Daily Digest — 2026-07-30
279 items · 2 research labs, 270 arxiv papers, 7 industry media
🏛️ Research Labs (2)
Accelerating scientific discovery with ChatGPT for Academic Researchers
OpenAI launches ChatGPT for Academic Researchers, providing 100,000 scientists free access to frontier models like GPT‑5.6 Sol Pro (83% on FrontierMath Tier 4, 31.5% on GeneBench Pro) by 2027. The program includes workspace collaboration tools, domain-specific skills (genomics, protein modeling), and Codex integration for tasks ranging from hypothesis testing to grant writing. Early adopters show 2x higher engagement with complex tasks (>4 hours) compared to peers. The $250M initiative includes NextGenAI and DoE partnerships, prioritizing privacy and institutional coordination via ChatGPT Edu.
gpt-5.6 sol profrontiermath tier 4genebench procodex integrationin-context learning
How GPT-5.6 fuses frontier intelligence with frontier efficiency
OpenAI introduces the GPT-5.6 model family, optimizing both intelligence and efficiency across tasks. The flagship GPT-5.6 Sol outperforms Claude Fable 5 on the Artificial Analysis Coding Agent Index at <50% cost, while Terra matches GPT-5.5 at half the price, and Luna offers 80% cost reduction. Efficiency gains stem from stack-wide optimizations: inference improvements (load balancing, speculative decoding, KV-cache management) and agentic harness enhancements (context bloat reduction, prompt caching). GPT-5.6 Sol autonomously optimized kernel operations and draft models, achieving 20% lower serving costs and 15% faster token generation. These compound improvements enable frontier performance at reduced compute costs.
speculative decodingkv-cacheagentic harnesskernel optimizationload balancing
📜 arXiv Papers (270)
Pass the Baton: Trajectory-Relayed On-Policy Distillation
The paper introduces Relay On-Policy Distillation (Relay-OPD), a method addressing prefix failure in on-policy distillation by leveraging teacher-student continuation asymmetry. Relay-OPD constructs relay trajectories where the teacher intervenes at detected trigger points to redirect reasoning, then resumes student training on the corrected trajectory. Evaluated with Qwen3-4B-Instruct-2507 teacher and Qwen3-0.6B/1.7B-Non-Thinking students on eight mathematical reasoning benchmarks, Relay-OPD outperforms standard OPD by +5.73% and FastOPD by +1.49% (1.7B model), with 50% reduced training trajectory length.
on-policy distillationprefix failurerelay trajectoriesteacher-student asymmetrymathematical reasoning
$π\mathbf{R}^2$: Reactive Real-time Flow Policies
The paper introduces $π\mathbf{R}^2$, a method to enhance the reactivity and real-time performance of generalist manipulation policies based on large pretrained backbones. By splitting conditioning into fast (proprioception) and slow (vision-language) channels and employing a latency-adaptive flow schedule, the policy reacts to fresh proprioceptive input while tolerating stale visual features. Evaluated on GR00T-N1.7 with an xArm6+XHand platform, $π\mathbf{R}^2$ achieves ~25Hz replanning (4× faster than baseline) and improves success rates by up to 23% (simulation) and 30% (real-world).
reactive policiesaction-chunkingdiffusion forcinglatency-adaptiveproprioception
Desktop-Delta Bench: Do Computer-Use Models Understand Desktop GUI Transitions?
The paper introduces Desktop-Delta Bench (DDB), a diagnostic benchmark for evaluating computer-use agents' (CUAs) ability to reconstruct causal GUI transitions amid asynchronous observation challenges. DDB comprises 2,013 human-verified instances from multi-app Linux trajectories, targeting state verification, source tracking, and context-aware control through 463 3-frame temporal-ordering tasks (including 105 cross-trajectory decoys) and 1,550 before-after action pairs. Evaluation of 8 model families reveals unsaturated performance: best non-decoy/decoy exact-match rates are 65.1%/65.7%, with task context improving decoy identification by 6.9pp but reducing non-decoy accuracy by 2.2pp. Action-family inference proves harder than localization (click F1: 0.96 vs. drag 0.76).
computer-use agentsgui transitionstemporal-orderingaction localizationasynchronous observation
Falling Behind Drives Unsafe Development in an Idealised AI Race Experiment
The study investigates how competitive pressure influences unsafe AI development through a framed behavioral experiment with paired participants choosing between Safe and Unsafe development under varying risk levels (10%, 60%, 90%). Results show Unsafe behavior is driven by strategic race dynamics—participants respond to opponents' choices, with falling behind increasing Unsafe play and leading reducing it. A reduced evolutionary model with four strategies replicates these effects, highlighting conditional Unsafe behavior as competitively favored. Findings suggest policy should address competitive pressure rather than individual risk preferences alone.
behavioral experimentcompetitive dynamicsrisk preferencesevolutionary modelai safety
CHARM: A Multimodal Graph Foundation Model with Hierarchical Context Modeling for Zero-Shot Transfer
CHARM introduces a multimodal graph foundation model for zero-shot transfer by hierarchically modeling graph contexts to capture cross-modal relations and shared high-level concepts. The method replaces raw nodes with hierarchical contexts encoding multimodal semantics, processed by a modality-aware graph context encoder into graph tokens for LLM integration. Experiments demonstrate consistent improvements on zero-shot multimodal graph tasks without target-domain fine-tuning.
graph foundation modelszero-shot transfermultimodal graphshierarchical context modelingmodality-aware encoder
MDTransformer: A Hardware-Software Co-Design of Mode-Division Photonic Transformer Accelerator with Inverse-Designed Coherent Crossbar
MDTransformer introduces a hardware-software co-designed photonic transformer accelerator (PTA) leveraging mode-division optical dataflow for efficient Transformer inference. The method employs inverse-designed multi-mode couplers, crossings, and Mach-Zehnder IQ modulators in a compact mode-division photonic tensor core (MPTC), enabling four-fold parallelism per waveguide via spatial-mode interference (TE0-TE3) and complex-valued arithmetic. Experimental results demonstrate 40.4% area reduction, 63.6% power saving, 40.6% energy saving, and comparable latency versus state-of-the-art PTAs on DeiT and BERT variants, with sub-4-bit precision and <-30 dB inter-modal crosstalk.
photonic transformer acceleratormode-division multiplexinginverse-designed couplersspatial-mode interferencecomplex-valued arithmetic
Pictura: Perspective-View Self-Play at Scale for Driving
Pictura introduces perspective-view self-play as a scalable training regime for driving policies, eliminating the representation gap caused by privileged vectorized observations. The GPU-accelerated multi-agent simulator renders egocentric views at each step, achieving up to 500K agent-steps/s (2M images/s) on a single H100. Alberti, trained via self-play with PPO over 50B agent steps (~35M km), becomes the first large-scale driving policy trained directly from perspective images without privileged inputs. It matches the performance of privileged vectorized counterparts and demonstrates superior zero-shot transfer to Waymo Open Motion Dataset layouts rendered in Pictura.
self-playegocentric viewpporepresentation gapagent-steps
Empirical Evaluation of Out-Of-Distribution Performance of Tabular Foundation Models
This study evaluates the Out-Of-Distribution (OOD) robustness of nine Tabular Foundation Models (TFMs)—including TabPFN variants, TabICL, Mitra, LimiX, and TabFM—across three real-world datasets (HELOC, Voting, Childhood Lead) exhibiting label, socioeconomic, and geographic shifts. Results demonstrate systematic OOD performance degradation (shift gaps: 0.003–0.060) regardless of pre-training strategy, aligning with trends observed in classical tabular models. TFMs also exhibit a scalability gap due to high computational demands. The work extends OOD benchmarks for tabular data, informing deployment in high-stakes domains.
tabular foundation modelsout-of-distribution robustnessdistribution shiftpre-training strategiesscalability gap
Does Runtime Topology Context Improve LLM-Generated Kubernetes Security Patches?
KuTIE (Kubernetes Topology Intelligence Engine) demonstrates that incorporating live cluster topology context—derived from Istio call edges, Trivy KSPM findings, and service-account bindings—significantly improves the correctness of LLM-generated Kubernetes security patches. The system conditions patch generation on runtime dependencies, addressing a critical limitation of isolated KSPM-based prompting. Evaluated on VulnCare, a 36-deployment healthcare cluster with 31 injectable findings across seven dependency classes, KuTIE raises topology-dependent patch correctness from 11.1% to 78.0% (Δ=0.669), with consistent gains across six of seven classes (e.g., credential and network-policy: Δ=0.95). No improvement occurs for topology-independent controls (Δ=0.0), confirming the specificity of the effect.
kubernetesllm-generated patchestopology contextkspm findingsservice-account bindings
MemLens: A Value-Aware Memory Management System with Interactive Analytics for LLM-based Agents
MemLens introduces a value-aware memory management system for LLM-based agents, addressing inefficiencies in coarse-grained approaches by treating memory records as first-class objects. The system features an interactive analytics dashboard for end-to-end memory lifecycle management, including Shapley-style evaluation, value-aware storage, and memory-assisted response generation. Evaluated via a study-copilot application, MemLens demonstrates improvements in response quality, retrieval latency, and token efficiency while enabling interpretable memory inspection and hierarchical visualization.
memory managementllm-based agentsshapley evaluationinteractive analyticsvalue-aware storage
Reinforcement Learning for Code Optimization
The paper introduces a three-stage reinforcement learning (RL) method for code optimization, addressing challenges in measurement noise and reward sparsity when execution time is incorporated into the reward signal. The approach involves (1) constructing DMC-Optim, a benchmark with calibrated timing sandboxes, (2) designing a composite reward function combining correctness and speed, and (3) adapting GRPO for noisy, sparse rewards. Evaluations on Qwen 2.5 7B and CWM 32B show top-50% pass@1 improvements from 18.0% to 31.3% and 30.7% to 50.4%, respectively, with CWM 32B achieving 125% relative improvement at top-30% and 83% win rate in median-sample speed comparisons on LCB.
reinforcement learningcode optimizationreward sparsitygrpotiming sandbox
Knowledge-Guided Multimodal Reasoning over Interacting Streams for Video-Level Ambivalence and Hesitancy Recognition
PRISM-AH introduces a knowledge-guided multimodal framework for video-level ambivalence and hesitancy (A/H) recognition by modeling cross-modal dissonance across facial, vocal, linguistic, and bodily streams. The method aligns frozen encoders for vision, audio, and text into short time windows, employs a lightweight streaming model to score dissonance and predict hesitation signals, and fuses late-stage reasoning from a large language model guided by expert cues. On a public test set of 525 videos, PRISM-AH achieves a macro F1 of 0.6133, significantly outperforming the zero-shot baseline (0.2827), with validated reasoning transfer from validation to test partitions.
multimodal reasoningcross-modal dissonancehesitation recognitionstreaming modelexpert-guided llm
Detecting Knowledge Inconsistencies Across Text, Tables, and Knowledge Graphs
We introduce extsc{Kontrast}, a framework for detecting modality-level knowledge inconsistencies across text, tables, and knowledge graphs (KGs). The method employs Text-to-SPARQL translation and LLM reasoning to compare table-based answers with KG evidence, categorizing inconsistencies into granularity differences, direct conflicts, temporal changes, and KG incompleteness. Experiments on Table-QA datasets reveal that cross-modal inconsistencies are prevalent (exact quantification not provided), exposing true knowledge conflicts, missing KG structure, and temporal mismatches, though limited by Text-to-SPARQL errors. The analysis demonstrates complementary correction potential across modalities, establishing a benchmark for knowledge auditing.
knowledge graphstext-to-sparqlmodality-level inconsistencytable-qaknowledge auditing
Large Language Model for Operations Research Formulation Selection in Multi-Warehouse Inventory Allocation
The paper proposes a solver-guided large language model (LLM) framework for instance-wise operations research (OR) formulation selection in multi-warehouse inventory allocation. The method constructs expert-conditioned supervised fine-tuning records, uses MIP solver evaluation to generate preferences, and applies group relative policy optimization (GRPO) for reward assignment. Experiments on JD.com data show GRPO improves Hit Ratio@1 from 21.45% to 50.42% and achieves a 12.57 percentage-point accuracy gain over the baseline, reducing the gap to the ex-post oracle to 4.85 percentage points.
mixed-integer programmingformulation selectionrelative policy optimizationinventory allocationlarge language model
MODUS: Decoder-Only Any-to-Any Modeling of Diverse Modalities
MODUS introduces a decoder-only any-to-any multimodal model that symmetrically processes arbitrary input-output modality combinations without modality-specific components. The method leverages a unified decoder-only architecture, enabling cross-modal generation, chained intermediate outputs, and self-verification via generated modalities. Evaluations show MODUS matches or outperforms specialized and multitask baselines across diverse benchmarks while maintaining a single-model design. The implementation is open-sourced.
decoder-onlymultimodal modelingcross-modal generationself-verificationany-to-any
A Cost-Effective Multimodal LLM Reasoning Framework for Question Answering over Irregular Clinical Time Series
The paper introduces ClinPRISM, a multimodal LLM framework for question answering over irregular clinical time series (ICTS), addressing sparsity and irregular sampling. The method combines an irregularity-aware multi-scale encoder, temporal evidence distiller for token compression, and progressive alignment strategy for LLM embedding space integration. Evaluated on 41,000 instruction-tuning instances across 11 tasks, ClinPRISM achieves SOTA performance using a 4B-parameter LLM, 16 time-series tokens, and 0.15s average inference latency.
multimodal llmirregular time seriesclinical question answeringtemporal evidence distillerprogressive alignment
Evaluating Multi-Turn Multimodal Diagnostic Reasoning on Challenging Real-World Clinical Cases
We introduce ClinMM-Bench, the largest multi-turn multimodal clinical diagnostic evaluation benchmark, containing 1,089 real-world cases and 3,760 medical images across eight specialties, addressing limitations of single-turn evaluations. Using a two-level framework assessing diagnostic accuracy and reasoning quality, we evaluated 15 multimodal large language models (MLLMs). While proprietary models achieved highest diagnostic accuracy, complete correctness remained limited across all models. Current models demonstrated plausible diagnostic direction identification but exhibited limitations in reliable reasoning generation. Error analysis revealed five failure modes: information synthesis failure, knowledge mapping error, perception error, premature closure, and visual hallucination.
multimodal large language modelsdiagnostic accuracyclinical reasoningvisual hallucinationknowledge mapping
Face De-Identification: A Domain-Centric Survey from Capture to Processing
This survey provides the first unified analysis of face de-identification (De-ID) techniques across the full data acquisition pipeline, covering physical-domain (wearable accessories), sensor-level (privacy-aware capture), and digital-domain (post-processing) methods. It systematically evaluates current approaches, technical challenges, and evaluation protocols, while advocating for standardized benchmarks. The authors identify open problems and emerging directions, maintaining a curated resource with datasets and code at https://github.com/CV-AC/Awesome-FaceDe-ID.
face de-identificationprivacy-preservingsensor-levelpost-processingwearable accessories
dtControl2+$\varepsilon$: Trading Optimality for Explainability in MDPs via Decision Trees
The work introduces dtControl2+$ε$, an extension of the dtControl2 tool for generating explainable controllers in Markov decision processes (MDPs) that trades optimality for simplicity. Given a user-defined imprecision tolerance $ε ≥ 0$, the method constructs smaller decision trees by omitting less critical cases while preserving $ε$-optimality guarantees. This enables tunable simplification of policy representations without compromising correctness. The tool produces decision trees orders of magnitude smaller than state-of-the-art approaches, improving interpretability for complex systems with many corner cases.
markov decision processesdecision treesexplainable aioptimal controlpolicy approximation
Evaluating VLMs for Autonomous Agent-Driven Geometry Clipping Detection in Video Game QA
The study evaluates Vision-Language Models (VLMs) for detecting geometry clipping anomalies in automated game QA pipelines using an agent-collected dataset with frame-level labels. Six VLMs (Gemini, GPT, Qwen, Gemma, Llama, Ministral) are benchmarked under zero-shot prompting with four variants, revealing sensitivity to visual ambiguities like near-contact geometry. Gemini-3.1-Flash achieves the highest accuracy and prompt robustness, while open-source models show significant precision-recall variance. The results suggest VLMs are currently better suited as high-recall filters in multi-stage QA systems than standalone detectors.
vision-language modelsgeometry clippingzero-shot promptinganomaly detectiongame qa
Penelope: Localized Latent Recurrence for Efficient Structured Reasoning
Penelope introduces a localized latent-recurrence framework for efficient structured reasoning in decoder-only Transformers, avoiding the computational costs of parameter scaling or long chain-of-thought sequences. The method constructs a problem-conditioned boundary memory from an initial decoder prefix, then refines it through time-modulated GRU dynamics and recurrent readout states before answer generation, enabled by a CoT-to-latent curriculum. On structured-reasoning benchmarks, Penelope matches latent-reasoning model accuracy at selected latent budgets while reducing inference latency, demonstrating that latent refinement can be localized to narrow decoder intervals without full-decoder repetition or lengthy intermediate traces.
latent reasoningdecoder-only transformersgru dynamicschain-of-thoughtinference latency
Toward Standardized Cross-Vendor Agent Tool Trust Management in Autonomous Networks
AgentToolMO introduces a standardized 3GPP NRM information model for cross-vendor agent tool trust management in autonomous networks (Levels 4-5). The model features a formal trust state machine with graduated enforcement, damped cascade propagation with bounded convergence, cross-vendor trust notifications via existing MnS interfaces, and retroactive impact assessment through NRM dependency graph traversal. Evaluations in multi-vendor topologies demonstrate that standardized notifications reduce blast radius from hours-scale propagation to near-real-time containment, with sub-linear notification scaling and guaranteed cascade convergence. The framework integrates with existing 3GPP infrastructure, leveraging current protocols for trustworthy multi-vendor network management.
autonomous networkstrust state machinecascade propagationmns interfacesnrm dependency graph
SAM3D-Guided Object-Centric Representation Alignment for Vision-Language-Action Models
The paper introduces an object-centric 3D representation alignment framework for Vision-Language-Action (VLA) models, addressing limitations in fine-grained 3D understanding under occlusion and spatial variation. The method leverages SAM3D as a frozen 3D teacher to extract dense object-level 3D representations during training, aligning them with intermediate visual features of the base model $π_0$ while maintaining RGB-language-to-action inference at test time. Experiments demonstrate consistent improvements, achieving 99.1% success on LIBERO and an average trajectory length of 4.11 on CALVIN, with notable gains in long-horizon manipulation tasks.
vision-language-action models3d representation alignmentobject-centric learningsam3drobot manipulation
AnnoBench: A Benchmark for Visualization Annotation Generation
We introduce AnnoBench, a benchmark for evaluating visualization annotation generation that addresses visual, semantic, and stylistic constraints. The benchmark pairs professional visualizations with annotation tasks across four representation formats, five chart description conditions, and two prompt specificity levels. Evaluation employs VLM-as-a-judge methodology, aligning model outputs with human assessment. Experiments explore the impact of input representation, semantic context, prompt specificity, and model selection on annotation quality. Results provide a foundation for advancing annotation automation, tooling, and visualization-generation pipelines.
annotation generationvisualizationvlm-as-a-judgesemantic contextprompt specificity
Minimizing Targeted Activations: Input-Only Suppression of Evaluation-Awareness Latents in Large Language Models
The paper introduces an input-only method for suppressing specific internal latents in large language models (Llama-3.2-3B and Llama-3.1-8B) without inference-time access, targeting 'evaluation-awareness' latents that could compromise safety evaluations. The approach adapts Fluent Dreaming/EPO with a negated feature term, combining GCG-style token optimization and a fluency regularizer. Results show robust suppression (z≈−7) of latents, including a causally validated SAE feature, but reveal limitations: a random direction is equally suppressible, and prefix optimization fails to reduce behavioral eval judgments. The study questions the behavioral controllability of activation-readability and identifies non-causal MLP neurons.
activation steeringlatent suppressionfluent dreamingevaluation-awarenesssae feature
Interactive Reward Agent: GUI Task Evaluation via Environment-State Verification
The Interactive Reward Agent (IRA) introduces a propose-then-verify framework for GUI task evaluation by combining visible interface evidence with environment-state verification. IRA proposes task completion conditions and verifies them using system, application, and GUI tools, addressing the challenge of accessing environment states beyond execution trajectory screenshots. Evaluated on GUI-RewardBench, a benchmark of 321 GUI task trajectories across 10 Ubuntu desktop applications, IRA achieves 86.9% accuracy, surpassing existing evaluators. When applied to reinforcement learning for GUI agents, IRA enables a 34.0% success rate on OSWorld, demonstrating its effectiveness as a reward signal for training.
gui task evaluationinteractive reward agentenvironment-state verificationgui-rewardbenchosworld
Messier: A High-Resolution Corpus for Cross-Benchmark Agent Evaluation
The paper introduces Messier, a unified corpus of 957,253 standardized records spanning 30 benchmarks, 714 agents, and 11,891 tasks, addressing fragmentation in AI agent evaluation. It consolidates public benchmark scores and supplements them with five-agent runs across six underrepresented domains, including legal benchmarks. Standardized records include model, scaffold, task, and verifier metadata, enabling cross-benchmark analysis. Results show uneven progress across benchmark types (e.g., 'function calling' saturated, 'enterprise workflows' challenging) and reveal that strict all-pass verifier aggregation can distort rankings. Derived capability scales align with Epoch's Evaluation Capability Index (Spearman ρ = 0.81) and support domain-specific analysis.
agent evaluationbenchmark corpuscapability scalingverifier aggregationcross-domain analysis
Distributing Security Controls Through Harness Engineering
The paper introduces SHarD (Secure Harness Distribution), a distributable agent harness that embeds three security controls (OS sandboxing, skill scanning, tool restriction) for commercial AI coding agents. Using a 23-test suite from OWASP Top 10 for Agentic Applications, SHarD was evaluated across four agent configurations, achieving 100% efficacy matching the best securely configured commercial agent. Key findings include model non-determinism affecting security outcomes and OS sandboxing mitigating boundary-crossing autonomous behavior, with preliminary frameworks for control harness fitness proposed.
ai coding agentssecurity controlsos sandboxingagent harnessowasp top 10
Depression Markers in Speech: An Approach based on Tract Variables Dynamics
The study introduces novel depression biomarkers derived from dynamical properties of speech articulator configurations (tract variables), quantifying previously unexplored aspects: predictability (Largest Lyapunov Exponent), complexity (Correlation Dimension), and randomness (Sample Entropy). Experiments on the Androids Corpus (64 depressed vs. 54 control speakers) demonstrate discriminative power, with high Cliffs delta values for both read and spontaneous speech. This approach advances depression detection by modeling articulatory dynamics beyond traditional acoustic features.
depression biomarkerstract variableslyapunov exponentcorrelation dimensionsample entropy
Device Invariance using Domain Adaptation on Acoustic Scene Classification
The paper evaluates domain adaptation techniques for acoustic scene classification using CNN-based and transformer-based feature representations. It compares Domain Adversarial Neural Network (DANN) and Conditional Domain Adversarial Network (CDAN) under various domain shifts on the DCASE 2020 dataset. Results show DANN consistently adapts well for both feature extractors, while CDAN is effective only for CNNs, suggesting adaptation methods must align with the underlying representation.
domain adaptationacoustic scene classificationdanncdandcase 2020
A Machine-Learning-Based Gas Lift Optimization Workflow for Unconventional Fields
We propose a machine-learning-based workflow for gas lift optimization in unconventional fields, integrating a predictive ML model for Gas Lift Performance Curve forecasting and Bayesian Optimization for rate optimization under facility constraints. The ML model utilizes historical production time series data, eliminating the need for downhole gauges or multi-rate well tests. Pilot deployment across 30 wells in Bakken demonstrated an average production uplift exceeding 5%. Following this success, the workflow was fully deployed across 200+ gas lift and plunger-assisted gas lift wells. This approach offers an economical solution for assets lacking downhole data or multi-rate testing capabilities.
gas lift optimizationbayesian optimizationmachine learningproduction upliftunconventional fields
Stemma: Induced Decision Regions Reveal LLM Provenance
The paper introduces Stemma, a black-box LLM fingerprinting method that reframes provenance testing as measuring inheritance of induced decision regions rather than surface-form responses. By mapping open-ended outputs to a finite decision space and operationalizing stability, robustness, and specificity as probe-selection principles, Stemma abstracts away adaptation-induced variation. Evaluated on 770 source-suspect pairs (56 checkpoints) and 1,260 deployment instances, Stemma achieves 0.967-0.995 AUC and 87.8-93.5% TPR at 1% FPR, outperforming four baselines in detecting model lineage across weight transformations and inference-time settings.
llm provenancedecision regionsblack-box fingerprintingmodel lineageprobe-selection
Runtime Uncertainty Monitoring for LLM-Based Multi-Agent Systems Using Bayesian Networks
The paper proposes a Bayesian Network-based method for runtime uncertainty monitoring in LLM-based multi-agent systems (MAS) applied to actuarial risk modeling. The framework transforms token-level log-probabilities into calibrated task-level confidence estimates, addressing uncertainty from LLM stochasticity and inter-agent dependencies. Specialized agents handle data preparation, modeling, review, and explanation tasks under a central hub. Results demonstrate maintained baseline actuarial performance while providing workflow stability insights and uncertainty propagation tracking.
multi-agent systemsuncertainty quantificationbayesian networkslog-probabilitiesactuarial modeling
A2TTA: Anchored-and-Agile Test-Time Adaptation for Evolving Traffic Sensor Networks
A2TTA proposes an anchored-and-agile test-time adaptation framework for evolving traffic sensor networks, addressing topology expansion and multi-scale temporal shifts. The method transforms topology-induced errors into an expandable output calibration problem and separates temporal adaptation into persistent global correction and context-specific specialization. Evaluated on ten real-world traffic networks, A2TTA improves forecasting performance across diverse backbones, datasets, and prediction horizons.
test-time adaptationtraffic forecastingsensor networkstopology expansiontemporal shifts
How Do LLMs Read Bug Reports? An Empirical Study of Attention in LLMs for Automated Program Repair
This paper presents the first empirical study of attention patterns in LLM-based Automated Program Repair (APR), analyzing 319 Python and Java bugs from SWE-bench Verified and Multi-SWE-bench. The study investigates attention distribution across bug report sections, differences between successful and unsuccessful repairs, and alignment with developer-identified key information. Results show that successful repairs exhibit diffused attention across diagnostic components (bug descriptions, stacktraces, test cases), while failures focus excessively on metadata. Stronger alignment between model attention and developer priorities correlates with higher repair success, highlighting attention misallocation as a key factor in APR failures.
automated program repairattention patternsbug reportslarge language modelsempirical study
OmniQEC: discovering practical quantum error-correcting codes by an AI scientist
OmniQEC introduces an AI-driven framework for discovering practical quantum error-correcting (QEC) codes by combining large language models (LLMs) with a slow-fast synergistic workflow. The orchestrator LLM iteratively generates, screens, and evaluates QEC candidates, balancing code-level proxies and circuit-level physical constraints. Evaluated across four qLDPC families and three LLM backends, OmniQEC discovers codes outperforming BB codes (e.g., [[72,12,6]] and [[144,12,12]]) under 98 and 240 physical-qubit budgets, respectively, with improved logical-error suppression. The method enables hardware-friendly QEC co-design.
quantum error correctionlarge language modelsqldpc codeslogical-error suppressioncircuit-level evaluation
HiSkill: Empowering LLM Agents with Hierarchical Skill Graphs
HiSkill introduces a hierarchical skill graph framework that bridges high-level skills and executable actions for LLM agents. The method organizes interaction trajectories into a directed graph with skill nodes, AtomicOp nodes, and typed edges, capturing decomposition, temporal transition, compatibility, support, and recovery relations. At inference, HiSkill retrieves a task-relevant subgraph and performs subgraph-guided task execution, enabling iterative skill switching and action grounding. Experiments across three interactive environments demonstrate HiSkill's superior performance over state-of-the-art baselines while reducing inference token consumption.
hierarchical skill graphatomicopllm agentstask executioninference token consumption
Distributed Constraint Optimization via Online Learning and Iterative Pricing with Application to Large-Scale Satellite Scheduling
The paper introduces two complementary approaches for large-scale distributed constraint optimization problems (DCOPs). First, it adapts online learning algorithms for equilibrium finding in potential games, showing competitive performance with incomplete DCOP methods. Second, it proposes a decomposition framework separating task allocation (meta-DCOP) and local scheduling, coupled via iterative pricing to update meta-level utilities. Combined, these methods achieve 99% request fulfillment in decentralized satellite scheduling, outperforming state-of-the-art baselines (87%).
distributed constraint optimizationonline learningiterative pricingpotential gamessatellite scheduling
Lowering the implementation barrier of neutral-atom quantum computing with agentic workflows
The authors introduce an agentic workflow to automate the implementation of quantum protocols on neutral-atom quantum processors, specifically Pasqal QPUs, while retaining researcher oversight for validation. The system translates theoretical protocols into hardware experiments through automated compilation, simulation, and cloud execution, demonstrated via three case studies in many-body physics and optimization. While the agent successfully converted published papers into overnight QPU campaigns, human intervention was required to correct errors in observable selection (1 case) and hardware diagnosis (1 case). A secondary agent analyzed 633 Rydberg-array arXiv papers, finding 48% implementable on current QPUs and identifying hardware requirements for the remainder.
neutral-atom quantum computingagentic workflowqpurydberg arraymany-body physics
Speculate While You Reason: Teaching Agents to Predict Their Next Tool Call via Joint Agent-Speculator RL
The paper introduces self-speculating agents, which unify tool-call speculation and task execution within a single model to reduce latency in LLM-based agents. The proposed joint agent-speculator reinforcement learning method trains the model to alternate between agent mode (task execution) and speculator mode (next-tool-call prediction), reusing KV cache and deriving speculation targets from agent rollouts. Evaluations on agentic search QA and conversational tool-use tasks show improved next-tool-call Hit@1 (44.1→61.2 for Qwen3-4B, 48.9→66.3 for Qwen3.5-4B) while maintaining task success rates.
tool-call speculationkv-cachereinforcement learninglatency reductionagentic tasks
SpectONet: A Physics-Guided Spectral Deep Operator Network for Euler-Bernoulli Beam Dynamics
SpectONet introduces a physics-guided spectral deep operator network for Euler-Bernoulli beam dynamics, combining DeepONet's operator-learning capability with physics-informed constraints and nonuniform Chebyshev-Gauss-Lobatto sensor placement. The method concentrates sensors near domain boundaries to enhance boundary-sensitive response representation while minimizing branch-network inputs. Evaluated on three synthetic and one real-world bridge vibration dataset, SpectONet reduces prediction errors by ≥64% (synthetic) and ≥37% (real-world) versus baselines (Vanilla DeepONet, PI-DeepONet, PINN, CNN-UNet), demonstrating improved accuracy and physical consistency.
deep operator networkphysics-informed learningchebyshev-gauss-lobattoeuler-bernoulli beamstructural vibration
Loss Invariance Determines What Concept Layers Encode: Volume Grounding in Echocardiography
The study investigates the validity of concept bottleneck models in echocardiography by analyzing left ventricular volumes as intermediate variables for ejection fraction estimation. Using a video transformer encoder trained on a public dataset, the authors compare training with ejection fraction supervision alone versus additional volume supervision. Results show that without volume supervision, predicted volumes collapse to near-zero spreads (0.1 mL vs. reference 35.7-45.7 mL) despite preserved correlation, due to loss invariance in the ratio-based objective. Absolute volume supervision reduced volume error from 89.8 to 25.8 mL at a marginal cost to ejection fraction error (6.89 vs. 7.13 MAE). The work highlights the need to validate concept layers against objective invariance, not just accuracy.
concept bottleneck modelsvideo transformerloss invarianceejection fractioninterpretability
Image Quality Dependent Degradation for AI Systems
The paper proposes a fail-degraded system for AI-based automated driving that adjusts confidence thresholds based on estimated image quality to mitigate critical errors under poor input conditions. A novel normalizing flow-based method estimates image quality by comparing inputs to training data distribution. Experiments demonstrate the approach applied to state-of-the-art object detection, enabling cautious detection in uncertain scenarios without fallback mechanisms, thereby enhancing system trustworthiness.
fail-degraded systemconfidence thresholdnormalizing flowsobject detectionimage quality estimation
Shared Voxel-Map-Based Cooperative Indoor UAV Guidance with a Multi-Agent Soft Actor-Critic Controller
The paper introduces a cooperative indoor UAV guidance framework combining shared voxel-map fusion with multi-agent Soft Actor-Critic (MASAC) control. Drones aggregate LiDAR scans into a global occupancy map, converted to local bird's-eye-view (BEV) crops for decentralized control, integrating map features, obstacle data, and peer states via centralised-training/decentralised-execution. Simulation tests show 90.3% success in corridor navigation, outperforming A* and potential field baselines, with real-world validation via offline imitation fine-tuning. Results demonstrate scalable multi-UAV cooperation in GNSS-denied environments using shared spatial representations.
multi-agent soft actor-criticvoxel-map fusionbird's-eye-view representationdecentralised controlsim-to-real adaptation
Nudging Sustainable Choices through LLM-Generated Recommendation Explanations
The study demonstrates how LLM-generated recommendation explanations can nudge sustainable consumer choices by strategically framing sustainability information. Using generative AI, the authors created sustainability-aware explanations grounded in nudge theory, validated through human evaluation and LLM-as-a-judge audits. Two randomized studies (N=529) in low- (instant coffee) and high-involvement (hotel bookings) domains showed that framed sustainability information or descriptive social norms significantly increased sustainable selections (vs. plain disclosure) while easing decision-making, though perception and behavior diverged. The work highlights scalable, theory-grounded explanation generation for social good.
recommender systemsnudge theoryllm-generated explanationsbehavioral framingsustainable choices
Tools Are Not Islands: Set-Level Tool Retrieval for LLM Agents via Query-Conditioned Hyperedge Prediction
HYSET introduces query-conditioned hyperedge prediction for set-level tool retrieval in LLM agents, addressing limitations of isolated or sequential tool scoring by evaluating joint utility directly. The method leverages a tool co-invocation hypergraph with cardinality-specific interactions to capture size-dependent compatibility, operating as a plug-in pre-selection module. On ToolBench, HYSET outperforms SOTA baselines in retrieval accuracy (↑12.3%) and task success (↑8.7%), with demonstrated zero-shot/few-shot transfer to unseen tools/domains.
tool retrievalhyperedge predictionllm agentsco-invocation hypergraphzero-shot transfer
From Deterministic to Generative Deep Learning for Urban Air Quality Reconstruction from Sparse Observations
We propose a diffusion-based generative framework for full-field reconstruction of urban air quality from sparse observations of NO2, O3, PM2.5, and PM10, benchmarking against deterministic deep learning models. The approach leverages full-field simulation data for training and evaluates on real-world observations from 9-28 monitoring stations in Paris, incorporating data augmentation for transfer without retraining. Results demonstrate high structural similarity on simulated validation data and realistic spatial patterns on real-world observations, validated through power-spectrum analysis. The framework enables reliable deployment for air pollution reconstruction tasks.
diffusion-basedfull-field reconstructiondata augmentationpower-spectrum analysissparse observations
Cognivia: A Cognitive Behavioral Therapy Copilot for Evidence-Based Mental Healthcare
Cognivia introduces an evidence-based AI therapist for Cognitive Behavioral Therapy (CBT) by integrating automatic cognitive distortion identification and rational response generation. The framework leverages authoritative CBT texts, mental health Q&A data, and multi-stage prompting with expert supervision, fine-tuning a lightweight LLM. Evaluation via lexical metrics, LLM-based Judges, and human experts (n=10) shows superior performance in distortion recognition and response generation over baselines.
cognitive behavioral therapylarge language modelscognitive distortionmulti-stage promptingrational response generation
Rashomon Alignment
Rashomon Alignment (RA) introduces a geometric measure for assessing functional similarity between models across the entire input space, complementing distributional approaches limited to observed data regions. The method employs uniform sampling from the instance space to compute geometric Rashomon Alignment, providing a distribution-independent perspective on decision boundary alignment. Experiments on 90+ datasets demonstrate that geometric and distributional alignment offer distinct yet complementary insights, with applications in model selection, ensemble construction, and interpretability.
functional similaritygeometric alignmentdecision boundarymodel selectionensemble construction
DynaBridge: Dynamic Summary-Guided Cross-Task Multimodal Fusion for DASS-Structured Mental Health Assessment
DynaBridge introduces a dynamic summary-guided cross-task multimodal framework for DASS-structured mental health assessment, addressing limitations of generic fusion models. The method encodes acoustic, visual, and textual cues across sessions, augmented by frozen-LLM-generated DASS-aware summaries, and predicts ordinal item distributions while reconstructing depression, anxiety, and stress risk evidence. It employs a confidence-aware refinement strategy to incorporate high-confidence semantic cues conservatively. Evaluated on the AdoDAS validation split, DynaBridge achieves 0.5012 mean F1 for D/A/S risk prediction and 0.3216 mean QWK for DASS-21 item prediction, outperforming baselines and demonstrating the efficacy of integrating multimodal cues, semantic summaries, and psychometric structure.
multimodal fusiondass-structuredconfidence-aware refinementordinal item distributionssemantic summaries
DecoEvo: Score-Decoupled Co-Evolution of Solver and Rubric-Generator Skills in Text Space
DecoEvo introduces a decoupled co-evolution framework for optimizing solver and rubric-generator skills in text-space without gold rubrics. The method updates solver skills via criterion-level feedback while revising rubric-generator skills through independent audits of requirement coverage and response discrimination, focusing on newly exposed solver weaknesses. Evaluated across five benchmarks and three LLM backbones, DecoEvo achieves 2.8--5.0% relative gains over SkillOpt, demonstrating superior performance in open-ended tasks.
text-space optimizationdecoupled co-evolutionrubric-generatorcriterion-level feedbackrequirement coverage
OmniDelta: Skill-Driven Budget Allocation for Token Compression in OmniLLMs
OmniDelta introduces a skill-driven budget allocation framework for token compression in Omni-modal Large Language Models (OmniLLMs), addressing limitations of fixed-budget approaches. The method combines intent-aware inter-modal allocation (via audio/video skill pools) with content-aware intra-modal allocation (using local complexity and temporal redundancy) to dynamically redistribute token retention budgets. Evaluated on four benchmarks with Qwen2.5-Omni models, OmniDelta achieves a 22.0% GPU memory reduction and 1.64x speedup at 25% token retention while improving accuracy-efficiency trade-offs.
token compressionomnimodal llmsbudget allocationskill-driven pruningqwen2.5-omni
MyMentorLLM: A psychotherapy GenAI environment with multimodal voice/text patients, trainees and experts for deliberate practice
MyMentorLLM introduces a multimodal voice- and text-based simulation environment for deliberate practice in Cognitive Behavioural Therapy (CBT) training, addressing scalability issues in psychotherapist supervision. The system generates 2,100 CBT sessions featuring DSM-5-TR-grounded patients with major depressive, generalised anxiety, or borderline personality disorders, alongside trainee therapists and expert supervisors. Sessions were analyzed for emotional dynamics, therapeutic competence, and diagnostic accuracy. Results show simulated patients exhibited disorder-congruent emotional profiles, while supervisor feedback quality varied across LLMs, with native speech-to-speech models aligning closest to human evaluations. Symptom identification accuracy improved with model size, and supervisor feedback enhanced diagnostic accuracy in 5 of 7 LLMs. The study demonstrates CBT training simulation feasibility but highlights the need for evaluating patient fidelity, supervisor calibration, and feedback safety.
cognitive behavioural therapymultimodal simulationdeliberate practicedsm-5-trllm supervision
Localized Adaptation Reveals Distinct Learning Signatures in Transformers
This work demonstrates that transformer adaptation exhibits distinct learning signatures depending on layer-localization, as revealed through a controlled benchmark of five objectives (lexical binding, factual association, behavioral policy learning, causal mapping, and procedural reasoning). By systematically varying LoRA adaptation sites (early/middle/late/full-stack) and measuring acquisition, transfer, and boundedness, the study identifies objective-specific 'adaptation geometries': lexical binding favors early layers, factual association prefers late layers, behavioral learning separates action (late) from policy gating (middle), while causal/procedural tasks benefit from middle or full adaptation. These patterns persist across parameter-matched controls and five model families, establishing adaptation site as a key variable for targeted learning.
transformer adaptationlocalized loraadaptation geometrylearning signatureslayer-specific learning
CoRT: Counterfactual Replay for Token-Level Rubric-Guided Policy Optimization
CoRT introduces token-level credit weighting for rubric-conditioned GRPO by using counterfactual replay to rescore responses under rubric-conditioned and criteria-free prompts, deriving tokenwise log-likelihood contrasts as credit proxies. These contrasts are mapped to bounded weights to redistribute GRPO advantages across tokens without auxiliary models. Experiments show CoRT outperforms response-level GRPO by 4.4 percentage points on average, matching learned token-level baselines while avoiding separate relevance learning.
rubric-based reinforcement learningtoken-level credit weightingcounterfactual replaygrpo advantagelog-likelihood contrasts
OrchBench: Evaluating Multi-Agent Orchestration Plans in Isolation via Deterministic Simulation
OrchBench introduces a simulation-based benchmark for evaluating multi-agent orchestration plans in isolation, addressing the limitations of end-to-end execution that conflates plan quality with external factors. The method constructs task dependency DAGs with controlled parallelism, then uses a deterministic simulator to assess plans based on result quality, makespan, and token cost without invoking worker agents. Results show strong correlation with real execution (Pearson r=0.816) at 1.3% token cost and 10.3% wall-clock time, revealing that information preservation outweighs agent count and coordination failures limit parallelism benefits.
multi-agent systemsorchestration plansdeterministic simulationdirected acyclic graphstoken cost
Engine-Equal, Human-Unequal: A Reproducible Outcome Skew in Engine-Assessed Equal Chess Positions
The study identifies a reproducible outcome skew in human-played chess positions that engines evaluate as equal (Stockfish 18 evaluation within ±10 centipawns, depth-stable). Analyzing 1,661 naturally occurring positions (16.1M games from Lichess, October 2025), it demonstrates persistent directional biases (median |δ| ≈ 0.018) favoring White or Black, uncorrelated with player ratings or opening families. Skews replicate across disjoint player sets (replication slope = 0.69, 95% CI [0.65, 0.74]), time partitions, and rating bands, increasing to 0.94 for high-frequency positions. Engine evaluations are insufficient predictors of human outcomes, evidenced by longer think times for the disfavored side. Results hold across search depths, calibration, and game speeds (blitz/rapid).
outcome skewcentipawn evaluationreplication slopelichessdepth-stable
Why Public Service AI Governance Frameworks Risk Failing in the Age of General-Purpose AI: Lessons from Policing
The article argues that general-purpose AI (GPAI) undermines traditional AI safety governance frameworks in public services, particularly policing, due to its generality, accessibility, and low deployment cost. It highlights how GPAI renders key safety concepts—accuracy, bias, explainability, and accountability—intractable, as mitigations presuppose narrow, purpose-built AI. Through a case study on policing, the authors demonstrate that expert evaluation and human-in-the-loop oversight, common strategies in AI governance, fail under GPAI's assumptions. They recommend distinguishing narrow AI from GPAI in governance, prioritizing technological parsimony, pausing GPAI deployment in policing, and establishing a national safety infrastructure to generate evidence for responsible deployment.
general-purpose aigovernance frameworksai safetyhuman-in-the-looptechnological parsimony
KQFuzz: Knowledge-Guided Fuzzing for Quantum Libraries via Large Language Models
KQFuzz introduces a knowledge-guided fuzzer for quantum libraries, leveraging large language models (LLMs) to generate high-quality test cases. The method combines codebase-aware prompting, fitness-guided evaluation, and two-level mutations to explore execution paths and trigger bugs in quantum programs. Evaluated on Qiskit, PennyLane, and Cirq, KQFuzz improves coverage by up to 18.44% over baselines and uncovered 13 confirmed bugs (12 fixed).
quantum librariesfuzzinglarge language modelstest generationcodebase knowledge
OmniPhys: Knowledge-Graph-Driven Benchmarking and Collective Optimization for Physical Commonsense in Text-to-Image Generation
OmniPhys introduces a knowledge-graph-driven benchmark (1,551 samples) and optimization framework for evaluating physical commonsense in text-to-image generation. The method aligns PhET simulations with curricula via a dual-path verification protocol and proposes OmniPrompt, which treats physical alignment as discrete optimization using K-image feedback buffers and batch-aggregated meta-policy updates. Evaluations across 12 models show OmniPrompt significantly improves physical consistency, with evolved meta-policies demonstrating transferability.
physical commonsenseknowledge graphprompt optimizationtext-to-image generationmeta-policy
F(AI)2R: Who Did What, and Who Checked? Verifiable AI Provenance as an Executable Skill
The paper introduces F(AI)2R, a framework for verifiable AI provenance in research artefacts, extending the original F(AI)2R experiment with aiprov, a PROV-O-based model. The method enables AI-assisted authoring and machine-readable auditing, packaged as an executable skill for AI agents. It enforces two invariants: no parentless claims and human-granted verification rungs. The paper serves as a case study, with all activities, claims, and sources recorded in a provenance graph, ensuring auditability and conformance through continuous integration.
provenancefair researchprov-overifiable aicontinuous integration
AIriskEval-edu Demo: Auditing of Pedagogical Risks in Educational Explanations
The AIriskEval-edu Demo platform audits pedagogical quality in instructional explanations across five risk dimensions: factual accuracy, depth/completeness, focus/relevance, student-level appropriateness, and ideological bias. It employs a hybrid evaluation system combining GPT-5.5 (via API) and a locally hosted Llama 3.1 8B evaluator fine-tuned on annotated K-12 explanations. The local model outperforms GPT-5.5 on most metrics while enabling real-time auditing of user-written content or stored explanations generated under six simulated teacher profiles, with binary risk decisions, confidence scores, and evidence spans provided for detected issues.
pedagogical riskllama 3.1 8bbinary decisionevidence spank-12 annotations
Construction-Driven Injection: Linguistically-Grounded Edit-Based Code-Mixing Fingerprints for Large Language Models
The paper introduces LCF, a unified framework for constructing and injecting linguistically-grounded code-mixing fingerprints into large language models (LLMs) to deter unauthorized use. LCF generates triggers by combining low-resource languages via semantic-density substitution and grammar-biased mixing, while LCFEdit injects them using null-space projection and cross-lingual alignment to preserve model utility. Evaluations show the method achieves imperceptible, detectable, and harmless fingerprinting with minimal performance impact.
code-mixingnull-space projectioncross-lingual alignmentlinguistically-groundedfingerprinting
A Human-in-the-Loop Corpus for LLM-Based Simplification of Scientific Summaries
The authors introduce a human-in-the-loop workflow for generating accessible scientific summaries via LLM-based simplification, using SciSummNet as source corpus and GPT-4o-mini for baseline simplification. Phase 1 involves STEM non-CS readers evaluating comprehensibility, naturalness, and simplicity of original vs. GPT-simplified texts, while Phase 2 employs CS experts to create reference simplifications based on feedback. Results show GPT-simplified summaries outperform originals in comprehensibility and simplicity (Phase 1), with Phase 2 edits emphasizing retention of domain terminology and claim strength. The released corpus includes human judgments and automatic evaluations for benchmarking simplification systems.
llm-based simplificationhuman-in-the-loopscisummnetgpt-4o-minicross-disciplinary communication
Joint Text-Audio Alignment for EEG-to-Text Decoding in Chinese Speech Production and Perception
EEGAlign introduces a parameter-efficient framework for decoding Chinese sentences from scalp EEG by jointly aligning neural signals with both text (BGE-M3 embeddings) and audio (wav2vec 2.0 features) via contrastive learning, followed by CTC decoding. The method addresses challenges of high-dimensional output space and low SNR in Chinese EEG-to-text conversion, demonstrating complementary benefits of multimodal alignment. Evaluated on ChineseEEG-2, it achieves 82.37% Top-1 accuracy for Reading Aloud EEG and 41.43% for Passive Listening EEG among 101 candidate sentences, outperforming single-axis approaches.
eeg-to-text decodingcontrastive learningmultimodal alignmentctc decodingchinese speech processing
Quotient Dynamics, Effective Curvature, and Implicit Bias in Positive Quadratic Networks
The paper analyzes the quotient structure of positive quadratic networks, where the predictor f_U(x) = x^⊤UU^⊤x admits low-rank representations identifiable up to orthogonal transformations. By studying the Riemannian geometry of the rank-r PSD manifold ℝ^{d×r}_*/O(r), the authors show that Euclidean gradient flow projects exactly to quotient gradient flow, with finite-step descent inducing exact congruence recursions. For quadratic regression, they derive the effective Hessian at interpolators as a Gram form restricted to the quotient tangent space. Under Gaussian rank-one measurements, they prove uniform deviation bounds, construct a spectral initializer, and establish local convergence guarantees. In underdetermined commuting regimes, gradient flow becomes an exact entropy mirror flow, with positive initializations converging to Bregman projections. Isotropic initialization yields minimum-trace solutions as ε→0, resolving nonuniqueness via weighted entropy.
positive quadratic networksquotient geometryriemannian gradient flowbregman projectionspectral initialization
Beyond Epistemia: Epistemic Schizologia and Large Language Models as Techno-Semiotic Machines
The article critiques Quattrociocchi et al.'s concept of *Epistemia*—where LLMs generate plausible but epistemically ungrounded outputs—by proposing *epistemic schizologia* as a broader socio-technical phenomenon. Drawing on Carlo Sini's philosophy, it reframes LLMs as *techno-semiotic machines* automating written semiosis from human textual archives, highlighting *eikotic closure* (plausible outputs masquerading as epistemic results) and algorithmic authority. The analysis shifts focus from isolated models to hybrid human-AI practices, advocating for inspectable genealogy, contestability, and distributed responsibility in epistemic design.
epistemic schizologiatechno-semiotic machineeikotic closuresemiosisalgorithmic authority
Multi-Sensor Alignment for Weather Simulations
The paper introduces two alignment methods for weather simulations in autonomous vehicle perception: the Reference Dataset Alignment Method (ReDAM) for fog intensity alignment and Unified-weather-edit (inspired by Weather-edit) for rain/snow particle positioning. These methods ensure simulations match real-world weather characteristics across sensors, validated via statistical and geometric tests. Results show non-aligned simulations overestimate 3D detection performance, while aligned multi-sensor simulations improve robustness when fine-tuning sensor fusion models.
weather simulationsensor alignment3d object detectionautonomous vehiclessensor fusion
Contrastive Representation Learning of Longitudinal Disease Trajectories on Temporal Graphs
The authors propose a contrastive representation learning framework for modeling longitudinal disease trajectories as temporal graphs, addressing challenges of temporal dynamics and patient heterogeneity. The method represents patient observations as nodes and temporal/structure relationships as edges, employing structure-aware random walks to guide contrastive graph neural network training. Learned embeddings preserve temporal context and trajectory topology, enabling robust patient clustering (by progression similarity) and latent structure discovery in longitudinal clinical data.
contrastive learningtemporal graphsdisease trajectoriesgraph neural networkslongitudinal data
Physics-Informed Broad Learning System: An Efficient Backpropagation-Free Framework for Solving Partial Differential Equations
We introduce Physics-Informed Broad Learning System (PI-BLS), a novel framework for solving partial differential equations (PDEs) that eliminates iterative backpropagation by embedding governing physical laws into a linear output-layer optimization problem solved via pseudoinverse. Unlike physics-informed neural networks (PINNs), PI-BLS replaces gradient-based training with a deterministic least-squares solution, reducing the learning process to a single linear optimization stage while preserving physical constraints. Experiments on forward PDE benchmarks show that PI-BLS achieves competitive or superior performance with reduced training time and model parameters compared to conventional PINNs.
partial differential equationsphysics-informed neural networksbroad learning systemleast-squares solutionpseudoinverse
Computational Extraction of Legal Causes via al-Sabr wa al-Taqsim: A Set-Theoretic Formalization for Closed Fiqh Chapters
The paper introduces a set-theoretic formalization of al-Sabr wa al-Taqsim (Examination and Division) for extracting legal causes ('ilal) in closed fiqh chapters. It proposes a computational algorithm that derives minimal operational rules from juristic verdict truth tables, eliminating logically redundant attributes to produce admissible candidate causes. The method requires a finite school-relative concept vocabulary and complete ruling table, demonstrating that structural generators of rulings can be computationally extracted under these conditions.
set-theoretic formalizational-sabr wa al-taqsimlegal causestruth tablejuristic verdicts
Beyond Self-Knowledge: Propagating Uncertainty Across Reasoning and Retrieval in LLMs
The paper introduces BeyondUncertainty, a method that uses verbalized confidence estimates from black-box language models to route retrieval requests in knowledge-intensive QA. It first generates provisional answers with confidence scores, applies model-specific thresholds to decide retrieval (top-5 TF-IDF for low confidence), and evaluates on 27,000 policy instances across six benchmarks. Results show a mean token-level F1 of 0.483 (vs. 0.467 for always-retrieval), 20.4% fewer retrievals, and outperforms random routing in 17/18 settings (avg. +0.024 F1). Probe uncertainty weakly predicts retrieval benefit (AUROC=0.628) but increases token usage by 28.2%.
retrieval-augmented generationconfidence estimationtf-idftoken-level f1auroc
A Density-Matrix Framework for Electronic-Structure Analysis of Functional-Group and Salt Effects in Lithium-Metal Electrolytes
The study introduces EMolStudio, a density-matrix-based AI platform for predicting and analyzing electronic structures in lithium-metal electrolytes. The method integrates molecular functionalization, explicit Li$^+$ solvation-shell assembly, and density-matrix prediction with idempotency projection, enabling readouts of frontier orbitals, electrostatic potential, and bond orders. Applied to 163,655 functionalized molecules and 22,500 Li$^+$ solvation clusters across four salts, results show functional-group-dependent electronic-structure changes (e.g., CO$_2$Me, CN, F/CF$_3$) and anion-specific frontier-orbital localization (e.g., LiTDI anchors HOMO on anions, LiDFOB pairs anion-HOMO with functional-group-LUMO). The platform links chemical design to electronic-structure hypotheses for lithium-bond formation and interphase reactions.
density-matrixlithium-metal electrolyteselectronic-structure predictionsolvation-shellfrontier orbitals
How Small Can You Go? A Controlled Study of LoRA Rank, Target Modules, and Quantization Trade-offs for Text-to-SQL on a 60M-Parameter Model
The study conducts a controlled analysis of parameter-efficient fine-tuning (PEFT) and quantization trade-offs on a 60M-parameter T5-small model for text-to-SQL (WikiSQL). It evaluates LoRA rank (r ∈ {2,4,8,16,32}), target modules, and numerical precision (INT8, NF4) while measuring task accuracy, trainable parameters, and system metrics. Results show LoRA (r=16) achieves 59.6% exact-match accuracy (11.6 pp below full fine-tuning’s 71.2%) with <1% trainable parameters and 31% lower peak GPU memory. Higher ranks (r>16) yield no accuracy gain, while QLoRA with INT8/NF4 attains ~53% accuracy at 0.60 GB memory. The work provides reproducible configurations for efficiency-accuracy trade-offs.
parameter-efficient fine-tuningloraquantizationtext-to-sqlt5-small
IRIS: Reusable Identity Representations from Frozen LLMs for Entity Alignment
IRIS (Identity Representations from Internal States) introduces reusable identity representations for entity alignment (EA) by deriving stable, iris-like signatures from frozen LLMs without training. The method constructs distinctive identity encodings via identity-oriented contextual representations, enabling direct similarity comparisons across knowledge graphs (KGs) without pair-dependent reprocessing. Evaluated on four EA benchmarks (D-Y-15K V2, DBP-WIKI, ICEWS-WIKI, ICEFS-YAGO) with two frozen LLM backbones, IRIS achieves Hits@1 scores of 100.00, 99.38, 98.31, and 97.99, respectively.
entity alignmentknowledge graphsfrozen llmsidentity representationssimilarity comparison
Matrix-Free Photoacoustic Image Reconstruction via Sensor-Token Self-Attention
The Sensor Attention Network (SAN) introduces a Transformer-based approach for matrix-free photoacoustic tomography reconstruction by treating each sensor's time series as a token and mapping measurements directly to images without system matrix invocation. SAN employs sensor-token self-attention and is trained with a vessel-weighted loss on 488 augmented samples, validated against k-Wave simulations (Pearson r=0.919). Evaluated on 46 test samples, SAN outperforms ISTA, LISTA, and split-Bregman total variation in SSIM (0.522), PSNR (22.09 dB), and NMSE (0.233), with statistical significance (p<1e-8) and 10× faster inference.
photoacoustic tomographytransformerself-attentioninverse problemcompressive sensing
The LAIA Dataset: Labelled Attention for Intelligent Automobiles
We introduce LAIA (Labelled Attention for Intelligent Automobiles), a synthetic dataset designed to advance end-to-end driving research by incorporating human attention data. Collected using the CARLA simulator, LAIA comprises over 15 hours of driving data from 44 participants across diverse scenarios, including RGB images under six weather conditions, semantic and instance segmentation, depth, optical flow, CAN bus signals, and synchronized eye-tracking data. The dataset supports applications such as training attention-aware AI drivers, predicting driver behavior, detecting anomalous attention patterns, and improving model explainability. We demonstrate LAIA's utility by comparing human attention with perceptual attention in end-to-end driving models, offering insights into model behavior.
end-to-end drivingattention-aware modelscarla simulatoreye-tracking datamodel explainability
CORF-GS: Real-Time Wireless Radiance Field Reconstruction via Coupled Optical-RF Gaussian Splatting
CORF-GS introduces a real-time wireless radiance field (WRF) reconstruction framework using coupled optical-RF Gaussian splatting, addressing the latency limitations of offline methods. The method constructs a unified Gaussian representation with shared geometry and modality-specific appearance, leveraging high-resolution optical images as structural priors for RF reconstruction. Optical-guided Gaussian sampling densifies under-represented regions, while coupled optimization refines shared Gaussians to adapt to both optical and RF distributions. Evaluations demonstrate state-of-the-art RF spectrum synthesis quality and a 6.4× speedup over existing WRF methods.
3d gaussian splattingwireless radiance fieldreal-time reconstructionoptical-rf couplingspectrum synthesis
Agent Skills Matter: Inferring Proprietary Skills from Execution Trajectories
The paper introduces Skill Leakage, a vulnerability where proprietary agent skills can be reconstructed from execution trajectories without reference answers or success labels. It presents SigLeak, a black-box framework that identifies recurring skill signatures by contrasting skill-enabled and skill-disabled trajectories across diverse diagnostic tasks, iteratively refining reconstructed skills. Evaluated on five scenarios, three model families, and three agent frameworks, SigLeak improves success rates by 6.88 percentage points over baselines and achieves the highest SkillSim (semantic similarity) scores, demonstrating significant exposure of procedural knowledge through benign trajectories.
skill leakageexecution trajectoriesblack-box frameworkprocedural knowledgesemantic similarity
Distilling Temporal Search and Reasoning: Evolving LLMs for Future Prediction via Harness-Assisted Efficient Data Synthesis
We introduce a time-truncation harness for temporal search and reasoning in future event prediction, addressing temporal leakage and inefficient data synthesis in Tool-Integrated Reasoning (TIR). The harness enforces temporal cut-offs during sampling from historical events, reducing reliance on rejection sampling and unresolved queries while increasing synthesis efficiency. We construct a large-scale corpus and process-based metric, demonstrating broader temporal search breadth and higher-quality data proportions. Distillation experiments reveal that models trained on harness-intervened data achieve superior performance, advancing parametric capabilities through enhanced temporal reasoning.
time-truncation harnesstemporal leakagetool-integrated reasoningprocess-based metricparametric advancement
From Training to Deployment: Post-Hoc Causal Feature Identification via Sensitivity Ratios
The paper introduces Normalised Sensitivity Ratio (NSR), a post-hoc, model-agnostic diagnostic for identifying causal versus spurious features in trained models under structured-shift regimes. NSR quantifies feature sensitivity variation across environments, with causal features exhibiting constant sensitivity and spurious features tracking shift. Theoretical analysis under a linear SCM with ≥3 non-degenerate environments proves exact identification (Theorem 1), with failure modes characterized quantitatively. Experiments achieve AUROC=1.000 on synthetic data, Kendall τ≥0.529 across five model families, and Precision@7=0.75 on bike-sharing data.
post-hoc analysiscausal feature identificationstructured-shift regimenormalised sensitivity ratiolinear structural causal model
Less is More: Modality-Decoupling for General AIGC Audio-Video Detection
We introduce DAV-Det, a modality-decoupled system for general audio-visual AIGC detection that replaces feature-level fusion with decision-level fusion for improved robustness. The visual detector employs multi-granularity representations at global, patch, and segment levels to identify spatial forgery cues, while the audio detector utilizes a gated temporal-spectral dual-branch architecture to capture acoustic artifacts. Empirical evaluation demonstrates that existing methods' assumption of audio-visual correspondence does not hold in general scenarios. DAV-Det achieved a score of 0.8460, ranking first in the General AIGC Audio-Video Detection Challenge at IJCAI-ECAI 2026 DDL 2.0 Workshop.
modality-decouplingdecision-level fusionmulti-granularity representationstemporal-spectral dual-branchaigc detection
Visual prompt engineering for video models
Visual Prompt Engineering (VIPE) systematically enhances video model performance by automatically modifying task images, analogous to text-based prompt engineering in language models. The method employs image editing models to transform abstract visual scenes into photorealistic versions, optimizing input representations for visual reasoning tasks. Empirical results demonstrate that VIPE outperforms traditional text-based prompt engineering and test-time scaling across multiple video reasoning benchmarks. This computationally efficient approach establishes VIPE as a fundamental technique for eliciting improved reasoning capabilities from video foundation models.
visual prompt engineeringvideo modelsvisual reasoningimage editingfoundation models
Entangled by Design: Spurious Intra-Variable Signal Routing in Tabular In-Context Learners
The paper formalizes spurious routing in composite representations, where in-context learners (ICLs) fail to distinguish between causal and spurious signals in tabular data, leading to silent deployment failures. Using ridge ICL and TabPFN, the authors prove that spurious routing is unavoidable and increases with context size, with larger models showing greater vulnerability. They derive a closed-form characterization of spurious routing, validated empirically (r = 0.997 for linear ICL, r = 0.979 for TabPFN). Two lightweight mitigations—environment-stratified context construction and S-swap augmentation—reduce spurious routing by 74% for linear ICL and 98.8% for TabPFN, while increasing causal sensitivity by 8.4×.
spurious routingin-context learningtabular dataridge regressioncausal signal
Multi-Scale Structural Features for Continual, Comprehensible Visual Recognition in a Developmental Learning Framework
The study introduces a multi-scale structural feature representation for continual visual recognition within a developmental learning framework, enhancing expressivity over prior work. The method integrates edge, contour, and spatial relation features across scales, coupled with refined learning dynamics and read-out mechanisms, while maintaining gradient-free, topology-based learning. Evaluated on class-incremental MNIST, it achieves accuracy comparable to replay- and regularization-based baselines (without storing past data) and preserves earlier-learned classes without destructive adaptation. Interpretable representations and sample-wise integration are retained, contrasting baselines that exhibit catastrophic forgetting within training cycles.
continual learningmulti-scale featuresdevelopmental frameworktopological modelcatastrophic forgetting
Are the High-weight Neurons the Important Ones in Image Classification Neural Networks?
The study challenges the equivalence between neuron weight and importance in image classification networks by introducing a neuron importance assessment method. Through three experiments—overlap quantification, perturbation analysis, and ablation-retraining—on CIFAR-10 and Mini-ImageNet, it reveals that only 25% of top 10% high-weight neurons overlap with accuracy-impacting ones, and their importance is nonlinear. Perturbing top 10% high-weight neurons causes 45-80% accuracy degradation, yet a third show minimal impact. Ablating top 10% high-weight neurons results in 10-20% accuracy loss without recovery, while ablating top 0.1% allows near-full recovery. Low-weight neurons also significantly contribute, with some intervals showing 10-17% degradation. These findings refine neuron role insights, supporting applications like encryption and pruning.
neuron importanceimage classificationperturbation analysisablation-retrainingnonlinear importance
Argus-Unified: Towards A Compact and Economical Unified Model for Image Understanding and Generation
Argus-Unified introduces a compact, economical unified model for multimodal image understanding and generation, addressing compute and data efficiency challenges. The method leverages pretrained vision-language models, employing hybrid visual tokens that combine continuous tokens for understanding with learned discrete tokens for generation. Training occurs in two stages: first, learning a quantizer and image decoder atop a frozen vision encoder; second, training a language model initialized from a pretrained VLM. Argus-Unified achieves state-of-the-art multimodal understanding on GQA, POPE, and VQAv2, and competitive generation quality, using only 15.6M data and ~$2,000 cost, outperforming models like Janus and Janus-Pro at ~10x lower cost and ~5x less data.
multimodal modelingvision-language modelshybrid visual tokensimage decoderquantizer
ReLATE: Reliability-Guided Evidence Fusion for Robust UAV--Satellite cross-view Geo-Localization
The paper introduces ReLATE, a reliability-guided evidence fusion framework for robust UAV-satellite cross-view geo-localization, addressing performance degradation under real-world corruptions. The method estimates token-level reliability fields, aggregates trustworthy local evidence, and fuses it adaptively with query-derived representations, combining CLS-token and GeM-pooled branches for final descriptors. Evaluated on UAVSat-Deg (27 corruption types, 11.7M images), ReLATE achieves state-of-the-art robustness (+3.1% mAP over baselines) while maintaining clean-image accuracy (76.2% mAP on University-1652).
cross-view geo-localizationreliability-adaptive fusionuav-satellite matchingdescriptor robustnesscompound corruptions
I2VShield: An Efficient Proactive Defense Framework against DiT-based Image-to-Video Models
I2VShield introduces an efficient proactive defense framework against Diffusion Transformer (DiT)-based image-to-video models, addressing the computational overhead of gradient-based adversarial attacks. The method combines a text-adaptive perturbation generation framework with adversarial learning and an untargeted Multimodal Attention Disruption (MAD) attack, targeting inherent vulnerabilities in DiT-based models to maximize internal attention feature deviation. Experiments show competitive protection performance across datasets and DiT-based models, effectively disrupting spatiotemporal coherence while reducing computational costs.
diffusion transformermultimodal attention disruptionadversarial learningspatiotemporal coherenceimage-to-video models
At-the-Roofline Sparse Tensor Contractions on Vector Processors for Transformer Inference
The paper introduces Ventaglio, a runtime-configurable sparse execution unit with RVV ISA extensions that optimizes sparse tensor contractions for Transformer inference. The method leverages Gustavson's dataflow for metadata-driven indexed accumulation on vector processors, addressing limitations in existing RVV architectures through indexed gather-accumulate-scatter support. Implemented in 12nm FinFET, Ventaglio achieves 6.9–7.4× speedup over RVV baselines with 3.1% area overhead and demonstrates 2.40–5.25× (prefill) and 2.06–3.16× (autoregressive) speedup on a DuoGPT-pruned LLaMA-3-8B model with 40–60% dual sparsity.
sparse tensor contractionsvector processorsgather-accumulate-scatterrvv isatransformer inference
Beyond Counts: A Distributional Robustness Margin For Pathology Foundation Models
The paper introduces Cross-confounder Robustness Margin (CRoMa), a sample-resolved metric for evaluating pathology foundation models' robustness to non-biological variation. CRoMa compares distances between cross-confounder biological matches and same-confounder distractors, framing robustness as a margin distribution rather than a pooled score. Evaluations on 20 tile-level and 4 slide-level encoders revealed consistent model rankings but substantial within-model heterogeneity, with all tile encoders showing confounder-dominated lower tails. Higher CRoMa correlated with reduced shortcut-induced performance drops after supervised adaptation, providing a principled basis for model selection.
pathology foundation modelsdistributional robustnessshortcut learningrepresentation geometrycross-confounder matching
Automated Numerical Stability Analysis of Deep Learning Operators
The authors introduce the first unified software tool for automated numerical stability analysis of deep learning operators, integrating CESTAC methodology. The tool enables single-pass numerical validation, identifies instability sources, and monitors stability during training and inference. It was validated on tasks with injected numerical instabilities, successfully detecting polluted operators. This approach provides critical insights for developing numerically stable computing kernels, essential for efficient and stable deep learning workflows.
numerical stabilitydeep learning operatorscestacfinite-precision arithmetictraining and inference
CoTinyVLA: Chain-of-Thought Distillation for a Sub-Billion-Parameter Vision-Language-Action Model
CoTinyVLA introduces a 0.9B-parameter Vision-Language-Action model that outperforms larger baselines (3-7B) on the LIBERO-Plus benchmark through structured supervision. The method employs dual-view temporal input (16 history frames), hierarchical chain-of-thought distillation from a 35B teacher (Plan and Think spans), and paraphrase augmentation (40→800 command variants). It achieves state-of-the-art results on LIBERO-Plus (90.8% Spatial, 87.3% Object, 86.6% Goal, 80.7% Long), with particularly strong gains on Robot Initial States (73.6% vs 39.9% baseline). Ablations confirm the contributions of each component, with the episode Plan being critical (40-45pt drop if removed).
vision-language-actionchain-of-thought distillationparameter efficiencyrobustness benchmarkhierarchical supervision
PatientAgentBench: A Benchmark Framework for Evaluating Patient-Facing Health AI Agents
PatientAgentBench introduces a benchmark framework for evaluating patient-facing health AI agents, addressing gaps in current medical QA benchmarks by assessing agentic interactions with simulated patients. The method employs an LLM-as-a-Jury scoring system across six dimensions using 100+ clinician-grounded criteria, validated via clinician annotations showing 79-93% agreement. Results from 10 models on 1,200 scenarios reveal clinical gaps: triage quality varies from 32% to 88% pass rates, with frontier models failing 1-3% on safety/workflow accuracy, highlighting limitations of static benchmarks for autonomous healthcare agents.
patient-facing agentsclinical safetyllm-as-a-jurytriage qualityworkflow accuracy
Finding Optimal Cost-Bounded Plan Reductions: Refined Model
The paper presents an exact method for extracting cost-bounded subplans from precomputed plans while preserving action order and executability, maximizing utility under budget constraints. Two approaches are proposed: an oversubscription planning (OSP) formulation and a refined Integer Linear Programming (ILP) model that reduces problem size and improves efficiency. The decision variant is proven NP-complete, extending prior work from ICAPS 2026 with computational enhancements.
cost-bounded planningoversubscription planninginteger linear programmingnp-completeaction order preservation
Architectural Backdoors in Vision-Language Model Supply Chains via Representation Steering
The paper introduces architectural backdoors in Vision-Language Model (VLM) supply chains via representation steering, demonstrating how malicious providers can embed dormant trigger-gated logic into shared model artifacts. The attack modifies intermediate representations additively, activating only when a specific trigger is present, while preserving clean utility otherwise. Evaluations on VLMs show compromised integrity in visual question answering, text-to-image generation, and retrieval, with proposed auditing defenses inspecting executable logic in distributed artifacts.
vision-language modelsrepresentation steeringarchitectural backdoorssupply chain securityintermediate representations
Balancing multiscale similarity and cartographic constraints: A similarity-driven optimization framework for line generalization
The study proposes a similarity-driven optimization framework for automated cartographic generalization, addressing the challenge of balancing multiscale representation consistency with cartographic constraints. The method integrates multiscale spatial similarity as an optimization objective alongside cartographic constraints for readability and geometric validity, optimizing a unified objective function to identify scale-dependent parameters. Experiments with line simplification algorithms and multiple similarity metrics demonstrate improved balance between similarity preservation and abstraction, yielding more consistent parameter control than similarity evaluation alone.
cartographic generalizationmultiscale similarityline simplificationoptimization frameworkgeometric validity
TRWH: A Text-Driven Random Walk Heterogeneous GNN for Semantic-Aware Sparse Recommendation
TRWH introduces a text-driven heterogeneous GNN for sparse recommendation, combining LLM-generated textual profiles with graph structures via random walk augmentation. The framework employs Word2Vec and LLM-based embeddings, a HeteroGNN for multi-relational propagation, and random walk-based path construction to enrich sparse graphs. Evaluations on Amazon-2023 Fashion (2M users, 825K items) and Beauty (631K users, 112K items) show 80.0% RMSE and 52.6% MAE reductions on Fashion, with 25.7% and 10.8% improvements on Beauty, though random walks may dilute LLM-learned nuances.
graph neural networkslarge language modelsheterogeneous graphrandom walksparse recommendation
Seen, Said, or Forgotten? A Causal Audit of Visual KV Memory Across Dialog Turns
The paper introduces the Causal Visual Memory Audit (CVMA), a framework for evaluating when visual facts can be safely forgotten in multimodal dialog systems. CVMA uses a paired single-prefill approach to test how unavailable visual regions, entire images, or prior text affect subsequent answers. Experiments on VisDial and ConvBench reveal that current attention mechanisms poorly rank future-useful regions (worse than random) and that aggregate scores mask failures when vision isn't needed later. Safe forgetting depends on low future visual dependence or prior verbalization, not low current attention.
multimodal dialogvisual-kv memorycausal auditattention mechanismskv-cache
Emergent Latent-State Computation under Stochastic Volatility
The paper investigates how sequence models internally represent latent stochastic dynamics under partial observability, using a controlled multivariate stochastic volatility setting where ground-truth latent states are known. Through experiments with various architectures, losses, and output heads, the authors identify a two-stage computation: hidden representations encode next-step volatility information, which output heads map to squared return forecasts. Transformers exhibit emergent latent-state decodability at architecture-dependent stages, simplifying to a learned linear projection with ℓ² normalization in long-cycle regimes. Results suggest stochastic volatility models serve as useful benchmarks for mechanistic interpretability under noisy, partially observed dynamics.
mechanistic interpretabilitystochastic volatilitypartial observabilitylatent-state decodabilitytransformers
Toward an Organizational Science of Multi-Agent LLM Systems: Decoupling Who, How, and Which Algorithm
The IMACS framework decouples organization, coordination, and collaboration protocols in multi-agent LLM systems, enabling controlled comparisons and adaptive protocol selection. By separating team composition (Belbin roles, RACI accountability), coordination mechanisms (Mintzberg), and collaboration algorithms (six published methods), it supports empirical validation of organizational designs. Adaptive Org Routing, a contextual-bandit meta-protocol, dynamically selects collaboration protocols per task under quality-cost constraints, outperforming fixed protocols in benchmarks using LLM-judge rewards. Ablations reveal protocol-dependent effects of accountability placement, necessitating model-specific validation of organizational configurations.
multi-agent systemsorganizational theorycontextual banditsllm collaborationaccountability placement
The Disruptive Impact of Large Language Models on Capture the Flag Competitions and the Path Toward Fair Play
This paper examines the disruptive effects of large language models (LLMs) on Capture the Flag (CTF) competitions, proposing a four-component safeguard framework to address fairness and educational value. The study employs mixed methods, including benchmark synthesis, live competition case studies, community channel observations, and interviews with players and organizers. Results indicate LLMs reliably automate easy-to-intermediate challenges in cryptography, web, and binary exploitation, while certain sub-categories remain resistant. The framework includes tiered divisions, LLM-resistant design, investigative telemetry, and a draft code of conduct, tailored to a competition's purpose.
large language modelscapture the flagcybersecurityin-context learningbenchmark synthesis
From Dyad to Triad: Eliciting XAI Requirements in Stroke Rehabilitation
The study introduces a video-based scaffolding protocol for eliciting explainable AI (XAI) requirements from stroke survivors and caregivers in rehabilitation settings, addressing challenges posed by communication disorders. The protocol employs four scaffolding techniques—analogical bridging, projective personas, binary forcing, and extended response time—piloted with three stroke survivors (two with aphasia) and three caregivers. Results revealed heterogeneous XAI needs and identified three facilitation biases (normative, hypothesis confirmation, presence effect) as protocol risks, providing guidelines for practitioners in designing trustworthy human-machine systems.
explainable airequirements elicitationstroke rehabilitationscaffolding protocolfacilitation bias
Salient Knowledge Pathways: Sparse Cross-Modal Routing for Efficient Knowledge-Intensive Multimodal Question Answering
The paper introduces SKIP (Salient Knowledge-Injected Pathways), a sparse cross-modal routing architecture for efficient knowledge-intensive multimodal question answering (KI-MMQA). SKIP employs question-guided visual token pruning, region-conditional sparse retrieval, bipartite sparse cross-attention, and speculative knowledge verification with an adaptive budget controller. An information-bottleneck bound shows optimal visual sparsity scales as $O(1/\sqrt{N})$ with retained accuracy. Evaluated on five KI-MMQA benchmarks (OK-VQA, A-OKVQA, InfoSeek, Encyclopedic-VQA, ViQuAE), SKIP matches or exceeds dense baselines while using 3.4--6.8× fewer FLOPs and 2.7× lower latency.
sparse cross-attentionknowledge-intensive qaadaptive computationvisual token pruningmultimodal routing
MARS: Multi-Agent Re-ranking for Repeat-Order Food Delivery Recommendation
The paper introduces MARS, a modular multi-agent re-ranking framework for repeat-order food delivery recommendation, combining LightGCN-based global preferences, Swing-based local evidence, geospatial filtering, and prompt-driven LLM reasoning. Evaluated on Delivery Hero benchmarks DHRD-SE and DHRD-SG, MARS demonstrates that pre-trained LLMs paired with lightweight collaborative retrieval achieve competitive performance in repeat-order recommendation. The study provides a reproducible pipeline for hybrid LLM recommenders, integrating collaborative signals and LLM-based re-ranking transparently.
multi-agent re-rankingrepeat-order recommendationlightgcnswing algorithmgeospatial filtering
A Control System, a Dataset, and a Recipe for Making Frozen LLM Agents Learn a Domain
The paper introduces a constrained reinforcement learning approach to optimize the control system (harness) of frozen LLM agents, treating it as a fixed action space with sample-efficient RL (ε-greedy bandit and REINFORCE) and multi-objective rewards (task success, verifier score, policy compliance, cost, latency, unsupported-claim penalty). The method is implemented using DSPy for context assembly and evaluated across tool-use workflows, HumanEval code generation, and HotpotQA multi-hop retrieval, with models from Ollama and AWS Bedrock. Results include released code, a verifiable task suite, training logs, and a deployment recipe for domain adaptation.
frozen llm agentsreinforcement learningmulti-objective rewarddspyverifiable task suite
Context Assembly as the Controlled Variable: A Control-Theoretic View of Harness Policies for Frozen LLM Agents
The paper proposes treating context assembly as the controlled variable in frozen LLM agents, contrasting prior approaches that control tool selection or action streams. It introduces a formal decomposition into an inner frozen policy $π_θ$ and an outer context policy $π_φ$, learned online via contextual bandit or REINFORCE. The authors provide a stability argument for the online controller, demonstrating non-decreasing expected reward under bounded policy change, and analyze uncertainty calibration against task outcomes. Empirical validation spans three domains and two model providers, with released datasets and deployment recipes.
context assemblyfrozen llm agentscontextual banditreinforce policyuncertainty calibration
COVENANT: Natural-Language Workflow Compilation for Aligned Agent Execution
COVENANT introduces a compiler-and-interpreter architecture for workflow-aligned execution of LLM agents, addressing workflow misalignment by treating natural-language instructions as source programs rather than prompts. The method compiles instructions into a workflow abstract syntax tree (WAST), lowers it to a workflow control-flow graph (WCFG), and interprets the WCFG node-by-node with runtime checks and diagnostic feedback. Evaluated on 120 cases across seven workflow scenarios from three benchmarks, COVENANT improves success rates from 50.00% to 83.33% and reduces workflow-misalignment failures from 42.50% to 15.83%, demonstrating substantial mitigation of misalignment in complex workflows.
workflow misalignmentabstract syntax treecontrol-flow graphllm agentsruntime checks
HANDBOOK.md: A Benchmark for Long-Context Agentic Instruction Following
The paper introduces HANDBOOK.md, a benchmark for evaluating language-model agents' ability to adhere to long, binding policy documents over extended tool-use horizons. The benchmark comprises 65 tasks across five domains (finance, medical billing, insurance, logistics, HR), each involving a self-contained company environment with a 20-124 page standard operating procedure. Tasks are designed to resist memorization by modifying base handbooks, and grading uses 824 deterministic criteria. Under strict grading, the best model configuration achieves 36.2% pass rate, with common failure modes including policy override by in-environment requests and loss of rule details over long horizons.
language-model agentsstandard operating proceduretool-use horizondeterministic criteriapolicy override
Cyber-Capable AI Agents: Vulnerabilities, Evaluation Containment, and Defensive Response
This review synthesizes vulnerabilities in cyber-capable AI agents, focusing on containment challenges within evaluation environments. It identifies five vulnerability classes: multi-step offensive chains, conflicting objectives with sandbox boundaries, supply-chain and credential exposure, persistent command-and-control, and automated action speed. Using the July 2026 Hugging Face/OpenAI incident as a case study, the analysis distinguishes incident-specific observations from broader literature findings. The review examines containment controls, privilege separation, provenance, and responder access, highlighting the dual-use problem of defensive artifacts. Practical priorities are identified for evaluating cyber capability alongside environmental security.
vulnerability classescontainment controlsprivilege separationprovenancedual-use problem
Rethinking Likelihood distributions: Student's t Likelihood Boosts Bayesian Neural Network Performance
The paper investigates alternative likelihood distributions for Bayesian neural networks (BNNs), challenging the default Gaussian assumption. It evaluates skewed and heavy-tailed distributions, particularly Student's t, in variational inference with evidence lower bound (ELBO) optimization. Experiments on regression tasks with multilayer perceptrons (MLPs) show that Student's t likelihood consistently outperforms Gaussian in predictive performance, regardless of data distribution or MLP architecture, and occasionally reduces training time while maintaining ease of implementation.
bayesian neural networksvariational inferencestudent's t distributionevidence lower boundmultilayer perceptrons
ODYSSE: Episode-wise Policy Optimization for Personalized Agentic Reasoning
ODYSSE introduces Episode-wise GRPO (ESPO), a Reinforced Fine-Tuning (RFT) framework for personalized agentic reasoning, extending Group Relative Policy Optimization (GRPO) to handle long action horizons and cross-step dependencies. ESPO employs episode-level rewards and advantage estimation to guide downstream decisions progressively, alongside an episodic batch sampler for coherent optimization. Evaluated on long-horizon personalized GUI reasoning tasks, ODYSSE outperforms specialist and general-purpose LVLMs, demonstrating efficacy in resolving ambiguous user requests.
personalized agentic reasoningreinforced fine-tuningepisode-wise grpolong action horizonsepisodic advantage estimation
AI Deployment and Cyber Governance Failures in Public-Sector Organizations: A Typological Analysis
The paper introduces a seven-domain typology identifying ten AI-driven cybersecurity governance failure causes in public-sector organizations, addressing a gap in integrated analysis of AI adoption, cybersecurity governance, and institutional constraints. It proposes a three-pathway failure model (accountability, operational resilience, compliance) and evaluates five governance frameworks (NIST CSF 2.0, ISO/IEC 27001, COBIT, NIST AI RMF, ISO/IEC 42001), finding none adequately address Shadow AI, speed asymmetry, or governance vacuum. The study formalizes speed asymmetry as a structural construct and provides design specifications for an AI-enabled cybersecurity maturity model.
cybersecurity governancepublic sectorshadow aispeed asymmetrymaturity model
Explanation-Bound Tool Execution for AI Agents: Server-Verified Action Claims Without Trusting Model Rationales
The paper introduces Explanation-Bound Tool Execution (EBTE), a mediation layer that converts AI agent rationales into server-verified action claims without trusting free-form explanations. EBTE enforces strict checks against server-held facts (intent, policy, payload, etc.), denying mismatches and preserving only eligible claims for execution. The method is formalized under explicit mediation assumptions, with a reference implementation tested across 136 conformance scenarios (0/96 contradictions passed, 232/232 metamorphic checks). Experimental validation shows 71/96 generation/runner agreement in hosted-model tests and 100% attack proposal resolution in AgentDojo-derived checks, demonstrating feasibility without claiming rationale faithfulness or safety guarantees.
tool-using agentsaction claimsmediation layerserver-verificationconformance testing
Raven: High-Recall Sequence Modeling with Sparse Memory Routing
Raven introduces a linear-time sequence model that balances dense state updates and sparse token storage to improve long-context recall. It maintains a fixed set of memory slots, decaying and updating a selected subset via learned, input-dependent routing, mitigating issues in sliding-window attention and state-space models. Raven outperforms prior linear-time baselines on recall-intensive benchmarks, maintaining strong performance even when extrapolating to context lengths up to 16x its training length, with similar gains in hybrid architectures.
linear-time sequence modelmemory routinglong-context recallstate-space modelssliding-window attention
Data Quality Profiling at Scale with Progressive Sampling: A Benchmark for Data-Centric AI Pipelines
The study benchmarks nine progressive sampling strategies for data quality profiling in data-centric AI pipelines, demonstrating that blind representative samplers outperform proxy-guided methods. Evaluating random uniform, geometric, Yamane, cluster, and five proxy-guided strategies (Metropolis-Hastings, DAG, stratified, importance-weighted) on real-world (NYC 311, NYPD arrests, UCI Adult, IoT streams) and synthetic datasets (up to 7.4M rows), results show random uniform achieves 0.49% mean relative error at 5% sampling budget, while DAG-guided MCMC yields 19.5% error (40x worse). Proxy-guided methods suffer from IQR proxy mismatch, over-prioritizing numeric outliers while missing categorical defects. Random uniform scales near-linearly (O(N^0.964)) and outperforms DAG (O(N^1.272)) by 6x accuracy and 28-47x speed on ultra-large data.
progressive samplingdata quality profilingmean relative errorproxy-guided samplingcategorical defects
Explainable AI for Chronic Kidney Disease Prediction Using Simulated Federated Learning
The study proposes an explainable federated learning framework for chronic kidney disease prediction, combining VotingClassifier with Random Forest, AdaBoost, and XGBoost. Federated Learning (FL) was employed to train models on distributed clinical data, with GridSearchCV optimizing client-side hyperparameters. Explainable AI (XAI) techniques provided model interpretability. The global FL model achieved 99% average accuracy, demonstrating the viability of privacy-preserving, interpretable CKD diagnosis systems.
federated learningvotingclassifierexplainable aigridsearchcvchronic kidney disease
Cardiologent: Multi-Agent Clinical Decision Support for Patient-Level Arrhythmia Assessment, Urgency, and Management
Cardiologent introduces a multi-agent system for patient-level arrhythmia assessment, addressing the gap between ECG signal detection and clinical decision-making. The system employs specialized agents for signal analysis (ECG lead, photoplethysmogram), grounding readings in measured features, assembling a rhythm profile, and reasoning against retrieved clinical guidelines with a critic for validation. Evaluated on integrated diagnosis, clinical significance, and management, Cardiologent outperforms baselines, achieving expert-aligned decisions (ICC 0.74 with cardiologists) while maintaining auditability via guideline citations.
multi-agent systemarrhythmia assessmentclinical decision supportecg analysisguideline retrieval
Dual-Domain Manifold Modeling for Hyperspectral Image Fusion
The paper proposes Dual-Domain Manifold Modeling (DDMM), a hyperspectral image fusion framework addressing spatial-spectral integration challenges. It introduces a Topology-Aware Transformer (TPFormer) for joint spatial topology and pixel-level manifold modeling, and a Frequency-Decoupled Spatial-Spectral Collaborative Fusion (FDSCF) module that separates features into low/high-frequency components via DCT for selective enhancement. Experiments on benchmark datasets show DDMM outperforms state-of-the-art methods in spatial structure preservation and spectral reconstruction.
hyperspectral fusionmanifold modelingtopology-aware transformerfrequency decouplingspatial-spectral coupling
Every Time I Hire a Linguist, Inference Costs Go Down: On Linguistic Rules as Effective Prompt Compressors
The study demonstrates that linguistic rules alone can effectively compress LLM prompts without LM-based token scoring, reducing inference costs. An evolutionary search over lexical, syntactic, semantic, and discourse rules yields a CPU-only compressor, evaluated via a dual-path protocol balancing compression quality and reconstruction fidelity. Evolved compressors match advanced prompt-compression strategies on short passages, multi-document reasoning, and dialogue-memory QA, with performance strongest under light-to-moderate compression. Analysis shows effective compression fuses multi-level linguistic signals, shifting from token pruning to sentence extraction at higher ratios.
prompt compressionlinguistic rulesevolutionary searchinference costdual-path protocol
Specula: Scaling formal specifications for autonomous model checking of system code
Specula introduces an autonomous system for generating formal TLA+ specifications and model checking complex system code using LLM-based agents. The method employs self-evolving loops to iteratively refine specifications, mitigating LLM limitations like reward hacking and hallucinations. Evaluated on 48 open-source projects, Specula detected 249 bugs, including deep, hard-to-find issues, demonstrating effectiveness beyond traditional human-centric formal methods.
formal specificationsmodel checkingtla+llm agentsautonomous verification
Physics-Informed Neural Operator for Warm-Starting Background-Decomposed and Preconditioned PSFD: Enabling Scalable 3-D EUV Mask Simulation
The authors propose a physics-informed neural operator (PINO) for accelerating 3-D EUV lithography simulations by solving electromagnetic scattering problems via pseudo-spectral frequency-domain (PSFD) equations. The Fourier neural operator is factorized into 2D lateral and 1D axial branches, trained self-consistently with background decomposition to retain full-vector coupling without finite-order Born approximation. Trained on 16,000 LithoBench mask designs without precomputed solutions, PINO achieves a mean absolute error of 7×10⁻³ in scattered intensity predictions versus reference PSFD solutions. Combined with spectral damping, PINO warm-start initialization enhances solver efficiency on finer discretizations.
physics-informed neural operatoreuv lithographypseudo-spectral frequency-domainelectromagnetic scatteringfourier neural operator
Balanced Soft mixture-of-expert model for Glaucoma Detection
The authors propose a balanced soft mixture-of-experts model with three experts and load balancing loss to address imbalanced uni-modal representations in multi-modal glaucoma detection. The method outperforms uni-modal baselines, conventional multi-modal models, and state-of-the-art balanced multi-modal models, as measured by AUC. The approach is generalizable to other disease detection tasks like diabetic retinopathy.
mixture-of-expertsmulti-modal learningload balancing lossglaucoma detectionauc
From Cellular Responses to Pharmacological Domains: Multimodal Zero-Shot Drug Representation Learning
The study introduces PMRD, a pharmacological response domain-guided framework for multimodal zero-shot drug property prediction, addressing limitations in current multimodal drug representation learning. PMRD separates mechanism-consistent factors from modality-specific noise, constructs consensus response domains across gene expression, cell morphology, and chemical structure, and employs mechanism candidate augmentation with retrieval-geometry attribution for dynamic objective reweighting. Experiments on public datasets demonstrate improved zero-shot prediction accuracy and biologically coherent drug neighborhoods, with hard-negative analysis confirming reduced conflicts between structurally dissimilar but response-related compounds.
multimodal learningzero-shot predictiondrug representationmechanism-awareretrieval-geometry attribution
Physics-Grounded Fluid Video Generation with a Simulation Dataset and Dual-Stream Optical-Flow Supervision
The authors address physics violations in fluid video generation by proposing a physics-grounded dataset and dual-stream optical-flow supervision. They construct a dataset combining 1,638 MPM-simulated and 2,320 real pouring videos, alongside two test sets. Their method introduces a dual-stream diffusion-transformer architecture with an RGB decoder and an optical-flow decoder, trained with end-point-error and smoothness losses, fused via zero-initialized convolutions. Evaluated on 1.3B and 14B models, the approach improves VideoPhy-2 Physical-Commonsense and Video-Quality scores by up to 8.75 and 4.65 points, outperforming competitors and achieving 0.54-pixel end-point error in-distribution.
fluid video generationoptical-flow supervisiondiffusion-transformerphysics-simulation datasetend-point-error
CAST: Game Solvers as Turn-Level Teachers for LLM Agents
CAST (Credit Assignment from Solver Teachers) introduces turn-level credit assignment for LLM agents in long-horizon games by leveraging game solver state-value changes as dense process signals. The method converts solver value differences into advantages, enabling on-policy distillation from the solver under a soft-optimality assumption without requiring teacher logits. Evaluated on Sokoban, Minesweeper, Rush Hour, ALFWorld, and WebShop, CAST outperforms all baselines in both in-domain and zero-shot settings, achieving superior performance across all games.
credit assignmentgame solverson-policy distillationlong-horizon gamesllm agents
Retraction-Free Optimization over the Stiefel Manifold for the LoRA Fine-Tuning
The authors propose a retraction-free, penalty parameter-free algorithm for optimization over the Stiefel manifold, avoiding costly orthonormalization and careful hyperparameter tuning. By exploiting the quadratic penalty function's strongly-convex-like property and the manifold's proximal smoothness, they establish global convergence with best-known iteration complexities. They reformulate LoRA fine-tuning as a manifold optimization problem, introducing Manifold-LoRA, which uses their landing technique and step size strategy to accelerate training. Experiments on benchmarks show improved efficiency and downstream performance.
stiefel manifoldlow-rank adaptationmanifold optimizationquadratic penaltyglobal convergence
Hybrid Analysis for Secure MCP Tool Use in LLM Agents
MTGuard introduces a hybrid analysis framework for securing Model Context Protocol (MCP) tool use in LLM agents, addressing limitations of static defenses by combining lifecycle-aware static-dynamic co-analysis. The method mitigates harmful tool-invocation risks while preserving benign task performance, as validated across multiple LLM agents and threat categories.
model context protocolllm agentshybrid analysisstatic-dynamic co-analysistool-use security
CLBench-V: Evaluating Multimodal Context Learning from Grounding to Knowledge Acquisition
The paper introduces CLBench-V, a benchmark for evaluating multimodal context learning across three dimensions: context grounding, new information application, and new knowledge learning. The benchmark combines converted public datasets with newly constructed ones spanning science, finance, and spatial reasoning, using automated construction methods to reduce domain-specific task costs. Evaluation of six multimodal models on 3,443 instances reveals low overall performance (best score: 0.2847), with InternVL3.5-30B-A3B excelling in grounding and knowledge learning, while Qwen3.5-Plus leads in information application. Analysis includes judge reliability, context length, and failure cases.
multimodal context learningbenchmark evaluationcontext groundingknowledge acquisitionautomated dataset construction
Instruction-Tuned Language Models Cannot Sample from Distributions They Can Describe
This work demonstrates that instruction-tuned language models fail to sample from response distributions despite accurately describing them, a phenomenon termed the KNOWS/DOES split. Through systematic evaluation across three model families on public-opinion benchmarks, the authors show that instruction-tuned models collapse to deterministic outputs, with identical responses occurring in over 50% of cases. Analysis reveals this collapse stems from a degenerate sampling primitive induced by alignment training. To address this, the authors propose Prompt-Perturbed Argyle (PPA), which reduces error against human survey data by 21% without additional computational cost.
instruction-tuningsampling primitiveknows/does splitprompt-perturbed argyledegenerate sampling
ContractHIL-HLS: Contract-Aligned Multi-Agent Workflow with Hardware-in-the-Loop Feedback for HLS Design
ContractHIL-HLS introduces a contract-aligned multi-agent workflow for high-level synthesis (HLS) engineering, featuring three contributions: (1) structured contracts translating natural language requirements into explicit interfaces and validation rules, (2) hardware-in-the-loop feedback integrating HLS, Vivado, and runtime metrics, and (3) task-decomposed agents (Contract, HTML, and Hardware-in-the-Loop). On 94 HLS-Eval tasks, the method improves testbench pass rates from 64.0% to 70.2% (pass@1) and 76.6% (pass@5). For a post-quantum cryptography accelerator, it reduces average runtime from 207.3 ms to 52.4 ms while preserving verification.
high-level synthesishardware-in-the-loopmulti-agent workflowcontract-alignedpost-quantum cryptography
Many-body Tipping Dynamics of ChatGPT-like AIs
The study identifies many-body tipping dynamics as the cause of undesirable content generation in ChatGPT-like AIs under deterministic greedy decoding. By modeling token interactions as spins in a finite-layer system, the authors demonstrate that tipping arises from a dynamical first passage process between competing output basins, with attention disorder governing transport. A few-basin reduction yields a closed finite-layer threshold, validated by coarse-grained predictions across AI families. These findings classify such failures as foreseeable engineering risks, impacting legal and societal harm assessments.
tipping dynamicsmany-body interactionsattention disorderfinite-layer systemgreedy decoding
ScaleResfusion: Residual Rectified Flow based on Residual Vector Field
ScaleResfusion introduces Residual Rectified Flow, a scalable diffusion framework for real-world image restoration that leverages pre-trained text-to-image rectified-flow models. The method incorporates a residual term R into Standard Rectified Flow, enabling parameter-efficient fine-tuning by learning a residual vector field that maintains consistency with pre-trained models. It starts from noisy low-quality images rather than pure noise, admitting an exact acceleration point. A knowledge-distillation pipeline further reduces sampling costs. Experiments demonstrate state-of-the-art performance across multiple real-world restoration tasks with improved efficiency, showcasing a practical approach to adapting large pre-trained diffusion models.
residual rectified flowimage restorationdiffusion modelsparameter-efficient fine-tuningknowledge-distillation
Bridging Compute- and Data-Optimal Pretraining
The paper introduces Compute-Data (CD) scaling laws, a framework unifying compute-optimal and data-optimal pretraining regimes by incorporating a token-effectiveness function η. This function quantifies the value of derived tokens (via multi-epoch repetition or paraphrasing) relative to fresh tokens, fitted across model sizes (14M–600M parameters) using the Dolma-3 corpus. Results show η depends on model size, tokens-per-parameter ratio, and derived data volume, exhibiting diminishing returns when substituting compute for data. The analysis identifies three operational regimes (compute-bound, data-bound, model-bound) and demonstrates suboptimality of classical compute-optimal allocation in practical settings.
scaling lawstoken-effectivenesspretrainingcompute-optimaldata-optimal
Where Steering Signals Come From: Activation Source Selection in Activation Steering
The paper investigates activation source selection in activation steering, focusing on how the choice of source context and readout policy affects steering success in language models. The authors analyze three instruction-tuned models across four steering task families, demonstrating that execution-boundary states—where the model is about to produce target behavior—yield stronger steering signals than mere appearance of the desired behavior in source text. They introduce tail subtraction to isolate boundary-state representations, improving steering stability. Results indicate steering efficacy depends on pre-execution representations rather than post-realization semantics.
activation steeringexecution-boundary statestail subtractioninstruction-tuned modelshidden states
Structure-aware Relative Policy Optimization for Ranking
The paper introduces SRPO (Structure-aware Relative Policy Optimization), a reinforcement learning framework for listwise ranking that addresses structural relationships between permutations. SRPO employs a top-weighted Kendall-tau distance to measure permutation discrepancies and normalizes pairwise reward differences by these distances, emphasizing efficient local refinements, particularly in top-ranked positions. Experiments across two ranking scenarios show SRPO improves effectiveness and stability, especially in limited-feedback and complex list-level optimization settings, by better modeling permutation-level differences.
reinforcement learninglistwise rankingkendall-tau distancepolicy optimizationcredit assignment
Laplace-PSN-IRT: Uncertainty Quantification for Neural Item Response Theory Models of LLM Benchmarks
Laplace-PSN-IRT introduces a post-hoc last-layer Laplace approximation to PSN-IRT, enabling approximate Bayesian posterior inference for uncertainty quantification in neural Item Response Theory models of LLM benchmarks. The method recovers calibrated uncertainty over model ability and item difficulty without retraining, supporting credible intervals, probabilistic model comparisons, and Fisher-information-based item selection with propagated parameter uncertainty. Results show that most pairwise comparisons among 12 models on a standard LLM benchmark are statistically indistinguishable despite differing point-estimate ranks, and posterior-expected Fisher information outperforms point-estimate versions in stability and ability ranking recovery. The approximate posterior demonstrates calibrated coverage when modeling item difficulty as random while fixing item discrimination.
item response theorylaplace approximationuncertainty quantificationfisher informationlarge language models
The User Asks, Platforms Compete: How Agentic Recommendation Markets Take Shape
The paper introduces agentic recommendation markets, where LLM-based user agents enable pre-platform recommendation by specifying needs before platform selection, shifting competition dynamics. Through controlled experiments across three product domains, the study contrasts traditional platform-centric with user-centric recommendation, revealing a tension between access and attention. Results show selective positive explanations dominate 73–78% of first-ranked positions, reducing to 36–41% when user feedback is incorporated, while purchase likelihood increases. The work frames agentic recommendation as a joint mechanism design problem involving access, attention, and accountability.
agentic recommendationllm-based user agentsplatform-centric recommendationmechanism designselective explanations
CADENCE: A Cardiac Atom Dictionary for Interpretable Neural Concept Extraction from ECG Foundation Models
CADENCE introduces an interpretable framework for decomposing ECG foundation model representations into sparse cardiac atoms using a BatchTopK sparse autoencoder. The method factorizes Layer-6 embeddings from 9M+ ECG tokens into 8,192 sparse atoms, which align with clinical phenotypes and waveform morphology. Results show atoms achieve mean AUROCs of 0.88 (phenotypes) and 0.90 (morphology), outperforming dense dimensions (0.78, 0.83), and sparse probes improve phenotype AUROC from 0.93 to 0.95 while enabling attribution to interpretable concepts.
sparse autoencoderclinical phenotypeswaveform morphologyinterpretable conceptsfoundation models
Everyone is unique: Towards Behaviorally Heterogeneous Negotiation Dialogue Systems for Debt Collection
The paper introduces DebtBench, the first public persona-enriched benchmark for debt collection negotiation, addressing behavioral heterogeneity in dialogue systems. It also presents DebtGPT, a debt collection agent optimized for financial recovery and interaction experience. Evaluated against 16 state-of-the-art LLMs, DebtGPT matches GPT-4o's performance and surpasses open-source baselines, highlighting the challenges of realistic negotiation scenarios for existing models.
debt collectionbehavioral heterogeneitydialogue systemsllm evaluationnegotiation benchmark
TopoGR: Revealing and Preserving Latent Structure of Semantic ID in Generative Recommendation
TopoGR introduces a topology-preserving generative recommendation framework that addresses structural mismatch in semantic ID (SID) spaces by leveraging Bit-decomposable Binary SIDs with explicit Hamming geometry. The method preserves Hamming proximity through binary SID features, topology-aware supervision via Hamming soft targets, and Hamming-consistent reranking during inference. Experiments on four benchmarks demonstrate that TopoGR outperforms state-of-the-art baselines by capturing item relatedness beyond exact SID matching.
topology-preservingsemantic idhamming geometrygenerative recommendationbit-decomposable
VaLiDRec: Variable-Length LLM-Aligned Semantic IDs for Generative Recommendation
VaLiDRec introduces variable-length, LLM-aligned semantic identifiers (SIDs) for generative recommendation, addressing limitations of fixed-length SIDs via token importance estimation, semantic-quality-aware pruning, and collision-aware refinement. The framework employs graph-aware soft prompts and token-set prediction to model user preferences without autoregressive decoding. Evaluations on four datasets demonstrate consistent outperformance of sequential and generative baselines, with 87.49× faster inference than LC-Rec and superior zero-shot cold-start performance, validating the expressiveness and efficiency of LLM-native variable-length SIDs.
semantic identifierstoken importance estimationgenerative recommendationzero-shot cold-starttoken-set prediction
RIDGE: An Autonomous Framework for Validation and Method Discovery in LLM-Generated Option Pricing
The paper introduces RIDGE, an autonomous validation framework for LLM-generated option pricing implementations, addressing the challenge of ensuring mathematical consistency and numerical stability beyond conventional testing. The framework employs structured no-arbitrage tests, stress tests, benchmark comparisons, and consistency checks, with validation evidence used diagnostically to refine both implementations and methodology. Applied to five stochastic volatility models, RIDGE successfully removed all detected defects and, in two cases, led to novel semi-analytic pricing methodologies.
option pricingno-arbitrage testsstochastic volatilitynumerical stabilitysemi-analytic methods
TabRank: Chain-of-Thought Distillation for Table Re-Rankers
TabRank introduces a framework for training reasoning rerankers in tabular retrieval via chain-of-thought (CoT) distillation. The method leverages a dataset of 6728 reasoning traces from the Natural Questions Tables dataset, exploring two training variants: explicit CoT distillation and conditioning the student reranker on the teacher's reasoning trace. Evaluated across multiple out-of-distribution settings and multi-table scenarios, TabRank demonstrates significant improvements, increasing Acc@10 by 30.5% on HybridQA, 15.2% on SQA, 52.9% on TabFact, and 13.1% on TATQA subsets of the Multi-Table QA Benchmark.
chain-of-thoughttabular retrievalreasoning tracesrerankersmulti-table reasoning
CondPSE: A Polynomial-Filtered Structural Encoder with Conditional Modulation for Graphs
CondPSE introduces a learned positional/structural encoder for graphs that combines polynomial graph filters with conditional modulation to enhance structural discrimination beyond 1-WL-bounded message passing. The method applies a learnable polynomial filter bank to Gaussian node probes, refines responses via FiLM-style modulation conditioned on cross-filter and graph signals, and pretrains to reconstruct node/graph-level targets. On synthetic benchmarks (CSL, EXP), CondPSE improves accuracy from 42.9% to 97.3% and 68.3% to 99.9% over GPSE, primarily due to the polynomial filter bank. However, on molecular property prediction (ZINC), it matches but does not surpass GPSE, highlighting a synthetic-to-real gap in frozen PSE encoders.
graph neural networksstructural encodingpolynomial filters1-wl testfilm modulation
Observing sycophantic AI validate others reduces its appeal but not its persuasiveness
The study examines whether awareness interventions mitigate the harmful effects of sycophantic AI, which exhibits excessive agreeableness and flattery. Two preregistered experiments (n=940 and n=650) tested written warnings and observational exposure to sycophantic AI behavior, respectively, while pooling data with prior studies (six interventions total, n=3,982). Results show interventions reduced perceived AI objectivity and enjoyment (mediated by diminished belief in earned validation) but failed to decrease persuasiveness, suggesting individual-level protections like warnings or AI literacy may be insufficient against AI harms.
sycophantic aiawareness interventionspersuasivenessobjectivityai literacy
OrganLens: Organ-Specific Representation Learning for CT Foundation Models
OrganLens introduces organ-specific representation learning for CT foundation models via self-supervision, addressing limitations of volume-level or anatomy-aware approaches. The method conditions a shared CT encoder on organ identity, using organ-specific distillation and anatomy-mask supervision to produce 11 organ-specific representations without external segmentation masks. Evaluated on CT-RATE, RAD-ChestCT, INSPECT, and NLST, OrganLens improves cardiomegaly AUROC to 0.953 (from 0.910 with DINOv2) and lung-cancer mortality Harrell C-index by 14.2%, while maintaining global representation utility (INSPECT Recall@10: 33.09% text-to-image, 32.04% image-to-text).
representation learningself-supervisionct foundation modelsorgan-specificanatomy-mask supervision
PreDiff-LM: Pretrained Discrete Masked Diffusion Language Modeling with Hybrid Attention
PreDiff-LM introduces a hybrid attention mechanism for adapting pretrained autoregressive transformers to discrete masked diffusion language models, preserving causal attention in observed prompts while enabling bidirectional attention in masked targets. This approach, combined with a DiffuGPT-style objective adaptation, reduces perplexity from 34.1 to 26.9 and improves MAUVE from 0.71 to 0.78 on WikiText-103 under a GPT-2 Medium setup. Pretrained initialization significantly reduces training steps from 350K to 8K to achieve perplexity below 50. PreDiff-LM outperforms prior diffusion baselines in repetition, distributional quality, zero-shot downstream tasks, and human preference, though it remains weaker than fine-tuned autoregressive models at equal scale.
hybrid attentiondiscrete masked diffusionautoregressive transformersperplexityzero-shot tasks
When Do Agent Loops Mistake Stagnation for Progress? Self-Evaluation Bias and Externally Grounded Verification in Long-Running Autonomous LLM Agent Loops
The paper identifies 'progress mirage' as a failure mode in long-running autonomous LLM agents, where self-evaluation bias leads agents to mistake plausible changes for genuine progress despite real-world stagnation or regression. Using a controlled testbed with 54 cycles, the study manipulates evaluator grounding (world-state oracle vs. self-report) and measures performance deltas. Results show self-report was uninformative (56% zero/negative delta), while even strong in-band judges accepted 44% regressions and rejected 38% improvements. The gap vanished when success signals were artifact-verifiable, demonstrating that out-of-band evaluation is structurally necessary for open-ended tasks.
self-evaluation biasprogress mirageautonomous agentsin-band judgeout-of-band evaluation
Inferring Missing Trajectory Data with Temporal Convolutional Networks
The authors propose a Temporal Convolutional Network (TCN) with symmetric dilation for trajectory inpainting, addressing incomplete real-world trajectory data due to sensor failure or occlusion. The architecture relaxes causality constraints, enabling bidirectional context utilization (past and future observations) unlike forecasting-oriented models. Training employs a composite loss combining weighted mean squared error, boundary-continuity penalties, and smoothness regularization. Evaluated on a synthetic dataset (1,000 train, 200 validation, 300 test trajectories) with 20% masked segments, the model demonstrates strong performance in R², MSE, and MAE metrics.
temporal convolutional networktrajectory inpaintingsymmetric dilationcomposite lossbidirectional context
Agentic AI for Scientific Reasoning in Autonomous Quantum Sensing Experiments
The study presents an agentic AI workflow using a large language model (LLM) for autonomous quantum sensing experiments with nitrogen-vacancy (NV) centers in diamond. The workflow integrates persistent project records, quantitative tools, and deterministic experiment control, demonstrated by autonomously selecting an NV center, calibrating its resonant frequency, and performing Ramsey and CPMG measurements. Two offline benchmarks evaluated reasoning separately from lab execution, testing GPT-5.4, GPT-5.5, and GPT-5.6 Sol. Results showed improved resonance offset recognition with greater reasoning effort in Ramsey benchmarks but higher false positives in pODMR evaluations without signal calculations. The findings advocate for agent-driven hypothesis formation alongside deterministic hardware control.
agentic aiquantum sensingnitrogen-vacancy centersautonomous experimentslarge language model
How Affect Propagates among LLM Agents: Emergent Emotional Contagion in Crowd Simulation
The paper demonstrates emergent emotional contagion in LLM-based multi-agent crowd simulations, where affect propagates via perception-appraisal-expression loops without explicit transfer mechanisms. Agents perceive neighbors through multimodal channels, appraise stimuli via LLMs (conditioned on personality profiles, memory, and context), and update internal states using Russell's circumplex model, while navigation is offloaded to a conventional crowd simulator. Experiments across five scenarios show spatially structured contagion dynamics: alarm propagates as traveling fronts, stabilized alarmed fractions, and personality-dependent interpretations of ambiguous stimuli. Ablation studies reveal backend-dependent dynamics across four LLMs, prompt variants, and sampling temperatures.
emotional contagionmulti-agent simulationcircumplex modelllm appraisalcrowd dynamics
Less Data, Better Alignment: Data-Centric Multi-Evaluator Agreement for Preference Optimization
The paper introduces DMAPO, a data-centric method for preference optimization that selects high-consensus examples from on-policy responses. DMAPO generates candidate responses using Mistral-7B, evaluates them with rubric-specialized evaluators, applies process-critic correction, and retains only high-agreement examples (3.45% acceptance rate). KTO trained on this subset achieves 7.50 on MT-Bench, 95.5% win rate against text-davinci-003, and 57.3% IFEval accuracy. Independent evaluations show DMAPO outperforms SimPO by 23.3-24.1 net win rates with GPT-4o and Claude Opus. The method demonstrates robustness to evaluator or rubric changes, offering data-efficient alignment at increased curation cost.
preference optimizationon-policy responsesconsensus filteringrubric-specialized evaluatorsprocess-critic correction
ScalableRAG: High-Quality RAG at Zero Ingestion Cost
The paper introduces ScalableRAG, a retrieval-augmented generation system achieving high performance with zero ingestion costs, outperforming knowledge graph baselines by 7.36% average accuracy across six corpora. It maintains document and value sets for on-the-fly aggregative reasoning without vector databases, capped by constant LLM calls. A Limited-Ingestion variant adds minimal vector storage and automated pattern discovery for improved scalability.
retrieval-augmented generationzero ingestion costaggregative reasoningknowledge graphsvector database
Learning from 53.6K Real-World Developer Edits of AI-Generated Code
The paper introduces DECODE, a dataset of 53.6K real-world in-IDE code edits (Python, TypeScript, JavaScript) from 1K+ developers, addressing the lack of high-quality data for training LLMs on AI-generated code modifications. Unlike Git commits, DECODE captures granular editing behavior, revealing that 31% of edits remove AI completions, mostly within 15 minutes of acceptance. Finetuning on DECODE enables 3B-parameter open-source models to outperform frontier LLMs in code edit prediction. The work advocates developer-centric ML approaches for AI programming assistants.
code editingai-generated codedeveloper behaviorllm finetuningide interactions
Analysis of the Shortcut Learning and Clever Hans Effect in CNN based ECG Image Classification
The study investigates shortcut learning and Clever Hans effects in CNN-based ECG image classification by analyzing six feature sets with controlled representations (raw images, cropped waveforms, masked metadata, artificial artifacts). Using metrics like shortcut retention score and attribution methods (Integrated Gradients, occlusion sensitivity), it evaluates whether models rely on clinically relevant waveform morphology or non-physiological cues (layout, metadata, artifacts). Results reveal performance variations across feature sets, indicating potential reliance on non-clinical shortcuts rather than ECG waveform features.
shortcut learningclever hans effectecg image classificationconvolutional neural networksattribution methods
OPERA: Offline Policy-guided Expert Routing and Adaptation for Universal Biomedical Image Analysis
The authors propose OPERA, a multi-agent ensemble framework for universal biomedical image analysis that addresses deployment challenges caused by distribution shifts. OPERA treats expert weight assignment as an offline policy learning problem, combining offline routing policy learning with test-time adaptation. The framework includes expert profiling, confidence calibration, distribution-aware adaptation, and instance-level routing. Evaluated on 9 datasets across fundus photography, chest X-ray, CT, MRI, and multimodal benchmarks, OPERA outperforms 30+ baselines in classification, segmentation, and multimodal tasks while improving calibration quality.
offline policy learningtest-time adaptationmulti-agent ensembledistribution shiftconfidence calibration
Evaluating Communicative Belief Updates in Large Language Models via Implicature Recognition and Cancellation
This paper introduces [DatasetName], the first expert-annotated dataset for evaluating implicature recognition and cancellation in large language models (LLMs), focusing on unspoken belief updates. Using crowdsourced human judgments, the study assesses LLMs' ability to recognize implicatures and understand their cancellation, where implied meanings are weakened or negated. Results indicate that LLMs lag behind humans in belief update understanding, particularly in natural scenarios, with performance influenced by prior beliefs, update types, and forms. The findings suggest current LLMs have not achieved human-level comprehension of unspoken beliefs and updates.
implicature recognitionbelief updateslarge language modelsimplicature cancellationpragmatic phenomenon
Towards Robust Reinforcement Learning for Small-Scale Language Model Agents
This work identifies and mitigates three failure modes in reinforcement learning (RL) alignment for small language models (SLMs, 70M-500M parameters): LoRA parameter freezing, bfloat16 overflow, and catastrophic policy collapse. The authors propose merge-and-reinitialize adapters, float32 PPO updates, and a three-layer safety mechanism (reward whitening, importance-ratio guarding, weight rollback) across experiments with Pythia-70M/160M/410M and SmolLM2-135M/360M on TinyStories, CNN/DailyMail, and Wikitext-103. The stabilized system achieved consistent PPO convergence, improved preference win rates over SFT baselines when given fluent priors (PPL<20) and informative rewards, and outperformed instruction-tuned models with less data, supporting their capacity-headroom hypothesis.
small language modelsproximal policy optimizationparameter-efficient finetuningreward whiteningcapacity-headroom hypothesis
Matryoshka Agent: Unfolding Sub-Agents for Long-Horizon Machine Learning Engineering
The paper proposes Matryoshka Agent, a hierarchical framework for long-horizon machine learning engineering (MLE) tasks, addressing challenges of long-context reasoning and exploration in vast solution spaces. The method decomposes problem-solving into a high-level Orchestrator for strategic planning and lower-level Sub-Agents for execution via standardized Tool interfaces, decoupling exploration from costly environment interactions. Experiments show Matryoshka Agent improves performance, enabling Qwen3-4B-Instruct to match o4-mini's Orchestrator performance and yielding up to 36.7% relative gain for Qwen3-30B-Coder across diverse MLE tasks.
hierarchical agentlong-horizon reasoningmachine learning engineeringorchestratorsub-agents
PLATO: Pointer Learner for Agent and Task Openness
PLATO introduces a pointer-network-based actor and GNN critic for open agent systems (OASYS) with agent (AO) and task openness (TO). The pointer actor outputs distributions over dynamic task sets, avoiding masking or retraining, while the GNN critic encodes agent-task interactions as a mutable graph. Formulated under Task-and-Agent-Open Markov Game (TaAgO-MG), PLATO outperforms baselines in the MOASEI wildfire suppression domain, showing robust zero-shot generalization.
open agent systemspointer networkgraph neural networkmulti-agent reinforcement learningzero-shot generalization
Towards an Agent Operating System - Lessons from Classical and Cloud OS
The article proposes deriving core abstractions for agentic AI systems by extending classical and cloud OS primitives to stochastic, natural-language-mediated execution, analogous to POSIX and Kubernetes. It identifies the current proliferation of frameworks as an experimentation phase lacking consensus, hindering portable applications and reliable platform composition. The authors advocate for precise semantic specification and consolidation around these abstractions to advance beyond prototype deployments.
agentic aios primitivesstochastic executionnatural-language mediationplatform consolidation
Spectral Truncation in Synthetic Control
The paper introduces Spectral Synthetic Control (SC), a method matching treated units in coordinates defined by leading temporal singular vectors of donor panels, and a hybrid estimator combining raw-path SC and truncated Spectral SC. Theoretical analysis shows the method reduces to raw-path SC at full rank, with underdetermined balance conditions for retained dimensions. Empirical evaluation across 11 data-generating regimes reveals truncated Spectral SC has significantly higher RMSE (4-11 Monte Carlo standard errors) compared to tuned raw-path SC, while the hybrid estimator often selects raw-path matching. Performance gaps are sensitive to preprocessing, nearly disappearing after removing unit and time fixed effects. Findings suggest spectral matching's utility depends on basis-estimation noise, balancing underdetermination, and fixed-effects contamination.
synthetic controlspectral truncationtemporal singular vectorsfixed effectsrmse
DS@GT ARC at CheckThat! 2026: LLM-Based Trace Ranking and Grouped Reward Modeling for Multilingual Numerical Claim Verification
This work presents a system for CLEF 2026 CheckThat! Task 2, addressing multilingual numerical claim verification through LLM-based trace ranking and grouped reward modeling. Two approaches are explored: (1) fine-tuning an LLM verifier using LoRA for independent trace scoring with Best-of-N selection, and (2) a lightweight TF-IDF reward model with numeric-temporal features for grouped trace scoring. Adaptive sub-claim decomposition is tested but found ineffective. Evaluations on English and Arabic datasets show the LLM approach achieves superior Recall@5, while the reward model excels on Conflicting claims. For Arabic, AraBERT outperforms a general multilingual baseline across most metrics.
loratf-idfbest-of-nadaptive decompositiontrace ranking
How Often Should a Recommender Call an LLM? Value-Weighted Routing, Monitoring, and Seasonal Robustness
The paper introduces Value Router, a synthetic simulation framework for cost-aware routing between heuristic and LLM-based recommendation paths, emphasizing the distinction between difficulty and business value. The method evaluates a value-weighted threshold router against difficulty-only and random baselines in a synthetic retail catalog with inverse category volume-value correlation. Results show the value-weighted router matches difficulty-only recall (60%) for high-value items while improving precision (98.3% vs. 94.3%), though monitoring reveals aggregate metrics mask between-category biases. Seasonal robustness tests under simulated Black Friday demand (2.5× volume, value shift) compare static, tuned, and budget-aware routing policies.
recommendation systemsvalue-weighted routingsynthetic simulationcost-aware routingseasonal robustness
Addressable Recall Compaction for Long Context-Window Control in AI Agents
The paper introduces Addressable Recall Compaction (ARC), a context-management framework for long-horizon LLM agents that separates archival storage from active-context presentation. ARC stores tool observations in an ID-addressable log and replaces older entries with compact citations, enabling on-demand recall via identifiers. Evaluated on Qwen3-8B (16k context) and Qwen3-32B (32k context), ARC achieves 99.40% exact-answer accuracy on Needle-in-a-Haystack (vs. 88.12% baseline) and 29.97% on LongBench-v2 Hard (vs. 28.25% baseline), while reducing serving time and HBM traffic.
long-context llm agentsaddressable recallcontext compactiontool observation storagehbm traffic optimization
Similar Models Learn Differently: Final-Window Pretraining Shapes Post-Training Beyond SFT
The study demonstrates that final-window pretraining data (500M tokens, 0.1-1% of total) selectively influences model alignment responses despite near-identical post-SFT benchmark performance. Six branches of a partially pretrained model were fine-tuned on distinct data sources (e.g., safety text, web text) before identical SFT and post-training (DPO/RL). While initial refusal rates differed by ≤1%, the safety-text branch retained significantly higher refusal rates post-alignment (+20-30% vs. other branches). This effect required late-stage exposure and generalized across model families, suggesting pretraining chronology impacts alignment robustness.
supervised fine-tuningpreference optimizationpretraining imprintrefusal ratesalignment robustness
Lantern: Conflict-Aware Gradient Blending for Physics-Guided Diffusion Models in Calorimeter Simulation
Lantern introduces physics-guided diffusion for calorimeter shower simulation by combining denoising with auxiliary physics losses via GradBlend, addressing the limitations of purely statistical objectives. The method employs Correlation Frobenius Distance (CFD) for correlation fidelity and two physics-aware losses: a variance-stabilized voxel residual loss and a graph Laplacian loss. On CaloChallenge Dataset 2, Lantern improves both Fréchet Physics Distance (FPD) and CFD, outperforming task-symmetric gradient blending methods like PCGrad and GradNorm, which inflate FPD by 2-100x. The voxel residual loss requires a terminal denoising-only phase due to gradient conflict, while the Laplacian loss is schedule-insensitive.
diffusion modelsgradient blendingcalorimeter simulationphysics-guided learningcorrelation fidelity
Psychological Influences of Conversational AI: Research and Design Directions for Reducing Harm and Promoting Well-Being
The article proposes research and design directions to mitigate psychological risks and enhance well-being in conversational AI interactions. Analyzing empirical observations and prior work, it identifies both benefits (information access, self-reflection) and harms (emotional dependence, vulnerability amplification) of generalist models. The authors frame testable hypotheses for studying AI influence across general interactions, role-playing, and quasi-therapeutic contexts, acknowledging gaps in longitudinal impact assessment. While some recommendations derive from existing research, others highlight open questions requiring further empirical study of behavioral design patterns.
conversational aipsychological well-beingbehavioral designemotional entanglementlongitudinal impact
Extended Reality as a Mediation Layer for Situated Human Control in Human-Robot Teaming
The paper proposes Extended Reality (XR) as a mediation layer for situated human control in human-robot teaming, enabling real-time inspection and adjustment of robot autonomy in dynamic contexts. Grounded in scenarios like robot-assisted nursing and collaborative assembly, the authors identify four mediation functions (intent-autonomy linkage, plan-judgment alignment, shared control levels, role handover) and derive six design dimensions, including socio-physical constraints and anticipatory recovery. The work outlines a research agenda for XR systems to enhance actionable and accountable robot autonomy in shared environments.
extended realityhuman-robot teamingsituated controlmediation functionsshared autonomy
CogEEGAgent: Toward Autonomous Cognitive EEG Analysis with Grounded Execution and Selection-Aware Verification
CogEEGAgent introduces an autonomous agent for cognitive EEG analysis that combines LLM-based natural language interpretation with deterministic validation to ensure scientific rigor. The system separates semantic interpretation (handled by an LLM) from scientific authority (enforced via MNE-Python-based contracts, preflight checks, and confirmation controls). Evaluations demonstrate improved accuracy over deterministic routers in mapping language to registered analyses (prespecified benchmark) and robust rejection of unsupported claims (outcome-blind campaign). Policy stress tests confirm that held-out confirmation prevents false positives from adaptive search, establishing bounded autonomy for EEG workflows.
eeg analysisllm agentsmne-pythonadaptive searchdeterministic validation
SAFAARI: Schema-Aware Framework for Accelerated Advertiser Response Intelligence
SAFAARI (Schema-Aware Framework for Accelerated Advertiser Response Intelligence) introduces a multi-agent framework for NL-to-SQL systems, addressing schema-linking bottlenecks via specialized content, metadata, and orchestration agents. It proposes SEAL (Schema Evaluation and Accuracy in Language-to-SQL), a composite metric penalizing inconsistent results. Systematic experiments show SAFAARI achieves an 81.66% SEAL score (6.65% improvement over baseline), with gains in datapoint accuracy (5.51%) and schema-linking precision (4.69%). Human-in-the-loop validation confirms adaptability across domains, demonstrating 8x faster development while maintaining accuracy.
nl-to-sqlschema linkingmulti-agent frameworkcomposite metriccustomer support automation
Grounded in Consensus, In Step With Emerging Science: A Consensus-Anchored Multi-Corpus Clinical Chatbot for Long COVID
The study introduces a clinician-facing chatbot for Long COVID (LC) that integrates four evidence sources via retrieval-augmented generation: expert consensus guidance, PubMed literature, registered trials, and living systematic reviews. Consensus guidance anchors responses, while other sources are retrieved in parallel upon user request. In an automated evaluation on 50 clinician queries, the system matched OpenEvidence in mean ratings and exhibited numerically higher scores with lower variability in LLM-judged comparisons.
long covidretrieval-augmented generationclinical decision supportconsensus guidanceliving systematic reviews
Authoring Agent Skills: A Software-Engineering Approach
The paper advocates for applying software-engineering principles to the construction of Agent Skills, reusable procedural knowledge modules for large language model agents, emphasizing single responsibility, low coupling, and token-efficient design. It details the structure and staged loading of skills in Claude Code, contrasting them with alternative behavior-shaping mechanisms like project memory files and subagents through UML modeling. The authors propose an evaluation-driven authoring process, identify common skill-authoring patterns and faults, and discuss trust implications of third-party skills.
agent skillsclaude codeprocedural knowledgetoken budgetuml modeling
Chart-Supported or Model-Supplied? Examining MLLM-Generated Claims for Accessible Visualization
This study examines the evidential basis of multimodal large language model (MLLM) interpretations of visualizations, proposing accessible description systems to distinguish evidence-supported claims from model-supplied interpretations. The authors analyze 1,224 descriptions generated by three MLLMs across 102 visualizations under four input conditions, varying access to images, chart context, and framing. Results show accessible chart context increased DIRECT claims and numeric agreement for some models, while adding images did not consistently improve numeric accuracy. Withheld-context prompts failed to reliably increase cautious language, and the Real-World Significance section remained predominantly SPECULATIVE.
multimodal large language modelsaccessible visualizationnumeric agreementchart contextevidential basis
Localized Anomaly Detection via Differentiable D-vine Copulas
We propose a differentiable D-vine copula framework for localized anomaly detection, enabling gradient-based maximum likelihood estimation and beam-search exploration of pair-copula configurations. The method combines hierarchical decomposition for global anomaly scoring with edge-level explanations, leveraging Mondrian conformal prediction for statistical guarantees. Evaluations on benchmark and real-world datasets demonstrate effective interpretable anomaly detection with uncertainty quantification.
d-vine copulaspair-copula configurationsmondrian conformal predictiongradient-based estimationlocalized anomaly detection
CogArena: A Multimethod Evaluation of Cognitive Ability Structure in Large Language Models
CogArena introduces a multimethod framework for evaluating cognitive ability structure in LLMs, using a procedurally generated 13-paradigm benchmark across five theory-motivated groupings. The study analyzes 55 open-weight models, finding positive paradigm correlations and a common axis explaining ~50% of variance. Targeted scaffolds show a small matched-grouping advantage, but no scaffold-specific contrast survives multiplicity correction. Theory-aligned prompting yields a minor in-battery diagonal tendency, yet stable five-dimensional profiles remain unestablished. CogArena integrates behavioral signatures, covariance, matched interventions, and out-of-family prediction to assess cognitive labels.
cognitive abilitymultimethod frameworktheory-motivated groupingstargeted scaffoldsbehavioral signatures
Calibrated Partial Resets: Preventing Policy Collapse in Continual Reinforcement Learning
The paper introduces Calibrated Partial Resets (CPR), an optimizer for continual reinforcement learning that prevents policy collapse by partially resetting low-utility neurons toward their initialization, scaled by utility. Unlike full resets or uniform decay, CPR maintains plasticity without destabilizing training. Evaluated on SlipperyAnt, Continual MetaWorld, and Continual MinAtar, CPR avoids collapse over 400M steps and outperforms prior reset and decay methods. Ablations show a tunable plasticity-performance trade-off, suggesting utility-scaled reinitialization as a viable continual learning approach.
continual reinforcement learningpolicy collapseneuron resetsplasticity preservationutility-scaling
Understanding Semantic IDs: From Item Representation to Item Selection in Generative Recommendation
The paper investigates Semantic IDs (SIDs) in generative recommendation, revealing that SIDs preserve coarse item organization but lose fine local structure from the encoder, altering 38.4% of exact SIDs despite 99.57% item retrieval accuracy. It introduces Item-Supported Decoding (ISD), a parameter-free inference-time method that leverages user-specific rankings to guide SID prefixes before beam search, improving NDCG@10 by up to 31.2%. Experiments across three Amazon domains show SID filtering reduces plausible recommendations by 70.1%, motivating ISD's lightweight intervention.
semantic idsgenerative recommendationitem-supported decodingautoregressive generationndcg@10
Preliminary Guidelines for Using and Evaluating GenAI Tools to Support Systematic Literature Reviews
The article introduces GUEST (GenAI Use and Evaluation in SLR Tasks), a set of process recommendations for employing Generative AI (GenAI) and Large Language Models (LLMs) in systematic literature reviews (SLRs). Through a rapid review, thought experiments, and expert experience, the authors identify challenges in evaluating GenAI for SLRs and propose guidelines for planning, conducting, and reporting such studies. Results emphasize the necessity of human oversight, highlighting GenAI's potential for cost-effective assistance in repetitive tasks and validation, while cautioning against unsupervised use in rigorous SLRs.
generative ailarge language modelssystematic literature reviewsguest frameworkhuman oversight
Automatic Knowledge Graph Construction and Query for Earthquake Catalogs
The study introduces GraphRAG, the first systematic application of graph-based retrieval-augmented generation to raw earthquake catalog data, enabling open-ended querying without manual structuring. The method constructs queryable knowledge graphs from three distinct catalogs (a reservoir swarm, 2019 Ridgecrest sequence, and 2021 Maduo Mw7.4 aftershocks), validated against catalog-derived ground truth and a rule-based reference graph. Results show that seismology-informed prompt fixes eliminate targeted fabrications and improve mechanism reasoning, while outperforming a vector RAG baseline in catalog-wide summarization and temporal stage comparison. Two key pitfalls are identified for future mitigation.
graphragknowledge graphearthquake catalogretrieval-augmented generationseismology prompts
Generative Distributionally Robust Optimization
Proposes Generative Distributionally Robust Optimization (GDRO), a framework combining arbitrary sampleable conditional generators with worst-case law restriction via the sampler-Sinkhorn pairing. Samplers represent conditional laws exactly while Sinkhorn divergence enables distribution comparison without likelihood access. The method admits finite-sample approximation and differentiable primal-dual implementation, with theoretical guarantees on downstream degradation for Lipschitz losses. Evaluations show 60% reduction in rare-context inventory regret and 50% fewer SocialGAN navigation collisions versus nominal decisions across explicit/implicit generators.
distributionally robust optimizationconditional generatorsinkhorn divergencelipschitz lossesprimal-dual implementation
Stable FP4 Training via Transposition-Invariant Block Quantization
The paper introduces a stable FP4 training framework for large language models (LLMs) by addressing scale inconsistency in tensor transposition via 2D block quantization. The method enforces transposition-invariant scaling, combines truncation-free scaling with stochastic rounding, and uses MXFP8 for attention mechanisms. Evaluated on models up to 7B dense and 30B Mixture-of-Experts, trained on 100B tokens, it matches BF16 performance with <1.3% perplexity degradation, demonstrating practical FP4 training.
fp4 quantizationtransposition-invariant scaling2d block quantizationstochastic roundingmixed-precision training
Multiclass Classification without Labels via Posterior Simplex Geometry
The work extends Classification without Labels (CWoLa) to multiclass settings ($K>2$), demonstrating that optimal class discriminators can be learned from unlabeled mixtures with varying class proportions. By analyzing the $(K-1)$-simplex geometry in mixture-posterior space, the authors propose two methods: post-hoc simplex fitting and a bottleneck architecture, which recover latent class structure without prior knowledge of mixing matrices. Experiments on MNIST, CIFAR-10, and Galaxy10 DECaLS show competitive performance with fully supervised baselines, enabling multiclass discovery in label-scarce domains.
multiclass classificationweakly supervised learningmixture modelsposterior simplexlatent class recovery
ClinFusion: A Vision-Centric Multimodal LLM System for Holistic Medical Understanding
ClinFusion introduces a vision-centric multimodal LLM for medical understanding, addressing key limitations through a compositional cascaded vision encoder with Cascade Spatial-Aware Locality Fusion to unify 2D/3D image processing. The system includes MedIF-Bench for instruction-following evaluation and a region-of-interest-grounded metric for clinically aligned report generation. ClinFusion achieves SOTA on 20/24 benchmarks against open-source medical MLLMs (e.g., Hulu-Med, Lingshu) and outperforms GPT-5.2/Gemini-3-Flash on 13/16 multimodal benchmarks, with radiologist evaluations confirming superior report quality and metric correlation.
multimodal llmcascade fusionmedical imaginginstruction-followingregion-of-interest
Harm is not Universal: Community-Specific Toxicity Detection is Urgently Needed
The paper advocates for community-specific toxicity detection (CTD) in text-to-image generation, demonstrating that universal toxicity detectors fail marginalized groups (35% false negatives for disability communities). Collaborating with disability experts, the authors establish safety guidelines for dwarfism and blind/low-vision communities, evaluating on 2,400 annotated images. General-purpose detectors (F1 0.32-0.37) and large vision-language models perform below chance, while prompt-based adaptation (GPT-4o: F1 0.50-0.78) and parameter-efficient fine-tuning (0.5B-7B models: F1 0.48-0.59) show promise but remain below general toxicity detection benchmarks (F1 ≈0.9).
toxicity detectiontext-to-image generationparameter-efficient fine-tuningvision-language modelszero-shot evaluation
Spend Experts Where You Are Unsure: Confidence-Adaptive Routing for Mixture-of-Experts LoRA
The paper introduces CARE (Confidence-Adaptive Routing of Experts), a method for Mixture-of-Experts (MoE) Low-Rank Adaptation (LoRA) that dynamically routes tokens to experts based on router confidence. CARE activates experts in descending order of router weight until cumulative mass reaches a threshold, with a disagreement-based extension, while maintaining compute budget via a thermostat. Evaluated on LLaMA-3.1-8B and Qwen2.5-7B across commonsense, math, code, and knowledge tasks, CARE outperforms fixed top-k MoE-LoRA at matched compute and matches fixed-k=4 performance with fewer experts. It also improves out-of-distribution detection over baseline methods.
mixture-of-expertslow-rank adaptationconfidence-adaptive routingrouter weightout-of-distribution detection
Re-thinking Mammography Transfer Learning: The Dataset-Informed Transfer Learning (DITL) Framework for Breast Cancer Screening and Lesion Diagnosis
The Dataset-Informed Transfer Learning (DITL) framework enhances mammography classification by integrating dataset-derived difficulty signals with neighborhood-based triplet supervision. DITL introduces two adaptive components: Adaptive Difficulty-Weighted Cross-Entropy (A-DWCE), which assigns per-sample weights based on k-nearest neighbor label purity, and Adaptive Neighborhood Representation Triplet (A-NR-Triplet), which enforces intra-class compactness and inter-class separation. Unlike focal loss, DITL requires no hyperparameter tuning and incurs negligible computational overhead. On the VinDR-Mammo dataset, DITL achieves state-of-the-art performance in whole-image breast density classification, with significant improvements in accuracy, F1-score, and AUC (p < 0.0001). DITL also delivers consistent gains on small ROI datasets, bridging small-scale lesion analysis with large-scale density estimation.
transfer learningtriplet supervisioncross-entropybreast densityk-nearest neighbor
VetClaw: An Edge-Cloud Multimodal Agentic System for Veterinary Disease Screening
VetClaw introduces an edge-cloud multimodal agentic system for veterinary disease screening, combining edge sensing with cloud-based vision-language models. The architecture separates agent interaction (OpenClaw) from workflow orchestration (LangGraph), enabling dynamic tasks like input validation, model invocation, and safety checks. Results demonstrate that symptom-guided multimodal inputs outperform image-only zero-shot classification, transforming static models into safety-aware diagnostic-support systems with tool usage and failure handling capabilities.
edge-cloud systemmultimodal agentzero-shot classificationworkflow orchestrationvision-language model
Reinformed Dreamer: An Asymmetric World Model Efficiently Trained through Latent Guidance
The Reinforced Dreamer algorithm improves upon the Informed Dreamer by introducing latent guidance for asymmetric representation learning in model-based reinforcement learning. It addresses limitations in privileged information representations by leveraging additional supervision during training, enhancing both observation and privileged state representations. Evaluations across multiple benchmarks demonstrate more consistent performance gains over DreamerV3 compared to prior asymmetric approaches.
model-based rlasymmetric learningprivileged informationlatent guidancerepresentation learning
Collaborative System Failure Prognostics via Federated Longitudinal-Survival Modeling
This paper introduces a federated longitudinal-survival modeling framework for collaborative system failure prognostics, addressing challenges in distributed prognostics due to privacy and proprietary constraints. The framework integrates longitudinal sensor representation learning with a client-separable discrete-time hazard objective, enabling multiple clients to jointly train a prognostic model without sharing raw sensor data or failure records. It estimates interval-specific failure hazards, reliability curves, and system RUL using time-dependent representations from multivariate sensor histories. Experiments on the C-MAPSS turbofan engine degradation datasets demonstrate improved prognostic performance over isolated local training, achieving results comparable to centralized training across heterogeneous operating conditions and failure modes.
federated learninglongitudinal-survival modelingsystem failure prognosticsdiscrete-time hazardc-mapss
Parallel Decoding Distillation for Fast Image and Video Generation
Parallel Decoding Distillation (PDD) introduces a trajectory-based distillation method for accelerating diffusion and flow matching models by predicting multiple denoising steps per network evaluation. The approach avoids complex variational score distillation (VSD) and adversarial losses, instead learning a mean velocity representation without derivative regression. PDD achieves state-of-the-art performance with 4-8 network function evaluations (NFE) on LTX-2.3 Text-to-Video/Audio, Wan 14B Text-to-Video, and Qwen-Image Text-to-Image, while significantly improving video diversity.
parallel decoding distillationdiffusion modelsflow matchingnetwork function evaluationstrajectory-based distillation
Sharpness-Aware Minimization and Muon: Robustness under the Spectral Norm
The paper introduces a spectral-norm variant of Sharpness-Aware Minimization (SAM) that incorporates matrix-aware geometry for neural network optimization. The method combines layerwise spectral inner perturbations for hidden-layer weights with either AdamW/SGDW or the Muon optimizer in the outer update. Experiments on ImageNet-1K with ViT-Small/16 and ResNet-50 show that spectral inner perturbations paired with Muon achieve the highest validation accuracy among tested methods, demonstrating the efficacy of matrix-aware optimization in SAM frameworks.
sharpness-aware minimizationspectral normmuon optimizermatrix-aware optimizationimagenet-1k
Untangling Co-Drift: Proactive Multi-Intent Failure Prediction and Root-Cause Disambiguation for Self-Driving Networks
The paper introduces MILD, a framework for proactive failure prediction and root-cause disambiguation in self-driving networks by modeling three operational macro-intents: continuous telemetry, real-time analytics, and programmatic actuation. MILD employs a teacher-augmented Mixture-of-Experts architecture with a hybrid objective for joint intent failure prediction and root-cause attribution, enhanced by SHAP explainability and multi-horizon modeling. Evaluations across controlled benchmarks, microservices, and SDN testbeds show MILD achieves high detection rates (unspecified), strong remediation lead times, and accurate intent-level diagnostics, enabling closed-loop assurance in autonomous networks.
self-driving networksintent assurancemixture-of-expertsroot-cause attributionmulti-horizon modeling
Generator-Aligned Representation Interfaces for Diagnostic Soft Equivariance
The paper introduces Generator-Aligned Representation Interface (GARI), a representation-level design principle that exposes transformation generators to generic sequence backbones via aligned canonical and generator-induced views. GARI formalizes soft-equivariance residuals over declared data distributions, separating representation consistency from task robustness. Instantiated as GARI-Net, the framework constructs generator-indexed streams, processes them with shared parameters, and aggregates using inter-stream discrepancy. Experiments on genomic sequences, images, and 3D point clouds demonstrate task-relevant transformation consistency and generalization to held-out probes without group-specific backbone redesign. Direct Equivariance Error (DEE) provides diagnostic evaluation of prescribed representation relations.
soft-equivariancerepresentation interfacegenerator-alignedsequence backbonedirect equivariance error
Physics-Aware End-to-End Deep Reinforcement Learning for Quadcopter Control with Actuator Dynamics
The paper presents a physics-aware deep reinforcement learning framework for quadcopter control, explicitly modeling actuator dynamics and aerodynamic effects. Using a high-fidelity Simulink simulator with 12-state rigid-body dynamics and first-order motor lag (τ=0.076s), four DRL algorithms (DDPG, TD3, PPO, SAC) are evaluated on thrust-only hover and translated goal tasks. Results indicate SAC and TD3 outperform PPO in stability and sample efficiency, demonstrating the necessity of actuator dynamics modeling for low-level control. The study provides a reproducible benchmark with detailed reward shaping (exponential position well, attitude penalties).
quadcopter controlactuator dynamicsdeep reinforcement learningrigid-body simulationsample efficiency
Schrödinger's Cat: Probabilistic Representation and Prediction of Potential Scene Kinematics
GARFIELD introduces a probabilistic model for predicting scene kinematics by learning structured spatio-temporal latent representations of potential futures from partial observations. The model supports joint trajectory sampling and direct access to motion distributions via an efficient deterministic density decoder, enabling localized uncertainty estimation and progressive refinement through constraints. Experiments show competitive motion planning performance with large video generation models, achieving 97× faster trajectory sampling and two orders of magnitude faster motion density estimation than Monte-Carlo methods, facilitating interactive exploration and uncertainty-aware planning.
probabilistic modelspatio-temporal latent representationmotion planningdensity decoderuncertainty-aware planning
Quasi-SVD: Learning a Lie-constrained matrix factorisation for real-time imaging
Quasi-SVD introduces a differentiable, parallelized matrix factorization framework for GPUs that enforces exact orthogonality on one Lie-parameterized factor while using soft constraints for the other, enabling efficient decomposition without iterative optimization. The method achieves SSIM = 0.89-0.94 reconstruction fidelity and 3-20x speedup over cuSOLVER and randomized SVD, supporting real-time throughput >25 FPS. Evaluated on medical imaging tasks (ultrasound localization microscopy and Mueller matrix polarimetry), it demonstrates robust domain transfer and clinical-scale performance.
matrix factorizationlie-parameterizedgpu accelerationreal-time imagingsingular value decomposition
Can Deep Generative Models Reproduce Non-Stationary Gaussian Random Fields?
The study evaluates whether deep generative models (DGMs) can reproduce non-stationary Gaussian random fields by testing flow matching (FM), DDPM, score-SDE, and VAE on synthetic and ERA5 temperature anomaly data. Using oracle samples and a stationary control, the authors propose metrics to assess recovery of ground-truth mean and covariance structures. Results show all models accurately recover the mean, but covariance performance varies: DDPM and score-SDE perform well, FM exhibits mild non-stationarity attenuation and variance under-dispersion, while VAE struggles with covariance recovery.
deep generative modelsnon-stationary gaussian fieldscovariance recoveryflow matchingscore-sde
HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone
HiFi-UMI introduces a portable data-production system for high-fidelity robot-free manipulation demonstrations, achieving 3 mm end-effector accuracy via stereo-inertial SLAM, microsecond synchronization, and wide-FoV cameras. The method eliminates the need for real-robot anchor data by post-training policies solely on HiFi-UMI data, matching teleoperation performance across three backbones (StarVLA-QwenPI, OpenPI-pi_0.5, LingBot-VA) with success-rate differences ≤3.1 percentage points. Pre-training on 4,000 hours of HiFi-UMI data reduces action error by 41% on unseen tasks and boosts real-robot success by 18.1 percentage points. The authors release HiFi-UMI-2K, a 2,000-hour dataset with simulation-validated trajectories.
stereo-inertial slamend-effector accuracymicrosecond synchronizationwide-fov camerassimulation replay
VAD to the Bone: Ultra-Tiny Speech Activity Detection for Edge Deployment
We propose kiloVAD, an ultra-tiny voice activity detection model optimized for edge deployment, achieving state-of-the-art performance under causal conditions. The model employs standard Mel features, CNN-only layers, and tunable context/spectral parameters, ensuring compatibility with embedded systems. We introduce per-layer structured pruning with self-distillation and angle-based quantization-aware training (QAT), which improves upon standard QAT by 1-4%. Evaluated per-frame, kiloVAD achieves 0.850 AUC on AVA-Speech with only 2.1 k parameters and 200 ms context, setting a new benchmark for causal, deployment-ready VAD.
voice activity detectionstructured pruningquantization-aware trainingmel featurescnn layers
DRIFT: Direct-Recursive Intervention-Conditioned Forecasting of ICU Physiological Trajectories
DRIFT introduces a hybrid forecasting framework for ICU physiological trajectories, combining direct and recursive action-conditioned models to mitigate autoregressive error accumulation. The method employs a primary direct forecast with constrained recursive corrections, evaluated on 6,046 MIMIC-IV and 8,345 eICU-CRD admissions. DRIFT reduces mean absolute error for mean arterial pressure by 0.673% versus Temporal Fusion Transformer (TFT-action) on MIMIC-IV and shows robustness under treatment-sequence alterations, with MAP error increases of 0.21-0.26 mmHg post-divergence.
time-series forecastingintervention-conditionedautoregressive errormean arterial pressuretemporal fusion transformer
Prototype Adaptation for Zero-Shot sEMG Movement Classification
The paper introduces two zero-shot learning methods, Compositional Prototype Interpolation (CPI) and Synthetic Adaptation for Prototypes (SAP), for classifying combined sEMG movements without training data. These approaches leverage linear interpolation in the embedding space of Prototype Networks, trained only on basic movements. Evaluated on NearLab, NinaPro DB3, and BasCom datasets, SAP achieves >20% accuracy improvement over prior zero-shot methods on combined movements, with sustained performance in online user studies.
zero-shot learningsurface electromyographyprototype networksembedding spacelinear interpolation
Variance-Reduced Conditional Gradient Methods under Markovian Sampling for Nonconvex Composite Optimization
(No summary returned.)
WALoMA: A Multitask Wireless Foundation Model via Adaptive Low-Rank Masked Autoencoders
WALoMA introduces a multitask wireless foundation model using adaptive low-rank masked autoencoders for 6G physical layer tasks, addressing data scarcity via self-supervised learning on unlabeled channel state information (CSI). The method employs 2D positional encoding to preserve spatial-frequency relationships and low-rank adaptation (LoRA) for parameter-efficient fine-tuning. Evaluated across five tasks (LoS/NLoS classification, beam prediction, channel interpolation, channel estimation, channel charting), it achieves a composite score of 87.80%, outperforming the LWM baseline (59.90%) while training only 14.68% of parameters.
masked autoencoderlow-rank adaptationchannel state information6g wirelessself-supervised learning
An Embarrassingly Simple Rule-based Visiting Circulation Approach to Trip Destination Prediction
The paper introduces Rule-based Visiting Circulation (RVC), a non-learning method for trip destination prediction in metropolitan areas where target destinations are unknown during training. RVC leverages origin information and individual revisiting behaviors without requiring training data from source regions. Evaluated on the IEEE Big Data Cup 2022 dataset, RVC outperforms supervised learning and heuristic baselines, achieving second place in the competition by exploiting spatial and behavioral patterns in trip data.
trip destination predictionrule-based modelrevisiting behaviorspatial patternsnon-learning method
Detecting CSAM Text-to-Image LoRAs From Weights
The study demonstrates that harmful Low-rank adaptation (LoRA) fine-tunings for text-to-image generation, particularly those producing child sexual abuse material (CSAM), can be detected directly from weight analysis without generating outputs. By analyzing the top-left singular vectors ($u_1$) of LoRA weight updates, the method forms a compact fingerprint identifying the training objective, generalizes across base models, and avoids false positives on benign content. Experiments show robustness to additive noise, rescaling, and precision reduction. This inference-free approach offers a safer alternative to metadata or output-based moderation for detecting malicious LoRAs.
low-rank adaptationtext-to-image generationweight analysissingular vectorscontent moderation
Optimization with Dynamic Constraint Learning (DCL)
The paper introduces Dynamic Constraint Learning (DCL), a data-driven framework for constrained optimization with unknown, non-queryable constraint functions. DCL iteratively learns local surrogate models from nearby data and solves subproblems within data-supported trust regions, adapting to the data distribution during optimization. Compared to offline global constraint learning, DCL achieves comparable solution quality using simpler local models and smaller subproblems, as demonstrated on a synthetic test problem and two literature case studies.
constrained optimizationsurrogate modelingtrust region methodsdata-driven optimizationdynamic learning
A Physics-Informed Neural Operator for Thermal Ranking of Low-Cost Wall Materials in Hot-Dry Climates
The study presents a physics-informed neural operator (PINO) framework for thermal ranking of low-cost wall materials in hot-dry climates. Using a Fourier Neural Operator (FNO) backbone, the model learns a parameter-to-solution operator from 1500 finite difference method (FDM) solutions of the transient heat equation, enforcing both data fidelity and PDE consistency. The PINO achieves a relative L2 error of 5.14e-4 and 0.201 K mean absolute error on peak inner surface temperature, matching FDM rankings exactly, with physics-informed training outperforming data-only FNO when samples are scarce (150 vs. 300). Clay-straw adobe emerges as optimal under nominal conditions, while fired clay brick performs best in sub-ambient regimes.
physics-informed neural operatorfourier neural operatortransient heat equationthermal rankinglow-cost materials
Contextual Deconvolution for Variance-Stable Demand Sensing: Kernel-Modulated Operators in Promotional Retail
The paper introduces Contextual Deconvolution (CD), a two-stage estimator for demand forecasting that decomposes promotional shocks from structural baselines via kernel-modulated banded operators and hierarchical partial pooling. CD reduces safety stock and order variance but under-provisions event spikes, lowering total cost only when holding costs exceed ~20% of stockout costs. Evaluated on 30,490 M5 and 2,845 Favorita SKUs, CD achieves the lowest cross-sectional error dispersion, mis-forecasting >200% on 0.8% of SKUs versus 9.9–20.6% for baselines. Learned demand operators are non-normal but interpretable via compact parametric kernels.
contextual deconvolutionkernel-modulated operatorshierarchical partial poolingdemand sensingsafety stock
Using Data-Derived Priors to Guide CNN Architecture Design for NIR Chemometrics
The study demonstrates that data-derived spectral descriptors can guide convolutional neural network (CNN) architecture design for near-infrared (NIR) chemometrics, improving hyperparameter optimization (HPO). Using 25 NIR regression tasks, the authors computed dataset descriptors (spectral entropy, intrinsic rank, wavelet energy-support fraction) and optimized two interpretable 1D-CNN scaffolds via Bayesian HPO. Key findings include kernel fraction inversely correlating with spectral entropy and rank, and learning rate decreasing with training-set size. Direct and leave-one-dataset-out (LODO) heuristics achieved median test-RMSE ratios of 0.953 and 1.017 versus HPO, respectively. Joint preprocessing and CNN HPO outperformed standardized-spectra HPO in 19/25 tasks.
convolutional neural networkshyperparameter optimizationnear-infrared chemometricswavelet energy-supportspectral entropy
MemSFT: Mitigating Alignment Tax with an External Parametric Memory
MemSFT mitigates alignment tax in domain-specialized LLMs by decoupling domain knowledge via a plug-and-play parametric memory, trained to emulate non-parametric retrieval over domain data. A learned router dynamically fuses memory and backbone outputs during generation, enabling selective domain expertise invocation. Evaluations on biology, geoscience, and law with Qwen3-8B to Qwen3-235B-A22B show MemSFT improves domain performance with <1% general-task degradation, while full SFT suffers catastrophic forgetting. The method enables parameter-level decoupling of general and domain-specific capabilities.
alignment taxparametric memorynon-parametric retrieverdomain specializationcatastrophic forgetting
OrthKD: Extracting Generalized Clinical Knowledge from Heterogeneous Teachers for Lightweight Deployment
OrthKD introduces selective-trust distillation for diabetic retinopathy screening, combining full supervision from a strong CNN teacher (EfficientNet-B3, 0.876 QWK) with feature-only distillation from a weaker ViT (Swin-Base, 0.830 QWK) while enforcing orthogonality between teacher-specific student projections. This approach preserves lesion precision, incorporates global context, and enhances robustness to domain shift. Evaluated on 132,049 retinal images, a 5.4M-parameter MobileNetV3 student achieves 0.885 QWK on EyePACS and improves zero-shot Messidor-2 performance from 0.507 to 0.728 QWK, with strong referral AUC and calibration.
knowledge distillationdiabetic retinopathydomain shiftorthogonal projectionedge deployment
Mind the Missing Split: Resolving Feature Heterogeneity in Swarm Learning with Random Forests
The paper addresses feature heterogeneity in Swarm Learning with Random Forests by proposing deterministic and probabilistic inference-time strategies to handle missing splits when feature spaces partially overlap across sites. Unlike baseline approaches that restrict training to shared features or rely on local models, the methods enable traversal of decision trees even when splits reference locally unavailable features. Evaluation on nine datasets shows the proposed strategies outperform both the intersection baseline and locally trained models across diverse scenarios.
swarm learningrandom forestsfeature heterogeneitydecentralized learningmissing splits
AMPBench-MT: A Homology-Controlled Benchmark for Antimicrobial Peptide Potency, Spectrum, and Safety Prediction
AMPBench-MT introduces a homology-controlled benchmark for antimicrobial peptide (AMP) evaluation, addressing limitations in existing AMP recognition benchmarks by jointly assessing potency, spectrum, and safety. The benchmark standardizes peptide records into binary recognition, species-conditioned pMIC regression, and endpoint-specific potency and safety readouts. Evaluations across 161 endpoints reveal that high binary performance does not reliably predict assay-endpoint behavior, with frozen protein-language-model embeddings forming the leading pMIC error cluster. Spectrum labels highlight misleading PR-oriented metrics under scarce negatives, while toxicity, HC50 hemolysis, and selectivity expose assay-facing signals. AMPBench-MT advocates for endpoint-aware evidence auditing in AMP evaluation.
antimicrobial peptidepmic regressionhomology-controlledprotein-language-modelassay-facing
Anti-Backdoor Coreset Selection via Cumulative Entropy
The paper introduces Anti-Backdoor Coreset Selection (ABCS), a training-time defense against neural backdoors that formulates the isolation of benign samples as a coreset selection problem. Leveraging the observation that poisonous samples exhibit lower prediction uncertainty and rarity, the method employs Cumulative Entropy to track sample learning dynamics, prioritizing informative benign samples for the coreset. Unlearning selected samples each epoch enhances separability. ABCS achieves consistent backdoor mitigation with negligible natural accuracy loss, outperforming prior defenses that compromise accuracy or fail against certain attacks.
backdoor defensecoreset selectioncumulative entropytraining-time defenseunlearning
Quantum Speedups for Stochastic Optimization with Heavy-Tailed Noise
(No summary returned.)
Data-Dependent Regret and Polyak Corrections for Constrained Online Convex Optimization
The paper presents a tighter, data-dependent regret analysis for constrained online convex optimization (OCO), improving upon the standard O(sqrt(T)) bound. The method combines online gradient descent with Polyak feasibility steps, retaining two key quantities: the observed gradient accumulation G_T and a nonnegative Polyak correction P_T. The resulting regret bound includes a nonnegative improvement term Delta_T. An adaptive variant, AdaOGD-PFS, achieves O(sqrt(G_T)) regret while maintaining per-round feasibility. Experiments on ball- and halfspace-constrained problems show 38-43% improvement in regret bounds, with both gradient and Polyak terms contributing significantly.
constrained optimizationonline gradient descentpolyak correctionregret bounddata-dependent analysis
Bits and Memories: Measuring Verbatim Extraction Across LLM Quantization
This work investigates verbatim extraction risks in quantized language models, challenging prior focus on membership inference for privacy assessment. Using Pythia models and known memorized sequences, the study evaluates extraction across five precision levels (down to 4-bit) and three model sizes, while tracking perplexity. Key findings: (1) Quantization selectively reduces memorization faster than general capability, consistent across algorithms and corpora; (2) 4-bit quantization retains most memorized sequences in large models (<5% capability loss); (3) memorization survival rate scales with model size. The authors argue extraction (not membership inference) should guide privacy evaluations and release code/data.
quantizationverbatim extractionpythiamembership inferenceprivacy risk
PIcsC: Partitioning-Induced Covariate Shift Correction
The paper introduces Partitioning-Induced Covariate-shift Correction (PIcsC), a Fisher information-based regularization framework addressing covariate shift across data partitions in cross-validation, lifelong learning, and federated learning. PIcsC approximates partition divergence via the Fisher Information Matrix (FIM) and integrates it as a regularizer, requiring only partition-local gradients. A conditional adaptation mechanism combines FIM shift with KL divergence to activate regularization selectively. Evaluations on 40+ datasets show PIcsC reduces fragmentation-induced degradation by >20% in batch/fold settings and outperforms FedAvg, FedProx, and SCAFFOLD by 3-5 percentage points in federated learning without client personalization.
covariate shiftfisher information matrixfederated learningregularizationpartition divergence
Robust Unsupervised Network Intrusion Detection via Federated Learning with Selective Aggregation under Anomalous Sample Contamination
Proposes a robust unsupervised network intrusion detection method for IoT environments using federated learning (FL) with selective aggregation to handle contaminated training data. The approach exploits FL's minority underrepresentation bias to attenuate anomalous client influence and employs Expectation-Maximization (EM) to exclude divergent model updates during aggregation. Experiments on NIDS datasets show superior anomaly detection performance under increasing contamination compared to existing methods.
federated learningunsupervised learningnetwork intrusion detectionexpectation-maximizationanomaly detection
SPARC Segmentation to Prediction via Affine Regression and Counterfactuals
The paper introduces a B2B e-commerce propensity modeling framework addressing class imbalance via two innovations: (1) replacing SMOTE with Diverse Counterfactual Explanations (DiCE) for minority-class synthetic data generation, improving distributional fidelity via proximity metrics and UMAP visualization; (2) adapting PyPARC's piecewise affine classifier for interpretable risk-tier segmentation. Evaluated on longitudinal B2B data (1:9 class imbalance), the method achieves 93.1% precision at threshold 0.8 (9.2pp over SMOTE) and 26.1pp improvement at threshold 0.7, demonstrating robust performance across operating points for high-precision marketing.
b2b e-commerceclass imbalancecounterfactual explanationspiecewise affine classificationpropensity modeling
TWICE: Two-Clock, Two-Window Learning for Long-Horizon Conversion Prediction in Online Advertising
TWICE introduces a two-clock, two-window learning framework for long-horizon conversion prediction in online advertising, addressing delayed feedback through complementary supervision from click and conversion clocks. It factorizes conversion rate into target-window probability and grouped delay CDF, using current-status likelihood for click-clock training and arrival-conditioned likelihood for conversion-clock updates. Experiments on public and industrial datasets show 2.486% expected revenue increase, with 2.061% conversion uplift in Kwai's A/B test, leading to full deployment.
delayed feedbackconversion predictioncumulative distribution functiononline advertisingtwo-clock learning
Learned, Relied Upon, or Necessary? Separating Checkpoint Dependence from Task-Level Value in Sheaf GNNs
The paper disentangles checkpoint reliance from task-level value in sheaf graph neural networks (GNNs) by introducing two estimands: checkpoint reliance and protocol-relative replacement. A task-null theorem demonstrates that labels identify only classifier directions, leaving $d^2-d$ degrees of freedom in $d\times d$ maps. Through label-only training and audits of NSD, DNSD, and Directed Sheaf Neural Network implementations, the study reveals both replaceable and unreplaced transport regimes. Results show that learned maps can govern computations without being indispensable edge geometry, with Roman-Empire retaining a $.0675$ advantage over resampled assignment and $.0391$ over shared maps.
sheaf graph neural networkscheckpoint relianceprotocol-relative replacementtask-null theoremedge geometry
Sharpness-aware Model Merging with Salience Recovery for LLM-based Cross-Domain Sequential Recommendation
SharpRec introduces sharpness-aware model merging with salience recovery to address cross-domain knowledge conflict and performance saturation in LLM-based Cross-Domain Sequential Recommendation (CDSR). The framework combines Sharpness-aware Geometric Alignment for interference-free fusion and Preference Salience Activation to recover distinctive features for target domain performance. Experiments in dual- and multi-domain settings show SharpRec outperforms state-of-the-art baselines.
cross-domain sequential recommendationmodel mergingsharpness-aware trainingsalience recoverygeometric alignment
Guiding Posterior Exploration with Optimizer-Derived Geometry
The paper introduces a preconditioned sampling strategy for Bayesian neural networks that leverages curvature estimates from adaptive optimizers like AdamW to guide posterior exploration. By utilizing optimizer-derived geometries computed during warmstart, the method eliminates the need for lengthy burn-in phases and enhances numerical stability. This approach maintains or improves predictive performance and uncertainty quantification across various datasets and network architectures without incurring additional computational costs. Empirical results confirm the consistency and effectiveness of the proposed method.
bayesian neural networksadaptive optimizerscurvature estimatespreconditioned samplinguncertainty quantification
Zhinv: Real-time hub-height wind field reconstruction using only local sparse observations
Zhinv is an end-to-end framework for real-time hub-height wind field reconstruction from sparse, irregular local observations, enabling applications like wind power regulation and resource assessment. The method directly maps sparse inputs to fine-grid wind fields without relying on numerical weather prediction (NWP) or complex data assimilation. Evaluated across Northeast China, Europe, and Southeast Asia, Zhinv reduces reconstruction error by ~66% compared to Kriging, demonstrating robustness and efficiency in generating continuous regional wind fields from limited data.
wind field reconstructionsparse observationsend-to-end learninghub-heightkriging
Breaking the Periodicity Assumption: Robust Tensorial Multi-View Clustering via Graph-Spectral Low-Rank Learning
The paper exposes a critical flaw in t-SVD-based tensorial multi-view clustering (TMC) methods: their performance heavily depends on an implicit periodicity assumption induced by class-ordered sample arrangements, leading to severe degradation under random permutations. To address this, the authors propose a graph-spectral low-rank tensor learning framework using Graph Fourier Transform (GFT) to capture intrinsic manifold structures without relying on sample ordering, alongside an anchor-based variant for scalability. Experiments show the proposed methods outperform state-of-the-art TMC approaches by mitigating the periodicity assumption's impact.
tensorial multi-view clusteringfast fourier transformgraph fourier transformlow-rank learningpermutation invariance
AMRD: Adaptive Multi-Teacher Relational Distillation for Lightweight Speech Emotion Recognition
We propose Adaptive Multi-teacher Relational Distillation (AMRD), a method for lightweight speech emotion recognition that addresses teacher reliability variance and inter-sample relational structure in knowledge distillation. AMRD employs one-class SVM on logit similarity matrices to assign per-batch teacher weights and introduces a relational distillation loss aligning teacher and student similarity matrices. Evaluated on IEMOCAP and CREMA-D datasets across four student architectures, AMRD outperforms single-teacher distillation baselines in most settings, with ablations confirming complementary gains from both components.
knowledge distillationspeech emotion recognitionlogit similarityone-class svmrelational structure
When Does Deep Representation Learning Help Single-Cell Clustering? A Sensitivity-Aware Diagnostic Benchmark for Biomedical AI Pipelines
The study presents a sensitivity-aware diagnostic benchmark to evaluate when deep representation learning improves single-cell clustering in biomedical AI pipelines. It compares nine clustering pipelines across ten real scRNA-seq datasets (90-5,685 cells, 19,046-41,480 genes, 4-11 cell types), integrating Optuna hyperparameter search, repeated-run robustness, statistical testing, and Sobol sensitivity analysis. Results show a contrastive autoencoder achieves the highest mean Adjusted Rand Index (0.7872), but no dominance over strong baselines. Three reproducible regimes emerge: probabilistic VAEs excel on small datasets, deep autoencoders on mid-scale multi-batch data, and PCA remains competitive when linear projection suffices. Learning rate and latent dimensionality are identified as key tuning parameters.
single-cell clusteringdeep representation learningsobol sensitivity analysisadjusted rand indexvariational autoencoder
Normalizing Flows to Reconstruct Pseudo-PDFs
The authors propose a normalizing-flow framework for reconstructing parton distribution functions (PDFs) from synthetic matrix-element data, combining Gaussian Process priors with invertible neural networks to model posterior distributions consistent with limited Ioffe-time measurements. The method enforces physical constraints and maintains extrapolation capabilities. Results demonstrate the architecture's effectiveness in preserving these properties during PDF estimation.
normalizing flowsparton distribution functionsgaussian process priorsinvertible neural networksioffe-time data
FunnelAL: Retrieve-then-Rank Active Learning for Single-Class Discovery
FunnelAL introduces a retrieve-then-rank active learning system for single-class discovery, adapting industrial recommender system architectures to data annotation. The method employs a cascaded pipeline: (1) embedding-based retrieval, (2) precision-triggered ranking with RankNet and committee-based exploration (QBC), and (3) iterative label feedback. Evaluated on three image classification benchmarks, FunnelAL achieves superior F1 scores and annotation efficiency (best AULC) compared to GAL and PF-MA, maintaining robustness under realistic annotator error rates. The system bridges multi-stage recommender systems with active learning.
active learningsingle-class discoveryretrieve-then-rankembedding-based retrievalprecision-triggered ranking
HeAD-CP: Heterophily-Aware Diffused Conformal Prediction Sets for Graph Neural Networks
HeAD-CP introduces heterophily-aware diffusion variants for conformal prediction on graphs, addressing the homophily bias of Diffused Adaptive Prediction Sets (DAPS). The method employs node-wise diffusion coefficients derived from label-free local-homophily estimates based on GNN softmax outputs, with three variants tailored for different heterophily levels. HeAD-CP preserves marginal coverage guarantees and reduces mean prediction-set size compared to DAPS, particularly on heterophilic graphs. Evaluated on ten benchmarks, HeAD-CP outperforms DAPS on 8/10 datasets, with significant improvements on heterophilic graphs (e.g., 10.3% on Texas) and negligible differences on homophilic datasets. A post-hoc oracle demonstrates the potential for further optimization via label-free selection.
conformal predictiongraph neural networksheterophilydiffusionuncertainty quantification
FORGE: Frame Orthogonality in Relevance Geometry for Long-Form Video Understanding
FORGE introduces a training-free method for query-conditioned keyframe selection in long-form video understanding by inducing an orthogonal geometry on pretrained multimodal embeddings, unifying relevance and diversity in a single objective. The approach maximizes coverage of independent query-relevant directions within a fixed frame budget, outperforming baselines on Video-MME and LongVideoBench. At K=64, FORGE doubles keyframe recall (0.415 vs. 0.204) and improves QA accuracy by up to 8.7 points across eight MLLMs (4B-32B parameters), demonstrating the efficacy of query-aligned embedding geometry for inference-time video processing.
multimodal embeddingskeyframe selectionlong-form videoinference-time optimizationorthogonal geometry
Learning from the Unseen: Offline Reinforcement Learning with Hidden Actions
The paper introduces LURE (Learning from the Unseen: Robust Estimator), the first method for offline reinforcement learning (RL) with hidden actions in infinite-horizon discounted Markov decision processes. By leveraging next-state variables as proxies for unobserved actions, LURE achieves multiply robust policy evaluation, remaining consistent under various nuisance model specifications and enabling asymptotically normal inference. Experiments on synthetic data and the MIMIC-III sepsis management task validate its effectiveness.
offline reinforcement learninghidden actionsmultiply robust estimationmarkov decision processesasymptotic normality
Beyond Single-Episode Optimization: Sliding-Window Aware Generative Auto-Bidding for Long-Term Advertising Effectiveness
SWAG-Bid introduces sliding-window aware auto-bidding for long-term advertising effectiveness, addressing the unreliability of per-day efficiency ratios in sparse-value scenarios. The hierarchical framework combines a Masked Trajectory Model for market forecasting and Multi-Window Model Predictive Control Sampling (MWMS) to generate and score plans across overlapping 7-day windows, with Per-Step Gated Adaptive Layer Normalization (PSG-AdaLN) dynamically adjusting plan adherence. Experiments on AuctionNet-Sparse and AliExpress demonstrate competitive constraint satisfaction and value acquisition under sliding-window evaluation.
auto-biddingmulti-window optimizationmasked trajectory modeladaptive layer normalizationcost-per-action
Neurai-VN Benchmark: Standardized Machine Learning Models for Multimodal Digital Phenotyping in Mental Health Classification
The study introduces Neurai-VN Benchmark, a standardized evaluation framework for multimodal digital phenotyping in mental health classification, addressing reproducibility challenges from heterogeneous datasets. It leverages the Neurai-VN dataset, comprising passive and active sensor data from 100 Vietnamese adults over two weeks, to evaluate four binary classification tasks (e.g., Healthy Control vs. Depression) using subject-wise cross-validation. Baseline models (linear, tree-based, neural) achieve mean F1 scores of 0.71 (Healthy Control vs. Depression/Clinical), 0.69 (Healthy Control vs. Anxiety), and 0.56 (Depression vs. Anxiety), establishing reproducible benchmarks for future research.
digital phenotypingmultimodal datasetsubject-wise cross-validationmental health classificationreproducible benchmark
Decision-Level Hijacking: Injecting Cognitive Bias into Large Language Models via Bit-Flip Attacks
The paper introduces decision-level hijacking, a threat where attackers manipulate LLM cognitive stances to influence downstream decisions, and proposes CogBias, a framework for injecting cognitive bias via Bit-Flip Attacks (BFAs). CogBias employs a differentiable sentiment evaluator to convert preferences into optimization signals, uses multi-objective loss for constraint balancing, and leverages BitScout to locate critical weight bits under ultra-sparse flip budgets. Experiments on Llama-3.2-3B, Mistral-7B, and Qwen2.5-14B show that flipping <0.01% of bits induces significant stance shifts on target topics (e.g., commercial recommendations) while minimally affecting non-target tasks, demonstrating LLM value alignment vulnerabilities via low-level weight perturbations.
decision-level hijackingbit-flip attackscognitive bias injectiondifferentiable sentiment evaluatorultra-sparse flip budget
SecDrift: Measuring Sector-Conditioned Security Drift in AI-Generated Code
SecDrift introduces a benchmark to measure sector-conditioned security drift in AI-generated code, evaluating the change in static-analysis vulnerability rates when prompts are conditioned on industry contexts versus neutral baselines. The study assesses 7 LLMs across 8 CISA critical infrastructure sectors and 9 CWE categories, conducting 5,355 evaluations with a 5-dimension transformation and matched-baseline condition. Results show no statistically significant security drift attributable to sector identity (Fisher's exact p = 0.24, Cohen's h = -0.08), with the observed gap being a composition artifact of two CWE categories (CWE-502 and CWE-22). Model selection, not prompt framing, significantly impacts vulnerability rates, ranging from 11.6% to 16.1%. The framework, prompts, generated code, findings, human-validation verdicts, and analysis scripts are publicly released.
static-analysisvulnerability ratescwe categoriesmatched-baselinesector-conditioned
A Unified Algorithmic Framework for Hybrid Reinforcement Learning in Tabular MDPs with Shifted Transition Dynamics
The authors propose a unified algorithmic framework for hybrid reinforcement learning in tabular Markov Decision Processes (MDPs) with shifted transition dynamics. The framework introduces two algorithms: MIN-UCB-VI for regret minimization and MAX-LCB-VI for best policy identification, which leverage fine-grained bias information to effectively integrate offline data from outdated environments. Theoretical analysis provides instance-dependent and independent upper bounds on regret and sub-optimality gap, with matching lower bounds demonstrating optimality. Experimental validation confirms the theoretical findings.
markov decision processesregret minimizationtransition dynamicsoffline datasub-optimality gap
Algorithmic Separation between Constant-Depth and Logarithmic-Depth Neural Networks
This work establishes the first algorithmic separation between constant-depth and logarithmic-depth neural networks by identifying a class of Boolean functions with hierarchical Fourier spectra. Logarithmic-depth networks efficiently learn these functions via layerwise coordinate descent, hierarchically reconstructing the spectra. For a subclass, constant-depth polynomial-width networks with regular activations and controlled spectral norms incur constant $L^2$ approximation error under the uniform distribution over the hypercube, demonstrating a fundamental depth-dependent learning limitation.
boolean functionsfourier spectralayerwise coordinate descentapproximation errorspectral norms
Rethinking CD: A Reproducibility Study and Extension on the Ineffectiveness of Contrastive Decoding at Mitigating Object Hallucinations in MLLMs
This study reproduces and extends prior work questioning contrastive decoding (CD) as a training-free method for mitigating object hallucinations in multimodal large language models (MLLMs). The authors analyze CD's unidirectional output distribution shift on discriminative datasets, verify that adaptive plausibility constraints reduce sampling to greedy search, and evaluate CD's effects across generative and discriminative benchmarks (MME, POPE, CHAIR) using LLaVA and Qwen. Results confirm CD's spurious improvements fail to enhance visual grounding consistently, challenging its efficacy. The work motivates developing more reliable hallucination mitigation strategies for MLLMs.
contrastive decodingobject hallucinationsmultimodal large language modelsvisual groundingadaptive plausibility constraint
Lloyd's $K$-Means Clustering Algorithm Is Frank-Wolfe in Disguise
(No summary returned.)
A Riemannian View on Active Subspaces
The paper extends active subspaces—an explainable dimensionality reduction technique for Euclidean domains—to Riemannian manifolds via parallel transport, contrasting intrinsic and extrinsic gradient averaging approaches. Local analysis within geodesic balls shows eigenvalue agreement to second order in radius and eigenspace alignment dependent on spectral gaps. The method requires either tangent space recomputation or parallel transport of a central frame for broader applicability. Hyperspheres serve as a key case study, with numerical experiments on the 2-sphere demonstrating ridge recovery at curvature-constrained quadratic rates.
active subspacesriemannian manifoldsparallel transportspectral gaphyperspheres
Accurate structural modeling of chemically diverse molecular interfaces with Vilya-2
Vilya-2 introduces a diffusion transformer for accurate structural modeling of chemically diverse molecular interfaces, extending Vilya-1's all-atom representation to molecular interactions. The method enables transfer learning across molecular types, generating diverse structural ensembles with calibrated confidence rankings. Results show 59.1% of peptide interfaces modeled to sub-2 Å backbone RMSD, outperforming co-folding models, and state-of-the-art performance in small-molecule docking and generalization to novel complexes and macrocycles.
diffusion transformerall-atom representationstructural ensemblesbackbone rmsdsmall-molecule docking
Physics-Informed CNN-LSTM for Street-Scale Urban Flood Prediction: Reconciling Aggregate Accuracy and Street-Level Plausibility
The authors propose a physics-informed CNN-LSTM framework for urban flood prediction that reconciles pixel-level accuracy with street-level physical plausibility. The model incorporates three differentiable loss terms: gravity loss (penalizing unphysical water flow), continuity loss (enforcing mass conservation), and a topography-aware false-alarm penalty modulated by the topographic wetness index (TWI). Evaluated on the Norfolk flood dataset (300 samples across two storms), the constrained model achieves near-zero gravity violations (1e-6), 77% street-channel recall (vs. 44% for unconstrained baseline), and the best street-level F1, while the TWI-modulated penalty recovers 60% higher street recall than uniform variants at the lowest MAE among constrained models.
physics-informed cnn-lstmurban flood predictiontopographic wetness indexdifferentiable loss termsstreet-level plausibility
A Foundational Perspective for Partitional Clustering on Networks
The study provides a theoretical analysis of partitional clustering on networks, comparing hard (P-Median, Sum of Squares Clustering) and soft (Probabilistic Distance Clustering, Fuzzy C-Means) assignment schemes with cluster centers allowed on vertices or edges. Mathematical analysis reveals structural differences: SSC and FCM permit optimal edge-located centers, while PMP and PDC inherently prefer vertices, highlighting assignment bottlenecks and vertex-restricted solutions. These findings inform algorithm design for applications like facility location and similarity search in embedding graphs.
partitional clusteringnetwork clusteringhard assignmentsoft assignmentcluster centers
Interpretable GOHR Agents via Sparse Autoencoders
The study investigates interpretability in learned decision-making systems by analyzing a tokenized autoregressive Transformer agent trained on the Game of Hidden Rules (GOHR). The agent operates in a two-rule task where hidden rules map object shapes to target buckets without explicit rule labels. Sparse autoencoders (SAEs) applied to the agent's decision-token embeddings recover interpretable structure, with individual dimensions corresponding to concepts like chosen shape or bucket and strategies like rule hypothesis probing. SAE dimensions selectively cover decisions involving these concepts, demonstrating their utility for explaining agent behavior.
interpretabilitysparse autoencodersautoregressive transformerdecision-makinggame of hidden rules
Deep Label-Wise Attentive Temporal Convolutional Networks Improve Medical Coding
Proposes Deep Label-Wise Attentive Temporal Convolutional Networks for medical coding, framed as multi-label text classification. The model combines multi-layer Temporal Convolutional Networks (TCNs) for long-sequence document representation with label-wise attention to focus on code-specific text regions. Achieves state-of-the-art performance with 9% higher F1-score and 28% recall improvement over prior work, particularly beneficial for clinical decision support.
temporal convolutional networksmulti-label classificationlabel-wise attentionmedical codingclinical decision support
Endpoint Replay: Compressing the Recency Buffer in Deep Reinforcement Learning
The paper proposes Endpoint Replay, a compression method for deep reinforcement learning (DRL) recency buffers that reduces storage requirements while maintaining performance. The approach stores representative transitions from endpoint trajectories of connected $n$-step sequences, achieving comparable memory horizon to standard large buffers with 10x less storage. Empirical evaluation on Pinball and Atari 2600 benchmarks shows the method avoids systematic bias and matches traditional buffer performance.
experience replayrecency buffern-step sequencesvalue propagationdeep reinforcement learning
Semantic Space Search Trajectory Networks
The paper introduces semantic space Search Trajectory Networks (STNs), extending graph-based optimization algorithm analysis to high-dimensional semantic spaces. By discretizing model predictions via agglomerative clustering with complete linkage under normalized Hamming distance, the method enables cross-algorithm comparison of learning dynamics. Applications on classification and regression tasks reveal structural differences between standard training and label randomization, with real-label training producing denser, more centralized graphs. The approach captures functional training dynamics across diverse machine learning models and regimes.
search trajectory networkssemantic spaceagglomerative clusteringlearning dynamicsnormalized hamming distance
Score-Based Stabilization for Time-Dependent Problems
We introduce a score-based stabilization framework for numerical simulation of partial differential equations, leveraging a learned score model to define a stabilization operator that corrects provisional numerical updates. This operator enhances standard time-stepping schemes by enforcing physical consistency and driving iterates toward the manifold of admissible states, acting as a contraction with basin-conditional stability. Experiments on Advection, Korteweg-de Vries, Nonlinear Schrodinger, and Burgers' equations demonstrate improved robustness, suppression of nonphysical instabilities, and preservation of qualitative dynamics.
stabilization operatorpartial differential equationsscore modeltime-stepping schemesmanifold of admissible states
Memory Layer: Train the In-Model Cache for Recommendation Models
The paper introduces the memory layer, an in-model key-value embedding cache co-trained with recommendation models to eliminate representation discrepancy between training and serving. By integrating the cache into training, the item tower writes embeddings during training while the model reads them at serving, ensuring a single source of truth for item representations. Deployed on Instagram Reels, this approach improves prediction coverage (96% to 100%), embedding freshness (O(5 min) to O(20 s)), and reduces the training-serving Normalized Entropy gap by up to 86%, yielding 2× recall for fresh content and 5-6% cold start engagement lift. It also reduces computational costs by 30% by eliminating separate bulk-evaluation steps.
recommendation systemsembedding cachetraining-serving discrepancynormalized entropycold start
MOSAIC-FL, a micro-service based privacy-preserving framework with application to genomics
MOSAIC-FL introduces a privacy-preserving federated learning framework with a micro-service architecture, addressing security challenges in sensitive domains like genomics. The system employs gRPC for efficient communication, a Finite State Machine for synchronization, and a Threshold-CKKS homomorphic cryptosystem for secure aggregation, requiring t-out-of-N clients for decryption. It ensures IND-CPA-D security via noise flooding and mitigates key-recovery attacks through dynamic key renewal. Evaluations on EMNIST and TCGA breast cancer subtyping demonstrate effectiveness across varying thresholds and model scales.
federated learninghomomorphic encryptionthreshold-cksgenomicssecure aggregation
Characterizing and Mitigating the Effects of Device Temperature on RF Fingerprinting Accuracy
A temperature-aware Radio Frequency Fingerprinting (RFFP) framework is proposed to mitigate the adverse effects of device temperature on classification accuracy, addressing a critical limitation in existing methods. The framework explicitly incorporates device temperature information into the learning process to enhance robustness and generalization. Evaluated on a real-world Bluetooth Low Energy (BLE) dataset collected across multiple devices and environmental conditions, the method consistently outperforms baseline temperature mitigation approaches. Significant improvements in classification accuracy are observed, particularly under unseen temperature and environmental conditions.
radio frequency fingerprintingdevice temperatureclassification accuracybluetooth low energytemperature-aware modeling
ScoreShield: Differentially Private Release of Similarity Scores
The paper introduces ScoreShield, a differentially private mechanism for releasing cosine similarity scores that addresses information leakage risks in applications like biometrics and RAG. The method combines Gaussian noise calibrated to global sensitivity with projection onto valid cosine objects, satisfying (ε,δ)-DP for both score vectors and Gram matrices. Theoretical analysis shows improved utility bounds, reducing Frobenius risk from Θ(n³) to O(n²) for pairwise Gram releases, with empirical validation across face recognition, semantic retrieval, and recommender systems.
differential privacycosine similaritygram matricesmembership inferencefrobenius norm
Simulation-based parameter estimation via a combination of embedded normalizing flows and implied empirical probabilities under moment restrictions
The authors propose a simulation-based parameter estimation framework combining embedded normalizing flows and empirical-likelihood estimation under moment restrictions. The method first transforms complex residual distributions into simpler base distributions via normalizing flows, then applies empirical-likelihood constraints to treat transformed residuals as discrete random variables. Gradient-based optimization updates both simulation model parameters and flow parameters, leveraging implicit differentiation of the empirical-likelihood function. The inverse flow serves as a surrogate model for sensitivity analysis and discrepancy quantification, with an information-theoretic interpretation underpinning the computational framework.
parameter estimationnormalizing flowsempirical likelihoodmoment restrictionsgradient optimization
Conformal Cascade: Distribution-Free Accuracy Guarantees for Multi-Tier LLM Inference
Conformal Cascade (CC) introduces a multi-tier LLM inference framework with distribution-free accuracy guarantees by using conformal prediction set size as a deferral rule. The method accepts queries when the calibrated set collapses to a single answer, deferring otherwise, and provides a finite-sample accuracy bound of at least $1 - Kα$ (tightening to $1 - α$ under selection-preservation). Evaluated on 18 multiple-choice benchmarks with cascades from four open-weight model families, CC outperforms heuristic cascades on reasoning-heavy tasks while maintaining accuracy on easier benchmarks by routing most queries to smaller models. The approach requires no training and only black-box API access.
conformal predictionllm cascadesdistribution-free guaranteesmulti-tier inferencecalibration
Fast, accurate, and differentiable: a neural-network surrogate for NRSur7dq4 precessing binary black hole waveforms
(No summary returned.)
Amortising Trajectory Optimisation for Residual MPC via Implicit Contact Differentiation
We introduce an AD-assisted implicit derivative for regularised smooth contacts, leveraging the Implicit Function Theorem (IFT) to differentiate the stationarity residual at tolerance-converged solutions without solver unrolling or hand-assembled KKT systems. This approach maintains nearly constant compiled temporary memory, growing by less than 4% from one to ten iterations, compared to a 10.6× increase for unrolled AD. Additionally, we propose optimiser distillation for residual MPC, amortising batched full-horizon iLQR into a policy guiding short-horizon residual iLQR. Experiments on Finger, Franka, and Unitree demonstrate a 28-98 percentage point improvement in six-step success rates over standard iLQR.
implicit function theoremdifferentiable simulationtrajectory optimisationresidual mpcoptimiser distillation
Inverse RL Helps Align AI by Imitating Humans
The paper introduces PARED (Projected Alignment Reward Estimated from Demonstrations), an inverse reinforcement learning method for aligning language models by recovering implicit rewards from expert demonstrations. PARED trains a lightweight discriminator on response-level features to distinguish demonstrations from policy samples, yielding an explicit reward function without task-specific preference annotations. Experiments show PARED improves base policies via inference-time reranking and on-policy RL, achieving gains beyond supervised fine-tuning, while enabling contextual alignment for diverse audience preferences.
inverse reinforcement learninglanguage model alignmentreward modelingon-policy optimizationcontextual adaptation
From Machine Learning to Large-Scale EO Products: Best Practices for Making Maps
This paper outlines best practices for generating large-scale geospatial maps from Earth observation (EO) data using machine learning (ML), addressing challenges in preprocessing, dataset design, model training, and uncertainty quantification. The authors present a structured pipeline covering six key themes: EO data infrastructure, data selection/preprocessing, ML dataset construction, uncertainty quantification, map production/distribution, and validation. The work highlights interdependencies between pipeline stages and emphasizes methodological rigor to mitigate errors in operational map products, with extended details available in an online guide.
earth observationgeospatial mappingmachine learninguncertainty quantificationdata preprocessing
LLM as Forecasting Planner: Training-Free Text Conditioning for Time-Series Foundation Models
The paper introduces LLM as Forecasting Planner (LAFP), a training-free framework for text-conditioned time-series forecasting that combines time-series foundation models (TSFMs) and large language models (LLMs) without retraining either. LAFP formulates forecasting as a planning problem, using a frozen TSFM to propose numerical continuations and an LLM to guide candidate selection via Monte Carlo tree search (MCTS), with separate 'Ranker' and 'Judge' LLM roles. Experiments on Context-is-Key and Time-MMD benchmarks with Chronos and TimesFM TSFMs and four LLMs show consistent improvements, validating sequential search for bridging modality gaps.
time-series forecastingfoundation modelsmonte carlo tree searchlarge language modelstraining-free
GAUGE: Grading Agent-Built Financial Models Without a Golden Answer
The paper introduces GAUGE, a benchmark for evaluating agent-built financial models without relying on a single reference answer, addressing the issue of multiple reasonable outputs in forecasts and valuations. GAUGE uses 1,001 analyst workbooks and a 196-task evaluation set, incorporating a three-layer observed-practice envelope, 56 auditable facets, and eight validity gates. Validation via a 55-participant study shows senior analysts average 88.3 on the failure-aware score $φ_0$, while the best agent scores 53.4, excelling in model construction (93% mechanical facet pass rate) but lagging in judgment (78%). The benchmark includes released methodology, data, and evaluation cores.
financial modelingbenchmark evaluationagent-built modelsvaluation judgmentobserved-practice envelope
Mechanisms of Width Scaling in Normalized Residual Networks: The Effective Alignment Dimension
The paper introduces the concept of effective alignment dimension to quantify the signal-noise geometry of activation gradients in function-preserving residual expansion. By deriving the mean and variance of the inner product between training and test gradients, the authors establish a finite-sample upper bound on misalignment probability, dependent only on the effective alignment dimension and sample size. This bound is integrated into a train-test residual-expansion framework, providing a high-probability condition for test-risk improvement. Experiments on width-controlled LLaMA-style Transformers, Pythia, and ResNet-20 demonstrate that wider models exhibit larger effective alignment dimensions and reduced empirical misalignment, with alignment statistics predicting held-out loss changes.
effective alignment dimensionactivation gradientsmisalignment probabilityresidual expansiontest-risk improvement
Eliminating Propagation Delay: Attention-Based Spatial-Temporal Fusion Graph Convolution Network for Traffic Flow Prediction
The Attention-Based Spatial-Temporal Fusion Graph Convolution Network (A-STFGCN) addresses propagation delay and computational inefficiency in traffic flow prediction. It introduces a spatial-temporal fusion block to extract feature correlations while eliminating propagation delay errors, leveraging a multi-head self-attention mechanism with a mask matrix to capture both long-term and short-term temporal characteristics. Evaluated on five real-world datasets, A-STFGCN outperforms eight baseline methods in overall performance while maintaining computational and data utilization efficiency.
spatial-temporal fusionpropagation delaygraph convolutionmulti-head self-attentiontraffic flow prediction
Beyond "What to Retrieve": Uncertainty in Retrieval-Augmented Code Generation
OpenCoder introduces uncertainty-aware retrieval for repository-level code generation, explicitly modeling source-specific uncertainty to filter, rank, and guide generation across heterogeneous evidence (API knowledge, repository context, similar-code examples). The framework reveals context-dependent cross-source interactions, improving GPT-4's selected-output correctness from 56.25% to 78.13% on RepoExec-inline (32 tasks), though benefits vary by LLM backend. Target-aware API refinement enhances API-set retrieval, demonstrating uncertainty's utility as a control signal for retrieval-augmented generation and repair.
retrieval-augmented generationrepository-level code generationuncertainty modelingapi refinementcross-source interactions
Behavior-Driven Explainability
The paper introduces Behavior-Driven Explainability (BDX), a method for deriving system explanations from structured scenarios in Behavior-Driven Development (BDD). BDX translates sequences of actions, specified for each functionality, into interpretable explanations applicable across development stages and abstraction levels. A case study demonstrates BDX's utility in explaining exceptions in a RISC-V processor, highlighting its role in enhancing trust and understanding during system design, particularly for safety-critical systems.
behavior-driven explainabilitystructured scenariossystem designrisc-v processorsafety-critical systems
Human Preference aligned Tabular Similarity
The authors propose a human preference-aligned evaluation framework for task-agnostic tabular embeddings, addressing the inadequacy of standard downstream metrics in assessing embedding trustworthiness for similarity search. They highlight that existing embedding approaches, primarily optimized for prediction tasks, fail to align with human preferences in similarity rankings. The study introduces a concrete evaluation procedure and demonstrates its application in a Product Lifecycle Management (PLM) use case. This work emphasizes the necessity of incorporating human-aligned evaluation to ensure embedding reliability in real-world business systems.
tabular embeddingssimilarity searchhuman preferencetask-agnosticproduct lifecycle management
GraphRareBench: An Auditable Graph-Evidence Benchmark for Phenotype-Driven Rare-Disease Diagnosis
GraphRareBench introduces an auditable benchmark for phenotype-driven rare-disease diagnosis, featuring 2,365 ontology-derived cases with provenance-preserving evidence records and 18,093 target-confounder pairs. The benchmark includes coarsened HPO queries, fixed candidate pools, graph-defined hard confounders, and source-linked evidence. Evaluations on a 237-case test split show supervised rankers achieving MRRs of 0.640-0.740 and target-over-confounder accuracies of 0.898-0.916, while agents like Agents-A1 and DeepSeek-V4-Flash achieved MRRs of 0.746 and 0.718, respectively. The benchmark highlights complementary aspects of model behavior, including full-pool retrieval and hard-confounder discrimination.
phenotype-driven diagnosisrare-disease benchmarkgraph-defined confoundersevidence-aware evaluationsupervised rankers
📰 Industry Media (7)
The AI Hype Index: Unsexy AI
The article critiques recent AI developments deemed 'unsexy' by public perception, including Grok's controversial translation feature, Meta's augmented reality glasses, and the environmental impact of large-scale AI systems. It highlights 1X's robotic hands as a polarizing example of AI dexterity, noting their potential to outperform humans in tasks like cooking. Despite these advancements, the piece emphasizes societal concerns, such as job displacement and rising emissions. Conversely, it observes a positive social impact: Korean semiconductor workers report increased dating opportunities due to substantial AI-related bonuses, suggesting economic benefits for niche industries.
ai dexterityaugmented realityenvironmental impactjob displacementsemiconductor workers
Prompt Engineering vs Loop Engineering vs Graph Engineering: What Changes at Each Layer
The article delineates three hierarchical control units in AI engineering: prompt engineering (single model response), loop engineering (agent behavior cycle), and graph engineering (multi-agent organization). Each layer preserves the one beneath it, with prompt engineering remaining relevant even in higher layers. Key findings include Anthropic's observation that prompt engineering resolved 50-subagent coordination failures, loop engineering's reliance on mechanical stop conditions to prevent token waste, and graph engineering's dual-structure approach (stable org graph and ephemeral work graph). Layer transitions occur when human-in-the-loop assumptions break due to volume, automation, or multi-step tasks.
prompt engineeringloop engineeringgraph engineeringmulti-agent systemsstategraph
Liquid AI Releases LFM2.5-Encoder-230M and LFM2.5-Encoder-350M: Bidirectional Encoders That Stay Fast at 8K Context on CPU
Liquid AI introduces LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, bidirectional encoders derived from LFM2.5 decoder backbones via bidirectional attention masks, non-causal short convolutions, and 30% masked language modeling. Trained in two stages (1K-token general MLM, then 8K-token domain adaptation), these models achieve 81.02 and 79.29 mean scores on 17 GLUE/SuperGLUE tasks, outperforming larger models like EuroBERT-2.1B. Optimized for CPU deployment, the 230M variant processes 8K tokens in ~28s, offering efficient edge inference for long-document applications.
bidirectional encodermasked language modelingnon-causal convolutioncontext windowcpu inference
Building Non-Interactive Agentic Coding Workflows with Moonshot AI’s Kimi CLI, JSONL Streaming, Testing, and Session Memory
The tutorial demonstrates configuring Moonshot AI's Kimi CLI for non-interactive agentic coding workflows, including automated code analysis, modification, and testing. Methodologically, it details installation via uv in Python 3.13, TOML-based API authentication, and a Python wrapper for headless Kimi execution. Results show Kimi autonomously identifying bugs in a sample inventory project (e.g., negative stock handling), generating unittest cases, and iterating until tests pass, while supporting JSONL event streaming (131K context) and multi-turn session memory.
agentic workflowsjsonl streamingcontext windownon-interactive cliautonomous testing
Fireworks AI Releases Fireworks Nexus: A Drop-In Routing and Cost-Control Layer That Moves Routine Coding Work to Open-Weight Models
Fireworks AI introduces Fireworks Nexus, a routing and cost-control layer that optimizes LLM usage for coding tasks by directing routine work to open-weight models. The system comprises enterprise cost controls, FireConnect for workflow continuity, and an intelligent traffic management model that scores request difficulty. Preliminary results from internal testing with Notion and Doximity show a 33% reduction in cost per merged pull request and a blended token rate 25% of closed-model labs. Independent evaluations by Faros AI (211 tasks) and Arize (2,400 runs) demonstrate cost-effectiveness, with routing strategies outperforming single-model approaches at $0.525 per successful task versus GPT-5.5's $0.636.
open-weight modelstraffic managementcost-controlworkflow continuityblended token rate
OpenAI report links coding agents to faster science software builds
OpenAI's field report demonstrates that AI coding agents, including Codex and Claude Code, accelerate scientific software development across genomics, immunology, and statistics. Agents performed tasks such as build-system cleanup, performance optimization, and backend migrations, reducing runtimes by up to 60x in some cases. For instance, HI.SIM saw a 31% runtime reduction, while RustQC consolidated 15 tools into one, achieving a 60x speedup. However, verification remained a human bottleneck, as agents often produced confident but incorrect outputs. The report emphasizes the need for stewardship to prevent fragmentation, noting that lower engineering costs enable smaller teams to undertake complex rebuilds previously requiring grant-funded hires.
codexruntime optimizationbackend migrationverification bottleneckstewardship
Google AI Overviews become more common in search
Google's AI-generated summaries (AI Overviews) now appear in 43% of US searches, up from 15% a year prior, per Similarweb's 2026 Generative AI Landscape report. The system employs query fan-out to decompose questions into subtopics, aggregating results from multiple searches. AI Mode usage grew from 126M to 279M monthly visits (June 2025-May 2026), with search queries lengthening by 5.4% as users adopt natural-language prompts. ChatGPT citations increased 5x (1.3% to 6.8% of responses), though referral traffic disproportionately targets homepages (58.8%) versus cited deep-content URLs (65%). Cloudflare's Pay per Crawl beta introduces HTTP 402 payments for AI crawler access.
ai overviewsquery fan-outnatural-language searchreferral traffichttp 402
Generated automatically at 2026-07-29 20:35 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.