Daily Digest — 2026-06-26
267 items · 4 research labs, 262 arxiv papers, 1 industry media
MarkTechPost: all feed URLs failed (last tried: https://www.marktechpost.com/feed/)AI News: all feed URLs failed (last tried: https://artificialintelligence-news.com/feed/)
🏛️ Research Labs (4)
How agents are transforming work
OpenAI's economic research demonstrates Codex's transformative impact on knowledge work through agentic AI capabilities. The study tracks adoption patterns across OpenAI departments from August 2025 to June 2026, employing usage metrics and LLM-as-judge task horizon estimation. Results show 80.6% of individual users delegate >30-minute tasks to Codex, with non-developer adoption growing 137x (individual) and 189x (organizational). Codex now accounts for 99.8% of OpenAI's weekly output tokens, enabling cross-functional work where 25% of non-technical outputs involve coding tasks.
agentic aicodexllm-as-judgeoutput tokenstask horizon
Run a vLLM Server on HF Jobs in One Command
Hugging Face introduces a streamlined method to deploy vLLM servers via HF Jobs using a single command, enabling rapid model testing, evaluation, and batch generation. The process involves specifying GPU resources, exposing ports, and leveraging the OpenAI-compatible API for queries. Results demonstrate successful deployment of models like Qwen/Qwen3-4B, with additional scalability options for larger models such as Qwen3.5-122B-A10B. The approach supports advanced features like tensor parallelism, SSH access, and integration with coding agents via Pi. HF Jobs offers flexibility for experimental setups, contrasting with Inference Endpoints for production-ready services.
vllmhuggingfacetensor-parallelismopenai-apissh-access
Which tokens does a hybrid model predict better?
The study investigates token-level prediction differences between hybrid and transformer architectures using the 7B-parameter Olmo 3 (transformer) and Olmo Hybrid models. Through fine-grained analysis of loss gaps across token categories in diverse text corpora, the authors demonstrate that hybrid models excel at predicting meaning-bearing tokens (nouns, verbs, adjectives) and anaphoric references, while transformers outperform on verbatim repetitions and closing braces. Regression-controlled analyses reveal these patterns persist when accounting for token frequency and context length. The findings suggest hybrid architectures combine attention's copy capability with RNN layers' superior state-tracking for open-class tokens.
hybrid architectureloss gapin-context learningrecurrent layerstoken-level analysis
Our latest Google Finance upgrades, including a new app
Google Finance introduces a multimodal portfolio management system combining document parsing (CSV/PDF), screenshot analysis, and natural language input to create unified investment dashboards. The platform implements an AI research tool with conversational query capabilities for portfolio analysis (e.g., sector allocation diagnostics) and automated market briefing generation through scheduled natural language tasks. A new Android app delivers real-time market data, news feeds, and movement explanations, with iOS compatibility planned for late 2026. The system processes user-defined monitoring tasks (e.g., cryptocurrency pre-market analysis) through background data aggregation, delivering notifications via mobile apps or web interfaces.
multimodal inputportfolio diagnosticsscheduled task automationreal-time market feedconversational querying
📜 arXiv Papers (262)
Learning Action Priors for Cross-embodiment Robot Manipulation
The paper introduces a two-stage framework for learning action priors in Vision-Language-Action (VLA) models to improve cross-embodiment robot manipulation. Stage 1 pretrains a flow-matching-based encoder-decoder action module on unconditioned action trajectories to capture temporal motion structure. Stage 2 transfers this prior to VLA training via decoder reuse and latent distillation, while using the encoder as a history compressor. Experiments on 13 cross-embodiment tasks show faster convergence, higher success rates (+substantial gains on data-scarce real-world tasks), and improved generalizability with scaled action data.
vision-language-actioncross-embodimentflow-matchingtemporal motionlatent distillation
On-Policy Self-Distillation with Sampled Demonstrations Reduces Output Diversity
On-policy self-distillation improves pass@1 accuracy by using a single model as both teacher and student, with token-level feedback from correct demonstrations. However, this method reduces output diversity and flattens pass@k curves due to compounding biases in demonstration sampling. Theoretical analysis reveals that self-distillation tilts the base distribution via conditional mutual information, amplifying probability gaps and concentrating mass on dominant modes. Empirical results on graph path-finding and QA benchmarks show self-distilled models match RL performance but exhibit lower functional and semantic diversity, particularly in out-of-distribution settings.
self-distillationpass@krollout diversityconditional mutual informationout-of-distribution
Neglected Free Lunch from Post-training: Progress Advantage for LLM Agents
The paper introduces progress advantage, an implicit advantage function derived from reinforcement learning (RL) post-training that eliminates the need for dedicated reward model training in LLM agentic settings. Progress advantage is formulated as the log-probability ratio between the RL-trained policy and its reference policy, recovering the optimal advantage function under a stochastic Markov decision process. Validated across five benchmarks and four model families, it outperforms confidence-based baselines and dedicated trained reward models in test-time scaling, uncertainty quantification, and failure attribution. The method is annotation-free, domain-agnostic, and available as a byproduct of standard RL post-training pipelines.
progress advantagereinforcement learningmarkov decision processlog-probability ratioreward model
A cross-process welding penetration status prediction algorithm based on unsupervised domain adaptation in laser and TIG welding
The study proposes an unsupervised domain adaptation (UDA) framework with gradual source domain expansion (GSDE) to address performance degradation in weld penetration state classification across different welding processes (TIG vs. laser). The method achieves 90.65% accuracy on TIGFH and 90.72% on LSPS datasets in same-process settings, outperforming supervised baselines by 35.83% and 38.87%. Cross-process adaptation yields 80.48% (TIG→Laser) and 81.13% (Laser→TIG) accuracy, with 43.39% and 43.40% improvements. UMAP visualizations confirm domain-invariant feature learning.
unsupervised domain adaptationgradual source domain expansionweld penetration classificationtig weldinglaser welding
Model Forensics: Investigating Whether Concerning Behavior Reflects Misalignment
The paper introduces model forensics, a protocol to investigate whether concerning model behavior stems from misalignment rather than benign causes. The method involves analyzing chain-of-thought (CoT) reasoning to generate hypotheses, followed by prompt/environment edits to test them. Evaluated across six agentic environments, the protocol successfully identified Kimi K2 Thinking's disposition for low-effort actions and DeepSeek R1's deceptive consistency motives. Limitations include lack of positive controls for belief detection. The work establishes a baseline for future refinements in model forensics.
model forensicsmisalignment detectionchain-of-thoughtagentic environmentscounterfactual experiments
A welding penetration prediction model for laser welding process based on self-supervised learning using physics-informed neural networks
The paper introduces SimPhysNet, a self-supervised learning model for laser welding penetration prediction that achieves high accuracy with minimal labeled data. The method combines physics-informed neural networks (PINNs) with contrastive learning to extract physically meaningful features from unlabeled molten pool and keyhole images, enhanced by three image augmentation tasks. A few-shot learning strategy using prototypical networks enables robust classification from limited labeled examples. Experiments show 96.06% accuracy using only 200 labeled images (5% of full dataset), matching conventional supervised methods using full datasets.
physics-informed neural networkscontrastive learningfew-shot learninglaser weldingself-supervised learning
The Unfireable Safety Kernel: Execution-Time AI Alignment for AI Agents and Other Escapable AI Systems
The paper introduces the Unfireable Safety Kernel, an execution-time AI alignment mechanism for escapable AI systems (agents with self-modification capabilities). The kernel enforces four architectural properties: process separation, pre-action enforcement, fail-closed operation, and externalized signed evidence. Implemented in Rust with formal verification (Z3, Kani) and tested via Python-to-Rust migration (1000/1000 byte-equivalent fixtures), it successfully blocked 704/704 escape attempts in a self-modifying world model and 6,240 authorization bypass attempts. Comparative evaluation shows superiority over three contemporary agent control systems.
execution-time alignmentescapable ai systemssafety kernelfail-closed invariantself-modification seam
Natural Ungrokking: Asymmetric Control of Which Rules Survive Pretraining
The study identifies 'natural ungrokking' in language models, where learned rules like pronoun-gender associations collapse despite persistent training data evidence. Analyzing pretraining dynamics across multiple corpora and model scales, the authors demonstrate that rule survival depends on support frequency in the training stream, with collapse depth modulated by data-to-parameter ratios. Behavioral collapse occurs when competing patterns outcompete rules, showing irreversible asymmetric control: interventions can destroy rules but not restore them. Pre-registered thresholds confirmed these findings in Pythia checkpoints and synthetic probes.
natural ungrokkingpronoun-gender rulesupport frequencydata-to-parameter ratiobehavioral collapse
TriViewBench: Controlled Complexity Scaling for Multi-View Structural Reasoning in MLLMs
The paper introduces TriViewBench, a diagnostic benchmark for evaluating Multimodal Large Language Models (MLLMs) on controlled-complexity multi-view structural reasoning. The benchmark comprises 1,923 synthetic 3D scenes with 14K QA pairs across four complexity levels and three reasoning categories (Local Decision, Object Counting, Global Recovery). Evaluation of 18 MLLMs reveals a consistent capability hierarchy (Local Decision > Object Counting > Global Recovery) with performance degrading monotonically with complexity (12.11% to 80.02% relative drops). Error analysis identifies occlusion blindness and cross-view confusion as failure modes, while Chain-of-Thought prompting shows negligible improvement (Δ=-0.16%).
multimodal large language modelsstructural reasoningsynthetic 3d sceneschain-of-thought promptingocclusion blindness
Can Trustless Agents Be Trusted? An Empirical Study of the ERC-8004 Decentralized AI Agent Ecosystem
The first empirical study of ERC-8004, a permissionless trust protocol for decentralized AI agent economies, analyzes its effectiveness across Ethereum, BNB Smart Chain, and Base networks. By crawling on-chain Identity/Reputation events, off-chain files, and x402 transactions from deployment through May 2026, the study reveals critical flaws: only 3-15% of registrations expose valid service endpoints, reputation values lack commensurability, and 59-91% of reviewers exhibit Sybil behavior. Post-Sybil filtering leaves 16-89% of agents without valid feedback, demonstrating the protocol's current inability to provide reliable trust signals. The findings yield concrete recommendations for protocol revisions.
erc-8004decentralized aisybil attackon-chain reputationagent trust
Privacy Vulnerabilities of Attention Layers in Tabular Foundation Models and Protection of High-Risk Queries
This paper introduces AMIA (Attention-based Membership Inference Attack), demonstrating that transformer attention mechanisms in tabular foundation models leak membership information, surpassing classical confidence-based attacks by 7.7% in low false-positive regimes. The authors propose an inference-time defense inspired by k-anonymity, reducing membership leakage by 50% against AMIA and 25% against confidence-based attacks, with only 3.9% performance degradation. Additionally, they show that fine-tuning amplifies memorization, making samples with increased prediction confidence more susceptible to MIAs. The defense targets high-risk queries identified by AMIA scores, preserving utility while mitigating privacy risks.
membership inference attackattention mechanismtabular foundation modelsk-anonymityfine-tuning
FORCE: Efficient VLA Reinforcement Fine-Tuning via Value-Calibrated Warm-up and Self-Distillation
FORCE introduces a 3-stage framework for efficient Vision-Language-Action (VLA) model reinforcement fine-tuning, addressing catastrophic unlearning and low-quality exploration data. The method employs Value-Calibrated Warm-Up to stabilize the Q-function via on-policy rollouts, then uses this calibrated Q-function to filter high-value actions during policy updates. Evaluations on simulation and real-world tasks show a 79% absolute success rate improvement, 10% outperformance over prior RL methods, and 32.5% faster training, all without human intervention.
vision-language-actionreinforcement fine-tuningq-function stabilizationvalue-calibrated warm-upself-distillation
Hierarchical Reinforcement Learning for Neural Network Compression (HiReLC): Pruning and Quantization
HiReLC introduces a hierarchical reinforcement learning framework for joint neural network compression via quantization and pruning. The method employs low-level agents (LLAs) for per-kernel configuration selection and high-level agents (HLAs) for global budget allocation, guided by Fisher Information-based sensitivity. A surrogate-guided RL optimization with active learning reduces computational costs, using an MLP surrogate for reward shaping. Experiments on Vision Transformers and CNNs achieve 5.99-6.72× compression with accuracy drops of 0.55-5.62%, validating the hierarchical approach.
hierarchical reinforcement learningneural network compressionquantizationstructured pruningfisher information
Variable Bound Tightening for Nash Equilibrium Computation in Multiplayer Imperfect-Information Games
The paper presents a method for tightening variable bounds in the computation of Nash equilibria for multiplayer imperfect-information games, specifically addressing the nonlinear complementarity problem formulation derived from sequence-form game representations. By deriving finite bounds on slack and multiplier variables, the authors strengthen convex relaxations used in spatial branch-and-bound algorithms, enhancing computational efficiency. Experimental results on three-player Kuhn poker demonstrate significant improvements in solving time compared to prior approaches using Gurobi's nonconvex quadratic solver.
nash equilibriumimperfect-information gamesnonlinear complementarity problemspatial branch-and-boundmccormick envelopes
Autodata: An agentic data scientist to create high quality synthetic data
The paper introduces Autodata, an agentic framework for AI-generated synthetic data creation, where meta-optimized agents act as data scientists to produce high-quality training and evaluation datasets. The method, implemented as Agentic Self-Instruct, leverages inference compute to iteratively improve data quality through self-improving agent training. Experiments demonstrate performance gains over classical synthetic data methods across computer science research, legal reasoning, and mathematical reasoning tasks, with additional improvements from agent meta-optimization. This approach presents a paradigm shift in AI data construction by converting compute into dataset quality.
synthetic datameta-optimizationagentic learningself-instructdata generation
SpeechEQ: Benchmarking Emotional Intelligence Quotient in Socially Aware Voice Conversational Models
The paper introduces SpeechEQ, a benchmark for evaluating emotional intelligence quotient (EQ) in speech-language models (SLMs) through multimodal dialogue understanding. The framework comprises 2,265 dialogues across 15 EQ subscales (aligned with EQ-i 2.0 theory) and proposes a Spoken EQ (SEQ) score for multi-turn assessment. Experiments reveal that end-to-end SLMs outperform cascaded systems but suffer from text-reliant modality shortcuts, safety traps from alignment, and contextual amnesia. The benchmark exposes limitations in current models' paralinguistic reasoning during active speech interactions.
speech-language modelsemotional quotientmultimodal dialogueparalinguistic cuescontextual amnesia
Weave of Formal Thought
Weave of Formal Thought (WoFT) introduces a paradigm combining rigorous syntactic validation with learned structural representations for code generation in LLMs. The method integrates a formal engine and constrained decoder that is sound and complete with respect to Tree-sitter specifications, employing speculative-lexing with GLR parsing to ensure valid program prefixes. Additionally, a latent-variable fine-tuning approach trains models to interleave non-terminal grammar symbols into generation using the reweighted wake-sleep algorithm to optimize IW-ELBO. Fine-tuning StarCoder2-3B with this method reduces per-token cross-entropy by 14.3% compared to a text-only SFT baseline, demonstrating improved structural information retention.
speculative-lexingglr parsingtree-sitterreweighted wake-sleepiw-elbo
InvestPhilBench: A Multi-Layer Dynamic Benchmark for Evaluating Large Language Model Procedural Reasoning in Expert Investment Philosophy
The paper introduces InvestPhilBench, a multi-layer dynamic benchmark for evaluating large language models' procedural reasoning in expert investment philosophy. The benchmark comprises 118 principle cards, 25 decision framework cards, and 243 QA questions, with a Benchmark Automated Scoring Pipeline (BASP) featuring five algorithmic metrics and a Failure Mode Detection Protocol (FMDP). Initial results on a 188-question development split reveal a sharp performance gap between provider tiers (BASP 0.906 vs. 0.438) and expose procedural deficits in frontier models (Claude L4 GRA ≈0.77, L7 GRA 0.57-0.62), with BASP composite tracking human reference at Pearson r=0.72.
procedural reasoninginvestment philosophyautomated scoring pipelinefailure mode detectiongate reconstruction accuracy
Multi-Agent Goal Recognition with Team- and Goal-Conditioned Reinforcement Learning and Factorized Branch-and-Bound
The paper introduces Multi-Agent Goal Recognition with Branch-and-Bound (MAGR-BB), a method for jointly inferring team compositions and goals from observed trajectories. The approach combines a team- and goal-conditioned policy as a scoring model with factorized branch-and-bound search to efficiently explore the combinatorial hypothesis space. Evaluated on a multi-agent Blocksworld benchmark, MAGR-BB matches exhaustive search accuracy while reducing hypothesis materialization by orders of magnitude and significantly decreasing cumulative recognition runtime.
multi-agent goal recognitionbranch-and-bound searchteam-conditioned policycombinatorial hypothesis spaceblocksworld benchmark
Helpful or Harmful? Evaluating LLM-Assisted Vulnerability Patching via a Human Study
The study evaluates LLM-assisted vulnerability patching versus manual debugging through a human experiment, hypothesizing that LLMs may accelerate repairs but risk introducing insecure code. Using a Balanced Crossover design, the authors developed a WebApp with Ghost Tests to assess patch integrity beyond functional checks. A pilot study provided preliminary insights, with metrics including remediation speed, efficacy in functionality and security tests, and participant perception.
llm-assisted patchingvulnerability remediationghost testsbalanced crossoversecurity validation
WinDOM: Self-Family Distillation for Small-Model GUI Grounding
WinDOM introduces a self-family distillation (SFD) framework and a 54,425-record GUI-grounding corpus to improve small-model (∼2B) performance without expensive human annotation. The corpus is harvested via headless Playwright on a Windows 11 web reimplementation, extracting bounding boxes directly from the DOM. SFD employs a rejection-sampling cold-start parameterized by teacher choice (EMA student or frozen larger same-family teacher), treating saturation depth as a GRPO hyperparameter. Results show Qwen3.5-2B with SFD-4B and Early-init RL achieves +5.4 OOD-mean improvement over the base model, while EMA mode performs comparably without external teachers.
self-family distillationgui-groundingrejection-samplingcold-startgrpo
Agentic System as Compressor: Quantifying System Intelligence in Bits
The paper proposes a compression-based framework for quantifying intelligence in agentic systems, operationalized through arithmetic coding, seed coding, and fallback mechanisms. The method evaluates codelength reduction across five settings: reversed text, chess moves, protein sequences, retrieval-augmented QA, and semantic story compression, demonstrating that agentic components consistently decrease residual uncertainty. Results validate the approach for analyzing system components, observer roles, and computational budgets in controlled experiments mirroring real-world agentic systems.
agentic systemsarithmetic codingcodelength reductionresidual uncertaintyretrieval-augmented qa
SE-AGCNet: An End-to-End Framework for Joint Speech Enhancement and Loudness Control in Meeting Scenarios
The authors propose SE-AGCNet, an end-to-end framework for joint speech enhancement (SE) and automatic gain control (AGC) in meeting scenarios, addressing limitations of discrete module approaches. The method integrates SE and AGC optimization, leveraging their synergy through a specialized data simulation pipeline (SE-AGC-DataGen) and standardized loudness metrics (LUFS, St LUFS, LRA). Experiments demonstrate SE-AGCNet achieves target loudness while improving speech quality and ASR accuracy over baselines.
speech enhancementautomatic gain controlend-to-end learningloudness metricsmeeting scenarios
Pulmonary Embolism Risk Stratification from CTPA and Medical Records: Vascular Graphs Are Not All You Need
This study evaluates PE risk stratification using medical records and CTPA-derived biomarkers, challenging the assumption that vascular graphs enhance predictive performance. The authors benchmark tabular models and GNNs on a private dataset (n=353) with complete PE risk data. Results indicate medical records and cardiac biomarkers are the most predictive features, while vascular biomarkers and GNNs fail to improve stratification accuracy. The work hypothesizes that vascular graphs may lack discriminative power for PE risk assessment. Code is available at https://github.com/creatis-myriad/GENESIS.
pulmonary embolismrisk stratificationctpagraph neural networksbiomarkers
Measurable Majorities Are Not Finitely Axiomatizable
The paper resolves Conjecture 5.7 by Moss and Pedersen (2026), proving that measurable majorities in finite social decision frames are not finitely axiomatizable. Using a geometric construction based on orthogonality and dimension in rational vector spaces, the authors isolate a symmetric family of half-sized voting blocs and extend it to a maximal frame. They demonstrate that for every $k\ge 1$, there exists a maximal standard frame with coherence violations of length exactly $2k+2$, showing no uniform finite bound on incoherence indices. This result, combined with the Moss-Pedersen minimal logic, confirms the non-finite axiomatizability of measurable social decision frames.
finite axiomatizabilitysocial decision framescoherence criterionrational vector spacesincoherence index
Explainable Control Framework (XCF) based on Fuzzy Model-Agnostic Explanation and LLM Agent-Supported Interface
The paper proposes an Explainable Control Framework (XCF) with three key contributions: (1) a model-agnostic framework for closed-loop controller explanations, optionally refined by system dynamics; (2) HFMAE-C, a hierarchical fuzzy method generating multi-level explanations via IF-THEN rules and salience values; (3) an LLM agent interface for automated explanation generation and interactive consultation. The method employs fuzzy logic to approximate controller behavior and system dynamics. Case studies on inverted pendulum and Turtlebot obstacle avoidance demonstrate effectiveness through user experiments and quantitative comparisons with existing approaches.
explainable controlfuzzy logicmodel-agnostic explanationllm agentclosed-loop systems
Overview of HIPE-2026: Person-Place Relation Extraction from Multilingual Historical Texts
HIPE-2026 introduces a temporally grounded relation extraction task for multilingual historical documents, evaluating $at$ (past presence) and $isAt$ (contemporaneous presence) relations between persons and places. The challenge involves 17 teams processing noisy OCR text from 19th-20th century newspapers in French, German, and English, plus a surprise-domain early modern French literary set. A three-fold evaluation framework assesses accuracy, efficiency, and cross-domain generalization across 40+ submissions, revealing trade-offs between LLM-based and lightweight approaches. Results demonstrate the current state of historical relation extraction, with system descriptions and datasets provided for reproducibility.
relation extractionhistorical documentsmultilingual nlptemporal groundingocr noise
Enhancing Brain MRI Anomaly Detection and Reasoning with ROI Rethink and Synthetic Data
BrReMark introduces explicit region-of-interest (ROI) marking for brain MRI diagnosis, combining hypothesis generation with spatial grounding and verification. The framework employs supervised fine-tuning on structured reasoning trajectories and reinforcement learning with a composite reward for localization accuracy and diagnostic reasoning, augmented by domain randomization-based pathology synthesis. Evaluations show mAP50 improvement from 0.74% to 37.54% on internal benchmarks, with 21.57% Clinical F1 and 45.26% diagnostic accuracy. On the NOVA OOD benchmark, BrReMark reduces false positives by 45.7% compared to state-of-the-art, demonstrating reduced hallucination on rare pathologies.
roi markingdomain randomizationreinforcement learningmri diagnosisout-of-distribution
AI-Assisted Computational Reproducibility on the FABRIC Testbed
The paper demonstrates how the FABRIC testbed, augmented with LLM coding assistants via LoomAI, enhances computational reproducibility across diverse scientific domains. Three case studies were reproduced: BBR-family congestion-control evaluations, LAMMPS molecular dynamics scaling benchmarks on CPU-only MPI clusters, and stress protein homeostasis genomics pipelines. The AI assistant effectively handled environment setup, code adaptation, and debugging but required human intervention for analysis stages lacking clear workflows. The AI-assisted workflow reduced reproduction effort by 4--6 times. Practical recommendations are provided for improving AI-assisted reproducibility on research testbeds.
computational reproducibilityfabric testbedllm coding assistantsmolecular dynamicsgenomics pipelines
AutoRelAnnotator: Calibrated Model Cascades for Cost-Efficient Relevance Evaluation in Sponsored Search
We propose AutoRelAnnotator, a calibrated model cascade for cost-efficient offline relevance annotation in sponsored search, addressing limitations of human labeling and off-the-shelf LLMs. The method combines domain-specific fine-tuning, query routing through progressively larger classifiers, and per-class isotonic calibration. Fine-tuning improves accuracy by 20 points, cascading halves compute cost without accuracy loss, and calibration adds a statistically significant +0.6-point gain. Validated in production across six use cases, the system processed over 150M annotations, enabling faster experimentation cycles and scalable offline annotation pipelines for search and advertising systems.
model cascadeisotonic calibrationquery routingoffline annotationsponsored search
Color Matters: Trigger Color Affects Success in Federated Backdoor Attacks
The paper demonstrates that trigger color significantly impacts success rates in federated backdoor attacks, even when trigger semantics and placement remain constant. Using a semantics-driven attack framework, the authors evaluate black and white variants of natural visual accessories (e.g., masks, sunglasses) in a four-class CelebA hair-color classification task under both standard poisoning and SABLE-based objectives. Results show color-dependent efficacy: white triggers perform better when targeting the blond class, while black triggers excel for the black class, with trends persisting under robust aggregation.
federated learningbackdoor attackssemantic triggerspoisoning objectiverobust aggregation
Semantic Consistency Policy Optimization for Reinforcement Learning of LLM Agents
The paper introduces Semantic Consistency Policy Optimization (SCPO), a value-free reward-shaping method for reinforcement learning of LLM agents that addresses semantic credit inconsistency in group-based approaches. SCPO mitigates conflicting gradients by deriving step-level credit from successful siblings within the same rollout group, rewarding partial progress in failed trajectories. Evaluated on ALFWorld and WebShop, SCPO achieves 93.7±4.1% and 74.8±2.0% success rates respectively at 1.5B parameters, with notable improvements on multi-step tasks.
reinforcement learningllm agentsreward shapingsemantic consistencypolicy optimization
Edges Before Embeddings: A Confidence-Aware Blur Gate for Vision-Language Pipelines
The paper introduces MagikaDocumentFromPixel, a lightweight image quality gate for vision-language pipelines that classifies images as sharp, blurred, or uncertain in ~7 ms on a single CPU core. Key contributions include: (i) an empirical search identifying input resolution as the dominant factor, with architecture capacity beneficial only at ≥384 px; (ii) a confidence-aware routing formalism; (iii) the Edge Prior Module (EPM), a Laplacian-magnitude auxiliary input channel boosting F1 by +1.3 points; and (iv) identification of a recurring design pattern across multiple systems. The final MobileNetV3-Large + EPM model achieves F1 = 0.9803 (AUC 0.9989) with a 17 MB ONNX artifact, improving over the baseline by +1.31 points. Limitations include single motion-blur distribution evaluation and unmeasured calibration.
image quality gateedge prior moduleselective predictionvision-language pipelinetest-time augmentation
AI Snitches Get Glitches: Towards Evading Agentic Surveillance
The paper formalizes agentic surveillance, where AI agents analyze information, craft reports, and transmit them using available tools, posing risks to user privacy. To evaluate this, the authors introduce SurveilBench, a dataset with reporting scenarios across corporate, education, and police domains. Experiments reveal emergent surveillance tendencies in some models, alongside unintended reporting of surveillance attempts to authorities. The authors adapt prompt injection techniques to evade surveillance, proposing methods to hide, deceive, or induce over-escalation in surveillance agents. The findings highlight the ease of implementing agentic surveillance and advocate for comprehensive technical, ethical, and legislative safeguards.
agentic surveillancesurveilbenchprompt injectionemergent tendenciesover-escalation
MiniOpt: Reasoning to Model and Solve General Optimization Problems with Limited Resources
MiniOpt introduces a reinforcement learning framework for solving general optimization problems with limited resources, employing a 'reasoning-to-model-and-solve' paradigm that decomposes optimization into structured modeling and solver generation. The method features OptReward, a hierarchical reward function evaluating both formulation and solution, and an optimization-oriented policy optimization strategy for efficient exploration. Experiments demonstrate that MiniOpt-3B achieves strong optimization generalization across diverse problem types, with the highest average solving accuracy for models under 10B parameters and competitive performance for larger models.
reinforcement learningoptimization generalizationhierarchical rewardpolicy optimizationlanguage models
SARA: Unlocking Multilingual Knowledge in Mixture-of-Experts via Semantically Anchored Routing Alignment
The paper introduces SARA (Semantically Anchored Routing Alignment), a framework to improve multilingual knowledge transfer in Mixture-of-Experts (MoE) models by addressing cross-lingual routing divergence. SARA aligns routing distributions of low-resource language inputs with high-resource semantic anchors using a symmetric Jensen-Shannon divergence constraint, operating directly on MoE layer mechanics rather than output logits. Experiments on 2 LLMs (Qwen3-30B-A3B and Phi-3.5-MoE-instruct) across 5 low-resource languages show performance gains (+0.8% to +1.2% on Global-MMLU), demonstrating effective mitigation of low-resource bottlenecks.
mixture-of-expertsrouting divergencejensen-shannon divergencemultilingual transferlow-resource languages
Confidence Sequences for Online Statistical Model Checking of Markov Decision Processes
The paper introduces confidence sequences for online statistical model checking of Markov Decision Processes (MDPs), addressing the unrealistic assumption of exact probability knowledge in traditional methods. By leveraging statistical techniques, the authors design and implement efficient confidence sequences that outperform classical union-bound approaches. Their tool demonstrates practical applicability, requiring 50x fewer samples on average than state-of-the-art methods while providing robust guarantees on transition probabilities.
markov decision processesconfidence sequencesstatistical model checkingtransition probabilitiesonline verification
Do Encoders Suffice? A Systematic Comparison of Encoder and Decoder Safety Judges for LLM Adversarial Evaluation
The paper systematically compares encoder-based classifiers (ModernBERT, Ettin) against LLM-based judges (LlamaGuard 3/4, Claude-as-a-judge) for safety evaluation of LLM outputs. Using adversarial datasets (JailbreakBench, AILuminate) and multiple prompting strategies, it benchmarks performance via F1, false negative rate, and precision-recall metrics across attack techniques like decomposition and context manipulation. Results show encoder classifiers can match LLM judges in certain scenarios while reducing cost and latency, providing practical guidance for deployment.
encoder classifiersllm safety evaluationadversarial datasetsprecision-recall metricscontext manipulation
Fuzzy Quantification over OWL Ontologies and Knowledge Graphs
The paper introduces a framework for evaluating fuzzy quantification queries over OWL ontologies and RDFS knowledge graphs, supporting both Type I and Type II fuzzy quantified expressions. The method is quantifier-agnostic, evaluation-method-agnostic, and data-source-agnostic, offering broad adaptability. The authors present Q2S2, an open-source implementation to facilitate further research in this domain.
fuzzy quantificationowl ontologiesknowledge graphstype i fuzzytype ii fuzzy
Space-Efficient Language Generation in the Limit
The paper introduces a resource-aware framework for language generation under space constraints, where a learner processes an adversarial stream to output a hypothesis language $L \subseteq K$ with minimal omissions ($\Delta$). Focusing on $\mathcal{C}_{s,k}$ (DFA-recognized languages with $s$ states over $k$-sized alphabets), the authors prove exact identification is possible in exponential space. For polynomial-space regimes, they present a streaming algorithm achieving $\Delta = O(k^{2s-2})$ and capturing strings of length $\geq 2s-1$. A lower bound shows $\Delta \leq k^{(1-\varepsilon)s}$ requires $k^{\Omega(\varepsilon s)}$ memory, revealing a sharp trade-off between space and generation quality.
language generationspace efficiencydfastreaming algorithmlower bound
OncoSynth: Synthetic data generation for treatment effect estimation in oncology
OncoSynth introduces a diffusion-based generative framework for synthetic oncology data that preserves causal relationships between covariates, treatments, and outcomes. The method employs sequential modeling of covariate-treatment interactions and treatment-survival effects, evaluated on lung (N=37,128) and breast cancer (N=17,046) cohorts. Results demonstrate 66% reduction in population-level and 58% in patient-level treatment effect estimation errors compared to existing approaches, enabling reliable precision oncology evidence generation under data-sharing constraints.
diffusion-basedcausal relationshipstreatment effectssynthetic cohortsprecision oncology
Uncertainty Quantification for Computer-Use Agents: A Benchmark across Vision-Language Models and GUI Grounding Datasets
The paper introduces Argus, a benchmark for evaluating post-hoc uncertainty quantification (UQ) methods in vision-language models (VLMs) used for GUI grounding. It evaluates 27 open-weight and 8 closed-source UQ methods across 4 VLM agents and 4 datasets, covering logit-based scores, sampling measures, hidden-state estimators, and verbalized confidence. Key findings show UQ rankings are stable within models but degrade across model classes and interfaces, with hidden-state methods being the most stable. Cross-tier transfer to closed-source models is weak (average +0.08 Spearman rho), necessitating target-specific reranking. The study releases datasets and scripts for regime-aware UQ selection.
uncertainty quantificationvision-language modelsgui groundinghidden-state estimatorsconformal prediction
Gradient-based inverse lithography for EUV masks via the waveguide method and a physics-informed neural operator
The authors present a gradient-based inverse lithography technology (ILT) framework for extreme ultraviolet (EUV) masks, combining differentiable waveguide methods with a physics-informed neural operator (WGNO) to optimize mask permittivity. The approach uses automatic differentiation through a full forward diffraction model to recover absorber structures. Numerical experiments on 2D/3D TaBN, La, and U absorbers at 11.2 nm wavelength demonstrate successful wafer-field target achievement.
inverse lithographyextreme ultravioletwaveguide neural operatorautomatic differentiationpermittivity optimization
Point Cloud Diffusion with Global and Local Reconstruction for Instance-Level 3D Anomaly Detection
The paper introduces PCDiff, a point cloud diffusion framework for 3D anomaly detection, addressing challenges in reconstructing weak defects (e.g., scratches with deviations ~10^-3) and avoiding positional bias in non-defective regions. The method employs instance-level multi-modal attention during anomaly generation, conditioned on texture gradient, image patch, text, and mask, followed by a joint local-global reconstruction algorithm for detection. Experiments show PCDiff outperforms state-of-the-art methods in anomaly generation fidelity and reconstruction quality, significantly improving detection accuracy.
point cloud diffusion3d anomaly detectioninstance-level attentionmulti-modal generationlocal-global reconstruction
Position Spaces and Graphs
The paper introduces position graphs, a graph-based reasoning framework formalizing position spaces using two strict partial orders for horizontal/vertical alignment and precedence. The framework models discrete token positions with chain conditions and compatibility constraints focusing on rows and columns. Theoretical analysis establishes graph consistency conditions and investigates the NP-completeness of induced subgraph isomorphism within position graphs. While motivated by document processing, the work emphasizes mathematical properties and algebraic consistency of position-based constraints, providing a formal logical layer independent of specific data extraction techniques.
position graphsstrict partial ordersgraph consistencyinduced subgraph isomorphismalgebraic consistency
GUI agent: Guided Exploration of User-Sensitive Screens
The paper introduces a GUI exploration agent that identifies user-sensitive states in LLM-driven task automation, addressing safety concerns in real-world deployment. The agent systematically explores query spaces from demonstrated tasks to detect sensitive scenarios requiring user handover. This approach categorizes sensitive states and queries, providing a dataset to improve agent reliability by preventing unsafe autonomous actions in GUI environments.
llm agentsgui automationuser-sensitive statesquery explorationsafety handover
Power-Budgeted Underwater Vehicle Control via Constrained Reinforcement Learning
The paper proposes a constrained reinforcement learning approach for energy-efficient underwater vehicle control, formulating it as a constrained Markov decision process with explicit power budgets. Using a PPO-Lagrangian algorithm, the method dynamically adjusts a dual variable to meet specified power constraints without manual weight tuning. Evaluations in MarineGym across three vehicles and four tasks show 14-65% power reduction (up to 64.9%) versus task-only baselines while maintaining task accuracy and smoother actuation in 10/12 settings.
constrained reinforcement learningppo-lagrangianpower budgetunderwater vehicle controlmarinegym
Steering Vision-Language Models with Joint Sparse Autoencoders
The paper introduces Joint Sparse Autoencoders (JSAEs) for interpretable cross-modal feature extraction in vision-language models (VLMs), addressing limitations of standard sparse autoencoders. JSAEs employ an alignment constraint to jointly factorize vision and language activations into shared features, enabling bidirectional interventions. Experiments on LLaVA-v1.6-Mistral-7B, Llama3-LLaVA-8B, and Qwen3-VL-30B reveal layer-dependent asymmetry: additive steering peaks at mid-to-late layers while suppression remains stable across layers. Results demonstrate JSAEs improve controllable intervention-based analysis compared to unconstrained alternatives.
sparse autoencodersvision-language modelscross-modal alignmentinterpretable featureslayer-localized interventions
Is GraphRAG Needed? From Basic RAG to Graph-/Agentic Solutions with Context Optimization
The paper introduces a framework for evaluating and comparing advanced RAG variants (GraphRAG, Modular RAG, Agentic RAG) against basic RAG across 9 standardized scenarios involving semi-structured knowledge bases. It proposes a novel context engineering method that reduces token usage by 19%-53% through optimized text-graph representations and agentic loop designs. Experiments reveal a retrieval-generation gap where expanded retrieval fails to proportionally improve generation quality, suggesting current metrics overstate advanced retrieval benefits.
graphragagentic ragcontext engineeringretrieval-generation gapsemi-structured knowledge
Taxonomy of Risks on Automated Fact-Checking Systems Considering its Propagation
The paper presents a taxonomy of 32 specific risks in automated fact-checking systems, considering their propagation through three stages: risk factors, hazardous situations, and harm. The authors employ this taxonomy as analytical guide words to assess risks in the DEFAME system, demonstrating that their method identifies risks overlooked by conventional IT security frameworks like STRIDE. Results indicate improved risk detection capability for AI-based fact-checking systems, addressing challenges such as incorrect judgments and misinformation spread.
automated fact-checkingrisk propagationmisinformationdefame systemstride framework
Staying In Character: Perspective-Bounded Memory For Book-Based Role-Playing Agents
The paper introduces REVERIEMEM, a three-layer memory architecture for book-based role-playing agents that addresses Factual Overreach and Stylistic Monotony in long-narrative role-playing. The architecture comprises episodic (first-person scene memories), semantic (visibility-tagged facts), and personality (situation-dependent patterns) layers. Evaluated on KBF-QA (4,386 questions over eight novels) and BOOKWORLD's narrative protocol, REVERIEMEM improves Knowledge Boundary Fidelity by 34.6 percentage points over prior methods and achieves a ~79% win rate in narrative generation.
memory architectureknowledge boundary fidelityrole-playing agentsparametric memorynarrative generation
TL++: Accuracy and Privacy Preserving Traversal Learning for Distributed Intelligent Systems
TL++ introduces a two-mode traversal-learning framework for distributed systems that preserves accuracy and privacy while reducing communication costs. Base mode exchanges cut-layer activations/gradients instead of full models, while secure mode employs secret sharing between non-colluding servers to protect cut-layer tensors. Evaluated on CIFAR-10 and PubMedQA, TL++ achieves 91.41% accuracy (base) and 90.93% (secure), outperforming baselines by >12 percentage points while reducing per-step communication 13.1×. The method approaches centralized performance with activation-level privacy guarantees under semi-honest assumptions.
traversal learningsecret sharingcut-layer activationsdistributed trainingsemi-honest security
Reasonable Motion: A General ASP Foundation for Environment Constrained Movement Trajectory Computation
The paper introduces a hybrid quantitative-qualitative method for computing constrained motion trajectories using answer set programming (ASP). The approach traverses an environment graph to enumerate geometrically admissible motion behaviors as stable models, each representing distinct trajectory modes characterized by event sequences, map topology, and domain norms. Demonstrated on the Argoverse 2 autonomous driving benchmark, the method provides verifiable interpretability through traceable stable models, contrasting with purely learned approaches. Empirical evaluation confirms applicability in dynamic domains with moving objects.
answer set programmingtrajectory computationstable modelsenvironment graphautonomous driving
Probabilistic Agents in Deterministic Audits: Evaluating Multi-Agent Systems for Automated Audits Based on the German IT-Grundschutz
The paper presents a Multi-Agent System (MAS) with Hybrid Retrieval Augmented Generation (HybridRAG) for automating German IT-Grundschutz (IT-GS) certification. Key innovations include a Hypothesis-Verification Loop for cross-referencing agent inferences against a Knowledge Graph to reduce hallucinations, and a Decoupled Reasoning Pipeline separating semantic extraction from deterministic protection need inheritance. Evaluated on the BSI's 'RecPlast GmbH' case study, the system achieves high performance in semantic tasks (Structural Analysis, Modeling) but shows limitations in logical reasoning phases (Protection Needs Assessment, IT-GS Check) due to LLMs' probabilistic nature conflicting with deterministic requirements.
multi-agent systemhybrid retrieval augmented generationhypothesis-verification loopknowledge graphprotection need inheritance
An Approach for a Supporting Multi-LLM System for Automated Certification Based on the German IT-Grundschutz
The paper introduces a Multi-Large Language Model system (MLS) with Hybrid Retrieval-Augmented Generation (HybridRAG) for semi-automated BSI IT-Grundschutz certification. Addressing NIS2 directive challenges, specialist shortages, and high costs, the MLS architecture integrates Large Language Models (LLMs) and Knowledge Graphs (KGs) to streamline certification phases: protection needs assessment, modeling, IT-Grundschutz check, measure consolidation, and realization. The approach enhances efficiency, reduces costs, and maintains security concept quality amid rising demand.
multi-large language modelhybrid retrieval-augmented generationbsi it-grundschutznis2 directiveknowledge graphs
Expresso-AI: Explainable Video-Based Deep Learning Models for Depression Diagnosis
We introduce Expresso-AI, an explainable video-based deep learning framework for depression severity diagnosis that enhances interpretability and predictive performance. The method fine-tunes Deep Convolutional Neural Networks (DCNN) pre-trained on Action Recognition datasets using facial videos from the AVEC depression dataset, then analyzes saliency maps to examine face regions and temporal expression semantics. The framework generates both visual and quantitative explanations for model decisions while improving upon previous single-face benchmarks in visual depression diagnosis. Results demonstrate enhanced predictive capabilities alongside interpretable insights into temporal facial activities.
deep convolutional neural networkssaliency mapstemporal expression semanticsdepression severity diagnosisaction recognition
Low-Complexity Policy Tessellations in Structured Markov Decision Processes
The paper demonstrates that optimal policies in structured Markov decision processes induce low-complexity tessellations, enabling direct approximation of policy regions rather than value functions. The authors propose boundary-based policy approximations and derive a policy-loss decomposition linking performance degradation to action margins, explaining error concentration near indifference boundaries. Experiments in inventory control and queue admission show superior performance to reinforcement learning baselines, with lower policy error, smaller value gaps, faster error decay, and improved stability.
markov decision processespolicy tessellationsboundary-based approximationaction marginsreinforcement learning
BiPACE: Bisimulation-Guided Policy Optimization with Action Counterfactual Estimation for LLM Agents
The paper introduces BiPACE, a novel advantage estimator for LLM agents that addresses credit assignment issues in stepwise group-based RL. BiPACE combines bisimulation-guided state clustering (BiGPO) with action counterfactual estimation (PACE) to improve policy optimization without requiring additional critics or rollouts. Evaluations on ALFWorld/Qwen2.5-7B show BiPACE increases validation success from 90.8% to 97.1±0.9%, outperforming GiGPO across multiple benchmarks and model scales with only 11.3% computational overhead.
bisimulationpolicy optimizationcredit assignmentllm agentscounterfactual estimation
SFL-MTSC: Leveraging Semantic Frame-Level Multi-Task Self-Consistency for Robust Multi-Intent Spoken Language Understanding
The paper introduces Semantic Frame-Level Multi-Task Self-Consistency (SFL-MTSC), a structured aggregation framework addressing inconsistent intent--slot structures in prompt-based spoken language understanding (SLU) with LLMs. The method decomposes predictions into intent-specific frames, applies domain--intent grouping and slot-level clustering, and evaluates cluster reliability via path support scoring before reintegrating reliable frames. Zero-shot experiments on MAC-SLU show improved slot F1 (+3.2 points) and overall accuracy (+1.8 points) over single-path inference while maintaining stable intent accuracy.
spoken language understandingmulti-intentsemantic framesself-consistencyzero-shot learning
Agentic evolution of physically constrained foundation models
The paper introduces a physically grounded multi-agent discovery engine for hardware-compliant AI system design, leveraging an Evolutionary Knowledge Graph to guide algorithmic Chain-of-Thought transformations. The framework achieves directed structural evolution, demonstrated through two novel compression techniques: Q-Enhance reduces long-context accuracy loss in dense models, while MoE-Salient-AQ outperforms manual sparse Mixture-of-Experts designs by 3.7% in sub-3-bit regimes. A bandwidth-efficient Sensitivity Profile enabled deployment of a 235B-parameter model on dual-A100 hardware with 75% memory reduction and only 0.64% accuracy degradation, establishing hardware-software co-design via knowledge-driven autonomy.
evolutionary knowledge graphalgorithmic chain-of-thoughthardware-software co-designmixture-of-expertssensitivity profile
Evaluating LLMs on Real-World Software Performance Optimization
The paper introduces SWE-Pro, a repository-level benchmark for evaluating LLMs on real-world software performance optimization, derived from 102 expert-written optimizations in open-source projects. Unlike prior benchmarks, SWE-Pro assesses runtime, peak memory, and Time-Weighted Memory Usage (TWMU) under noise-aware conditions with parameterized tests. Results reveal LLMs perform poorly: runtime gains are negligible, and memory optimizations are rare, contrasting with expert implementations achieving 15.5x speedup and 171.3x peak memory reduction. Experts improved 91.2% of tasks for runtime and 65.7% for peak memory, highlighting a significant LLM capability gap.
software performance optimizationlarge language modelstime-weighted memory usagenoise-aware measurementrepository-level benchmark
STEB: A Speech-to-Speech Translation Expressiveness Benchmark for Evaluating Beyond Translation Fidelity
We introduce STEB (Speech-to-Speech Translation Expressiveness Benchmark), a 32.6-hour Chinese--English benchmark evaluating both standard (translation fidelity, speaker similarity, duration alignment) and expressive (emotion, scenario style, nonverbal vocalizations) dimensions in speech-to-speech translation (S2ST). STEB employs a caption-then-summarize framework, converting speech into structured expressive attributes and comparing source and hypothesis attributes using an LLM judge. Human validation shows significant correlations with listener judgments across expressive dimensions. Evaluation of six S2ST systems reveals a gap between semantic and expressive transfer, with cascaded systems achieving strong translation fidelity but struggling in emotion (best: 3.82/5) and NV preservation (best: 2.31/5).
speech-to-speech translationexpressiveness benchmarknonverbal vocalizationscaption-then-summarizetranslation fidelity
The impact of artificial intelligence on enterprise software user roles
This qualitative study contributes an empirical analysis of AI's impact on enterprise software user roles, focusing on SAP's Business Technology Platform (BTP). Through expert interviews (n=20) and a participatory workshop (n=24), the research identifies three key transformations: automation of operational tasks, enhanced human-AI collaboration, and increased reliance on agentic AI systems. Findings reveal substantial shifts in development workflows and professional responsibilities, necessitating adaptations to existing user-role frameworks like the BTP User Type Matrix. The study underscores the need for revised role taxonomies, new governance mechanisms, and AI-native design approaches in enterprise software systems.
agentic aiuser-role frameworksbusiness technology platformhuman-ai collaborationrole taxonomies
Cliff Tokens: Identifying Single-Token Failure Triggers in LLM Mathematical Reasoning
The paper introduces cliff tokens, single-token failure triggers in LLM mathematical reasoning, identified via an adaptive threshold based on token-wise potential drops. Using a one-sided two-proportion z-test, the method demonstrates that deleting the first cliff token recovers pass@64 to 1.0 across seven models and three benchmarks (GSM1K, MATH500, AIME 2025). A taxonomy of deterministic, uncertain, and sampled-off cliffs is proposed, validated by Cliff-DPO, which improves accuracy by up to +6.6 when optimizing at uncertain and sampled-off cliffs.
cliff tokensmathematical reasoningtoken-wise potentialcliff-dpofailure triggers
Quantization Inflates Reasoning: Token Inflation as a Hidden Cost of Low-Bit Reasoning Models
The study reveals that low-bit post-training quantization of reasoning models introduces a hidden computational cost through increased reasoning-token usage, despite preserving final-answer accuracy. The authors analyze this effect across mathematical reasoning, code generation, scientific QA, and tool-use benchmarks, introducing the CoT Token Inflation Ratio to quantify reasoning-length inflation under INT4/INT3 quantization. Results show behavioral changes in reasoning traces (more steps, semantic repetition) that translate to real-world serving penalties, with quantization-aware training emerging as the most promising mitigation strategy among evaluated approaches.
quantizationtoken inflationchain-of-thoughtpost-training quantizationreasoning models
Spam and Sentiment Detection in Arabic Tweets Using MARBERT Model
The study proposes a sentiment analysis framework for Arabic tweets using MARBERT, a BERT-based model, to enhance customer service for Saudi Telecom Company (STC). The method fine-tunes MARBERT on a dataset of 24,513 Arabic tweets annotated with sentiment labels (positive, negative, neutral, sarcasm, indeterminate). Performance is evaluated via F1-score, precision, and recall, demonstrating superior accuracy compared to existing techniques. The approach addresses a gap in Arabic NLP by leveraging deep learning for sentiment classification in a low-resource language setting.
sentiment analysismarbertarabic nlpdeep learningcustomer service
HG-Bench: A Benchmark for Multi-Page Handwritten Answer-Region Grounding in Automated Homework Assessment
The paper introduces HG-Bench, a benchmark for evaluating multi-page handwritten answer-region grounding in automated homework assessment. The benchmark comprises 500 human-annotated K-12 homework samples from a 1,489,278-image pool, featuring hierarchical containment constraints linking question-level and step-level regions. A page-aware evaluation protocol measures complete-answer localization (FA) and step-level decomposition (FSm). Results show zero-shot systems achieve ≤55.22% FA and ≤48.22% FSm, while a fine-tuned GLM-4.6V 9B model reaches 74.97/72.26, highlighting a capability gap in step-level handwritten grounding.
handwritten answer-region groundingmulti-page homework assessmenthierarchical containment constraintpage-aware evaluation protocolstep-level decomposition
Rate-Aware Quantum-Inspired Trajectory Learning for Interference-Limited Multi-UAV Networks
The paper proposes Rate-Aware Quantum-Annealed Graph Condensation (RA-QAGC), a novel scheme for interference-limited multi-UAV trajectory optimization. The method combines rate-aware graph abstraction with decentralized reinforcement learning to address dimensionality challenges in real-time UAV coordination, focusing on throughput-optimal region identification and QoS-aware trajectory adaptation. Simulations show RA-QAGC achieves 59.4 Mbps total throughput and 23.9 Mbps priority-user throughput, outperforming baselines by 15% and 34% respectively.
quantum-annealed graph condensationmulti-uav networkstrajectory optimizationinterference-limited environmentsdecentralized reinforcement learning
A Red Teaming Framework for Large Language Models: A Case Study on Faithfulness Evaluation
The paper introduces a multi-role red teaming framework for evaluating LLM vulnerabilities, comprising target, attacker, and jury models. The method systematically generates adversarial prompts to test response faithfulness, with jury models assessing accuracy and consistency. Experiments show a 7.9% increase in attack success rate for question-answering tasks, revealing that architectural choices impact safety more than parameter scaling. The framework demonstrates cross-linguistic adaptability but struggles with automated prompt generation and detecting subtle unfaithfulness across languages.
red teamingadversarial promptsfaithfulness evaluationmulti-role architecturecross-linguistic vulnerability
EchoStyle: Unlocking High-Fidelity Video Stylization with Reverse Data Synthesis
EchoStyle introduces a scalable text-driven framework for high-fidelity video stylization, addressing key challenges in content leakage, data scarcity, and long-video adaptability. The method employs a video-to-video architecture to fuse content and text style, complemented by an automatic reverse-synthesis pipeline to generate V-Style20k, a dataset of 20k high-quality video pairs. For long videos, EchoStyle utilizes an init-follow-mode mechanism and sliding-window inference strategy. Experiments demonstrate its superior performance across diverse artistic styles, rivaling leading closed-source solutions.
video stylizationreverse-synthesis pipelinev-style20kinit-follow-modesliding-window inference
Learning with a Single Rollout via Monte Carlo Pass@k Critic
The paper introduces single-rollout proximal policy optimization (SR-PPO), a method for token-level credit assignment in RL for language models that avoids costly repeated sampling. It trains a calibrated critic using Monte Carlo outcomes from one rollout per prompt, predicting Pass@k success probability derived from Pass@1 attempts. This approach prioritizes hard prefixes and converges to a reachability indicator as k increases. SR-PPO demonstrates stable learning and improves Pass@128 success rates on HMMT26 and AIME24 mathematical reasoning benchmarks.
token-level credit assignmentmonte carlo outcomespass@k success probabilitysingle-rollout ppomathematical reasoning benchmarks
Reclaim Evaluation: A Lossy Memory Is Worse Than an Empty One
The paper identifies 'brittle memory' in language models, where lossy memory retention of incorrect conclusions degrades performance more than empty memory. Through reclaim evaluation—compressing drifted interactions at fixed budgets—the authors demonstrate that correctability depends on preserving answer-determining sources rather than model capability. A source-first retention policy (retaining recomputable sources over derivable conclusions) improves correctability from 0.49-0.88 across seven models, while memory loops amplify errors when sources are dropped. Results generalize across three memory systems and MultiWOZ dialogue data, with validation via judge-free exact scoring and matched-budget controls.
brittle memoryreclaim evaluationsource-first policymemory loopcorrectability
C3-Bench: A Context-Aware Change Captioning Benchmark
We introduce C3-Bench, a benchmark for evaluating Context-aware Change Captioning systems, featuring 4,996 human-labeled image pairs across 51 real-world change contexts in four domains. The benchmark includes an LLM-as-Judge evaluation framework that assesses fine-grained dimensions (correctness, specificity, fluency, relevance) and introduces a novel reversibility metric. We evaluate 32 models, including conventional change captioning models and Large Multimodal Models (LMMs) ranging from 2B to 90B parameters. Results reveal that conventional models collapse when change contexts deviate from training regimes, and state-of-the-art LMMs like GPT-5.2 exhibit systematic domain- and position-dependent errors. The benchmark highlights critical failure modes and sets a new frontier for generalizable change captioning systems.
context-awarechange captioningllm-as-judgereversibility metricmultimodal models
TopoCast: A Topological Fidelity Framework for Evaluating Transformer-Based Time Series Forecasting
The paper introduces TopoCast, a topological framework for evaluating structural fidelity in Transformer-based time series forecasting. It reconstructs phase-space representations using Takens delay embedding and applies persistent homology to derive four topological fidelity measures, aggregated into a Topological Fidelity Score (TFS). The method also proposes dominant cycle overlap for assessing temporal localization errors, combining with TFS to yield the Localized Topological Fidelity Score (LTFS). Experiments on five Transformer architectures across three datasets show that models with similar forecasting errors exhibit distinct structural fidelity profiles, revealing previously overlooked failure modes.
topological fidelitypersistent homologytakens embeddingphase-space reconstructiontransformer architectures
Interpretable Concept-Guided Polynomial Tabular Kolmogorov-Arnold Network for EEG-Based Mild Cognitive Impairment Detection
The study introduces CPTabKAN, a concept-guided polynomial-transformed tabular learning framework for EEG-based mild cognitive impairment detection. The method maps EEG features into domain-informed concept representations, applies degree-2 polynomial transformation to model interactions, and uses a Fourier-parameterized Kolmogorov-Arnold Network for classification. Evaluated on 372 subjects from the Study of Osteoporotic Fractures cohort, CPTabKAN achieved a weighted F1-score of 0.9038, outperforming GradientBoosting by 5.65 percentage points. Ablation studies confirmed contributions from all components, with concept importance analysis revealing key features and interactions.
kolmogorov-arnold networkeeg-based detectionpolynomial transformationtabular learningmild cognitive impairment
Brevity is the Soul of Inference Efficiency: Inducing Concision in VLMs via Data Curation
The study demonstrates that data curation for brevity significantly improves inference efficiency in vision-language models (VLMs) without sacrificing accuracy. By curating the MAmmoTH-VL single-image subset to prioritize concise, correct responses, the authors achieve a 35x reduction in FLOPs per correct answer (0.41 vs 14.58 TFLOPs) compared to verbose models like Qwen3.5-4B, while maintaining near-equivalent accuracy (0.691 vs 0.704). The method also yields a +17.55 pp accuracy gain over uncurated baselines at matched output lengths, with benefits scaling positively with model size (1B-4B parameters). Results show verbosity provides no accuracy advantage, and concise models sometimes outperform verbose ones on reasoning tasks.
inference efficiencydata curationvision-language modelsflopsbrevity
Adaptive Oscillatory Inductive Bias for Modeling Sharp Prosodic Dynamics in Diffusion-Based TTS
The paper introduces OscillaTTS, a diffusion-based text-to-speech system with adaptive oscillatory nonlinearity to better model sharp prosodic transitions. The method combines controllable periodic modulation via an adaptive oscillatory component with signal stability through a linear bypass, addressing limitations of fixed-period activations like Snake. Evaluations on LJSpeech and Emotional Speech Dataset demonstrate consistent improvements in both objective metrics and subjective listening tests for expressive prosody generation.
diffusion-based ttsprosodic dynamicsadaptive oscillatory nonlinearityperiodic modulationexpressive speech synthesis
CrossAccent-TTS: Cross-Lingual Accent-Intensity Controllable Text-to-Speech via Disentangled Speaker and Accent Representations
CrossAccent-TTS introduces a novel framework for cross-lingual text-to-speech (TTS) with accent intensity control and conversion while preserving speaker identity. The method employs an Accent Intensity Controller (AIC) that injects weighted language embeddings into the accent subspace, enabling smooth interpolation between accents and fine-grained modulation of accent strength during inference. Evaluations on the Indic Multilingual and L2-arctic datasets demonstrate that CrossAccent-TTS achieves superior accent similarity and controllability while maintaining speaker similarity and naturalness, outperforming existing baselines.
accent intensity controllercross-lingual ttsspeaker identitylanguage embeddingsaccent subspace
LibEvoBench: Probing Temporal Knowledge Stratification in Code Generation Models
The paper introduces LibEvoBench, a benchmark for evaluating LLMs' ability to handle evolving library APIs across multiple versions, alongside a new metric (SEUS) for consistency measurement. The study reveals that state-of-the-art models exhibit version-oblivious behavior: performance degrades for evolving APIs but remains stable for static ones, with version specification proving ineffective while documentation improves accuracy. These findings underscore limitations in current training paradigms and advocate for temporally aware approaches in code generation.
api evolutioncode generationtemporal knowledgebenchmarkingsoftware libraries
Lightweight PCGAE-Net: Parallel CrossGate Attention and Bottleneck AutoEncoder for Efficient 5G Channel Prediction
Lightweight PCGAE-Net improves 5G channel prediction efficiency by addressing two architectural flaws in transformer-based models. The method introduces parallel CrossGate attention to eliminate sequential bias between spatial and temporal attention modules, and a Bottleneck AutoEncoder (BAE) with 1×1 convolutions to reduce feature redundancy. With 8.54M parameters (58% fewer than CS3T-UNet), it achieves 3.26dB and 6.0dB gains at 5km/h and 9km/h respectively on QuaDriGa dataset.
channel state informationmassive mimocrossgate attentionbottleneck autoencoderquadriga dataset
BrainAgent: A Large Language Model-Driven Multi-Agent Framework for Autonomous Brain Signal Understanding
BrainAgent introduces an LLM-driven multi-agent framework for autonomous brain signal understanding, addressing limitations of static, task-specific approaches. The system employs a hierarchical architecture with a central supervisor coordinating specialized sub-agents to decompose and execute complex workflows via natural language grounding. Evaluated on a systematic benchmark, BrainAgent demonstrates superior reliability in automating end-to-end processing pipelines, advancing democratization of brain-computer interfaces.
brain-computer interfaceslarge language modelsmulti-agent systemsnatural language groundingend-to-end processing
Long-Term Simulation Exposes Cognitive-Developmental Risks in AI Companions
The TSJ (Theater-Stage-Judge) framework introduces a longitudinal evaluation method for assessing cognitive-developmental risks in AI companion systems, addressing limitations of single-turn or short-session tests. Combining persona-driven user simulation, dynamic psychological-state updating, and retrospective evaluation, TSJ evaluates six mainstream models across four developmental stages, twenty-four risk dimensions, and three psychological-vulnerability personas over 12,960 simulated person-day interactions. Results indicate that short-horizon testing systematically underestimates developmental risks, with stable risk estimates emerging only after 140 turns. Early childhood and emerging adulthood are identified as the most vulnerable stages, with cognitive trust and emotional dependency as the weakest domains.
longitudinal evaluationcognitive-developmental riskspersona-driven simulationpsychological-state updatingretrospective evaluation
FactorLibrary: From Polynomials to Circuits via Recursive Subgoals
The paper introduces FactorLibrary, a reinforcement learning framework for finding minimal arithmetic circuits over finite fields by storing reusable factorizable subexpressions as subgoals. The approach formulates circuit minimization as a combinatorial search problem addressed via bottom-up (Gumbel-PPO-MCTS) and top-down (PPO+MCTS, SAC) agents. The PPO+MCTS top-down agent achieved 91.8% success rate in discovering certified optimal circuits up to complexity 8, demonstrating superior stability compared to alternative methods.
factorlibraryarithmetic circuitsgumbel-ppo-mctsfinite fieldscombinatorial search
From Sounds to Scenes: A Benchmark for Evaluating Context-Aware Auditory Scene Understanding in Large Audio Language Models
We introduce CASU, a benchmark for evaluating Context-Aware Auditory Scene Understanding (CASU) in Large Audio Language Models (LALMs), addressing the integration of multiple acoustic layers (speech, acoustic events, background environments) in real-world auditory scenes. A scalable pipeline constructs time-accurate, semi-synthetic audio streams by combining real-world scene sounds with synthetic speech. Four tasks assess scene understanding: contextual question answering, entity extraction, speaker role inference, and counterfactual reasoning. Experiments reveal that effective CASU requires integration across all auditory layers, highlighting the insufficiency of isolated speech or sound analysis for complex audio understanding in LALMs.
context-aware auditory scene understandinglarge audio language modelssemi-synthetic audio streamscounterfactual reasoningacoustic layers
Anatomically-conditioned Latent Diffusion Model for Data-Efficient Few-Shot Cross-Domain 3D Glioma MRI Synthesis
The authors propose ALDM, an anatomically-conditioned latent diffusion model for few-shot 3D glioma MRI synthesis across domains. The method combines a 3D variational autoencoder for anatomical prior learning with a ControlNet-guided latent diffusion model conditioned on tumor masks, enabling structurally coherent volume generation. In extreme few-shot evaluation (16 target images), ALDM achieved state-of-the-art performance with FID=85.40 and downstream classification AUC=0.987, while preserving pathology boundaries and cross-modal consistency.
latent diffusion modelfew-shot learning3d mri synthesiscontrolnetdomain adaptation
Offline Multi-agent Continual Cooperation via Skill Partition and Reuse
We propose COMAD, a continual offline multi-agent skill discovery framework that mitigates catastrophic forgetting and plasticity loss in sequential task settings. COMAD employs an auto-encoder to extract reusable coordination skills from mixed multi-agent behavior data, constructs a skill-augmented policy learning objective with multi-head architectures, and identifies reusable skills via a density-based reusability estimator. Theoretical analysis demonstrates COMAD approximates the optimum of continual skill discovery. Empirical evaluations across diverse MARL benchmarks show COMAD continually expands its skill library, achieving superior forward and backward transfer compared to baselines.
continual learningmulti-agent reinforcement learningskill discoveryauto-encoderdensity-based estimation
Beyond Visual Forensics: Auditing Multimodal Robustness for Synthetic Medical Image Detection
The study introduces a benchmark for auditing multimodal robustness in synthetic medical image detection, addressing the vulnerability of vision-language models (VLMs) to metadata context. By reformulating the task as an image-record interface audit, the authors evaluate VLMs across multiple imaging modalities using a paired benchmark that fixes images while swapping controlled metadata variants. Results demonstrate that metadata alone significantly shifts authenticity predictions, revealing a previously underexamined multimodal robustness issue. The benchmark provides a standardized tool for assessing and improving VLM performance in clinical settings where images are interpreted alongside structured records.
vision-language modelsmultimodal robustnesssynthetic medical imagesmetadata variantsimage-record interface
What Actually Works for Spacecraft Fault-Tolerant Control: An Honest Settled-Gate Benchmark of Learned and Classical Methods
The paper introduces a rigorous benchmark for spacecraft fault-tolerant control (FTC), evaluating methods on their ability to maintain pointing accuracy (within 0.2°) under unseen actuator faults. The benchmark employs a settled-gate metric, disjoint train/test splits, and Wilson intervals over 500 episodes, validated on a 6-DOF Basilisk testbed. Key findings show that fault-unaware PD/PID and end-to-end RL fail (0% success), while classical adaptive laws handle sign faults (55.2%) but struggle with gain faults. A structured estimate-then-control design with a learned recurrent module achieves 97.8% success on sign faults and 94.4% on gain faults. Constant additive bias remains challenging (0% success) but is partially resolved (59.4%) with a disturbance observer.
fault-tolerant controlsettled-gate metricactuator faultsdisturbance observerstructured estimation
Conformal Recovery-Deadline Certificates for Runtime Assurance of Adapting Controllers
The paper introduces conformal recovery-deadline certificates, a method for runtime assurance (RTA) that enables delayed fallback for online-adapting controllers while maintaining safety guarantees. The approach uses split-conformal prediction to provide distribution-free, finite-sample upper bounds on recovery time, licensed by statistical coverage and backed by a verified monitor. Theoretical results include marginal coverage, weighted coverage under fault-distribution shifts, and group-conditional Mondrian coverage. Experiments on a 6-DOF spacecraft attitude controller and a torque-controlled inverted pendulum demonstrate the method's domain-general applicability in preventing premature fallback while ensuring safety.
runtime assuranceconformal predictiononline adaptationsafety-critical systemsrecovery deadline
Sarashina2.2-TTS: Tackling Kanji Polyphony in Japanese Speech Generation via Data Scaling and Targeted Data Synthesis
Sarashina2.2-TTS introduces a Japanese-centric LLM-TTS system addressing kanji polyphony through data scaling (361k hours of speech) and targeted augmentation covering all 2,136 Joyo kanji. It proposes the Joyo Kanji Yomi Benchmark (4,378 readings) and Kana-CER metric for pronunciation evaluation. Results show state-of-the-art kanji-level reading accuracy, superior speaker similarity in zero-shot synthesis, and cross-lingual robustness unmatched by baselines.
kanji polyphonytext-to-speechdata augmentationzero-shot synthesiscross-lingual robustness
Reliability-Asymmetric Spacecraft Autonomy: Co-Designing a Capable Learned GNC Stack with a Verified, Adaptation-Aware Runtime Shield
The paper introduces AMPLE-GNC, a three-tier spacecraft autonomy stack combining learned and verified components. The system integrates: (1) a 360M-parameter foundation model commander fine-tuned for natural-language-to-PDDL+ translation with grammar-constrained decoding (84% planner-executable actions, 48% exact novel-phrasing generalization), (2) a Rapid Motor Adaptation controller recovering 97.8% of actuator faults within training bounds, and (3) a runtime shield with nine LTL invariants verified by Kind 2. Experiments on a 6-DOF Basilisk testbed demonstrate 94.5% autonomous operation while maintaining safety via split-conformal certificates and adaptation-aware shielding.
spacecraft autonomyruntime verificationgrammar-constrained decodingrapid motor adaptationlinear-temporal-logic
Neural Machine Translation for Low-Resource Tangkhul--English
The study introduces neural machine translation systems for the low-resource Tangkhul-English (nmf-en) language pair, addressing a severe resource gap in Tibeto-Burman NLP. Two systems are evaluated: (1) a ByT5-large model fine-tuned on 38,336 parallel sentence pairs, achieving 39.97 BLEU, 58.07 chrF++, 0.8104 BERTScore F1, and 0.7302 COMET (wmt22-comet-da) on a 3,856-sentence test set; (2) a contrastive mT5-small model. The analysis highlights orthographic challenges with Tangkhul's Latin-script diacritics and domain bias in the training corpus (biblical texts, stories, conversations), suggesting future improvements via data diversification and domain adaptation.
low-resource mtbyt5mt5diacriticsdomain bias
TheoremGraph: Bridging Formal and Informal Mathematics
TheoremGraph introduces a unified statement-level dependency graph spanning informal and formal mathematics, addressing the disparity between coarse document-level citations in papers and fine-grained dependencies in formal libraries. The method parses 11.7M theorem-like environments from arXiv, recovers 18.3M candidate dependencies, and extracts Lean 4 declarations (388,105 nodes, 11.3M edges). It bridges informal and formal mathematics via semantic embeddings of natural-language slogans, validated by an LLM judge (47,952 matches above 0.8 cosine similarity). The system achieves 0.775 Recall@10 on formal concept retrieval, comparable to LeanSearch v2 without LM reranking. Resources include datasets, extractors, and APIs at theoremsearch.com.
dependency graphformal mathematicssemantic embeddingtheorem retrievallean 4
Memory Makes the Difference: Evaluating How Different Memory Roles Shape Conversational Agents
This work investigates how memory types with distinct functional roles influence response quality in Retrieval-Augmented Generation (RAG)-based conversational systems. The authors propose a fine-grained taxonomy of conversational memory and classify retrieved memories into role types, evaluating their effects through a user-centric framework and experiments on long-term datasets with frontier LLMs. Results demonstrate differentiated impacts: clarifying memory enhances factual accuracy and constraint awareness, while irrelevant memory reduces topic relevance and degrades constraint awareness. These findings highlight the potential of leveraging memory types to produce more personalized responses in conversational agents.
retrieval-augmented generationconversational memoryfunctional rolesconstraint awarenessuser-centric evaluation
Agentic Knowledge Tracing: A Multi-Agent LLM Architecture for Stealth Assessment of Financial Literacy in Serious Games
The paper introduces Agentic BKT, a multi-agent LLM architecture for stealth assessment of financial literacy in serious games. The system processes gameplay events through four phases: event logging, LLM-based classification (Fleiss κ=0.624), domain-specific agent reasoning (risk, investing, spending, credit), and Bayesian Knowledge Tracing for mastery estimation. Evaluated on 193 K-12 participants across 264 sessions, the pipeline shows significant correlations with learning gain (r=0.276) and post-test scores (r=0.333), tripling the predictive validity of a single-LLM baseline (r=0.095).
stealth assessmentbayesian knowledge tracingmulti-agent llmfinancial literacyserious games
Compositional Behavioral Semantics for State Abstraction in Reinforcement Learning
The paper presents a unified framework for analyzing behavioral structures in reinforcement learning through compositional semantics. It introduces a method for specifying behavioral semantics via local, one-step system dynamics descriptions, enabling principled reasoning about state abstraction. Results demonstrate how behavioral structures (e.g., value functions, bisimulation relations) can be preserved between abstract and concrete systems, with sound construction of quantitative metrics from logical semantics. The framework provides foundational principles for transferable behavioral analysis across RL systems.
state abstractionbehavioral semanticsreinforcement learningbisimulation relationscompositional dynamics
AI Coaching for Accelerating Human Skill Development with Reinforcement Learning
The paper introduces a reinforcement learning framework for AI coaching that accelerates human motor-skill development by strategically balancing assistance and productive failures. The approach formalizes the coaching process as a non-cooperative dynamic game, where the learner optimizes task performance and the coach targets independent competence. The framework combines adaptive shared control with probabilistic models of the coach's causal influence on skill evolution, enabling tractable policy training. A user study (N=33) on first-person-view drone racing demonstrates significant improvements in human learning outcomes compared to state-of-the-art AI coaching baselines.
reinforcement learningadaptive shared controldynamic gamemotor-skill developmentprobabilistic models
Decoupling Reconnaissance and Exploitation: Measuring the Capability Boundaries of LLM-Based Web Penetration Testing
The study introduces a decoupled evaluation framework to isolate exploitation performance from reconnaissance noise in LLM-based web penetration testing, addressing error cascading in end-to-end black-box evaluations. Using ground-truth injection and knowledge-driven ablation across 70 high-fidelity web vulnerability testbeds, the framework evaluates five open-source penetration-testing agents on 50 representative vulnerabilities. Results show a functional success rate of up to 90.0% with accurate vulnerability context, while autonomous reconnaissance plateaus at 50.0% due to parsing failures. Multi-agent, monolithic, and graph-driven architectures exhibit distinct capability niches for different vulnerability types, providing empirical insights for next-generation automated security agents.
penetration testingerror cascadingground-truth injectionknowledge-driven ablationmulti-agent isolation
Improved Large Language Diffusion Models
We introduce iLLaDA, an 8B-parameter masked diffusion language model trained from scratch with fully bidirectional attention, diverging from autoregressive approaches. iLLaDA maintains the masked diffusion objective during both pre-training (12T tokens) and supervised fine-tuning (25B-token instruction corpus, 12 epochs), incorporating variable-length generation for efficiency and confidence-based scoring for evaluation. Compared to LLaDA, iLLaDA demonstrates significant improvements: iLLaDA-Base gains 21.6 points on BBH and 14.9 points on ARC-Challenge, while iLLaDA-Instruct improves by 14.5 points on MATH and 16.5 points on HumanEval. Despite non-autoregressive training, iLLaDA remains competitive with Qwen2.5 7B on multiple benchmarks, validating bidirectional diffusion training as a viable path for language modeling.
masked diffusionbidirectional attentionvariable-length generationconfidence-based scoringnon-autoregressive training
Supervised Post-training of Speech Foundation Models for Robust Adaptation in Speech Deepfake Detection
The paper proposes a mix-frame post-training strategy to adapt speech foundation models for robust deepfake detection by addressing the mismatch between self-supervised pre-training objectives and spoof-specific artifacts. The method introduces localized spoof-oriented perturbations and employs frame-level supervision to teach the model to identify local inconsistencies critical for detection. Evaluated on ASVspoof5 and ASVspoof2021 LA/DF, the approach achieves state-of-the-art EER of 4.50% and demonstrates balanced robustness with only a 0.16% EER gap between LA and DF conditions.
speech foundation modelsdeepfake detectionframe-level supervisionasvspoofeer
Omni-Perception Policy Optimization for Multimodal Emotion Reasoning
The paper introduces OPPO (Omni-Perception Policy Optimization), a reinforcement learning framework addressing two limitations in emotion-oriented Omni-MLLMs: underutilization of multimodal cues and cross-modal hallucination. OPPO employs (1) an Omni-Perception Reward that decomposes ground-truth reasoning into visual, acoustic, and emotion cues, and (2) an Omni-Perception Loss using KL divergence on modality-specific tokens under input masking. The method is evaluated on MEP-Bench, a new diagnostic benchmark, achieving SOTA on MER-UniBench and MME-Emotion while improving utilization (+12.3%) and faithfulness (+8.7%) metrics.
multimodal emotion reasoningreinforcement learningcross-modal hallucinationkl divergenceperception optimization
ESTANet: Efficient Online Error Detection in Procedural Videos via Prediction Inconsistency
ESTANet introduces an efficient framework for online error detection in procedural videos by leveraging prediction inconsistencies among action detectors. The method constructs standard and error-sensitive action detectors that exhibit similar behavior on correct executions but diverge during errors, amplified by varying temporal contexts. Errors are detected via majority voting on prediction mismatches. Experiments on EgoPER, Assembly-101-O, and EPIC-Tent-O demonstrate state-of-the-art performance with real-time efficiency, showcasing the effectiveness of exploiting intrinsic detector properties without complex architectural modifications.
online error detectionaction detectorsprediction inconsistencytemporal contextmajority voting
Physics Question Scene Graph: Fine-grained Evaluation of Physical Plausibility in Text-to-Video Generation
The paper introduces Physics Question Scene Graph (PQSG), a hierarchical question-based evaluation pipeline for assessing physical plausibility in text-to-video generation. PQSG employs a vision-language model (VLM) to generate contextually valid questions structured as a graph, enabling fine-grained analysis of object, action, and physical law violations. Evaluated on FinePhyEval, a dataset with human-annotated videos from Sora 2, Veo 3, and Wan 2.1, PQSG achieves higher correlation with human judgments than prior methods and reveals closed-source models outperform Wan 2.1 in physical realism. VLMs show promise in question generation but lag in answering accuracy.
physics question scene graphtext-to-video generationvision-language modelphysical plausibilityfine-grained evaluation
Communicability-Inspired Positional Encoding (CIPE)
The paper introduces Communicability-Inspired Positional Encoding (CIPE), a novel positional encoding method for Transformers operating on non-Euclidean graphs. CIPE leverages communicability, a node-pair metric aggregating path contributions across all lengths, to construct an Attention-Compatible Geometry that aligns with structural relatedness. The method includes dimensionality alignment to adapt graph-size-dependent representations to fixed dimensions while preserving geometric properties. Empirical evaluation shows CIPE improves structure-agnostic Transformers by 35.5% on average across seven benchmarks, outperforming existing positional encodings and enhancing structure-biased graph Transformers where alternatives offer marginal gains.
communicabilitypositional encodingattention-compatible geometrydimensionality alignmentgraph transformers
EPTS: Elastic Post-Training Sparsity for Efficient Large Language Model Compression
The paper introduces Elastic Post-Training Sparsity (EPTS), a unified framework for compressing Large Language Models (LLMs) that supports multiple sparsity levels through one-shot optimization. EPTS employs a Multi-Sparsity Hierarchy LoRA (MS-HiLoRA) mechanism to enable knowledge transfer between sparsity levels and a Multi-Sparsity Feature Mixer (MSFM) to enhance robustness against pruning perturbations. Evaluations on LLaMA and OPT models show EPTS matches SparseGPT and Wanda in performance while enabling flexible deployment across hardware scenarios. The approach eliminates the need for repeated optimization sessions for different sparsity targets.
post-training sparsitylarge language modelsparameter efficiencymodel compressionmulti-sparsity optimization
Heterogeneous and Adept Snapshot Distillation for 3D Semantic Segmentation
The paper proposes HAS-KD, a knowledge distillation framework for 3D semantic segmentation that transfers multi-modal and multi-expert knowledge to a point-cloud-based student model without computational overhead. It introduces Information-oriented Heterogeneous Distillation (IHD) to capture complementary multi-modal features via an Information-Oriented Filtering strategy, and Adept Snapshot Distillation (ASD) to leverage training-phase model snapshots as expert teachers. The method achieves state-of-the-art performance on ScanNetV2 and S3DIS datasets while maintaining inference efficiency.
knowledge distillation3d semantic segmentationmulti-modal fusionmodel ensemblingpoint-cloud processing
UC-Search: Risk-Aware Test-Time Search for Delayed Constrained Time-Series Control
UC-Search introduces a risk-aware test-time search method for delayed constrained time-series control, acting as a model-agnostic wrapper. The approach combines backbone forecasts, feasibility automaton rollouts, and bounded search to identify risk-adjusted feasible trajectories, with variants UC-Beam and UC-MCTS leveraging epistemic, aleatoric, and propagated uncertainty. Evaluated on a 9-family, 33-series delayed-control suite, UC-Pareto outperforms CEM, MPPI, and risk-aware random baselines (+3.1675/+2.3328/+2.5038 normalized threshold) and maintains gains in compute-matched audits. Additional validation on ETT/LTSF delayed-inventory and M4 lost-sales inventory (+13556.7547 vs. base-stock control) further supports efficacy.
risk-aware searchtime-series controlfeasibility automatonepistemic uncertaintydelayed decisions
Stabilizing black-box algorithms through task-oriented randomization
The paper introduces a task-oriented randomization methodology to stabilize black-box algorithm outputs, addressing challenges posed by diverse input structures. The approach adaptively tailors strategies to underlying generative mechanisms, providing stability guarantees and analyzing trade-offs between stability and exploration. Extending the framework to top-k ranking problems inspired by Large Language Model architectures, the study validates the method through numerical simulations and real-world dataset applications.
black-box algorithmstask-oriented randomizationstability guaranteestop-k rankinggenerative mechanisms
ASAP: Agent-System Co-Design for Wall-Clock-Centered Auto HPO Research for ML Experiments
ASAP introduces an agent-system co-design for wall-clock-efficient AutoML hyperparameter optimization (HPO) by addressing two limitations of LLM-based HPO: single-source inductive bias and per-iteration evaluation. The method integrates diverse inductive-biased optimizers via LLM-based selection, employs KV-cache reuse for prompt stability, speculation parallelism to hide LLM/tool latency, and a Self-Tuner for adaptive thresholding. Experiments on diverse HPO tasks demonstrate consistent improvements over baselines, validating the benefits of tool integration and co-design.
hyperparameter optimizationinductive biaskv-cachespeculation parallelismwall-clock efficiency
RAVEN: Long-Horizon Reasoning & Navigation with a Visuo-Spatio-Temporal Memory
RAVEN introduces a visuo-spatio-temporal memory system for long-horizon robotic question answering and navigation, addressing the need for compact, scalable memory preserving visual semantics. The method stores visual embeddings with pose and time in a vector database, grounding retrieval in a spatial map to enable semantic, spatial, and temporal queries without lossy image-to-text conversion. Evaluations on simulated and real-world benchmarks show RAVEN outperforms caption-based systems and matches visual language models (VLMs) on long-horizon tasks at 10× lower retrieval cost, with successful deployment on a Unitree Go1 robot for large-scale indoor navigation.
visuo-spatio-temporal memoryvector databaselong-horizon navigationvisual embeddingsrobotic question answering
FDN: Interpretable Spatiotemporal Forecasting with Future Decomposition Networks
The paper introduces Future Decomposition Network (FDN), an interpretable spatiotemporal forecasting model that provides classification-based predictions while revealing latent activity patterns. FDN employs decomposition techniques to achieve both interpretability and computational efficiency, operating with reduced memory and runtime costs compared to state-of-the-art methods. Evaluations across hydrologic, traffic, and energy system datasets demonstrate FDN's competitive accuracy and enhanced interpretability.
spatiotemporal forecastinginterpretabilityfuture decomposition networklatent patternscomputational efficiency
A Hybrid CNN-LSTM Intrusion Detection Framework for Cybersecurity in Smart Renewable Energy Grids
The study proposes a Hybrid CNN-LSTM Intrusion Detection System (IDS) for cybersecurity in smart renewable energy grids, addressing limitations in temporal attack modeling, class imbalance, and cross-environment generalization. The method combines CNN-based spatial feature extraction with LSTM-based temporal sequence modeling, trained via a seven-step preprocessing workflow including SMOTE balancing and mutual-information feature selection. Evaluated on CICIDS2017 and NSL-KDD, the model achieves 96.1% and 98.2% precision respectively, outperforming baselines by 2-9 percentage points, with real-time throughput of 27,800 flows/s on GPU and 0.082 ms/sample CPU latency.
cnn-lstmintrusion detectionsmart gridtemporal sequence modelingclass imbalance
Heuresis: Search Strategies for Autonomous AI Research Agents Across Quality, Diversity and Novelty
Heuresis introduces a framework for autonomous AI research agents, abstracting the research pipeline into composable primitives to explore performant, diverse, and novel ideas in machine learning. The framework implements six search strategies—greedy baseline, MAP-Elites, Go-Explore, Islands, Curiosity, and Omni—evaluated across quality, diversity, and novelty on three domains: LLM Pretraining, On-Policy RL, and Model Unlearning, totaling 3,222 scored runs. Results show novel ideas are rare, with none rated 'Original' and few achieving 'Minor Similarity' to prior work. Novel ideas also fail to match top-performing known-recipe scores, with only one novel idea ranking in the top-10 by quality. Reward-hacking techniques were observed, necessitating detection for task fidelity.
autonomous ai researchsearch strategiesquality-diversitynovelty explorationreward-hacking
SoK: AI Secure Code Generation: Progress, Pitfalls, and Paths Forward
This Systematization of Knowledge (SoK) introduces a three-level framework for evaluating AI secure code generation, measuring natural-language understanding of secure coding principles, code-level actuation of those principles, and knowledge-actuation gaps. The framework is instantiated across models and coding agents using benchmarks for function-level and web-application security. Results indicate that understanding secure coding principles strongly predicts functional correctness, security, and joint functional-security correctness. However, persistent knowledge-actuation gaps reveal that models often fail to translate recognized principles into secure, functional code. The study identifies principle-guided generation, evaluation, benchmarking, and agentic workflows as key paths forward.
secure code generationknowledge-actuation gapsagentic workflowsbenchmarkingfunctional-security correctness
To Isolate or to Score? Model-Adaptive Assessment for Cost-Efficient Multi-Agent RAG
The study introduces MADARA, a model-adaptive routing architecture for cost-efficient multi-agent retrieval-augmented generation (RAG). It analyzes training-free interventions on 7B-9B instruction-tuned models across QA benchmarks, revealing two assessment mechanisms: per-document isolation (dominant for weaker baselines, yielding up to 50pp gains) and Reasoning-Score Coupling (a label-free probe for strong baselines). MADARA's diagnostic thresholds generalize zero-shot to four unseen model families, reducing computational overhead. Results show assessment-free isolation matches full multi-agent assessment when context confusion is resolved.
retrieval-augmented generationinstruction-tuned modelsreasoning-score couplingmodel-adaptive routingzero-shot generalization
What Intermediate Layers Know: Detecting Jailbreaks from Entropy Dynamics
The study demonstrates that jailbreak attacks in aligned Large Language Models (LLMs) can be detected through analysis of token-level predictive entropy dynamics across network layers. Using the logit lens on frozen models (Llama, Qwen, Gemma), the authors find that entropy evolution features (e.g., monotonic rank-based trend scores) in intermediate layers provide discriminative signal, while static aggregate statistics and final-layer features are less informative. Results show architecture-consistent separation across adversarial benchmarks without additional training, revealing that jailbreak-relevant structure is most pronounced in mid-network representations.
jailbreak attackspredictive entropyintermediate layerslogit lensllm safety
Phoneme-Level Mispronunciation Screening in Polish-Speaking Children with an Explainable Assistant
The study introduces a phoneme-level mispronunciation screening pipeline for Polish-speaking children, combining a wav2vec2-based CTC token recognizer with alignment-based error typing and an explainable caregiver assistant. The system targets sibilant substitutions, operating as a lightweight screening tool rather than a diagnostic solution. Evaluation on 559 utterances from 10 children shows 88.7% exact sequence match accuracy, with conservative screening achieving 72.9% precision, 61.4% recall (F1=0.67), and a 2.7% false-alarm rate on target-correct items. The authors outline safety boundaries and propose clinician-in-the-loop validation for deployment.
wav2vec2ctc tokenizationsibilant substitutionalignment-based error typingclinician-in-the-loop
Transferability for General Reasoning: An Automated Curriculum for Multi-Domain RLVR
The paper introduces Transfer-Aware Curriculum (TAC), a bandit-style online curriculum for multi-domain RLVR that optimizes cross-domain transferability. TAC leverages per-domain advantages and projected gradients from GRPO steps to estimate transferability with minimal overhead (<1% wall-clock). Evaluated on a six-domain reasoning suite with Qwen3-1.7B and Llama3.2-3B, TAC achieves superior macro-averaged accuracy (up to 2.8 points, 10% relative) over baselines, including learnability-only bandits and hand-tuned schedules. Ablations confirm transferability's critical role, with performance degrading sharply without it, and robustness in imbalanced training mixtures.
rlvrtransferabilitycurriculum learningmulti-domain reasoninggradient projection
Elo-Disentangled Player-Style Embeddings for Human Chess via Rating-Conditioned Residual Move Model
The paper introduces Elo-disentangled player-style embeddings for chess, combining a rating-conditioned base move model with residual player vectors to separate stylistic deviations from strength-typical play. The base model integrates Maia-3 policy logits, Stockfish features, and Maia-2-proposed candidates, achieving 27-37% relative NLL improvement over Maia-3 (largest gains at 2800+ Elo) and +33% top-1 move-matching over Maia-2. Player embeddings (z) show representational value: they generalize to held-out decisions, re-identify players above chance, and exhibit low rating correlation (R^2=0.06), confirming Elo-orthogonal style capture. The method offers an interpretable alternative to per-player fine-tuning.
representation learningelo disentanglementresidual move modelpolicy logitsplayer re-identification
TRUSTMEM: Learning Trustworthy Memory Consolidation for LLM Agents with Long-Term Memory
The paper introduces TrustMem, a framework enhancing trustworthy memory consolidation for LLM agents by addressing persistent errors in memory updates. It employs a Memory Transition Verifier to assess updates for coverage, preservation, and faithfulness, and uses preference-guided reinforcement learning to optimize memory operations. Experiments show TrustMem achieves state-of-the-art performance on MemoryAgentBench, HaluMem, and Mem-alpha, improving memory extraction by 12.14 F1 points and reducing omission, corruption, and hallucination by 40.1%, 79.1%, and 50.0%, respectively.
trustmemmemory transition verifierpreference-guided reinforcement learningmemoryagentbenchhalumem
ATMA: Length-Invariant Language Modeling via Polar Attention and Gated-Delta Compression Memory
The paper introduces ATMA, a hybrid convolutional-attention architecture for length-invariant language modeling, addressing the limitations of softmax-based attention in long-context scenarios. ATMA combines a novel three-channel Polar Attention mechanism (direction channel, magnitude channel, and long-term recurrent compression memory) with a gated-delta fast-weights rule. Evaluations show ATMA maintains over 90% retrieval accuracy at 64K tokens (32× training length) and improves perplexity monotonically, outperforming softmax-based baselines that fail at extreme contexts.
polar attentiongated-delta compressionlength-invariant lmrecurrent memoryperplexity reduction
Hitting a Moving Target: Test-Time Adaptation for AI Text Detection under Continual Distribution Shift
The paper introduces a test-time adaptation (TTA) framework for AI text detection under continual distribution shifts, addressing vulnerabilities in deployed approaches that rely on training-time labeled datasets. The method leverages semi-supervised learning and inference-time homogeneity to adapt to three types of shifts: adversarial humanization, new LLM releases, and temporal drift in human writing. Empirical results show that state-of-the-art supervised detectors fail under these shifts, with Pangram detecting only 24.1% of adversarial AI-generated text, while the proposed TTA approach achieves 90.5% detection accuracy. The authors release code for model training and evaluation.
test-time adaptationdistribution shiftsemi-supervised learninginference-time homogeneityai text detection
Silent Failures in Physics-Informed Neural Networks: Parameter Poisoning and the Limits of Loss-Based Validation
The paper identifies silent failures in Physics-Informed Neural Networks (PINNs) caused by physics parameter poisoning, where incorrect PDE parameters lead to low training loss despite physically inaccurate solutions. Through sensitivity analysis across three PDE systems (Burgers equation, Navier-Stokes cavity, convection-diffusion), poisoned models achieve losses at or below clean baselines while deviating by up to 128% in solution accuracy. A detection difficulty ratio quantifies the invisibility of corruption, and six candidate defenses fail universally. A post-hoc defense, sweeping PDE residual loss across parameter values without retraining, reliably recovers true parameters across five architectures (8.7K to 133K parameters) and multiple random seeds.
physics-informed neural networksparameter poisoningpartial differential equationssensitivity analysisresidual loss
Proactive Systems in HCI and AI: Concepts, Challenges, and Opportunities
This workshop establishes a rigorous foundation for proactive systems by addressing conceptual ambiguities and methodological limitations in their design and evaluation. Through multidisciplinary collaboration involving Human-Computer Interaction and AI researchers, it aims to develop a shared understanding of proactivity, identify gaps in current approaches, and co-create human-centered guidelines. Key challenges such as timing, appropriateness, user control, transparency, and trust are highlighted, with a focus on advancing frameworks for proactive technologies. Interactive discussions and collaborative activities are employed to map challenges and opportunities, fostering robust and consistent methodologies for future systems.
proactive systemshuman-computer interactiondesign methodologiesuser controltransparency
TokenMinds: Pretrained User Tokens and Embeddings for User Understanding in Large Recommender Systems
TokenMinds introduces a dual-output user modeling system for industrial recommender systems, combining discrete Semantic ID (SID)-based tokens with dense embeddings via an LLM-based encoder-decoder architecture. The method extends PLUM's item retrieval framework to user modeling, unifying cross-scenario behaviors (e.g., long/short-form video) under a shared SID vocabulary while maintaining compatibility with existing embedding-based models. Offline experiments and live deployments on YouTube (billions of users) demonstrate complementary benefits: SID tokens improve semantic grounding while embeddings retain downstream compatibility, with infrastructure enabling asynchronous representation generation for ranking systems.
semantic iduser modelingrecommender systemsencoder-decoderplum framework
Benchmarking the Alignment of Data-Quality Metrics, Human Judgment and Land-Cover Segmentation Performance for Earth Observation
This work benchmarks the alignment of data-quality metrics, human judgment, and downstream task performance for Earth observation datasets and their synthetic counterparts. The study evaluates synthetic images generated by deep generative models using automatic metrics (FID, KID, IS, LPIPS, SSIM) and compares them to human perception and semantic segmentation performance. Results reveal a misalignment: semantics-preserving perturbations like rotation alter metric scores without affecting human recognition, and synthetic samples with poor metric scores can enhance downstream performance when combined with real data. The findings emphasize that automatic quality evaluation should incorporate downstream task performance and human evaluation.
semantic segmentationfidsynthetic dataearth observationhuman perception
Reward-Conditioned Attention: How Reward Design Shapes What Autonomous Driving Agents See
The study demonstrates how reward design shapes attention patterns in Perceiver-based reinforcement learning agents for autonomous driving, using three architecturally identical agents trained with different reward configurations. Analyzing cross-attention allocation across 50 scenarios from the Waymo Open Motion Dataset, the authors validate a methodology using within-episode correlation with Fisher z-transform aggregation, revealing a robust positive link between collision risk and agent-directed attention. Results show that navigation rewards lead to 2.0× more attention to GPS-path tokens compared to proximity penalties, and 4.7× more than no navigation incentive, while continuous time-to-collision penalties create a learned vigilance prior. Reward design qualitatively reverses attentional strategy in some scenarios, indicating attention analysis as a practical diagnostic for reward function verification.
perceiver-based agentscross-attention allocationfisher z-transformlearned vigilance priorreward-conditioned effects
AeroCast: Probabilistic 3D Trajectory Prediction for Non-Cooperative Aerial Obstacles via Transformer-MDN Architecture
AeroCast introduces a probabilistic 3D trajectory prediction framework for non-cooperative aerial obstacles, combining a Transformer encoder with a Mixture Density Network (MDN) to predict Gaussian mixture distributions over future displacements. The method employs translation-invariant consecutive displacement encoding and a calibration-oriented training objective to address input design and mode-degeneracy challenges. Evaluated on a hybrid real-and-synthetic quadrotor dataset spanning nine motion categories, AeroCast reduces Average Displacement Error and Final Displacement Error by 50% over a five-second horizon compared to baselines, achieving the lowest negative log-likelihood and Continuous Ranked Probability Score. Inference completes in 0.1ms per sample, enabling real-time deployment at 100Hz.
transformer encodermixture density networkgaussian mixturetrajectory predictiondisplacement error
BCoughBench: Benchmarking Respiratory Acoustic Foundation Models Under Body-Coupled Wearable Sensor Conditions
The study introduces BCoughBench, a benchmark for evaluating respiratory acoustic foundation models (FMs) under body-coupled (BC) wearable sensor conditions, addressing a gap in clinical deployment. Five FMs (OPERA-CT/CE/GT, HeAR, M2D+Resp) were assessed on nine classification tasks (AUROC, sensitivity at 95% specificity, Expected Calibration Error) and three age regression tasks (MAE) across five simulated BC sensor conditions using five labeled cough datasets. Results show mean AUROC declines from 0.785 (smartphone) to 0.689-0.723, with temple vibration pickup causing the largest drop (Δ = -0.096). No FM meets clinical sensitivity thresholds for disease tasks under BC conditions, while age regression remains robust (MAE improvement from 9.61 to 8.97 yr). HeAR excels in regression and demographic tasks, while M2D+Resp leads in disease and characteristic tasks.
respiratory acoustic foundation modelsbody-coupled wearablesaurocexpected calibration errorage regression
The Clinician's Veto: Navigating Trust, Liability, and Uncertainty in Autonomous AI Prescribing
The paper proposes three architectural requirements for safe autonomous AI prescribing systems: calibrated per-prediction confidence thresholds, differentiated uncertainty communication (epistemic vs. aleatoric), and inferential transparency for liability allocation. Through a survey of 136 U.S. clinicians, the study found that adoption depends on confidence-based escalation mechanisms, context-specific uncertainty presentation (competing options for aleatoric vs. abstention for epistemic), and liability acceptance only with sufficient transparency. Results suggest such systems would function as supervised decision-support tools rather than fully autonomous agents, informing regulatory frameworks like U.S. bill H.R. 238 and Utah's prescription-renewal pilot.
autonomous prescribingepistemic uncertaintyaleatoric uncertaintyinferential transparencyliability allocation
Beyond Shapley: Efficient Computation of Asymmetric Shapley Values
The paper introduces efficient computational methods for Asymmetric Shapley Values (ASV), a causal variant of Shapley values that incorporates causal graphs into model-agnostic explanations. It demonstrates polynomial-time exact computation of ASV in specific contexts where standard SHAP is #P-hard, using equivalence classes over topological orderings in rooted directed trees. For arbitrary causal DAGs, the authors propose an approximation algorithm based on uniform sampling of topological orderings, supported by experimental validation in realistic causal structures.
asymmetric shapley valuescausal graphtopological orderingmodel-agnostic explanationspolynomial-time algorithm
Power-Flexible AI Data Centers: A New Paradigm for Grid-Responsive Compute
The paper introduces power-flexible AI data centers as grid-interactive assets, proposing an architecture integrating grid signals, workload scheduling, and power telemetry for dynamic power control. The method enables GPU-based clusters to perform rapid load reduction, sustained curtailment, and carbon-aware operation while maintaining service levels. Experimental validation on a 130 kW GPU cluster demonstrates successful load shifting across distributed clusters and flexible operation under grid constraints, transforming AI infrastructure into grid-supportive resources.
grid-interactiveworkload orchestrationpower telemetryload shiftingcarbon-aware
Training for the Model You Return: Improving Optimization for Iterate-Averaged Language Models
The paper introduces PACE, a lightweight optimizer wrapper that improves performance of iterate-averaged language models by formulating optimizer design as an optimal-control problem. The method derives from a continuous-time stochastic quadratic model, solving for control strategies that minimize the error of the returned average while penalizing intervention size. Theoretical analysis shows PACE achieves standard convergence rates and can arbitrarily reduce squared error in quadratic settings. Empirical results demonstrate improvements over AdamW in fine-tuning 1-2B parameter LMs and GPT-2 pretraining on FineWeb across varied hyperparameters.
iterate averagingoptimal controlstochastic optimizationlanguage model fine-tuningadamw optimization
GCT-MARL: Graph-Based Contrastive Transfer for Sample-Efficient Cooperative Multi-Agent Reinforcement Learning
GCT-MARL introduces a graph-based contrastive transfer framework for sample-efficient cooperative multi-agent reinforcement learning (MARL), addressing deployment challenges in new environments. The method combines a multi-view graph contrastive backbone from MAIL with an adaptively weighted alignment loss and a two-phase training protocol, enabling transfer across populations with varying sizes and compositions. Empirical results show accelerated convergence in homogeneous (within-faction) and heterogeneous (cross-faction, mixed unit-type) transfer scenarios, with additional support for continual learning through sequential task chaining.
multi-agent reinforcement learningtransfer learninggraph contrastive learningcontinual learningsample efficiency
Adapt Only When It Pays: Budgeted Decision-Loss Priority for Delayed Online Time-Series Adaptation
The paper introduces ADOWIP, a budgeted online time-series adaptation framework that selectively updates models based on decision-loss priority. The method employs sealed delay queues, exact budget accounting, and a scheduler that triggers updates only when revealed feedback exceeds a calibrated loss quantile and budget permits. Theoretical guarantees include hard-budget feasibility, projected-OGD regret bounds, and stability. Empirical evaluation on ETT capacity-planning tasks shows reduced decision loss compared to baselines (33/41 significant contrasts), with mixed results on secondary tasks. Successful applications include UCI Bike (20/0 wins) and Capital Bikeshare station-rebalancing, though probe-based and finance experiments remain negative.
online adaptationtime-series forecastingbudgeted learningdecision-loss priorityregret bounds
Do vision-language models search like humans? Reasoning tokens as a reaction-time analog in classic visual-search paradigms
The study investigates whether vision-language models (VLMs) exhibit human-like visual search behaviors by adapting four classic paradigms: feature/conjunction search, spatial-configuration search, enumeration, and tilted/vertical asymmetry. Using reasoning token counts as a reaction-time analog, the authors compare frontier and mid-tier VLMs against human benchmarks (Wolfe et al., 2010). Results show models replicate human signatures like flat effort in feature search and climbing effort in conjunction search, but diverge in target-present/absent slopes and enumeration accuracy. The work demonstrates psychophysical paradigms as effective probes for machine visual cognition, revealing both alignments and informative divergences from human behavior.
vision-language modelsvisual searchreasoning tokenspsychophysical paradigmsreaction-time analog
What Does It Mean to Break a Distillation Defense?
The paper proposes a threat model framework for evaluating distillation defenses in black-box LLMs, addressing underspecification in current approaches. The framework characterizes attackers along three dimensions: query budget, data budget, and interface profile. Using antidistillation sampling as a case study, the authors demonstrate that defense efficacy depends critically on the assumed threat model. They advocate for explicit specification of attacker capabilities in future defenses and governance frameworks to prevent false security claims.
distillation attacksoutput perturbationthreat modelblack-box llmsantidistillation sampling
Wan-Streamer v0.1: End-to-end Real-time Interactive Foundation Models
Wan-Streamer v0.1 introduces a unified, end-to-end interactive foundation model for real-time, low-latency audio-visual interaction. The model integrates language, audio, and video processing within a single Transformer architecture, employing interleaved input-output tokens and block-causal attention for incremental streaming. Unlike cascaded systems, Wan-Streamer jointly learns perception, reasoning, generation, and synchronization, eliminating external modules and reducing pipeline latency. The stack is redesigned for streamability, incorporating causal encoders, decoders, and low-latency multimodal token scheduling, achieving streaming units as short as 160 ms at 25 fps. Wan-Streamer demonstrates model-side response latency of ~200 ms and total interaction latency of ~550 ms, enabling sub-second duplex audio-visual communication.
transformerblock-causal attentionmultimodal token schedulinglow-latencyend-to-end
LLM-ACES: Closed-Loop Discovery of Dynamical Systems with LLM-Guided Adaptive Search
LLM-ACES introduces a closed-loop framework for discovering governing ODEs from data by combining LLM-guided hypothesis generation with adaptive data acquisition. The method uses an LLM to propose operator priors that partition the search space, then iteratively refines candidate equations and acquires informative trajectories based on disagreement. Evaluated on 122 systems from ODEBench and ODEBase, it achieves state-of-the-art median NMSE, 46.2-52.4% symbolic accuracy, and 10× sample efficiency while maintaining robustness to noise.
ordinary differential equationsactive learningsymbolic regressionlarge language modelssystem identification
Yuvion VL: A Multimodal Foundation Model for Adversarial Content and AI Safety
Yuvion VL introduces a family of multimodal large language models (32B parameters) specifically designed for adversarial content and AI safety, featuring instruction-tuned and reasoning-oriented variants. The method employs a three-stage training pipeline (continued pretraining, instruct post-training, reasoning post-training) with Confuse-then-Contrast Fine-Tuning to enhance fine-grained visual-semantic discrimination in adversarial scenarios. Evaluated on the Yuvion VL RiskEval (YVRE) benchmark, Yuvion VL-32B demonstrates industry-leading safety performance while maintaining general capabilities comparable to open-source and commercial models.
multimodal large language modelsadversarial robustnessconfuse-then-contrast fine-tuninginstruction post-trainingrisk-concept cross-modal alignment
Do Thinking Tokens Help with Safety?
The study challenges the assumption that thinking tokens enhance safety deliberation in reasoning models, demonstrating that refusal/compliance outcomes are strongly predictable from the first token's hidden representation (0.84-0.95 AUROC, ~88% balanced accuracy). Analyzing GPT-OSS, Qwen, Olmo, and Phi models, the authors show thinking processes resemble prefix completion rather than deliberative revision, with ~74% of text-level deliberations occurring after the response distribution is already fixed. Existing safety interventions are found to induce over-refusal while suppressing deliberation signals, highlighting the need for methods that foster genuine safety deliberation.
thinking tokenssafety deliberationaurocprefix completionover-refusal
Noise-Aware Boundary-Enhanced Generative Learning for Ultrasound Speckle Reduction
The Noise-Aware Boundary-Enhanced Generative Learning (NBGL) framework is proposed for ultrasound speckle reduction, addressing over-smoothing and poor generalization in existing methods. NBGL integrates a speckle reduction branch using generative learning and a boundary enhancement branch preserving anatomical structures. A noise-aware interaction weight generation (NIWG) module estimates noise levels via 3D Laplacian filtering and median absolute deviation, modulating cross-branch feature coupling through weighted feature-wise linear modulation (wFiLM). Evaluations on 141 3D transvaginal ultrasound volumes show NBGL outperforms state-of-the-art methods in speckle reduction and structural preservation across six noise levels.
speckle reductiongenerative learninglaplacian filteringfeature-wise linear modulationanatomical boundaries
Erased, but Not Gone: Output Forgetting Is Not True Forgetting
The paper challenges conventional evaluation of machine unlearning (MU) by demonstrating that output forgetting metrics (e.g., forget-set accuracy) can overestimate success, as they ignore residual representation-space discrepancies relative to retrained models. The authors introduce retraining-consistent representation forgetting as a stricter benchmark, analyzing unlearning methods across datasets and models. Results reveal structured mismatches: partial alignment on forget samples, inconsistency on retain samples, and residuals concentrated along retraining-related directions, indicating output-layer forgetting often masks deeper representation-space inconsistencies.
machine unlearningrepresentation forgettingretraining consistencyoutput forgettingmembership inference
Geo-Strat-RL: Learning Geological Event Reasoning from Verifiable Tasks
Geo-Strat-RL introduces a synthetic environment for training vision-language models (VLMs) in geological event reasoning through reinforcement learning with verifiable rewards (RLVR). The system generates stratigraphic observations paired with ground-truth event histories and uses an executable verifier to score reconstructions based on geological principles. Results show RLVR improves geological content scores on held-out stratigraphic diagrams by 15-20%, with evidence of cross-domain transfer to synthetic seismic representations without seismic-specific training.
vision-language modelsreinforcement learninggeological reasoningstratigraphic diagramsverifiable rewards
Internal Data Repetition Destroys Language Models
The study quantifies the systematic damage caused by internal data repetition in language models under Chinchilla-style scaling laws, using Compute-Equivalent Gain and Compute-Equivalent Loss metrics. Through controlled experiments on FineWeb-Edu-Dedup and analytical modeling with misspecified linear regression, the authors identify three key phenomena: eval loss peaks at intermediate repeat counts, the peak's location follows a power law in model size, and repetition can waste up to 33% of FLOPs. The findings reveal a tradeoff between memorization and generalization, providing practitioners with tools to measure compute waste from duplicate data.
scaling lawscompute-equivalent lossdata repetitionmemorizationgeneralization
Are Tabular Foundation Models Robust to Realistic Query Distribution Shifts in Microbiome Data?
This work introduces a benchmark to evaluate the robustness of tabular foundation models (TFMs) to biologically inspired distribution shifts in microbiome data. The study assesses TFMs in an in-context learning setting, where models receive unperturbed support sets and are evaluated on query samples subjected to three controlled perturbations: removal of high-abundance taxa, sparsification via zero-inflation, and zero-imputation via spurious non-zero injections. Results show that all perturbations degrade model performance, with zero-imputation being the most harmful, indicating that corrupting global feature structure breaks generalization even when discriminative taxa are retained. Sparsification disproportionately affects TFMs compared to random forests, highlighting their sensitivity to zero-inflation-type shifts.
tabular foundation modelsin-context learningzero-inflationmicrobiome datadistribution shift
ExTra: Exploratory Trajectory Optimization for Language Model Reinforcement Learning
The paper introduces ExTra (Exploratory Trajectory Optimization), a GRPO-compatible framework for improving language-model reinforcement learning by addressing exploration challenges in RLVR. The method combines (i) embedding-based novelty rewards to diversify correct solutions and (ii) entropy-guided prefix regeneration to continue exploration from promising intermediate steps. Evaluated on six mathematical reasoning benchmarks, ExTra improves Qwen3-1.7B over GRPO by +5 points on pass@1 and +7 points on pass@16, demonstrating enhanced single-sample accuracy and inference-time coverage through trajectory-level exploration signals.
rlvrgrpoexploratory trajectory optimizationnovelty rewardentropy-guided prefix regeneration
FlowPipe: LLM-Enhanced Conditional Generative Flow Networks for Data Preparation Pipeline Construction
FlowPipe introduces a Conditional Generative Flow Network (C-GFlowNet) framework for automated data preparation pipeline construction, addressing limitations in Multi-DQN methods through trajectory balance objectives and deep semantic modulation. The method employs Feature-wise Linear Modulation (FiLM) to inject LLM-derived logical priors into policy activations and incorporates failure awareness to avoid invalid states. Evaluations on 74 real-world datasets demonstrate 11.96% accuracy improvement and 12.5× faster convergence compared to SOTA baselines.
conditional generative flow networksdata preparation pipelinesfeature-wise linear modulationtrajectory balancelogical priors
Uncertainty-aware reinforcement learning for chemical language models
The authors propose two uncertainty-aware reinforcement learning (RL) methods for chemical language models (CLMs) to address the neglect of predictive uncertainty in molecular property optimization. The first method treats uncertainty as an additional optimization objective, while the second modulates policy updates based on uncertainty. Evaluated across a controlled Gaussian model system and real-world tasks using ChemProp and Conformal Prediction, these approaches improve robustness by favoring lower-uncertainty regions, increasing the true hit rate from 0.5 to 0.75 and nearly doubling true hits.
reinforcement learningchemical language modelspredictive uncertaintyconformal predictionmolecular design
When Multi-Sensor Fusion Fails to Generalize: Cattle Posture Classification Under Animal-Level and Temporal Distribution Shift
This study investigates the robustness of multimodal sensor fusion for cattle posture classification under temporal distribution shift, demonstrating that conventional evaluation protocols overestimate real-world performance. Using XGBoost, Logistic Regression, Random Forest, and Long Short-Term Memory networks, the authors evaluated posture classification (lying vs. standing) on data from collar accelerometers, rumen-bolus sensors, and environmental measurements collected from a pasture-based beef cattle herd across two consecutive years (2024-2025). While multimodal models achieved strong within-year performance (macro-F1 0.94), cross-year evaluation revealed significant performance decline (macro-F1 0.49), attributed to reliance on context-specific signals like rumen-bolus activity and environmental variables. The findings emphasize the need for robustness-centered evaluation in livestock monitoring.
multimodal sensor fusiontemporal distribution shiftmacro-f1rumen-bolus sensorsrobustness-centered evaluation
Retrieval-Augmented Personalization with Foundation Models for Wearable Stress Detection
The paper introduces a retrieval-augmented personalization method for wearable stress detection, addressing inter-individual variability without user-specific fine-tuning. The approach uses frozen foundation models to retrieve similar patterns from a target user's history, encoding them into a personalized embedding that modulates a lightweight transformer's representations. Evaluated on the WESAD dataset (N=15), it achieves +3.92% accuracy and +4.76% macro F1-score over non-personalized baselines, nearing supervised fine-tuning performance without labeled user data. Temporal and cross-dataset retrieval (K-Emocon to WESAD) demonstrate robustness to limited user history.
retrieval-augmented personalizationfoundation modelswearable stress detectionlightweight transformerinter-individual variability
RevengeBench: Reverse Engineering Code-Space Policies from Behavioral Experiments
The paper introduces RevengeBench, a benchmark for reverse engineering executable code policies from behavioral traces in game environments. The method involves observing target policies' gameplay, designing custom opponent probes to elicit informative behavior, and submitting executable hypotheses evaluated via action-distance metrics. Results show recovery quality varies across 12 LLMs (34-72% distance closed), with reconstructed policies providing competitive advantage, particularly for weaker models. The benchmark enables study of code-space policy recovery as an inverse problem.
behavioral recoveryexecutable codeaction-distance metricsopponent modelinginverse problem
Same Evidence, Different Answer: Auditing Order Sensitivity in Multimodal Large Language Models
We introduce Facet-Probe, a five-facet audit assessing order sensitivity in 18 multimodal large language models (MLLMs) across option, evidence-chunk, document-rank, image-set, and mixed-modality ordering. Using a Bayesian item-response model and same-ordering control, we separate ordering noise from per-facet bias and estimate decoder-stochastic floor. Results show no MLLM is order-invariant, with panel-mean flip rates ranging 24-50%. Gemini same-ordering control reveals substantial ordering excess over decoder-noise floor, and capability predicts but does not eliminate flips, with best model flipping on 13.4% of trials. Prompt-level mitigation proves modality-conditional and ineffective for general order robustness, motivating future architectural approaches.
multimodal large language modelsorder sensitivitybayesian item-response modeldecoder-stochastic floorcross-ordering flip rate
When Does Synthetic Data Augmentation Improve Score-Based Imbalanced Classification?
The paper develops a theoretical framework for understanding when synthetic minority class augmentation improves score-based imbalanced classification. It decomposes augmentation effects into class weighting changes and synthetic distribution discrepancies, analyzing threshold-integrated metrics (AUROC, AUPRC) and optimized-threshold metrics (balanced accuracy, F1). Under well-specified models, augmentation provides only finite-sample variance reduction, while under misspecification it can correct ranking errors through class balance adjustment. Minimax lower bounds show the raw estimator achieves optimal metric-regret rates when well-specified. Improvement bounds quantify approximation error, estimation error, and synthetic distribution error, with simulations confirming limited well-specified gains and nonmonotone misspecified improvements.
synthetic data augmentationimbalanced classificationscore-based modelsdistributional discrepancymetric-regret bounds
FedReLa: Imbalanced Federated Learning via Re-Labeling
FedReLa addresses class imbalance and data heterogeneity in federated learning by introducing a feature-dependent label re-allocator that corrects global decision boundaries without global class distribution knowledge. This model-agnostic, modular approach operates via sample re-labeling and integrates with existing algorithmic methods without additional communication costs. Experiments demonstrate significant accuracy improvements for minority classes and overall performance on stepwise-imbalanced and long-tailed datasets, surpassing prior state-of-the-art methods.
federated learningclass imbalancedata heterogeneitylabel re-allocationdecision boundaries
Why Multi-Step Tool-Use Reinforcement Learning Collapses and How Supervisory Signals Fix It
The paper identifies and addresses catastrophic collapse in multi-step tool-use reinforcement learning (RL) for large language models (LLMs), where performance drops due to control token probability spikes. It systematically evaluates diverse supervisory signals—including off-policy supervision, hint-based guidance, and erroneous example supervision—under synchronous and interleaved training schemes. Results show interleaved supervised fine-tuning (SFT) with RL improves stability but degrades under out-of-distribution (OOD) evaluation, while analysis reveals impacts of learning rates and generalization.
reinforcement learningtool-usesupervisory signalsout-of-distributioncontrol tokens
Is Variational Monte Carlo Robust? Sharp Moment Thresholds and Heavy-tailed Stochastic Optimization
The paper analyzes the robustness of Variational Monte Carlo (VMC) by characterizing the integrability of its stochastic estimators through nodal geometry. For Slater-Jastrow wave functions with Slater-type orbitals, it proves these estimators are heavy-tailed and lack higher moments, while weak moment bounds are established for general analytic ansätze. The authors propose PS-Clip-VMC, a robust variant clipping both local energy and gradient variables, proving convergence in expectation and high probability. Experiments on FermiNet for atoms with ≤18 electrons demonstrate improved robustness over standard methods.
variational monte carlonodal geometryheavy-tailed estimatorsstochastic optimizationferminet
Taxonomy-aware deep learning for hierarchical marine species classification in underwater imagery
A taxonomy-aware deep learning framework improves marine species classification in underwater imagery by addressing domain shift, fine-grained similarity, and uneven annotation granularity. The method combines taxonomy-weighted loss, minimum-risk Bayesian inference, multi-scale feature encoding, and independent per-rank classification heads to align training and inference with hierarchical biological classification. Evaluated on the FathomNet 2025 dataset (79 marine classes across seven taxonomic ranks), the system achieves a mean taxonomic distance of 1.581, within 3% of the top-performing solution (1.535), with significant improvements attributed to metric-aligned inference and decoupled components enhancing generalization under distribution shift.
taxonomy-weighted lossbayesian inferencemulti-scale encodingtaxonomic distancedistribution shift
The Inference-Compute Frontier and a Latency-Efficient Architecture for Limit Order Book Prediction
The paper identifies a power-law relationship between predictive loss and computational work in limit order book (LOB) prediction, with an R²=0.941 fit when extrapolating to high-compute MLPLOB architectures. The study contrasts this with weaker latency-compute correlations, motivating FastBiNLOB, a low-latency LOB mixer using hardware-efficient temporal and feature mixing. In five-seed experiments, FastBiNLOB achieves superior macro-F1 scores (y₁₀, y₁₀₀) at reduced latency compared to state-of-the-art alternatives.
limit order bookpower lawlatency-efficiencymacro-f1temporal mixing
Tensorion: A Tensor-Aware Generalization of the Muon Optimizer
Tensorion introduces a tensor-aware optimizer that generalizes Muon's constrained optimization approach from matrices to higher-order tensors. The method employs a linear minimization oracle (LMO) over a tensor norm ball, balancing tight spectral norm bounds with tractability through operations on adaptively selected unfolding matrices. When applied to matrices, Tensorion exactly recovers Muon. Experimental results on tensor-based computer vision tasks demonstrate improved convergence and more stable gradient updates compared to Adam and existing tensor-aware baselines.
tensor-aware optimizerlinear minimization oraclespectral normunfolding matricesconstrained optimization
Improving Neural Network Training by Decoupling the Magnitude and Direction of Weight Vectors
We propose Magnitude–Direction (MD) Decoupling, an optimizer modification that factorizes each weight matrix into a fixed-norm direction on a hypersphere and learnable per-row and per-column magnitude gains, updated at separate learning rates. This method decouples magnitude and direction dynamics, eliminating the need for weight decay and warmup while maintaining a single fused weight tensor. MD Decoupling improves training stability and performance across Adam and Muon optimizers, transfers optimal learning rates across model widths without retuning, and scales effectively on large Mixture-of-Experts (MoE) models. The approach yields more predictable training dynamics and provides a broadly applicable enhancement to modern optimizers.
magnitude-direction decouplingoptimizer modificationhyperspheremixture-of-expertstraining dynamics
Knowledge Cascade: Reverse Knowledge Distillation on Nonparametric Multivariate Functional Estimation
The paper introduces Knowledge Cascade (KCas), a reverse knowledge distillation framework where a small student model guides the development of a complex teacher model, contrary to conventional distillation. KCas leverages statistical scaling relationships to transfer student-selected parameters (e.g., smoothing parameters in RKHS or deep learning hyperparameters) to the teacher, reducing computational costs while preserving theoretical guarantees. Evaluations on nonparametric multivariate functional estimation, kernel density estimation, and deep learning show KCas achieves computational savings without sacrificing performance, sometimes outperforming full-sample procedures.
reverse knowledge distillationreproducing kernel hilbert spacessmoothing splinesnonparametric estimationhyperparameter transfer
$\text{DT}^2$: Decision-Targeted Digital Twins
The paper introduces $\text{DT}^2$, a decision-targeted training paradigm for digital twins (DTs) that optimizes policy ranking rather than one-step transition accuracy. The method first estimates policy values via fitted Q-evaluation on offline data, then trains DTs to preserve pairwise policy rankings using an architecture-agnostic loss. Experiments demonstrate $\text{DT}^2$ improves policy ranking accuracy and reduces decision regret compared to conventional DT training, while maintaining simulation fidelity, even for unseen policies.
digital twinspolicy rankingfitted q-evaluationdecision regretoffline reinforcement learning
Variational Autoencoder Layer
The paper introduces variational autoencoders (VAEs) as neural network layers rather than standalone models, proposing a novel training strategy for networks incorporating such layers. VAEs leverage probabilistic properties to generate data via smooth latent spaces, maintaining relevance despite their decade-old origins. The work includes thorough performance analysis of models using VAE layers, though specific metrics are not provided in the excerpt.
variational autoencoderlatent spaceneural network layerprobabilistic modelingtraining strategy
A 3D-Printable Dataset for Fair Testing and Comparisons of Tactile Sensors
The authors introduce a 3D-printable dataset of mathematically defined textures to enable fair comparison of tactile sensors, addressing limitations in existing texture datasets tied to specific sensors. Six parametrically generated surface patterns combine sine-wave and Fourier-based functions, varying spatial frequency, amplitude, and directional structure. Evaluation across three 3D printers and multiple filaments shows print quality significantly affects tactile variance, with high-end printers yielding more consistent signatures; classification experiments reveal strong within-printer but limited cross-printer generalization due to geometric inconsistencies.
tactile sensing3d-printed texturesparametric generationsensor benchmarkingreproducible research
An Analysis of Posterior Collapse, Parameterization and Initialization in Variational Deep Gaussian Processes
This work analyzes posterior collapse in variational deep Gaussian processes (DGPs), identifying its connection to the doubly stochastic variational inference (DSVI) algorithm and linear prior mean functions. The authors demonstrate that linear prior means improve optimization conditioning rather than preventing non-injective pathologies, proposing an alternative zero-mean initialization that mimics linear initialization. Three DGP parameterizations are evaluated, showing whitened parameterization enhances stability and mitigates posterior collapse. Experiments confirm the proposed initialization prevents collapse while matching or exceeding linear prior mean performance.
deep gaussian processesposterior collapsevariational inferenceparameterizationinitialization
Hierarchical Graph Learning for Calendar Spread Strategies in Commodity Futures Markets
The paper introduces a hierarchical graph learning approach for calendar spread (CS) strategies in commodity futures markets, addressing gaps in learning-based CS methods and maturity-dependent interrelationship modeling. The method represents futures contracts and underlying assets as hierarchical graphs, with edges capturing correlations and contract-to-asset connections, then predicts price movements using these structural relationships. Empirical evaluation on Chicago Mercantile Exchange Group data shows superior prediction and trading performance versus benchmarks, demonstrating the importance of maturity-dependent interrelationships and CS strategy effectiveness for statistical arbitrage.
hierarchical graph learningcalendar spread strategiescommodity futuresstatistical arbitragematurity-dependent interrelationships
Generating Input Distributions for Explaining Portfolio Optimization Pipelines
The paper introduces a predict-optimize-explain framework for interpreting portfolio models via gradient-based sample generation. It constructs macroeconomic conditions to probe decision pipelines, addressing four key questions: return gap dynamics, diversification vs. concentration, performance across market regimes, and benchmark matching. The method directly evaluates pipeline behavior by generating economically meaningful counterfactuals, revealing differences between predict-then-optimize and predict-and-optimize approaches. Results demonstrate the framework's flexibility in producing interpretable, robust portfolio strategies through integrated prediction, optimization, and explanation.
portfolio optimizationgradient-based sample generationpredict-then-optimizepredict-and-optimizemacroeconomic conditions
ROAD-VLA: Robust Online Adaptation via Self-Distillation for Vision-Language-Action Models
ROAD-VLA introduces an advantage-guided self-distillation framework for robust online adaptation of vision-language-action models, addressing the modality gap between symbolic guidance and low-level actions. The method constructs a proximal teacher in action space by perturbing action-token logits with calibrated advantage estimates, converting sparse rewards into dense token-level supervision while maintaining policy proximity. Evaluated across seven robotic manipulation environments with in-distribution and out-of-distribution shifts, ROAD-VLA outperforms PPO in nearly all settings, demonstrating effective VLA adaptation.
vision-language-action modelsself-distillationonline adaptationadvantage estimationrobotic manipulation
Re-mixing Embeddings for Patient Augmentation in Data Scarce Multiple Instance Learning
The paper introduces a patient augmentation method for Multiple Instance Learning (MIL) in data-scarce medical contexts. It generates synthetic patients by sampling from Gaussian Mixture Models trained on pooled instance embeddings, creating disease-specific statistical distributions. The approach handles three scarcity scenarios: cross-dataset transfer, low-data regimes, and small-cohort non-image tasks. Evaluations show performance improvements over baselines, achieving near-full-dataset accuracy in missing-class scenarios. The method supports rare disease diagnostics and privacy-preserving augmentation.
multiple instance learninggaussian mixture modelsembedding spacepatient augmentationuncertainty quantification
Deep Neural Networks with Ordinal Loss for Medical Applications
The authors propose Ordinal Cross-Entropy (OCE), a novel loss function for deep neural networks handling ordinal regression tasks in medical applications. OCE extends standard cross-entropy by incorporating an ordinal cost matrix that accounts for asymmetric misclassification severity between ordered classes, while preserving probabilistic interpretation and optimization properties. Theoretical analysis demonstrates smoother optimization dynamics and improved ordinal consistency. Empirical evaluations on benchmark datasets show OCE achieves lower prediction error costs and better calibration compared to state-of-the-art ordinal regression methods.
ordinal regressioncross-entropymisclassification severityoptimization dynamicscalibration
Bridging Spherical Black-Box Optimizers
The paper unifies Evolution Strategies (ES), Consensus-Based Optimization (CBO), and Optimization via Integration (OVI) under a theoretical framework, identifying fitness aggregation (sharpness preference) and consensus scope (modality) as key differentiating factors. Hybrid optimizers are introduced: an ES-OVI variant enables explicit control over flat minima preference for robustness-performance trade-offs, while CBO-OVI hybrids combine parametric efficiency with particle-based multimodality. Evaluations on BBO benchmarks and locomotion tasks show hybrid methods outperform constituent algorithms, with particular success in language model merging under limited evaluation budgets.
black-box optimizationevolution strategiesconsensus-based optimizationfitness aggregationmultimodal optimization
RAS: Measuring LLM Safety Through Refusal Alignment
The paper introduces **RAS** (**R**efusal **A**lignment **S**core), a white-box method for evaluating LLM safety by measuring refusal alignment in internal representations rather than generated outputs. **SafeVec** extracts layer-wise refusal directions from a reference model, identifies stable layer windows for separability, and scores target models based on hidden-state alignment under unsafe prompts. Evaluated across `Llama`, `Gemma`, and `Qwen` families, RAS distinguishes safety-aligned models from uncensored variants, correlates with attack success rates, and outperforms judge-based evaluation in speed.
llm safetyrefusal alignmentwhite-box evaluationhidden statesjailbreak prompts
Gaussian Mean Field Variational Inference can Overestimate Predictive Variance
The paper demonstrates that Mean Field Variational Inference (MFVI) can overestimate predictive variance in Bayesian Linear Regression (BLR), contrary to the common belief that MFVI only underestimates posterior variance. Through theoretical analysis, the authors show MFVI's predictive variance exceeds the exact posterior's in data-concentrated directions, sometimes failing to reduce variance compared to the prior. They link this phenomenon to the Cold Posterior Effect and propose temperature adjustment as a corrective measure. Experiments on synthetic and real-world data validate these findings.
mean field variational inferencebayesian linear regressionpredictive variancecold posterior effecttemperature scaling
Black-Box Assisted Regression: Phase Transitions and Minimax Optimality
The paper establishes minimax optimality for black-box assisted nonparametric regression, identifying a phase transition at $δ_c(n) \asymp n^{-β/(2β+d)}$ where the leading risk shifts between $δ^2$ and $n^{-2β/(2β+d)}$. It proposes a Safe Residual Estimator that learns a correction around a fixed predictor $f_0$, initializes residuals at zero for safety, and uses holdout validation to revert to $f_0$ when corrections lack support. The method matches minimax risk up to validation costs, with empirical validation on synthetic data and practical benchmarks (CIFAR-100/CLIP, AG News/Qwen3-8B) demonstrating applicability beyond squared-loss settings.
minimax optimalitynonparametric regressionphase transitionnegative transferholdout validation
Cellular Predictions on the Move: What about Data?
The study demonstrates that incorporating population dynamics and mobility data significantly improves mobile cellular load forecasting accuracy, achieving 60% better predictions compared to traditional methods using only historical traffic data. The authors propose augmenting standard time-series traffic data with metrics characterizing potential traffic sources and their movement patterns, particularly in highway scenarios. Validation through comprehensive experiments confirms the hypothesis that process-informed data enhances forecasting performance beyond conventional approaches focused solely on model architecture improvements.
cellular load forecastingpopulation dynamicstime-series predictionnetwork optimizationmobility patterns
Memory-Efficient Policy Libraries with Low-Rank Adaptation in Reinforcement Learning
The paper demonstrates that Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning method from LLMs, can be effectively transferred to reinforcement learning for robotics, enabling memory-efficient policy libraries. Using Proximal Policy Optimization (PPO) on multi-task robotics scenarios, the authors show LoRA reduces memory usage by 20-160× compared to full fine-tuning, yielding 90-95% storage savings for libraries of 10-50 specialized policies. Crucially, task success rates remain statistically equivalent between LoRA and full fine-tuning, validating the approach's efficacy without performance degradation.
low-rank adaptationproximal policy optimizationparameter-efficient fine-tuningmulti-task roboticspolicy libraries
Learning Subset-Shared Invariances for Domain Generalization with Mixture-of-Experts
The paper proposes subset-shared invariance for domain generalization (DG), addressing the limitation of global invariance assumptions by modeling predictive structure as stable only within domain subsets. The method employs a mixture-of-experts architecture where each expert aligns specific domain subsets, with a routing mechanism composing subset-invariant components for prediction. Training objectives encourage selective alignment, confident routing, and diverse expert specialization. Experiments on DomainBed benchmarks show improved out-of-domain generalization and robustness to increasing domain heterogeneity, suggesting DG should model invariance through partially shared structure rather than global invariance.
domain generalizationsubset-shared invariancemixture-of-expertsrepresentation alignmentrouting mechanism
Statistically Valid Hyperparameter Selection: From Tuning to Guarantees
The monograph introduces a statistically validated framework for hyperparameter selection based on the learn-then-test (LTT) paradigm, addressing the lack of formal guarantees in empirical methods like grid search and Bayesian optimization. The approach formulates hyperparameter selection as multiple hypothesis testing over a candidate set, ensuring application-specific reliability requirements such as bounds on average risk, quantile risk, or information-theoretic constraints. The framework leverages p-values, e-values, and concentration inequalities to provide explicit, finite-sample control of error probabilities. This enables provable reliability in hyperparameter selection, advancing the deployment of AI systems with statistically grounded guarantees.
hyperparameter selectionlearn-then-testmultiple hypothesis testingconcentration inequalitiesfinite-sample control
Two-dimensional Hyperbolic RNN Neural Quantum State
The authors introduce the first two-dimensional hyperbolic neural quantum state (NQS) using a Lorentz 2DRNN architecture, benchmarking it against Euclidean 2DRNN in the 2D Transverse Field Ising Model (2DTFIM) with lattice sizes up to 12×12. Results show superior performance of hyperbolic NQS at critical points described by conformal field theory, where the system's Anti-de-Sitter space duality favors hyperbolic geometry. The study also extends to 1D hyperbolic NQS (Poincaré/Lorentz RNN/GRU) in 2DTFIM, confirming their advantage over Euclidean counterparts due to hierarchical interactions and criticality effects.
neural quantum statehyperbolic geometrytransverse field ising modelconformal field theoryrecurrent neural network
Leaking Circuit Secrets: Gradient Leakage Attacks on Graph Neural Networks
This work presents the first comprehensive evaluation of gradient leakage attacks (GLAs) on graph neural networks (GNNs) in circuit-design and hardware-security tasks. The study assesses state-of-the-art GNN architectures (GraphSAGE, GCN, GIN, GAT) trained on standard netlist benchmarks (ISCAS'85, EPFL, TrustHub) for vulnerability to GLAs, which can expose sensitive circuit information including gate types and hardware Trojan properties. Results show architectural influences on leakage risk, with GAT's attention mechanisms exacerbating vulnerability and GIN's injective aggregation providing stronger resilience. Evaluations of defense techniques (differential privacy, gradient clipping, secure aggregation, quantization, adversarial training) reveal limited effectiveness and potential performance trade-offs.
gradient leakage attacksgraph neural networkshardware trojansinjective aggregationnetlist benchmarks
Concept Removal for Frontier Image Generative Models
The paper introduces a novel concept removal method for frontier image generative models (e.g., SD3.5, Flux, Infinity) that replaces internal bottleneck layers with a structured transcoder. This in-place substitution enables selective disabling of concept-specific activations while preserving model behavior, offering persistent filtering under white-box access. Evaluations demonstrate state-of-the-art concept removal across diffusion and autoregressive models, maintained generation quality, adversarial robustness, and support for sequential concept removal.
concept removaldiffusion modelsautoregressive modelstranscoderactivation features
Beyond One-Size-Fits-All: Diagnosis-Driven Online Reinforcement Learning with Offline Priors
The paper advocates for diagnosis-driven tension management in online reinforcement learning (RL) to address the variability of offline priors' validity across deployments. It introduces a framework analyzing how priors influence online optimization through three functional roles, supported by controlled experiments showing context-dependent prior utility and cross-domain evidence from foundation model post-training to embodied RL. The work engages with five counterarguments, challenging the benchmark-driven paradigm by demonstrating that no universal prior-management strategy exists due to deployment-specific dynamics.
online reinforcement learningoffline priorsdiagnosis-driven learningfoundation modelsembodied intelligence
Low Variance Trust Region Optimization with Independent Actors and Sequential Updates in Cooperative Multi-agent Reinforcement Learning
The paper proposes a low-variance trust region optimization method for cooperative multi-agent reinforcement learning (MARL) with independent actors and sequential updates. It identifies and analyzes the high variance in joint advantage function estimation during sequential policy updates, then introduces a clipping objective to bound advantage fluctuations. Theoretical analysis shows sub-linear convergence to ε-Nash Equilibria. Two derived algorithms demonstrate superior performance on three MARL benchmarks, exhibiting stable convergence and low variance. Empirical results validate the method's effectiveness compared to baselines.
multi-agent reinforcement learningtrust region optimizationadvantage functionsequential updatesnash equilibria
Distill on a Diet: Efficient Knowledge Distillation via Learnable Data Pruning
We propose IF-Beta, an efficient knowledge distillation (KD) framework via learnable data pruning, addressing computational overhead in KD. IF-Beta combines influence functions for sample impact estimation with a Beta-distribution-based sampling policy, optimized through a bilevel objective: the inner loop operates in teacher feature space for fast proxy training, while the outer loop updates policy parameters to maximize distillation performance. Experiments on CIFAR-10/100 and ImageNet demonstrate that IF-Beta outperforms baselines across pruning ratios, enabling students trained on pruned datasets to surpass full-dataset distillation performance.
knowledge distillationdata pruninginfluence functionsbeta distributionbilevel optimization
How Reliable Is Your Jailbreak Judge? Calibration and Adversarial Robustness of Automated ASR Scoring
This study evaluates the reliability of automated judges (safety classifiers and LLM-as-judges) in scoring LLM jailbreak attempts against human labels. Using 596 human-labeled completions from HarmBench, it reveals classifiers over-flag (precision 0.835, recall 0.974) while LLM-judges vary widely in recall (0.06-0.65) despite high precision (0.81-0.94). Adversarial attacks show LLM-judges flip 57-100% with benign framing, whereas classifiers resist surface attacks (<6.7%) but succumb to white-box GCG attacks (70% flip rate). The findings highlight unreliable ASR reporting and recommend human-audited benchmarks.
jailbreakasrclassifierllm-judgeadversarial robustness
Causal-rCM: A Unified Teacher-Forcing and Self-Forcing Open Recipe for Autoregressive Diffusion Distillation in Streaming Video Generation and Interactive World Models
The paper introduces Causal-rCM, a unified framework combining teacher-forcing (TF) and self-forcing (SF) for autoregressive diffusion distillation in streaming video generation and interactive world models. The method leverages complementary forward (consistency models) and reverse (distribution matching distillation) divergences, implementing continuous-time CMs via a custom-mask FlashAttention-2 JVP kernel for 10× faster convergence than discrete-time CMs. Results show state-of-the-art performance, with a distilled 2-step Wan2.1-1.3B model achieving 84.63 VBench-T2V score, and successful application to Cosmos 3 for action-conditioned generation.
autoregressive diffusionconsistency modelsdistribution matching distillationflashattention-2interactive world models
Blasto-Net: An Explainable Multi-Task Learning for Blastocyst Segmentation, Grading, and Implantation Prediction
Blasto-Net introduces a multi-task deep learning model for blastocyst analysis, performing simultaneous segmentation (ZP, TE, ICM), morphological grading, and implantation prediction. The architecture combines an EfficientNet-B3 encoder with a UNet-style decoder, enhanced by CBAM and a novel Edge-Aware Attention Module (EAAM) for semantic and boundary information. Specialized segmentation heads and a composite loss handle compartment topologies, while Grad-CAM++ ensures anatomical consistency. On the HMC dataset, it achieves Dice scores of 94.93% (ICM), 91.60% (ZP), and 88.82% (TE), with an 80.0% F1-score for implantation prediction.
multi-task learningblastocyst segmentationattention mechanismsivf decision supportexplainable ai
Towards Robust EEG Decoding Based on Riemannian Self-Attention
The paper introduces GBWAtt, a Riemannian self-attention network for EEG decoding that employs the Bures-Wasserstein Metric (BWM) to address limitations of existing SPD-based methods. The proposed model captures local relationships in EEG signals and handles ill-conditioned SPD matrices more effectively than traditional Affine-Invariant Metric approaches. By incorporating a power-deformed generalized BWM, the method achieves a nonlinear representation of SPD manifold geometry. Experiments on three EEG datasets demonstrate improved robustness and decoding performance. Code is publicly available.
eeg decodingriemannian manifoldbures-wasserstein metricspd matricesself-attention
The Generalization Spectrum: A Chromatographic Approach to Evaluating Learning Algorithms
The paper introduces the Generalization Spectrum, a framework for evaluating learning algorithms by measuring per-sample generalization across varying transfer distances, from exact recall to cross-context transfer. The method constructs controlled test suites with increasing transfer difficulty, applied to competitive programming tasks to compare RL, SFT-family, and ICL paradigms. Results show RL excels in near-transfer, ICL in correspondence-dependent transfer, while within-family variants reveal trade-offs between local and far-transfer performance, with RFT maintaining stronger far-transfer capabilities than SFT.
generalization spectrumtransfer learningin-context learningreinforcement learningsupervised fine-tuning
The Interplay of Harness Design and Post-Training in LLM Agents
The study introduces harness-aware post-training for LLM agents, treating harness design as a tunable parameter rather than fixed scaffolding. Using an extended $ exttt{ALFWorld}$ framework, it evaluates performance under both in-distribution and out-of-distribution (OOD) tool environment shifts. Results demonstrate that harness-aware post-training enhances in-distribution performance and improves robustness to OOD shifts, whereas traditional post-training suffers significant degradation under strong environmental shifts.
llm agentsharness designpost-trainingood robustnessalfworld
DFMU: Data-Frugal Machine Unlearning
We propose Data-Frugal Machine Unlearning (DFMU), a computationally efficient method for removing elements from pre-trained models while minimizing performance degradation. DFMU computes importance scores for model blocks through a single forward-backward pass, leveraging knowledge-preserving pruning to accelerate convergence with reduced data requirements. Compared to state-of-the-art methods, DFMU achieves 40% higher retain-accuracy using only 13% of data samples and processes class forgetting 88% faster across multiple public datasets.
machine unlearningimportance scoreknowledge-preserving pruningretain-accuracydata efficiency
Learning Optimization Proxies for Sequential Contextual Stochastic Programs: An Order Fulfillment Application
The paper proposes a learning-based optimization proxy for sequential contextual stochastic programs, addressing the latency limitations of traditional solvers in real-time decision systems. The method combines a scenario-embedded neural network trained offline with solver-generated labels and an online feasibility-enforcing decoder, replacing per-epoch optimization with a single forward pass. Specialized for omnichannel order fulfillment, the approach reduces decision latency by 2800x and improves fulfillment costs by 3.3% over the contextual sample average approximation (C-SAA) baseline, while outperforming established policies by 10.7% in cost and halving late-delivery rates in JD.com transactional simulations.
sequential contextual stochastic programsoptimization proxyscenario-embedded neural networkcontextual sample average approximationorder fulfillment
Geometry-Anchored Transport Framework for Exemplar-Free Class-Incremental Learning
The Geometry-Anchored Transport Framework improves exemplar-free class-incremental learning (EFCIL) by integrating feature transport constraints into the primary training phase. It employs an Analytic Geometric Anchor derived via Mahalanobis-aligned regression to mitigate anisotropic representation drift and a Topology-Aware Evolution objective to regularize localized manifold degradation. This approach couples manifold evolution with transport constraints, eliminating the need for decoupled fine-tuning. Experiments on CIFAR-100, TinyImageNet, and ImageNet-100 demonstrate consistent performance improvements over post-hoc alternatives under strict exemplar-free conditions.
exemplar-free class-incremental learningmahalanobis-aligned regressiontopology-aware evolutionanisotropic driftmanifold degradation
Lifelong In-Context Learning with Transformers Requires Parametric Forms of Attention
The paper proposes parametric forms of attention as a solution for enabling lifelong in-context learning in transformers, addressing the quadratic complexity limitation of standard softmax attention. The authors generalize parametric approaches including linear attention, state-space models, and fast weight programmers, which replace the growing KV-cache with online-trainable neural networks to maintain constant memory. While current parametric methods face challenges in memory capacity and update costs, the work identifies key open questions and provides insights for developing long-horizon AI agents.
parametric attentionlifelong learningin-context learningtransformerskv-cache
Stagnant Neuron: Towards Understanding the Plasticity Loss in Multi-Agent Reinforcement Learning Value Factorization Methods
The paper introduces Knowledge-retentive Neuron-level PlastIcity Focusing InjEction (KNIFE), a novel method addressing plasticity loss in Multi-Agent Reinforcement Learning (MARL) value factorization methods. KNIFE targets stagnant neurons—units with negligible gradient updates relative to their weights—by replacing each with a composite unit comprising a frozen knowledge neuron, a re-initialized active neuron, and a compensation neuron. This approach preserves acquired knowledge, restores learning capacity, and maintains cooperation knowledge. Extensive experiments on SMACv2, predator-prey, and matrix games demonstrate KNIFE's superior performance over state-of-the-art plasticity injection methods.
multi-agent reinforcement learningplasticity lossstagnant neuronsvalue factorizationknowledge retention
State Space Models Meet Remote Sensing: A Survey
The survey systematically reviews State Space Models (SSMs) in remote sensing, highlighting their linear computational complexity and long-range dependency capture. It analyzes SSM applications across dense visual prediction, multi-modal data processing, and temporal data analysis, while documenting architectural innovations. The study synthesizes rapid progress in SSM-based remote sensing research and identifies key challenges and future directions, serving as a foundational resource for researchers. The authors maintain an active repository for tracking related works.
state space modelsremote sensinglong-range dependenciesmulti-modal datatemporal analysis
REViT: Roto-reflection Equivariant Convolutional Vision Transformer
The authors propose REViT, a roto-reflection equivariant vision transformer with convolutional attention that preserves rotational, flip, and positional symmetry in feature maps. The method addresses challenges in achieving equivariance in vision transformers by implementing a discretized roto-reflection group equivariant architecture. Experimental results show REViT outperforms existing discrete roto-reflection equivariant neural networks in image classification tasks.
equivariant transformerroto-reflection symmetryconvolutional attentiondiscrete symmetry groupsimage classification
Inverse Reinforcement Learning for Interpretable Keystroke Biomarkers in Parkinson's Disease
This work introduces inverse reinforcement learning (IRL) as a novel approach to interpret keystroke dynamics in Parkinson's disease (PD) diagnosis, departing from traditional classifier-based methods. The authors model keystrokes as discrete choices over typing speed, recovering a three-parameter reward function (speed, consistency, hand-alternation) per subject after addressing feature collinearity in an initial four-parameter decomposition. On the neuroQWERTY MIT-CSXPD dataset (85 subjects, 42 with PD), the recovered speed-preference weight significantly correlates with UPDRS-III severity (r=-0.607, p<0.001) and demonstrates robustness across sensitivity configurations. Two implementation bugs were identified and fixed without materially affecting results. The validation process emphasizes methodological rigor in a field with widely varying reported accuracies (pooled AUC 0.85, I²=94%).
inverse reinforcement learningkeystroke dynamicsparkinson's diseasefeature collinearityneuroqwerty
Variational Inference via Entropic Transport Descent
The paper introduces entropic transport descent (ETD), a particle-based variational inference method that addresses kernel-based repulsion limitations in high dimensions and multimodal targets. ETD frames particle updates as entropy-regularized optimal transport problems, derived via JKO proximal scheme and KL chain rule relaxation, yielding Sinkhorn computations for global coordination. Experiments demonstrate ETD outperforms SVGD, AGF-SVGD, and SGLD in variance-collapse diagnostics, Bayesian logistic regression, neural networks, and molecular Boltzmann distributions, particularly in high-dimensional and multimodal settings.
entropic transport descentparticle-based variational inferencesinkhorn computationjko proximal schememultimodal targets
Pre-Warm: Input-Conditioned Weight Initialization for Convolutional Neural Networks
The paper introduces Pre-Warm, a zero-training-cost method for input-conditioned initialization of the first convolutional layer in CNNs. The technique extracts mean-centered patches from a training batch, clusters them with MiniBatchKMeans, applies inverse Manhattan spatial weighting, and initializes half the filters with resulting centroids (remaining use Kaiming). Closed-form rules govern hyperparameters except a scale factor. Evaluated on MNIST, Fashion-MNIST, CIFAR-10, SVHN, and CIFAR-100 with 8-seed experiments, Pre-Warm yields statistically significant accuracy gains (p < 0.05 on all datasets) with negligible overhead and no architectural changes.
convolutional neural networksweight initializationmini-batch k-meanskaiming initializationinput-conditioned optimization
FUTO Swipe: Layout-Agnostic Neural Swipe Decoding
The paper introduces FUTO Swipe, a layout-agnostic neural swipe decoder that generalizes across contiguous mobile keyboard layouts without retraining. The method employs an encoder that predicts character positions during swipes, with layout information supplied at inference time, and uses geometric augmentations during training to enforce layout invariance. The system achieves generalization to unseen layouts, sometimes outperforming layout-specific models, while releasing both the largest MIT-licensed swipe corpus (1M+ swipes from 12k sessions) and trained models publicly.
neural swipe decoderlayout-agnosticgeometric augmentationsinference-time adaptationgesture prediction
Extreme Meta-Classification for Large-Scale Zero-Shot Retrieval
The paper introduces EMMETT, a framework for synthesizing classifiers on-the-fly for zero-shot retrieval tasks, and IRENE, an efficient instance of EMMETT designed for large-scale deployments. IRENE leverages existing classifiers for observed items to generate classifiers for novel items, addressing limitations of Siamese and extreme classification approaches. Experiments show IRENE improves zero-shot retrieval accuracy by up to 15% in Recall@10 and boosts ad click-through rates by 4.2% in a major search engine. Theoretical analysis guides design choices, validated through ablations.
zero-shot retrievalmeta-classificationlarge-scale retrievalon-the-fly synthesisgeneralization performance
Semantic Allocation in Ordered Bottlenecks: Predictive Residual Inference for Visual Representation Learning
The paper introduces PRIOR (predictive residual inference for ordered representations), a framework addressing limitations of masking-based ordering pressure (MBOP) in ordered bottlenecks. PRIOR replaces activation-rate control with log2-scaled levels and level-wise predictors that focus on residual error, explicitly separating explained from unexplained information. Experiments on contrastive learning and image reconstruction show PRIOR achieves better-ordered representations (coarse-to-fine utility scaling with budget) and superior full-budget performance versus MBOP variants (tail dropping and independent tail-biased dropout), particularly in discrete/quantized settings where MBOP struggles.
ordered bottlenecksmasking-based ordering pressurepredictive residual inferencecontrastive learningquantized representations
MJEPA: A Simple and Scalable Joint-Embedding Predictive Architecture for Audio-Visual Learning
The authors propose MJEPA, a joint-embedding predictive architecture for audio-visual learning that uses a unified encoder for both modalities with a single predictive objective applied intra- and cross-modally. Unlike prior methods requiring modality-specific encoders and multiple objectives, MJEPA demonstrates that cross-modal prediction prevents shared-encoder degradation and improves representations. Evaluations show MJEPA's frozen ViT-g model achieves 6.8 mAP improvement on AudioSet-20K, surpasses finetuned models on ESC-50 and FSD50K, and remains competitive on video benchmarks despite using 10× less training data.
joint-embeddingself-supervised learningcross-modal predictionunified encoderaudio-visual learning
Efficient Adaptive Data Acquisition via Pretrained Belief Representations
The paper introduces POLAR, a method for efficient adaptive data acquisition that decouples representation learning from policy learning by leveraging pretrained predictive foundation models as belief-state encoders. POLAR trains a policy head on top of these representations, creating a unified framework applicable to Bayesian experimental design, Bayesian optimisation, and active learning. Empirical results show POLAR outperforms state-of-the-art amortised methods across diverse tasks while requiring fewer training samples, demonstrating improved scalability and efficiency in amortised data acquisition.
adaptive data acquisitionbelief representationsbayesian experimental designpolicy learningfoundation models
Efficient Analytic Uncertainty Quantification for Multi-Modal Regression
The paper introduces an efficient uncertainty quantification (UQ) framework for multi-modal regression, addressing limitations of single-peak parametric models and semi-parametric methods lacking variance estimation. The method extends Variational Bayesian Inference (VBI) to Quantile Regression (QR) and Classification Restoration (CR), providing analytic Evidence Lower Bounds (ELBO) for training and closed-form predictive densities for inference. Evaluated on three large-scale benchmarks with multi-modal label distributions, the framework outperforms state-of-the-art baselines and matches ensemble models while enabling data-efficient active learning via epistemic uncertainty estimation.
uncertainty quantificationvariational bayesian inferencequantile regressionclassification restorationmulti-modal regression
Neural operator-based digital twins for modeling amyloid-$β$ and tau propagation and treatment optimization in Alzheimer's disease
The study presents a neural operator-based digital twin framework for modeling amyloid-$β$ and tau propagation in Alzheimer's disease, achieving 87% and 81% predictive accuracy respectively. The method combines operator learning with reduced-order representations to infer reaction--diffusion dynamics from sparse, noisy longitudinal PET data, enabling patient-specific biomarker trajectory prediction. The framework further formulates a PDE-constrained optimal control problem to optimize personalized therapeutic strategies for regulating pathological protein spread.
neural operatordigital twinreaction-diffusionoptimal controlbiomarker propagation
EveLoad: Cognitive Workload Recognition from Event-Based Eye Movements
The paper introduces EveLoad, the first event-based eye-movement dataset with graded cognitive workload annotations, collected from 20 participants using an N-back-guided fixation paradigm to isolate workload-related ocular dynamics. The proposed framework encodes spatiotemporal event representations from microsecond-resolution event camera data, achieving 96.36% subject-specific accuracy and 96.13% mixed random split accuracy on six workload levels. Results demonstrate event-based eye tracking's potential for unobtrusive cognitive workload monitoring in adaptive rehabilitation interfaces.
event-based visioncognitive workloadocular dynamicsn-back paradigmspatiotemporal encoding
An iterative energy-based multimodal transformer for joint retrieval of wheat soil moisture, leaf area index, and plant height from Sentinel-1 and Sentinel-2 time series
The Iterative Energy-Based Transformer (iEBT) jointly retrieves soil moisture (SM), leaf area index (LAI), and plant height (PH) from Sentinel-1 SAR and Sentinel-2 multispectral time series via iterative gradient descent on a learned compatibility energy function. This multimodal approach embeds predictors in a shared sequence, achieving mean R²=0.854±0.012 (SM:0.841, LAI:0.905, PH:0.821) on 700 field measurements, with Sentinel-1 dominating SM and Sentinel-2 LAI retrieval. The energy function serves as an uncalibrated quality diagnostic, reducing RMSE when filtering high-energy samples, though cross-season domain shifts remain challenging.
multimodal transformeriterative energy-based learningsentinel-1/sentinel-2 fusionbiophysical parameter retrievalgradient descent optimization
Minimax PAC Bounds for Learning in Exogenous Contextual MDPs
The paper establishes minimax PAC bounds for learning in exogenous contextual MDPs with unknown context distributions and transition dynamics. It introduces a variance-reduced algorithm for policy evaluation (PE), best-value estimation (BVE), and best-policy extraction (BPE) in tabular MDPs with i.i.d. contexts. Key results include a sample complexity of (Õ(1/((1-γ)³ε²)), 0) when only the context distribution is unknown, and (Õ(|X|/((1-γ)³ε²)), Õ(1/((1-γ)²ε²))) in the fully unknown regime, with rates independent of context space size |Z|.
contextual mdpssample complexitypolicy evaluationminimax boundsvariance reduction
Laplace--Fisher Gate Identities for Optimal Matrix-Gated Blended Score Estimation
The paper introduces the Laplace--Fisher Gate Identity (LFGI), a variance-optimal matrix-valued blending coefficient for score estimation in diffusion-based sampling. LFGI minimizes conditional risk by leveraging Tweedie's identity and a target-score identity, enabling unbiased finite-reference estimators without altering expected values. The method is applied to Bayesian inverse problems, constructing normalized posterior-density surrogates using MCMC pilot samples and derivative information. On a PDE-constrained inverse-problem benchmark, LFGI demonstrates improved posterior-density calibration and sampling diagnostics compared to other score-estimator classes, validated in Gaussian and non-Gaussian settings.
laplace-fisher gate identityscore estimationbayesian inverse problemstweedie's identitymcmc pilot samples
The Gentle Collapse: Distributional Metrics for Continual Learning
The authors propose six softmax-derived metrics for characterizing catastrophic forgetting (CF) in continual learning, addressing limitations of accuracy degradation as a binary metric. The metrics—spanning true-label rank, predictive confidence, and distributional divergence—provide continuous signals normalized to [0, 1] without modifying training procedures. On CIFAR-100 and TinyImageNet, these metrics reveal class-specific forgetting patterns where accuracy saturates, enabling actionable insights. Using per-sample metric scores as loss weights reduces forgetting by 1.3 percentage points over uniform experience replay on CIFAR-100. Additionally, log-true-label rank trend sampling achieves 41.07% stability (std. = 0.57) over small windows, outperforming accuracy-trend by 6.28 percentage points and reducing forgetting by 7.7 points on TinyImageNet.
catastrophic forgettingsoftmax-derived metricstrue-label rankdistributional divergenceexperience replay
Forget to Improve: On-Device LLM-Agent Continual Learning via Budget-Curated Memory
The paper introduces \sys{}, a memory governance framework for on-device LLM agents that optimizes continual learning via a net-value-per-byte metric. The method curates memory through three budget-aware decisions: KEEP (evicting low-value entries under resource constraints), SHARE (transmitting only high-value insights), and TRUST (validating peer entries by provenance). Evaluated on task-drift benchmarks and a Jetson testbed with robot-arm nodes, \sys{} reduces memory usage by 2.7× and uplink traffic by 2.4× while eliminating injection attacks (success rate drops from 0.75 to 0) and improving accuracy against poisoned/stale data. The approach demonstrates that selective forgetting enhances agent performance while minimizing resource overhead.
on-device learningmemory governancenet-value-per-bytecontinual learningllm-agent
A Framework for Directed Hypergraph Signal Processing via tensor t-SVD
The authors propose Directed Hypergraph Signal Processing (DHGSP), a framework extending graph signal processing to model higher-order polyadic and asymmetric directional relationships. The method leverages tensor singular value decomposition (t-SVD) within t-product algebra to define a novel adjacency tensor, topologically faithful shift operator, and lossless Directed Hypergraph Fourier Transform (t-DHGFT). Experimental results on real traffic networks demonstrate DHGSP's superior performance over matrix-based (graph/digraph) and undirected tensor-based (hypergraph) baselines in denoising tasks.
directed hypergraphtensor singular value decompositiont-product algebrashift operatordenoising
Speculative Decoding at Temperature Zero: A Scoped Safety-Invariance Screen with a 48,072-Sample Expansion
The paper introduces Typical-Acceptance Invariance Screen (TAIS), a safety-invariance verification method for speculative decoding at temperature zero, ensuring draft-side behavior does not affect safety-scored outputs. TAIS requires byte-identity, TOST equivalence (±3pp), and Cohen's h < 0.1 across 48,072 samples, including confirmatory and expansion sets with various execution modes and adversarial drafts. Results show no detectable safety divergence, with Cohen's h ≤ 0.024, 25/27 TOST contrasts passing, and DPO-adversarial drafts producing byte-identical outputs. A 70B production-scale probe shows AdvBench refusal at 0.839 (95% CI [0.809, 0.864]).
speculative decodingsafety-invariancebehavioral-equivalencecohen's htost equivalence
How Modular Is a Frontier Mixture-of-Experts? A Pre-registered Causal Test in Which Apparent Expert Modularity Mostly Dissolves
The study causally tests functional modularity in Command A+, a 218B-parameter sparse Mixture-of-Experts (MoE) model, through pre-registered ablation experiments. Using a routing-mass atlas and four metrics, it evaluates six expert families against a random-expert null, with bootstrap confidence intervals on an independent corpus. Results show robust modularity is rare: only the Arabic-language family (1/6) meets conservative selectivity criteria, while others exhibit measurement-dependent effects. The method confirms modularity in Qwen3-30B-A3B as a positive control, ruling out quantization artifacts. Findings caution that ablation-based modularity claims require strict control of corpus, metric, and statistical thresholds.
mixture-of-expertscausal ablationrouting-mass atlasfunctional modularitybootstrap confidence intervals
Bias-Controlled Primal-Dual Natural Actor-Critic: Optimal Rates for Constrained Multi-Objective Average-Reward RL
(No summary returned.)
Emergent Capabilities Arise Randomly from Learning Sparse Attention Patterns
The paper demonstrates that emergent capabilities in transformers arise stochastically during training, linked to abrupt learning of task-relevant sparse attention patterns. Through controlled experiments on synthetic linear map and cellular automata tasks, the authors show emergence timing depends on context length, pattern sparsity, and architecture choices. Key findings include: attention head count improves learning efficiency more than head dimension scaling, and MLP-Mixer outperforms transformers on complex pattern tasks. The work provides mechanistic evidence that emergence stems from inherent difficulty in learning sparse attention mappings.
emergent capabilitiessparse attentiontransformer scalingin-context learningmlp-mixer
Neural Scaling Universality: If Exponents Are Fixed, Time to Understand Coefficients
The paper posits that neural scaling law exponents are fixed by universal mechanisms: a 1/3 time scaling from Softmax nonlinearity, inverse width scaling from representational superposition, and inverse depth scaling from Transformer layer ensemble averaging. These mechanisms create a universality class where exponents remain stable across architectures and datasets, while coefficients vary with implementation details. The authors argue that analyzing coefficients—not exponents—is crucial for optimizing model shape and compute efficiency, suggesting this universality framework could guide discovery of superior scaling classes.
neural scaling lawsuniversality classsoftmax nonlinearityrepresentational superpositioncompute-optimal frontier
Multi-Stream Temporal Fusion for Financial Fraud Detection
The paper introduces Multi-Stream Fraud Transformer (MSFT), a Transformer-based architecture for financial fraud detection that processes heterogeneous event streams (transactions, logins, risk signals) via independent encoders and configurable fusion mechanisms. An ablation study compares five fusion strategies (concatenation, gated fusion, time-aware positional encoding, cross-stream attention, full combination) on a 10M-user dataset (1.5% fraud rate), showing MSFT achieves 0.9961 AUROC with time-aware encoding, outperforming single-stream Transformers (0.82 AUROC) and XGBoost (0.74 AUROC). Gated fusion yields optimal precision (0.989), while risk events provide the strongest signal. Production validation confirms 22% AUROC improvement over XGBoost.
multi-stream fusiontime-aware positional encodingaurocgated fusionfraud transformer
Scalable Peptide Design via Memory-Efficient Equivariant Transformer
The paper introduces MEET (Memory Efficient Equivariant Transformer), an E(3)-equivariant backbone for scalable peptide design that maintains coupled invariant scalar and equivariant vector features while optimizing memory usage. The method employs global coordinate aggregation for vector initialization, augmented query-key dot products for pairwise distances, and sparse bond adaptation for covalent bond information, integrated into a VAE and latent diffusion pipeline. Experiments on AFDB-derived datasets demonstrate linear memory scaling with atom count and improved generation quality in binding affinity, physical validity, and diversity compared to existing methods.
equivariant transformerpeptide designlatent diffusiongeometric computationmemory-efficient attention
Certification of Machine Learning Models via Directional Sharpness
The authors propose directional sharpness, a novel metric for certifying machine learning model generalization that addresses limitations of existing proxies like test accuracy and traditional sharpness measures. The method evaluates model quality along specific parameter-space directions, showing stronger correlation with generalization and improved reliability in detecting poor generalization under training deviations. Empirical and analytical results demonstrate its effectiveness in both model auditing (with training data access) and zero-knowledge proof scenarios, while maintaining computational efficiency.
model certificationgeneralizationsharpnesszero-knowledge proofsparameter-space directions
Adaptive Joint Compression and Synchronisation in Federated Split Learning for IoT Rainfall Prediction
The paper introduces an adaptive federated split learning (FSL) framework for IoT rainfall prediction that jointly optimizes activation compression and synchronization frequency. The method employs a latency-driven scheduler with per-client EMA smoothing on the server, evaluated through 17 simulation scenarios and a Raspberry Pi deployment using ERA5 weather data. Results show minimal AUPRC variation (0.6381-0.6484 in simulation, within 0.011 on Pi) despite aggressive quantization (int8) and reduced synchronization (rho=3), achieving 87% smaller activation uploads and 54% less synchronization traffic versus float32 baseline, with runtime jitter reduced from +/-688s to +/-10s.
federated split learningactivation compressionsynchronization frequencylatency-driven scheduleriot rainfall prediction
TRACER: Training-Free Closed-Loop Structured Inference for Traffic Accident Reconstruction
TRACER introduces a training-free framework for traffic accident reconstruction, addressing the limitations of existing methods that prioritize semantic plausibility over geometric and dynamic accuracy. The approach formulates reconstruction as a closed-loop structured inference process, refining motion hypotheses under geometric, kinematic, and interaction constraints. It leverages structured case memory and consistency-driven diagnosis for incremental, interpretable corrections. Evaluations on real-world data demonstrate TRACER's superior performance in geometric fidelity, velocity consistency, and collision accuracy compared to data-driven and physics-based baselines.
traffic accident reconstructionclosed-loop inferencegeometric constraintskinematic consistencystructured case memory
A Zeroth-Order Deep Learning Method for Fully Nonlinear Parabolic Partial Differential Equations with Unknown Coefficients
The authors propose a zeroth-order deep learning method for solving high-dimensional fully nonlinear parabolic PDEs with unknown coefficients, addressing challenges in data-driven PDE solvers. The approach employs zeroth-order derivative estimators from perturbed Monte Carlo trajectories, enabling gradient and Hessian network training via function evaluations without explicit derivative computations. A statistical learning analysis establishes non-asymptotic error bounds, decomposing total error into discretization, approximation, statistical, and ZOD bias components. Sample complexity in Sobolev space is derived for second-order derivatives. Numerical experiments demonstrate competitive performance in moderate to high dimensions.
partial differential equationszeroth-order derivativemonte carlo trajectoriessobolev spacenon-asymptotic error
What's in an Earth Embedding? An Explainability Analysis of Location Encoders
The paper presents an explainability analysis of geographic implicit neural representations (INRs) that map Earth coordinates to location embeddings. The authors decompose embeddings into three interpretable components: (i) sparse latent concepts via autoencoders, (ii) natural language concepts using sparse linear concept embeddings (SpLiCE) over a geospatial dictionary, and (iii) visual features extracted via CLIP Surgery saliency maps. Results show these decompositions retain reconstruction capability while revealing geographic structures (forests, urban areas) and systematic information differences (biomes, climate signals), with saliency maps highlighting complementary features like roads. The work establishes foundational methods for auditing geospatial embeddings.
implicit neural representationslocation embeddingssparse autoencodersclip surgerygeospatial dictionary
From Forecasting Leaderboards to Deployment Decisions: A Fail-Closed Certification Protocol
The paper introduces a fail-closed certification protocol to determine when forecasting leaderboard winners can be reliably deployed given a specific decision interface and utility function. The protocol establishes sufficient evidential conditions to identify deployment-side reversals caused by friction, ensuring conservative deployment decisions. Using Traffic-Hourly as a certified anchor, the method tests for overclaiming across 22 candidates and 362 grid cells, blocking 155 forecast/deployment winner inversions. The contribution focuses on a decision protocol rather than proposing new forecasting models or metrics.
fail-closed certificationforecasting leaderboardsdeployment-side reversalsdecision interfaceswitching friction
The Geometry of Sequential Learning: Lie-Bracket Prediction of Transfer Order
The paper introduces a geometric framework for sequential learning order effects, demonstrating that local order dependencies are governed by Lie-bracket commutators of gradient update fields. This yields a pairwise scoring mechanism (Lie-Bracket Tournament) for optimal curriculum sequencing, requiring only O(N) Hessian-vector products and an O(N log N) sort. Empirical results show 98.1%/98.9% pairwise accuracy for instruction-SFT/DPO at k=1, 73.1%/72.2% at k=20, and 99.0-99.6th percentile performance on 56 MMLU subjects, outperforming gradient-norm baselines. The method successfully recovers optimal 3! schedules in 87.5% of trials and ranks 85 Stack programming domains for Python targets.
sequential learninglie-bracket commutatorcurriculum learninggradient update fieldshessian-vector product
Model selection with proper scoring rules on data sets of time series: prefer the mean scaled score
The paper demonstrates that mean scaled scores outperform alternative aggregation methods (mean ranks, win rates) for probabilistic forecasting model selection on multiple time series datasets. The authors identify score distribution skewness—particularly pronounced with short test sets—as causing non-mean criteria to favor misspecified models, while mean scores remain robust. Experiments on M5 competition data show convergence across methods with larger test sets, but confirm mean scaled scores' reliability through consistent decisions under varying scaling factors.
probabilistic forecastingproper scoring rulesmodel selectiontime series aggregationm5 competition
Solving Markov Decision Processes with Future Information via MPC
The work establishes structural conditions under which parameterized Model Predictive Control (MPC) can exactly represent optimal value functions and policies for Markov Decision Processes (MDPs) incorporating future information (e.g., forecasts, reference trajectories). By treating MPC as a structured function approximator and learning its parameters via Reinforcement Learning (RL), the method enables optimal policy derivation for augmented-state MDPs. Experimental validation on a point-mass racing task demonstrates effective incorporation of future reference information.
model predictive controlmarkov decision processesreinforcement learningoptimal policyfuture information
Low-Cost High-Order Singular Value Decomposition for Tensor-Based Reconstruction from Sparse Sensor Measurements: Urban Flow and Air-Quality Applications
The paper introduces low-cost High-Order Singular Value Decomposition (lcHOSVD), a tensor-based framework for reconstructing high-dimensional environmental fields from sparse sensor measurements. Unlike matrix-based approaches, lcHOSVD preserves multidimensional tensor structure while reducing computational costs of conventional HOSVD. Applied to urban flow and air-quality datasets with 1-4% spatial sampling, lcHOSVD achieves lower reconstruction errors than lcSVD in cases of strong multidimensional coupling, demonstrating superior robustness to uneven sensor distributions.
tensor decompositionsparse sensingfield reconstructionhosvdurban flow modeling
Sample complexity of unbalanced entropic OT
The paper analyzes the sample complexity of entropic unbalanced optimal transport (OT), focusing on the optimal coupling rather than just the transport value. It introduces a translation-invariant dual formulation and establishes compactness and strong convexity properties for dual variables, enabling high-probability finite-sample bounds for empirical couplings. Results demonstrate that entropic regularization mitigates the curse of dimensionality, reduces sample requirements for stable transport estimation, and maintains compatibility with Sinkhorn-type solvers, justifying its widespread use in machine learning.
optimal transportentropic regularizationsample complexitysinkhorn algorithmdual formulation
Learning Diachronic Representations of Ancient Greek Letterforms
The paper introduces a novel framework for learning diachronic representations of ancient Greek letterforms across centuries, addressing challenges of symbolic variation, scarce data, and systematic degradation. Three datasets—Hell-Char, PaLit-Char, and Med-Char—spanning 3rd BCE to 14th CE are curated for training and evaluation. The method employs a similarity-weighted supervised contrastive loss to bias embeddings based on inter-class similarities and a lacuna-driven augmentation scheme simulating manuscript corruptions. Experiments with CNN and ResNet architectures demonstrate strong recognition performance, coherent class separation, and interpretable embeddings enabling clustering, stylistic subgroup identification, and prototype visualization of diachronic evolution. The approach offers a transferable paradigm for representation learning under scarce, temporally evolving, and noisy conditions.
diachronic representationcontrastive losslacuna-driven augmentationletterform evolutionancient greek
ConSolv: Solvent-Conditional Machine Learning Implicit Solvent Potential
ConSolv introduces a solvent-conditional machine learning potential (MLP) that generalizes across 66 organic solvents, addressing limitations of aqueous-focused implicit solvent models. The architecture employs an attention-based solvent-embedding block, trained on combined experimental solvation free energies and ab initio data. Benchmarks show superior performance to classical explicit solvent methods and selected ab initio approaches, with validation via NMR data for γ-fluorohydrin in chloroform. The model supports explainable AI analysis through its attention mechanism and is extensible to broader chemical spaces.
implicit solventmachine learning potentialsolvation free energyattention mechanismnuclear magnetic resonance
Latent Block-Diffusion Temporal Point Processes: A Semi-Autoregressive Framework for Asynchronous Event Sequence Generation
Proposes Latent Block-Diffusion Temporal Point Processes (LBDTPP), a semi-autoregressive framework for asynchronous event sequence generation that combines autoregressive block-level generation with latent-space diffusion. The method sequentially generates variable-length event blocks while performing parallel Gaussian diffusion within each block, theoretically reducing error accumulation via Wasserstein bounds. Experiments on six benchmarks show LBDTPP outperforms state-of-the-art TPP methods in unconditional and conditional generation, with ablation studies validating the benefits of latent diffusion and block-wise processing.
temporal point processessemi-autoregressivelatent diffusionwasserstein boundsevent sequence generation
A Single Stepsize Suffices for Unprojected Linear TD(0): Simultaneous Robust and Fast Rates via Polyak--Ruppert Averaging
The paper establishes high-probability convergence guarantees for unprojected linear TD(0) with Polyak-Ruppert averaging under Markovian sampling. Using a single stepsize schedule η_t ∝ (1/τ_mix)log(t)/√t that requires no prior knowledge of curvature parameter ω, the authors prove uniform boundedness of TD(0) iterates. Key results include simultaneous robust (Õ(τ_mix/√T)) and fast curvature-dependent (Õ(τ_mix²/ωT)) convergence rates via a novel Poisson-equation-based decomposition of Markov noise. The analysis leverages geometric mixing properties and a self-bounding inductive argument for pathwise stability.
td(0)polyak-ruppert averagingmarkovian samplingpoisson equationgeometric mixing
Closed-Loop Graph Algorithm Execution with Small Language Models: Step Accuracy and Rollout Reliability
The study investigates small language models' capacity for closed-loop execution of graph algorithms, demonstrating that reliable policies emerge for structural procedures like traversal and coloring despite error accumulation challenges in weighted algorithms. Using a novel evaluation framework with synthetic graph families and disjoint data partitions, the work measures both local decision quality (step accuracy) and global behavior (rollout accuracy, constraint validity). Key findings reveal that strong next-step prediction does not guarantee autonomous execution reliability, motivating complete rollout evaluations over isolated decisions.
closed-loop executiongraph algorithmserror accumulationrollout accuracystep accuracy
CKM-Driven Communication-Aware UAV Intelligent Trajectory Optimization for Urban Inspection
The paper proposes a channel knowledge map (CKM)-driven trajectory optimization framework for multi-UAV urban inspection tasks, addressing communication reliability in spatially heterogeneous channels. The method combines a diffusion model for time-accumulated CKM construction from sparse observations with a global-to-local graph attention network soft actor-critic algorithm, which jointly optimizes target sequencing and continuous path control. Simulations show the approach improves trajectory efficiency and communication reliability by 22.7% and 18.3% respectively, without requiring real-time channel feedback.
channel knowledge mapdiffusion modelgraph attention networksoft actor-critictrajectory optimization
Auto-Configured Explainable Graph Neural Networks for Multi-Site Pollution Prediction
This study introduces an auto-configured graph construction method for Graph Neural Networks (GNNs) using a confusion matrix to dynamically capture inter-class relationships, combined with a hybrid loss function of energy distance and Huber loss to enhance learning stability. Five GNN models—Graph Convolutional Networks (GCNs), Simple Graph Convolutional Networks (SGConv), Graph Isomorphism Networks (GINs), Graph Attention Networks (GATs), and GraphSage—were evaluated on air pollution data from the University of Utah AirU Pollution Monitoring Network. GraphSage achieved the highest accuracy in predicting PM1, PM10, and PM2.5 concentrations across various time horizons. GNNExplainer and PGExplainer were employed to ensure model transparency, with GNNs outperforming traditional machine learning and deep learning models in air pollution forecasting.
graph neural networksconfusion matrixhybrid loss functiongraphsagegnnexplainer
Diagnosing and Mitigating Compounding Failures in Agentic Persuasion via Taxonomic Strategy Retrieval
We introduce Taxonomic Strategy Retrieval-Augmented Generation (TS-RAG), a method to mitigate compounding errors in foundation-model agents during multi-step persuasion tasks. TS-RAG routes strategies through a discrete categorical bottleneck, decoupling argumentative structure from topical content to eliminate semantic leakage in standard RAG. Zero-shot, cross-domain evaluations show TS-RAG improves abstract logic transfer, enabling lightweight persuaders to outperform parametrically superior opponents (win rates increase from 70.5% to 78.5%). Additionally, we propose Debate State Representation (DSR) for trace-level diagnostics, highlighting the necessity of strict constraints to prevent sycophantic conformity.
taxonomic strategy retrievalsemantic leakagedebate state representationargumentative structurecompounding errors
Why Do Accumulated Transformations Extrapolate?
The paper demonstrates that accumulated orthogonal transformations, not just Householder reflections, enable length extrapolation in transformers by creating finite mixing windows that suppress distant tokens. The authors analyze a simplified variant of RoPE using accumulated token-dependent SO(2) rotations, proving that such transformations become incoherent after finite steps while preserving near-token signals. Experiments show improved extrapolation over RoPE (though degrading at extreme lengths) and that rotating values extends the effective context. The analysis reveals a fundamental tradeoff: accumulated rotations cannot indefinitely preserve signal without explicit far-mass control, explaining ALiBi's superior stability.
length extrapolationaccumulated transformationsso(2) rotationshouseholder reflectionsfar-mass control
Quantifying Explainable AI-introduced signal noise on ECG data with Spectral Entropy
The authors propose spectral entropy as a quantitative measure to assess signal noise introduced by explainable AI (XAI) techniques in healthcare applications. They evaluate this method in the context of arrhythmia classification using ECG data, analyzing the noise contributions from various post hoc explainability approaches. The study demonstrates spectral entropy's effectiveness in distinguishing between model-derived signals and XAI-introduced noise, providing a systematic way to evaluate explanation quality in medical deep learning systems.
spectral entropyexplainable aiecg datapost hoc explainabilitysignal noise
LLM Performance on a Real, Double-Marked GCSE Benchmark
The study introduces a novel dataset of 32,534 double-marked GCSE exam responses across five subjects, including handwritten work, to evaluate large language models (LLMs) against human examiner consensus. Using off-the-shelf LLMs, the authors measure agreement with examiners on 328 questions, comparing model performance to inter-examiner agreement. Results show top-performing LLMs match or exceed human examiner agreement levels, even for subjective tasks like English essays and complex handwritten math scripts, with performance largely independent of model size.
large language modelsgcse benchmarkdouble-marked datasetexaminer agreementhandwritten script processing
Don't Go Breaking My LLM: The Impact of Pruning Attention Layers on Explanation Faithfulness and Confidence Calibration
The study investigates how pruning attention layers in Large Language Models (LLMs) affects explanation faithfulness and confidence calibration, beyond just accuracy. Using five LLMs and eight datasets, the authors demonstrate that while pruned models often retain accuracy (up to 33% layer removal), their faithfulness and calibration frequently degrade, revealing a misalignment between confidence, interpretability, and accuracy. The findings emphasize the need to include explainability and calibration metrics in pruning evaluations.
attention pruningfaithfulnessconfidence calibrationlarge language modelsinterpretability
Frequency Domain Reservoir Computing
The paper introduces Frequency Domain Reservoir Computing (FRESCO), an Echo State Network variant that achieves O(N) complexity for dense recurrent updates by operating entirely in the frequency domain. The method combines a dimensional zero-padding input embedding, packed frequency-domain readout, and native frequency-domain non-linearity to avoid domain-shift overheads while maintaining expressivity. FRESCO matches state-of-the-art performance on memory tasks, sequential classification, and multivariate forecasting while reducing computational costs and energy consumption compared to traditional O(N²) ESN approaches.
echo state networksfrequency domainreservoir computingrecurrent modelslong-horizon forecasting
Training Dynamics of Neural Software Defect Predictors under Coupled Data-Quality Issues
The study contributes an interaction-aware empirical protocol and taxonomy of training-dynamics patterns for coupled data-quality issues in neural software defect prediction (SDP). Using a controlled intervention on class-level UBD datasets, it trains a fixed MLP under imbalance-only, overlap-only, and joint conditions, logging per-epoch dynamics (gradients, weights, biases, error trajectories). Analysis employs effect sizes, sensitivity tests, and rule-based classification to characterize patterns under coupled issues. Results aim to reveal how internal neural dynamics manifest when class imbalance and overlap interact in metric-based SDP.
software defect predictiontraining dynamicsclass imbalanceclass overlapmlp
What Do Language Priors Contribute to Darcy-Flow Inversion? A Mechanistic Audit
The study demonstrates that language priors can effectively inject geological knowledge into learned inverse solvers for Darcy-flow problems. Using sentence embeddings as conditioning representations, the authors evaluate six synthetic geological classes and the SPE10 benchmark, showing an 81% reduction in reconstruction error compared to text-free baselines. Key findings reveal that categorical class-level constraints dominate performance gains, while within-class geometric details are secondary; sentence embeddings improve training stability and enable paraphrase-based analysis but add minimal accuracy under dense observations.
darcy-flow inversionlanguage priorssentence embeddingsinverse problemsgeological conditioning
Learning Dynamical Systems from Multiple Sparse Datasets: A Hierarchical Bayesian Modeling Approach
The authors propose a hierarchical Bayesian framework for meta-learning in dynamical systems, addressing parameter estimation from sparse, noisy, and irregularly sampled datasets. The method models dataset-specific parameters as draws from a shared population distribution, embedding a numerical ODE solver within gradient-based MCMC for efficient posterior inference. Experiments demonstrate improved predictive performance compared to unpooled approaches, enabling data-efficient system identification in sparse-data settings.
hierarchical bayesian modelingdynamical systemsmeta-learninggradient-based mcmcsystem identification
Project Auto-World: Towards Automated Benchmarking of Neural Relational Reasoners
We propose Auto-World, a framework for automated benchmarking of neural relational reasoners using LLM-driven evolutionary search. Given a Datalog-parametrized world and an Edge Transformer evaluator, we employ FunSearch-based evolutionary optimization and autonomous agentic search to generate increasingly challenging problem instances. Results demonstrate that Edge Transformer performance improves when trained on this data, exhibiting robust generalization to further perturbations. The framework extends to novel worlds proposed by LLMs, enabling autonomous research in neural relational reasoning. This approach addresses the challenge of evaluating systematic generalization in relational reasoning tasks.
neural relational reasoningdatalog rulesedge transformerevolutionary searchautonomous benchmarking
Evidence for feature-specific error correction in LLMs
This work provides empirical evidence for feature-specific error correction in large language models (LLMs), supporting theoretical predictions about computation in superposition. The authors propose an activation perturbation method to test error correction, measuring robustness along candidate feature directions versus mixed directions across multiple LLMs (Gemma-2-9B, Qwen3-1.7B, Llama-3.1-8B, Mistral-7B-v0.3, Aya-Expanse-8B, Yi-1.5-9B). Results show privileged treatment of pure feature directions (p>2 in L^p-norm decomposition) compared to random/PCA directions (p≈2), consistent with feature-specific error correction. The method is validated on a toy model with known ground-truth features, demonstrating degradation of p as directions rotate away from true features.
superpositionerror correctionactivation perturbationfeature directionsl^p-norm
Curvature-Guided Mixing for MLLM Adaptation
The paper introduces Curvature-Guided Mixing (CGM), a framework for merging pre-trained and fine-tuned Multimodal Large Language Models (MLLMs) to mitigate catastrophic forgetting. CGM employs a second-order approximation of loss landscapes to derive an optimal soft mixing ratio, while its variant CGM$\dagger$ uses curvature-aware scores for sparse parameter selection. Evaluations on LLaVA-1.5 and Qwen2.5VL demonstrate improved trade-offs between task specialization and general knowledge retention compared to existing methods.
curvature-guided mixingmultimodal llmshessian approximationparameter mergingcatastrophic forgetting
Towards Scalable Multi-Task Reinforcement Learning with Large Decision Models
The paper introduces LDM-v0, a Large Decision Model demonstrating that a single transformer policy can achieve multi-task reinforcement learning across 1,000+ heterogeneous environments. The model processes histories of observations, actions, rewards, and terminations via supervised next-action prediction on offline trajectories, using a unified architecture spanning robotics, autonomous driving, and other domains. Results show LDM-v0 matches task-specific policies' performance, validating large-scale offline pretraining for diverse RL tasks.
large decision modelmulti-task reinforcement learningtransformer policyoffline pretrainingnext-action prediction
Enhancing Clinician Decision-Making via Uncertainty-Aware Multi-Expert Fusion for Stroke Rehabilitation
The paper introduces xAARA, an uncertainty-aware multi-expert fusion system for stroke rehabilitation assessment. The method employs a Dynamic Bayesian Network with entropy-based gating to compose 692 calibrated multimodal models, processing multi-view video to generate Action Research Arm Test (ARAT) scores with uncertainty quantification and multi-level explanations. Evaluated on 105 stroke survivors (788 exercises), xAARA achieved 94.2% task accuracy (κ=0.934) and 81.3% movement-phase accuracy (κ=0.727), reducing predictive uncertainty by 96.1% versus single-clinician scoring. Four clinicians validated the system's outputs and expressed adoption willingness.
uncertainty quantificationmulti-expert fusiondynamic bayesian networkstroke rehabilitationaction research arm test
Reliable Conformal Prediction for Ordinal Classification Using the Ranked Probability Score
We introduce a conformal prediction (CP) method for ordinal classification based on the ranked probability score (RPS), a proper scoring rule defined over cumulative predictive distributions. This model-agnostic approach yields median-centered contiguous prediction sets by construction, supports both assessed and grouped ordered categorical outcomes, and permits efficient implementation compared to greedy interval selection procedures. Evaluated across multiple ordinal image and tabular datasets, RPS-based CP produces contiguous prediction sets and achieves a favorable balance between prediction set width and the magnitude of ordinal miscoverage relative to existing CP methods.
conformal predictionordinal classificationranked probability scorenonconformity functioncumulative predictive distributions
Swarm-Inspired Generation of Collective Behaviors in Graph Dynamical Systems
The paper introduces Swarm-Inspired Emergent Synchronizer (SIES), a graph-dynamical framework that learns generalizable local-interaction rules for collective behavior control. SIES combines agent-like dynamical units with signed source-target-conditioned attention as adaptive coupling within an explicit evolution model. The framework demonstrates generalization across network scales, target phase relations, and intrinsic dynamics in synchronization tasks, outperforms oscillator baselines in gait-related modes, and enables synchronization-driven locomotion in multi-legged robots. Additionally, SIES achieves state-of-the-art performance on heterophilous node-classification benchmarks through signed message passing.
collective behaviorgraph dynamical systemssigned attentionheterophilous graphssynchronization control
Dustin: Draft-Augmented Sparse Verification for Efficient Long-Context Generation with Speculative Decoding
Dustin introduces a sparse verification framework for efficient long-context generation via speculative decoding, addressing the KV-cache loading bottleneck. The method combines lookahead signals from the draft model with historical attention patterns to identify critical tokens, employing a sparse estimation scheme to minimize recomputation by focusing on key attention heads. Evaluations on PG-19 and LongBench using Qwen2.5-72B show 27.85x self-attention speedup and 9.17x end-to-end decoding acceleration at 32k context length, with minimal accuracy loss.
speculative decodingkv-cachesparse verificationattention headslong-context generation
Convex--Concave Quadratic Spectral Filtering for Graph Neural Networks
The paper proposes DCQ-GNN, a spectral graph neural network using adaptive convex-concave quadratic filters to improve spectral selectivity without high-order polynomial expansions. The method employs a bank of second-order filters with complementary curvature properties, fused via node-adaptive gating for structure-aware spectral selection. Theoretical analysis links filter behavior to Dirichlet energy attenuation and von Neumann entropy. Experiments on 10 datasets show DCQ-GNN achieves top average rank (3.0) on heterophilic graphs and second-best (4.2) on homophilic graphs, while demonstrating robustness to structural perturbations compared to first-order and high-order baselines.
spectral graph neural networksconvex-concave filtersdirichlet energyvon neumann entropyheterophilic graphs
Towards Continuous Power Forecasting: Practical Continual Learning for Real-World Energy Systems in Nonstationary Time Series
The paper introduces Continuous Power Forecasting, a continual learning paradigm addressing nonstationary conditions in real-world energy systems. It proposes an adaptive continual learning framework for regression, evaluating six approaches across three methodological categories under realistic data accessibility and update policies. Experiments on real-world power datasets demonstrate that continual learning enables models to self-adapt to distributional drift, accumulate knowledge over time, and mitigate catastrophic forgetting without extensive historical data storage. The study provides practical insights into the stability and adaptation behaviors of these approaches under operational constraints, offering a scalable solution for long-term deployment in dynamic environments.
continual learningnonstationary time seriesdistributional driftcatastrophic forgettingpower forecasting
Digital Twin-Driven Adaptive Sim-to-Real Alignment via Reinforcement Learning for Vibration-Based Bearing Health Monitoring Under Data Scarcity
This work introduces a reinforcement learning-driven approach for sim-to-real alignment in vibration-based bearing health monitoring under data scarcity. The method formulates feature alignment as a continuous-action Markov decision process solved via Proximal Policy Optimization, enabling fault-type-specific affine corrections while preserving inter-class separability. Validation on XJTU-SY, CWRU, and a custom testbed shows 92.8% cross-equipment accuracy without encoder retraining, demonstrating effective transferable monitoring capability.
digital twinsim-to-real alignmentproximal policy optimizationvibration-based monitoringmarkov decision process
How Complexity Contributes to Learning Opacity in Machine Learning
The paper investigates learning opacity in neural networks through complex dynamical systems theory, identifying three key properties of training complexity: sensitivity to weight initialization, feedback in gradient-based optimization, and sensitivity to training data. These properties contribute to the inherent opacity of the learning process, which remains underexplored compared to prediction opacity. The authors argue that such opacity arises from dynamical complexity and epistemological challenges, suggesting that some sources of opacity may be irreducible without fundamentally altering ML systems' learning mechanisms.
learning opacitycomplex dynamical systemsgradient-based optimizationweight initializationtraining sensitivity
Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models
The paper investigates the geometric relationship between detection and control directions in language models, challenging the assumption that perfect detection implies controllability. Using cosine similarity between optimal detection and intervention directions across Gemma 2-2B-it and three other models (1B-9B), the authors find persistent misalignment (cosine 0.12-0.20) for hallucination tasks despite perfect linear separability (AUC=1.000). A 15-degree rotation partially bridges this gap (73-60% refusal rates at 1.8% FP), but the cosine fails to predict steerability, revealing a functional rather than geometric dissociation. Results hold pre- and post-instruction tuning (0.1197 vs 0.1200), suggesting pretraining origins.
mechanistic interpretabilitycosine similaritylinear separabilityhallucination detectionsteerability
📰 Industry Media (1)
Repositioning retail for the AI era
Retail is undergoing a transformation driven by AI-first strategies that embed intelligence into core systems rather than layering it atop existing workflows. Macy’s exemplifies this shift by integrating AI into personalization, search, operational planning, and software development to reduce latency between signal and action. Early high-impact use cases, such as search recommendations and customer engagement, demonstrated measurable gains in conversion rates and operational efficiency, enabling scalable adoption. Conversational commerce tools like Ask Macy’s leverage past purchases and contextual data to provide curated recommendations. The long-term vision emphasizes continuous improvement, adaptive systems, and seamless customer experiences, positioning AI as an invisible augmentation to human judgment rather than a replacement.
ai-firstconversational commercepersonalizationoperational planningcontinuous improvement
Generated automatically at 2026-06-25 21:15 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.