapley Explanations
RelShap introduces a framework for computing Shapley value-based feature attributions that preserves relational constraints in machine learning pipelines, addressing the limitations of traditional methods that rely on feature independence. By incorporating data provenance and restricting evaluations to relationally valid configurations, RelShap maintains estimator-agnostic compatibility with Kernel SHAP, Monte Carlo, and Leverage SHAP while exploiting functional dependencies to reduce runtime through coalition equivalence classes. Empirical evaluations across multiple datasets and models demonstrate that RelShap produces more faithful explanations, correctly identifying dominant features in controlled settings where existing methods like Conditional SHAP and ManifoldShap fail.
shapley valuesrelational constraintsfeature attributiondata provenancefunctional dependencies
Language-Structured Relational Q-Learning for Threat-Aware Control in Safety-Critical Driving
The paper introduces Language-Structured Relational Q-Learning, implemented via an Ego-Centric Relational Q-Network (ERQ-Net), to improve threat-aware control in safety-critical driving. ERQ-Net jointly learns inter-vehicle relevance and action values from dynamic traffic graphs, using language descriptions during training but hiding prompts and semantic roles during inference. Evaluated across 2,500 scenarios, language-structured training boosts test success from 49-52% to 55-58% and adversary-focused attention by 1.2x to 2.1x, though adaptive control remains limited (76% solved by simple policies). The work identifies a recognition-control gap, showing reward reweighting and margin shaping fail to prevent policy collapse, and analyzes transfer to CARLA.
relational q-learningego-centric networksafety-critical drivinglanguage-structured trainingrecognition-control gap
Defending against Model Extraction for GNNs with Model Reprogramming
The paper introduces GraphRP (Graph Reprogramming Protection), a proactive defense against Model Extraction (ME) attacks for Graph Neural Networks (GNNs) in MLaaS settings. Unlike Euclidean-based defenses, GraphRP employs a Structure-Aware Gating Mechanism with learnable topological prototypes to dynamically modulate decision boundaries, preserving utility for benign queries while maximizing Fisher Information for adversarial ones. Theoretical analysis under bounded loss and optimal attacker assumptions shows increased attacker estimation error with structural sensitivity. Experiments on hard-label and soft-label ME attacks confirm GraphRP's effectiveness in degrading attack performance without compromising benign utility.
graph neural networksmodel extractionmodel reprogrammingfisher informationtopological prototypes
Cross-Corpus Evaluation of Generalizable Vulnerability Detection in IoT Firmware
The paper introduces IoTVulBench, a human-verified benchmark for cross-corpus firmware vulnerability detection, addressing limitations in existing synthetic or general-purpose datasets. The benchmark, constructed from GitHub repositories and validated by expert reviewers, was evaluated across five model architectures, two tuning methods, and three curriculum strategies. Results show that models trained on IoTVulBench achieved a Matthews Correlation Coefficient (MCC) of 0.58, outperforming PrimeVul (0.44) and D2A (0.39), with staged curriculum learning further improving MCC to 0.69 and a diversity-optimized ensemble reaching 0.73. The model demonstrated strong generalization, retaining 86% performance under identifier renaming and missing only 21% of vulnerabilities at a 0.5% false-positive rate.
iot firmwarevulnerability detectioncross-corpus evaluationmatthews correlation coefficientcurriculum learning
Forward Trajectory Steering for Hamilton-Jacobi Reachability Analysis
STEER2REACH (S2R) introduces a lightweight adaptive collocation sampling method for Physics-Informed Neural Networks (PINNs) in Hamilton-Jacobi (HJ) reachability analysis, eliminating complex training pipelines. The method steers forward trajectories using optimal control and disturbance signals with stochastic noise, requiring minimal modification to standard PINNs. Evaluated on reachability benchmarks, S2R achieves competitive or improved safety metrics and reduces relative L2 error compared to state-of-the-art MPC-guided HJ solvers, without multi-stage training or MPC supervision.
hamilton-jacobi reachabilityphysics-informed neural networksadaptive collocation samplingoptimal controlsafety value functions
Convergence Guarantees of Gradient Descent for Neural Networks via Generalized Lipschitz Smoothness
The paper establishes convergence guarantees for gradient descent in general feedforward neural networks of arbitrary width or depth, without restrictive initialization or dataset assumptions. Key requirements include Lipschitz smoothness, continuity, and linear boundedness of activation functions (e.g., tanh, softplus), and Lipschitz smoothness of the loss (e.g., mean-squared error). By introducing a generalized Lipschitz smoothness condition, the authors show that gradient descent decreases the loss with appropriately small learning rates, provided parameter norms remain bounded. They prove convergence of the minimum squared gradient norm to zero at a rate of $O(1/T^{1/L})$ for $L$-layer networks.
gradient descentlipschitz smoothnessfeedforward networksconvergence guaranteesparameter norms
Probing and steering biology across Boltz-1s trunk-diffusion boundary
The study investigates how biological information propagates between the representational trunk and diffusion module in Boltz-1, an AlphaFold3-class structure predictor. Using linear probes, sparse autoencoders (SAEs), and causal interventions on per-residue activations, the authors find that geometric features (e.g., secondary structure) transfer robustly across the boundary, while sequence chemistry (e.g., amino-acid identity) attenuates. Causal interventions show that decodable directions for helices and coils steer predictions, but a highly predictive beta-strand direction lacks causal influence. Probes outperform SAE features on existing labels but underperform on sparse annotations due to false positives. The release includes trained SAEs, activations, and analysis code.
alphafold3diffusion modulesparse autoencoderlinear probecausal intervention
Dual-Primal Graph VAEs for Noisy Label Aggregation
The authors propose Dual-Primal Graph VAEs, a novel graph variational autoencoder architecture for noisy label aggregation in crowdsourcing tasks. The model employs GAT-based message passing on both the adjacency graph of crowdsourced datasets and its dual, treating ground-truth labels as latent variables to enable unsupervised representation learning without requiring separate classifier training. This approach eliminates the need for pseudo-labels or synthetic data while outperforming state-of-the-art methods on crowdsourcing benchmarks. The framework's generality is demonstrated through graph augmentation incorporating side information from neural network classifiers, substantially boosting test-time classification performance.
graph vaenoisy label aggregationgat-based message passingunsupervised representation learningcrowdsourcing benchmarks
Gaussian Meta-Space Augmentation for Stacking Ensembles in Multimodal IPMN Risk Stratification
The paper introduces cUPMI, a class-conditional Gaussian augmentation method for log-probability meta-features in ensemble stacking, applied to ordinal risk stratification of intraductal papillary mucinous neoplasms (IPMNs). The approach combines multisequence MRI (T1W/T2W), radiomics, and anatomically decomposed pancreas analysis via regularized ensemble stacking, testing L2-logistic regression, random forests (RF), and XGBoost. Results show cUPMI improves RF (+0.015 AUC) and XGBoost (+0.024 AUC) in binary classification, with clearest ordinal benefits for XGBoost on 8-stream radiomics (+0.022 quadratic weighted kappa, QWK). Fold-locked fusion of radiomics and 2.5D CNN streams yields the strongest model (RF stack, QWK 0.595, AUC 0.839), outperforming radiomics, 2.5D ResNet, and 3D DenseNet-121 baselines.
ensemble stackinggaussian augmentationradiomicsmultisequence mriordinal classification
XGBoost "is all you need": the case of forecasting transmitted heat energy in District Heating Systems
The study demonstrates XGBoost's superiority over Long-Short Term Memory (LSTM) in forecasting transmitted heat energy in District Heating Systems (DHS), highlighting scenarios where conventional ML outperforms deep learning. Using a real-world DHS dataset, the authors analyze error distributions and computational efficiency. Results show XGBoost achieves higher accuracy, particularly in data-scarce intervals, while significantly reducing computational costs and carbon footprint compared to LSTM.
xgboostlong-short term memorydistrict heating systemstime series forecastingcarbon footprint
Large-scale AI-Ready Data for Anti-Cancer Drug Response Modeling
This work introduces a large-scale expansion of the IMPROVE benchmark for anticancer drug response prediction (DRP), integrating pharmacogenomic data primarily from PharmacoDB and additional sources. The expanded dataset includes millions of drug response measurements, broader multi-omics coverage, and over 50,000 compounds, significantly enhancing chemical diversity. DRP models trained on the expanded dataset were evaluated using drug-blind, cancer-blind, and disjoint data splits, demonstrating consistent improvements in drug-blind and disjoint settings compared to the original benchmark. These results indicate enhanced generalization to unseen compounds, positioning the dataset as a valuable resource for advancing DRP models in anticancer drug discovery.
drug response predictionpharmacogenomicsmulti-omicschemical diversitygeneralization
Variational Parameter Calibration with Physics-Aware Latent-Space Surrogates
The authors propose a physics-aware neural-network-based latent-space framework for reduced-order forward modeling and variational parameter estimation in parametric dynamical systems. Their autoencoder-based approach yields a differentiable surrogate that maps physical parameters to predicted flow fields via a latent representation, with observable supervision during training to retain parameter-correlated information. Evaluated on two computational-fluid-dynamics benchmarks, the method demonstrates improved case-level separability and temporal organization of latent representations. Experiments with noisy, low-resolution, and partial observations show reduced calibration error and variability compared to standard surrogate models, highlighting the insufficiency of reconstruction accuracy alone for inverse modeling.
latent-space surrogatesvariational parameter estimationreduced-order modelingphysics-aware learningdifferentiable surrogate
AutoGrable: What Is a Good Graph for a Table?
AutoGrable introduces a criterion for constructing graphs from tabular data without requiring trained graph models, based on label-alignment risk and occupancy terms. The method selects columns to form graph nodes and edges by evaluating how well the resulting partition separates rows with different labels while keeping similarly labeled rows together. Experiments demonstrate that AutoGrable efficiently prunes poor graph candidates, recovers label-generating columns in controlled tasks, and outperforms fixed, random, and task-aware constructors under a fixed predictor, while also identifying when graph construction is unnecessary.
graph learninglabel-alignment risk1-wlmessage-passing gnntabular-to-graph
Three Tokens Force Exponential Feature Rank in Nonnegative Kernel Attention
The paper establishes an exponential feature rank requirement for nonnegative kernel attention in three-token contexts. Analyzing Min-IP over Boolean inputs, it proves that any normalized nonnegative kernel-attention head achieving error <1/2 on all three-token sequences necessitates 2^Ω(m) features, even with arbitrary tokenwise values and query-dependent readouts. Dense softmax solves the task with m-dimensional scores, while position-dependent maps and causal queries preserve the bound. For deterministic multihead models with finite-alphabet cross-token channels, a linear transcript lower bound is shown relative to answer count and logarithmic in alphabet size.
kernel attentionmin-ipfeature ranknonnegative kernelstranscript complexity
Click2Poly: A VLM for vector mapping buildings and walls
Click2Poly introduces a human-in-the-loop AI assistant for accelerating the manual editing of building and wall vector layers in geospatial applications. Built as a QGIS plugin, it extends the Florence-2 Vision Language Model (VLM) to enable direct vector layer editing in response to user clicks. This approach reduces labor-intensive quality control and corrections required to meet cartographic standards, particularly in complex cases. The tool demonstrates practical utility in real-world production environments, streamlining the vector mapping process while maintaining accuracy.
vision language modelvector mappingqgis pluginhuman-in-the-loopgeospatial applications
Analysis of Federated Aggregation under Model Poisoning and Backdoor Attacks: A Reconstructed Cross-Dataset and Cross-Architecture Benchmark
The study presents a reconstructed benchmark for evaluating federated aggregation methods under model poisoning and backdoor attacks, considering predictive performance, threat definitions, metric semantics, and execution provenance. A 500-cell evaluation matrix was constructed across five aggregation methods, five datasets, five architectures, and four conditions (clean, sign-flipping, Gaussian, BadNets). Trimmed Mean achieved the highest clean macro-mean accuracy (76.02%) and lowest mean within-task rank (1.70), while Krum performed best under sign-flipping and Gaussian attacks. The audit revealed discrepancies in metric implementation and aggregation weight application, emphasizing the need for precise interpretation within recorded configurations.
federated aggregationmodel poisoningbackdoor attackstrimmed meankrum
Diffusion-Based Data-Driven Assortment Optimization
The paper introduces a diffusion-based framework for model-agnostic assortment optimization, addressing limitations of parametric choice models like MNL. The method represents assortments as binary vectors and performs stochastic search via a learned reverse diffusion process, augmented by a reward-guided mechanism to bias transitions toward high-revenue solutions. Empirical results demonstrate robustness to model misspecification, near-optimal performance in high-dimensional settings, and the ability to generate diverse high-quality assortments. The approach highlights generative modeling's potential for scalable combinatorial optimization in data-driven decision-making.
assortment optimizationdiffusion modelscombinatorial optimizationstochastic searchmodel-agnostic
Unmasking Toxic Mimicry in Medical Offline Reinforcement Learning for ICU Sepsis Management via Counterfactual Clinical Audits
We introduce Counterfactual Clinical Audits (CCA), a framework for detecting Toxic Mimicry in medical offline RL systems by stress-testing agents through physiological perturbations aligned with Surviving Sepsis Campaign guidelines. Evaluating a Medical Decision Transformer (MedDT) and Historical Causal Transformer (HCT-RL) on MIMIC-III ICU sepsis data, CCA reveals MedDT's paradoxical vasopressor reduction during lactate escalation, violating resuscitation protocols, while HCT-RL maintains physiological consistency. Results demonstrate systemic misalignment between statistical fit (MSE, FQE) and clinical safety, advocating CCA as a necessary evaluation standard for medical RL systems.
counterfactual clinical auditstoxic mimicryoffline reinforcement learningmedical decision transformersurviving sepsis campaign
Generative Learning for Quantum Measurement Design
FlowMeas introduces a generative learning approach for quantum measurement design, employing a generative flow network to sample shallow Clifford measurement circuits under resource constraints. The method balances statistical efficiency against implementation resources like circuit depth and entangling-gate count. At zero entangling depth, FlowMatches or outperforms leading product-measurement methods on molecular benchmarks, reducing energy estimation error by up to 27% with one or two entangling gate layers. The learned policy is reusable across related Hamiltonians, accelerating retraining along molecular potential-energy surfaces. Results extend to molecular Hamiltonians with up to 20 qubits and a 54-qubit interacting fermionic model, demonstrating scalability beyond prior benchmarks.
generative learningquantum measurementclifford circuitsentangling gatesmolecular benchmarks
Mechanism Design for Generative Engines: From Exploitation toward Win-Win Outcomes
The paper proposes VCR, a verifiable-content reward mechanism to align incentives between generative engines and content providers, addressing strategic tensions in citation allocation. Modeling the interaction as a repeated Stackelberg game with partial monitoring, the authors identify conditions where citation competition leads to stationary outcomes. Experiments on three benchmarks demonstrate VCR's superiority, achieving a 12.1 percentage point improvement in Net defense-utility score over baselines while enabling win-win outcomes under empirical equivalence criteria.
generative engine optimizationstackelberg gameverifiable-content rewardscitation warsmechanism design
Reoptimization Algorithms for Contextual Bandits with Knapsack Constraints
We propose a novel reoptimization-based algorithm for Contextual Bandits with Knapsack (CBwK) that achieves $O(\frac{(\ln T)^3}{T})$ average regret, improving upon the $O(\frac{1}{\sqrt{T}})$ bound in prior dynamic-pricing literature. Our method extends the Upper-Confidence-Bound (UCB) family by incorporating re-optimization techniques to jointly learn the mean reward function and make online assignments of products to customers. The CBwK problem involves finite customer types, products, and resources, where each product consumes resources with finite capacity, and rewards are linear functions of customer-product features plus noise.
contextual banditsknapsack constraintsreoptimizationupper-confidence-boundregret bound
Towards an approach to multivariate outlier detection for District Heating System data
The study evaluates multivariate outlier detection methods for District Heating System data, focusing on transmitted heat energy and ambient temperature. Methods tested include Z-score (benchmark), Mahalanobis distances, Principal Component Analysis (PCA), Isolation Forest, and Hotelling's T-squared test, aiming to identify irregular plant operation and reduce gas consumption. Domain-specific adjustments, such as excluding zero transmitted energy timepoints, were incorporated. PCA, Isolation Forest, and Hotelling's method yielded relevant results, leading to an ensemble approach based on consensus among these three methods for final outlier detection.
multivariate outlier detectiondistrict heating systemprincipal component analysisisolation foresthotelling's t-squared test
PAIR: Pairwise-Aware Inclusion Reweighting for Adaptive Rollout Allocation in RLVR
PAIR (Pairwise-Aware Inclusion Reweighting) introduces a pairwise-adaptive rollout allocator for RLVR, addressing statistical mismatch in gradient estimation by modeling rollout prefixes as vertices and pair-gradient terms as edges in a contrast graph. The method combines a prefix-only predictor for correctness and token cost, convex design for continuation probabilities, and inverse-weighted edge terms by joint inclusion probability. Evaluated on Qwen3-1.7B/4B, PAIR improves accuracy by +1.2 and +1.4 over pointwise allocators while reducing token generation by 51-52% versus full-group GRPO, with unbiased gradient estimation confirmed via frozen-population audit.
reinforcement learningadaptive allocationgradient estimationcontrast graphrollout prefixes
Adaptation of Generalist Robot Policies with Minimal Data
The paper introduces MiDAS, a method for adapting generalist robot policies to new tasks with minimal human data. MiDAS combines behavior cloning on single/few demonstrations with value-based online RL on a residual policy parameterization, enabling autonomous improvement from sparse initial guidance. Evaluated on LIBERO, RoboCasa, and a bimanual YAM platform, MiDAS achieves robust task performance from one demonstration, outperforming baselines and generalizing beyond demonstrated conditions. Notably, it improves a fragile policy to reliability over ~6 hours of online interaction, marking the first reliable single-demonstration adaptation for robot policies.
minimal-data adaptationresidual policybehavior cloningonline reinforcement learninggeneralist robot policies
Lifecycle-Optimal Tokenization: Vocabulary Size as a Deployment-Regime-Dependent Infrastructure Parameter
The study demonstrates that tokenizer vocabulary size in LLMs should be treated as a deployment-dependent parameter rather than a fixed design choice, with lifecycle cost minimization as the optimization objective. By formalizing total cost as $C_{lifecycle}(V) = C_{train}(V) + λ\cdot C_{infer}(V, B)$ and conducting experiments across GPU architectures (A10G, A100) and batch sizes, the authors show inference-optimal vocabulary shifts 16x (32k to 524k) with batch size due to unembedding matrix read amortization. Results indicate quality (BPB) is optimized at 65k for 1.3-2.3B models, with lifecycle-optimal vocabularies diverging up to 16x from training-optimal without quality penalty (<2% BPB spread), recommending 32k for on-device and 131-262k for datacenter deployments.
tokenizervocabulary sizelifecycle costunembedding matrixbatch size
Market-Information-Aware Gated-LoRA of Foundation Models for Transferable Day-Ahead Electricity Price Forecasting
The paper introduces a market-information-aware adaptation framework for transferable day-ahead electricity price forecasting, leveraging the Chronos-2 time-series foundation model. The method constructs a multi-source market information interface aligning price context with pre-clearing variables and trains a gated low-rank adapter (LoRA) that updates approximately 1% of model parameters without target-market labels. The gate scales the frozen source adapter based on reserve-tightness and operating-state signals. Evaluated via a leave-one-market-out protocol across four Chinese provincial day-ahead spot markets, the framework reduces average MAE/RMSE by 6.24%/7.99% compared to market-information-aware zero-shot Chronos-2 and by 3.05%/3.52% relative to vanilla Source-LoRA.
low-rank adapterelectricity price forecastingchronos-2multi-source market informationtransfer learning
ODE-Based Transformer Decoders for Iterative Sign Language Translation
The paper proposes an ODE-inspired Transformer decoder for iterative sign language translation, replacing residual refinement updates with higher-order Runge-Kutta methods (RK-2, RK-4) to enhance representation dynamics without increasing parameters. This approach improves translation accuracy while maintaining parameter efficiency, achieving 22.96 BLEU-4 on PHOENIX-2014-T and 19.34 BLEU-4 on CSL-Daily, outperforming the IPSLT baseline with fewer decoder layers and iterations.
transformer decoderordinary differential equationrunge-kutta methodssign language translationiterative refinement
Weightless Fine-Tuning: Personalizing LLMs via Logit-Space Transport
Weightless Fine-Tuning (WFT) is proposed as a training-free decoding-time method to personalize LLMs without weight updates, addressing the prohibitive costs of supervised fine-tuning (SFT) in multi-author settings. WFT approximates SFT's distributional effect by computing supervised residuals on training sequences and transporting them to prompts via a cross-prefix transport operator derived from dropout-induced cross-covariance, replacing gradient updates with logit-space corrections. On LaMP benchmarks, WFT achieves the best average performance, matches or exceeds SFT on individual tasks, and uses <7% of SFT's computation. Logit shifts show 0.875 cosine similarity with SFT over 95% of next-token probability mass.
weightless fine-tuninglogit-space transportcross-prefix transport operatordropout-induced cross-covariancesupervised residuals
Better, Faster, Stronger: Programmatic Skill Learning Best Reduces Agent Cost
The paper proposes programmatic skill learning as the most cost-effective method for adapting LLM agents to novel domains, arguing that deterministic program execution reduces trial-and-error costs and degenerate behaviors. SpeedRunner, a coding agent, analyzes past trajectories to refactor skills without replay or validation, demonstrating robust performance across three embodied environments. Results show SpeedRunner achieves state-of-the-art cost reduction and learning efficiency while maintaining robustness to distribution shifts and environmental randomness.
programmatic skill learningllm agentscost reductionembodied environmentstrajectory analysis
Long-Horizon Forecasting of Complete Financial Statements with Forma
ProForma-20Q introduces a benchmark for long-horizon forecasting of complete financial statements, predicting 78 line items 1-20 quarters ahead from anonymized firm data and industry codes. Forma, a transformer model, processes statements as (account, quarter, value) tuples and optimizes a masked-tuple Gaussian likelihood. It outperforms classical machine learning, gradient boosting, zero-shot time-series models, and large language models, with increasing accuracy at longer horizons crucial for valuation. Forma's forecasts maintain accounting identities and support scenario analysis without retraining, demonstrating that pinned revenue paths enhance statement precision.
financial statementstransformer modelmasked-tuple gaussian likelihoodaccounting identitiesscenario analysis
Spectral graph clustering with inhomogeneous latent geometry
The authors present DBSPEC, a density-based spectral clustering algorithm that recovers community structures in graphs with inhomogeneous latent geometries. The method analyzes the spectral properties of adjacency matrices through a limiting integral operator, focusing on eigenvectors deeper in the spectrum rather than the leading ones dominated by latent geometry. DBSPEC requires only approximate localization of informative eigenvalues and demonstrates robustness to poor eigenvalue separation. Theoretical predictions for informative eigenvalue locations align with empirical observations in real-world experiments, overcoming limitations of prior work restricted to homogeneous toroidal models.
spectral clusteringlatent geometryadjacency matrixintegral operatoreigenvalue separation
Dueling Deep Q-Learning for Intrusion Detection
The study introduces a dueling Deep Q-learning model for intrusion detection, addressing limitations of supervised methods in adapting to novel attack types. The architecture employs separate value and advantage streams, enhancing learning efficiency and stability. Evaluated on the CIC-IDS2018 dataset, the model achieves 99.68% average accuracy across diverse attack classes (DDoS, botnets, brute-force). SHAP-based Explainable AI (XAI) provides interpretability for predictions.
dueling deep q-learningintrusion detectioncic-ids2018shapexplainable ai
Uncertainty-Aware Compositional Localization and Placement Assessment of Catheters and Tubes in Chest X-Rays
The paper introduces UCompCXR, an uncertainty-aware compositional framework for localizing and assessing catheter/tube placement in chest X-rays. The method detects local fragments, clusters them into device instances via graph algorithms, fuses tip predictions through Gaussian estimation, and classifies placement per device. Evaluated on RANZCR CLiP (30,083 images, 5-fold CV), it detects 26% more devices than a MobileNetV3 baseline with 75% fewer false positives, achieves well-calibrated tip uncertainty (0.948 coverage at 95% CI), and reduces catastrophic failures, while maintaining deployability (2.27M parameters).
compositional localizationgraph-based clusteringprecision-weighted gaussian estimationchest x-ray analysisuncertainty calibration
Benchmarking Cyberattack Detection in Electric Vehicle Charging Infrastructure with Benign User Updates
The paper introduces a leakage-controlled benchmark for cyberattack detection in electric vehicle charging infrastructure, addressing the challenge of distinguishing malicious manipulations from benign user updates. The method employs a Dual-Branch Masked-Autoencoder (Masked-AE) Transition Boost model, combining masked reconstruction with radial-basis-function one-class support and shrinkage covariance distance to evaluate request normality and transition similarity. Evaluated on six attack types and their coordinated variants, the model achieves robust validation performance while preserving legitimate user updates, outperforming 22 baseline models in source-grouped five-fold cross-validation.
cyberattack detectionelectric vehicle chargingmasked-autoencodertransition boostone-class support
What Iterated Self-Feeding Probes of Language Models Measure, and a test that separates the construction from the model
The study introduces a method to distinguish between properties inherent to probing constructions versus those attributable to language models when using self-feeding probes like self-consistency or iterated refinement. By employing a ring of token cells resampled via Glauber dynamics and advancing two rings under common random numbers, the method isolates kinematic effects (e.g., damage light cone, token-space Lyapunov exponent λ_ca scaling) from model-specific behaviors (e.g., λ_ca zero-crossing during training). Validation includes reproducing a Domany-Kinzel damage field exactly and identifying four retracted measurements. The approach requires fixing either the construction or model to isolate effects.
glauber dynamicslyapunov exponentself-feeding probesdamage spreadingtoken-space
📰 Industry Media (7)
Flock is tightening its rules in response to a growing surveillance backlash
Flock, a police-tech firm operating 120,000 license plate readers (LPRs), implements mandatory case-number entry and automated auditing to curb officer misuse, following reports of stalking and unauthorized searches. The system flags suspicious activity but lacks independent validation. Data retention is reduced from 30 to 7 days, and agencies can restrict cross-departmental searches by purpose (e.g., kidnapping vs. immigration enforcement). Despite these measures, critics argue safeguards rely on self-reporting and demand third-party oversight. Backlash stems from surveillance scale, with 30+ cities reportedly dropping contracts, though Flock attributes losses to misinformation about facial recognition or data sales.
license plate readersautomated auditingdata retentionsurveillance backlashpolice-tech
How kids feel about AI, in their own words
This study explores adolescent perspectives on AI through qualitative interviews with individuals aged 10-18, revealing nuanced attitudes and usage patterns. Participants expressed concerns about AI's societal impact, environmental costs, and potential to undermine creativity, while acknowledging its utility in education, coding, and personal projects. Methods included direct interviews and a 'Cringe-o-meter' rating system for AI applications. Results indicate that 57% of teens use chatbots for information retrieval, 54% for schoolwork, and 47% for entertainment, with only 12% seeking emotional support. Despite widespread adoption, many participants emphasized maintaining agency over AI systems and expressed skepticism about their current capabilities.
chatbotsqualitative interviewsenvironmental impactcreative agencyinformation retrieval
Google AI Just Released Gemini 3.7 Flash: A Coding and Agent Model at $0.75/1M Input Tokens
Google AI released Gemini 3.7 Flash, a multimodal coding and agent model optimized for software engineering, document-heavy workflows, and web development. The model refines Gemini 3.6 Flash with algorithmic improvements, supporting a 1M-token context window, 64K output tokens, and customizable configurations for cost-latency tradeoffs. It achieves 43.6% on FrontierCode 1.1, 65.3% on DeepSWE v1.1, and 1588 Elo on WebDev Arena, outperforming its predecessor. Priced at $0.75/1M input tokens and $3.75/1M output tokens until December 2026, it offers a cost-effective alternative to Claude Sonnet 5 and GPT-5.6 Terra. Deployment is API and enterprise-only, with no open weights or self-hosting options.
multimodalcontext windowalgorithmic improvementscost-latency tradeoffsenterprise-only
Liquid AI Releases LFM2.5-VL-3B: A 3B Vision-Language Model That Reads Screens, Grounds Objects, and Calls Tools On-Device
Liquid AI introduced LFM2.5-VL-3B, a 3.1B-parameter vision-language model optimized for on-device deployment, featuring screen understanding, object grounding, and tool-calling capabilities. The model combines a 2.6B-parameter language backbone (LFM2.5-2.6B) with a 400M SigLIP2 NaFlex vision encoder, trained on 34T tokens and supporting 16 languages with a 32K context length. It achieves a 69.4 average across 28 vision benchmarks, matching 4.7B-class models like InternVL-3.5-4B, while excelling in ScreenSpot-v2 (80.7) and RefCOCO-avg (87.9). With 228 tokens/s decoding on Apple M5 Max and ~3GB memory footprint, it ships in multiple runtime formats under a revenue-capped open license.
vision-language modelon-device deploymentsiglip2 naflexfunction callingobject grounding
Dyna Robotics Introduces Dyna-2: A World-Action Model Pre-Trained on 1 Million Hours of Human Video
Dyna Robotics introduces Dyna-2, a world-action model pre-trained on 1 million hours of egocentric human video, demonstrating scalable transfer to robot manipulation tasks. The model employs a video-diffusion backbone with separate DiT stacks for video and action, trained via flow matching with joint video-action denoising. Scaling laws show monotonic improvement: held-out MSE = 0.0691·D^-0.0184 (R²=0.919), accuracy@0.5 = 0.357·D^+0.0203 (R²=0.865). Zero-shot transfer to unseen robot data yields MSE = 0.306·D^-0.0713 (R²=0.884), with post-training on 14 tasks achieving up to 53% mean normalized score. Video prediction drives cross-embodiment generalization.
world-action modelvideo-diffusionflow matchingscaling lawscross-embodiment
SpaceXAI Releases Grok 4.6: A 500K-Context Frontier Model Tuned for Long-Running Agents, Coding, and Knowledge Work
SpaceXAI released Grok 4.6, a 500K-context frontier model optimized for long-running agents, coding, and knowledge work. The upgrade retains Grok 4.5's base architecture but improves through extended supplemental training, regenerated supervised fine-tuning trajectories, and RL in agentic environments. Key enhancements include a new xhigh reasoning-effort level and improved self-verification on long trajectories. Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index (tying GPT-5.6 Sol Max) but trails in coding benchmarks like DeepSWE v1.1 (65.9%). The model is API-accessible with tiered pricing but lacks open weights or self-hosting options.
500k-contextagentic environmentsreasoning-effortsupervised fine-tuningself-verification
Okta targets AI agent token costs with MCP scoping
Okta proposes identity-scoped Model Context Protocol (MCP) tool filtering to reduce token costs in AI agent workflows. The method filters tools before prompt construction using OAuth scopes and user roles, removing unauthorized tools from agent prompts. Internal modeling showed >90% reduction in visible tools and proportional token cost savings, though absolute figures were not provided. The approach also limits attack surfaces by enforcing least-privilege access at the tool level.
model context protocoltoken costoauth scopesleast-privilege accessprompt overhead
Generated automatically at 2026-08-13 20:20 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.