Daily Digest — 2026-08-07
263 items · 4 research labs, 250 arxiv papers, 9 industry media
AI News: all feed URLs failed (last tried: https://artificialintelligence-news.com/feed/)
🏛️ Research Labs (4)
Improving GPT‑5.6 Sol in ChatGPT—and expanding access to GPT-5.6 Luna for free users
OpenAI introduces updates to ChatGPT, enhancing GPT-5.6 Sol for Plus and Pro users with improved factual reliability and a slider for response depth. Free users gain access to GPT-5.6 Luna with unlimited text chats and a Think button for complex queries. Internal evaluations show GPT-5.6 Luna and Sol reduce factual errors by 62% and 68%, respectively, compared to GPT-5.5 Instant. The update integrates Instant and Thinking experiences for consistency across conversation types. Safety measures include age-appropriate content boundaries and reinforced training for users under 18.
gpt-5.6 solgpt-5.6 lunafactual reliabilitythink buttonsafety training
Working with the American Psychological Association on youth mental health and AI
OpenAI collaborates with the American Psychological Association (APA) to integrate psychological science into responsible AI development for youth mental health. The partnership focuses on evidence-based approaches to address gaps in support, design developmentally appropriate AI tools, and equip caregivers and clinicians with practical resources. Key initiatives include family-facing guidance, clinician resources, and convening stakeholders to translate insights into actionable safeguards. This builds on OpenAI’s existing efforts, such as distress recognition in ChatGPT, localized crisis resources, and parental controls. The collaboration aims to ensure AI strengthens, rather than replaces, human relationships and support systems for young users.
psychological sciencedevelopmentally appropriatedistress recognitionparental controlsevidence-based approaches
From asking to doing: How the world is putting ChatGPT to work
OpenAI's country-level analysis reveals shifting ChatGPT usage patterns from information-seeking to task completion, based on message data from 1B+ users across Free, Go, Plus, and Pro accounts. The Economic Research Team's Signals dataset shows a 2x higher likelihood of productive use (writing, coding, analysis) in professional contexts versus exploratory queries elsewhere. Global adoption gaps are narrowing, with Latin America (Peru, Uruguay, Costa Rica), Africa, and Oceania showing the fastest per-capita growth (Q2 2026). Multimedia interactions (7.8% globally, >10% in Brazil/Colombia post-ChatGPT Images 2.0) and usage among users >35 (+10pp in France/Czechia) demonstrate expanding demographics and modalities.
in-context learningper-capita adoptionmultimodal generationtask completionusage analytics
Baseten on Hugging Face Inference Providers 🔥
Hugging Face integrates Baseten as a supported Inference Provider, enabling serverless inference for open-weight LLMs like DeepSeek-V4-Flash and GLM-5.2. The platform supports both custom API key routing and Hugging Face-managed billing, with SDK integration for Python and JavaScript. PRO users receive $2 monthly inference credits, while free-tier users have limited quotas. The system prioritizes providers based on user preferences and supports conversational and text-generation tasks initially.
serverless inferenceinference provideropen-weight llmsapi routinghugging face sdk
📜 arXiv Papers (250)
Argus: A General-Purpose Agentic Runtime for Long-Horizon Reasoning
Argus introduces a general-purpose agentic runtime for long-horizon reasoning, featuring a persistent, self-evolving architecture with Manager, Planner, Engineer, and Reviewer roles. The system separates user intent from operational objectives, using fixed model weights while evolving through runtime state and control policy. Evaluated across GPT-5.5 benchmarks, Argus achieves 78% on SWE-Bench Pro (vs. 59% for Direct Copilot) with 1.41x token efficiency, and shows 21% fewer solve-input tokens after self-evolution. It also attains 76.8% on AARRI-Bench and improves mathematical data synthesis by 28.0 points, with verified kernel optimizations and multi-day math campaigns.
agentic runtimelong-horizon reasoningself-evolutionverification-gatedfixed-weight harness
OctoLong: Mid-Training On Cross-Repository Code Contexts Enhances Long-Context Modeling
OctoLong introduces a context engineering pipeline combining AST parsing, language server backend, and package management to recursively retrieve code references, curating dependency-rich contexts up to millions of tokens. The authors train OctoLong-Instruct, a suite of long-context open LMs (600M–14B parameters), via mid-training on 50B tokens (12% OctoLong data) and instruction tuning. Evaluations against 18 SOTA models show gains in long-range retrieval, state tracking, repository-level code understanding, and API usage, despite minimal traditional context-extension data.
context engineeringast parsinglong-context lminstruction tuningrepository-level understanding
Teaching Nemotron Greek: Mining a Corpus, Adapting Retrieval, and Grounding Generation for Modern Greek across Specialist Domains
The authors adapt NVIDIA's Nemotron retrieval stack for Modern Greek, addressing its absence in multilingual retrieval benchmarks and retrieval-augmented generation (RAG) applications. They mine a corpus, generate synthetic supervision, and fine-tune retrieval models, rerankers, and readers. A Nemotron 1B embedder improves nDCG@10 from 0.362 to 0.835 after training on 65,773 Greek retrieval pairs, outperforming multilingual dense retrieval models. LoRA-tuning a Nemotron 30B-A3B mixture-of-experts reader increases answer correctness from 29.4% to 66.9%. They introduce HERA, the first large-scale Greek RAG benchmark, and release adapted models and datasets.
retrieval-augmented generationmodern greeknemotronlora-tuningndcg@10
OPD-V: Visual On-Policy Self-Distillation with Modality Balance
OPD-V introduces a visual on-policy self-distillation (OPSD) paradigm addressing modality imbalance in multimodal large language models (MLLMs). The method leverages positive and negative teachers—constructed via zoom-in images and masked images respectively—to derive modality-balance trust regions for token selection during distillation. Evaluations across 6 benchmarks, 4 MLLM backbones, and 5 post-training methods demonstrate consistent reasoning improvements with reduced training costs.
on-policy self-distillationmodality imbalancemultimodal large language modelstrust regionvisual reasoning
SSTQ:Privacy-Preserving Vector Quantization via Subsampled Stochastic TurboQuant
(No summary returned.)
Chained Recursive Language Models for Multi-Iteration Reasoning
The paper introduces Chained Recursive Language Models (Chained RLM), an inference-time architecture addressing long-context reasoning challenges in LLMs. The method involves repeatedly invoking the same model as fresh reasoning roots, each receiving the original problem, a compact summary, a blackboard, and task-specific artifacts from predecessors. This staged computation allows intermediate artifacts to be inspected, corrected, and extended, mitigating error propagation. The authors detail the system model, handoff mechanism, artifact workspace, and evaluation protocol, demonstrating measurable accuracy gains over direct LLM answering with recursive tool-calling.
chained recursive language modelsinference-time architecturelong-context reasoningartifact workspacerecursive tool-calling
Robust and Efficient Motion Reasoning for Privacy-Aware Classroom Incident Recognition
The study introduces a privacy-aware, computationally efficient framework for classroom incident recognition from CCTV-style observations, addressing the lack of specialized benchmarks and methods. The approach constructs hierarchical kinematic representations of human actions and employs knowledge distillation to transfer multi-order motion reasoning from a large teacher model to a compact student model. Evaluations demonstrate superior performance (outperforming larger baselines at <10% computational cost) and robustness in out-of-domain motion reasoning and synthetic-to-real generalization. A hybrid benchmark combining synthetic videos with real-world pose data will be released alongside the codebase.
privacy-aware recognitionkinematic representationsknowledge distillationsynthetic-to-real generalizationcomputational efficiency
Representational separation between unitary and channel quantum generative models via shared classical randomness at shallow depth
The work establishes a strict scalable representational separation between shallow unitary and channel quantum generative models using shared classical randomness. By augmenting bounded-connectivity shallow unitary circuits with local Pauli operations controlled by a single random bit, the channel model generates long-range correlations unattainable by unitary models at the same depth. For 1D nearest-neighbor architectures, unitary models require Ω(N) depth in the worst case to replicate these distributions. Measurement-based quantum computation naturally implements the randomness via adapted measurement outcomes, supported by numerical experiments.
quantum generative modelsshared classical randomnessunitary circuitsmeasurement-based quantum computationlong-range correlations
CoPlan: A Trustworthy Co-Intelligence Interface for Care Planning through Role-Based Contestable Argument Graphs
The paper introduces CoPlan, a co-intelligent interface for contestable AI-assisted care planning that preserves human agency through role-based argument graphs. The system employs multi-agent workflows where specialized AI agents generate interventions with supporting/challenging arguments, while human planners can modify or reject proposals before final plan generation. Demonstrated in aging-in-place scenarios, CoPlan integrates adaptive team recruitment, role-based argument review, and scheduling agents. The work contributes both a practical interface and a design framework for trustworthy human-AI collaboration in clinical decision-making.
co-intelligencecontestable aimulti-agent workflowcare planninghuman-ai collaboration
ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment
The paper proposes Answer-Backtracked Credit Assignment (ABC), a framework for training long-horizon search agents via fine-grained step-level supervision. ABC performs Answer-Backtracked Clue Recovery to identify intermediate clues from ground-truth answers, then applies Clue-Anchored Step Scoring to convert sparse trajectory-level outcomes into dense step-level rewards. The authors develop ABC-SFT (supervised fine-tuning) and ABC-GRPO (reinforcement learning) variants, training ABSeeker on Qwen3.5-4B with 8.5k examples. ABSeeker achieves 37.3-39.1% on BrowseComp benchmarks, improving to 52.9-55.3% with context management, outperforming same-scale models and matching 30B-scale performance.
credit assignmentlong-horizon searchclue recoverystep-level supervisionbrowsecomp
Hierarchical Graph Memory for LLM Agents with Path-level Localization and Rewrite
HiGram introduces a hierarchical graph memory framework for LLM agents, addressing inefficiencies in flat graph memories by organizing memories into coarse-to-fine upper-level nodes and MemoryUnits. The method employs MicroGraph-based path-level localization to identify relevant subgraphs and evidence paths before rewriting, alongside coordinated rewriting for intra-unit and inter-unit dependency updates. Evaluations on long-term conversational QA and conflict-aware memory benchmarks show improvements in answer quality (accuracy), token efficiency, and evidence selection under dynamic, static, and conditional conflicts.
hierarchical graph memorypath-level localizationmicrographcoordinated rewritingevidence selection
Item Response Theory for AI Safety
This work applies Item Response Theory (IRT) to analyze safety benchmarks for 192 language models, addressing issues of benchmark redundancy and model sandbagging. The authors fit IRT models to eight benchmarks, identifying three interpretable latent factors—refusal strictness, truthfulness, and contextual harm—that explain most variance. Psychometrically selected items achieve comparable accuracy to full benchmarks with 97-99% fewer evaluations (~10 items per benchmark) and enable detection of sandbagging or API model changes. The study demonstrates IRT as a scalable toolkit for safety evaluation, recommending adoption by frontier labs.
item response theorysafety benchmarkslanguage modelspsychometricssandbagging
Capability-Gated Planning: Cost-to-Goal Discovery and the Limits of Myopic Experiment Selection
The paper identifies a structural limitation in myopic experiment selection for automated scientific discovery, where constructive actions (e.g., acquiring epistemic capabilities) are undervalued due to bounded planning horizons. It formulates goal-directed discovery as a stochastic shortest-path problem in belief space, proving that myopic planners exhibit unbounded approximation ratios or fail to reach goals when capabilities are gated. The authors introduce CG-Plan, a replanner with a capability-aware heuristic (h = h_cap + h_exp), demonstrating in controlled experiments that performance gaps arise specifically under capability gating and persist across fixed horizons.
myopic planningepistemic capabilitystochastic shortest-pathcapability gatingapproximation ratio
MultiPathFormer: Towards a Foundation Model for Multipath Wireless Propagation
MultiPathFormer introduces a wireless foundation model that pretrains on multipath propagation sequences rather than channel tensors, using autoregressive next-path prediction with an Environmental RAG mechanism and first-path codebook. The method improves path statistics estimation (delay, power) by up to 59% by leveraging environment knowledge, and transfers to unseen users after pretraining on 27 environments. Fine-tuned on new scenarios, it outperforms channel-based foundation models, achieving 5.57 m mean localization error, 0.914 top-3 beam accuracy, 0.994 LoS classification accuracy, and 0.561 channel estimation NMSE.
wireless foundation modelmultipath propagationautoregressive pretrainingretrieval-augmented generationpath statistics estimation
VQ-VAD: Vector-quantized Motion Representation Learning for Human-centric Video Anomaly Detection
VQ-VAD introduces a human-centric video anomaly detection framework using vector-quantized motion representation learning to address limitations of continuous latent space modeling in pose-based approaches. The method adapts VQ-GAN to keypoint sequences, constructing a discrete codebook of normal motion patterns and detecting anomalies via reconstruction error. Evaluated across in-domain (81.83% on HR-SHT), cross-domain (76.69% transfer from CMU Panoptic to HR-SHT), and cross-dataset settings, VQ-VAD demonstrates robust performance on four benchmarks.
video anomaly detectionvector quantizationmotion representationpose-basedreconstruction error
Provable Limits and Certified Deferral for Verbalized Uncertainty in Small Language Models
The paper establishes theoretical limits and certified deferral procedures for verbalized uncertainty in small language models (0.5B-14B parameters). It evaluates 11 instruction-tuned models on ARC-Challenge and TruthfulQA (25,168 predictions), proving that strictly monotone calibration preserves risk-coverage frontiers, temperature scaling fails when confidence exceeds 0.5 while accuracy falls below it, and Clopper-Pearson provides finite-sample risk certificates. Empirical results show 8/22 model-task pairs hit temperature-scaling infeasibility bounds, Platt scaling reduces ECE to 0.02, but only 3 pairs achieve certified autonomy at 20% risk. The work also corrects an answer-ordering artifact in TruthfulQA.
verbalized uncertaintyrisk-controlled deferraltemperature scalingclopper-pearsoninstruction-tuned models
Hardware Design and Security in the Era of Chiplets and LLMs
This paper analyzes security challenges in modern semiconductor systems, focusing on heterogeneous 2.5D chiplet architectures and LLM-integrated EDA pipelines. It examines attack vectors across architectural, logical, and physical levels, including hardware stacks for LLM acceleration. For chiplet security, the authors propose 2.5D split manufacturing and active interposers to implement physically isolated Root of Trust architectures. For LLM-driven EDA pipelines, they identify native threats and review state-of-the-art defense techniques. The study concludes by exploring how LLM systems can enhance hardware security in chiplet-based designs.
chipleteda pipelinesroot of trustsplit manufacturingactive interposers
RepairFormer: Automated Repair of Structured Inputs Using Transformers
RepairFormer introduces a transformer-based framework for repairing structured inputs (JSON, DOT, OBJ, etc.) by formulating repair as supervised sequence generation. The method employs format tags, oracle validation, and boundary-localized repair to preserve content while generating valid outputs, focusing generation on fault regions to handle longer files. Evaluation shows 88% repair and 94% recovery rates, with 97.57% repair and 94.29% recovery on a benchmark, outperforming prior work by 5x in runtime while better preserving original content.
structured input repairsequence generationtransformerboundary-localized repairoracle validation
MarsCast: Transfer Learning of AI Weather Foundation Models to Planetary Atmospheres
The study demonstrates transfer learning of Earth's GraphCast weather foundation model to Mars, achieving state-of-the-art performance adaptation. Using the Mars Climate Database (MCD), researchers evaluated zero-shot and fine-tuned GraphCast predictions for Martian temperature and wind fields. Zero-shot forecasts initially captured current conditions but lacked diurnal variability, while fine-tuning with MCD variables and solar radiation enabled rapid learning of thermal cycles, achieving accurate 10-day forecasts within 10 epochs.
graph neural networkstransfer learningplanetary atmospheresweather forecastingfine-tuning
The Effect of Perceived Race and Gender on Police Language Use: Experimental Evidence from VR Simulations
This study investigates the impact of perceived race and gender on police language use in VR simulations, employing causal inference to measure average treatment effects (ATE) on conversational deference. Using virtual characters depicted as Black adult males, the analysis reveals that most officers exhibit less deferential language, except for White, biracial, and multiracial female officers, particularly in suspect scenarios. Marginal ATEs indicate significant tone differences (2+ points on a 0-10 scale), potentially escalating conversational breakdowns. The study also evaluates LLM-assisted ATE estimation methods, recommending mixed effects models with inverse propensity treatment weighting (iptw) for multilevel text data, while noting LLM finetuning requires further refinement.
average treatment effectvirtual realitycausal inferencelarge language modelsinverse propensity treatment weighting
Gradient Immunity: Null-Space Resistance to Malicious Fine-Tuning
The paper introduces Gradient Immunity, a defense mechanism against malicious fine-tuning in partially protected open-weight (PPOW) releases. The method employs a Unidirectional Safety Gate (USG) comprising a Null Space Cubic Layer and an Inverse Adapter, which blocks harmful gradients during fine-tuning while preserving forward behavior. Evaluated across six model-dataset settings, USG maintains attack success rates near pre-release levels (fixed threshold) and shows a clear safety-utility trade-off on BeaverTails unsafe samples. Results indicate representation-space blocking at release time can mitigate malicious adaptation without downstream cooperation.
gradient immunityunidirectional safety gatenull space cubic layerinverse adapterpartially protected open-weight
From Score Matrices to Football-Aware Match-State Simulation: An Auditable LLM Harness for Exact-Score Reranking
The paper presents an auditable hybrid architecture combining statistical football score forecasting with LLM-based contextual reasoning, documenting four iterative improvements. Methodologically, it integrates dynamic Poisson models (estimating team strength and score probabilities) with constrained LLM reasoning about tactical matchups and in-game dynamics, progressing from scalar corrections (V2) to goal-by-goal simulations (V3) and time-aware cascades (V4). On 2025-26 EPL matches, V4 achieved 14.7% Top-1 and 30.7% Top-3 exact-score accuracy (vs. V1's 10.0%/26.7%), with improved candidate coverage (84.7% vs. 77.3%) but no tail-candidate gains.
dynamic poisson modelsexact-score rerankinggoal-by-goal simulationauditable harnesscontextual reasoning
ArtAnno: Annotating Implicit Semantics in Artworks through LLM Agent-Driven Bidirectional Human-AI Augmentation
The paper proposes ArtAnno, a bidirectional human-AI augmentation (BiHAA) framework for annotating implicit semantics in artworks using LLM agent-driven interaction. The system features a Proactive Agentic Support Module for semantic mining and label suggestion, and an Interaction-Driven Evolution Module that distills annotation trajectories into reusable experience. Evaluations with 20 annotators show improved efficiency, knowledge accumulation, and reduced verification effort for non-experts. The multi-agent architecture enables real-time skill evolution through bidirectional HAI augmentation.
bidirectional human-ai augmentationimplicit semanticsllm agentannotation trajectoriesproactive agentic support
Short-term load forecasting under EU-AI Act Requirements in Safety-Critical Environments: Results from a 41-day live challenge on the aggregated German transmission-grid load
The paper presents a 41-day live challenge evaluating a short-term load forecasting (STLF) pipeline compliant with EU-AI Act requirements for safety-critical environments. The pipeline, implemented via the spotforecast2-safe Python library, predicts 24-hour aggregated German transmission-grid load using ENTSO-E data, incorporating anomaly detection, gap-aware data preparation, calendar and weather covariates, recursive multi-step forecasting, and hyperparameter tuning. It outperforms the official ENTSO-E day-ahead forecast baseline. Notably, transparent and low-cost local models (macl2l) achieve competitive performance against large pre-trained foundation models like chronos-2. The challenge infrastructure, submission history, and final leaderboard are publicly available.
short-term load forecastingeu-ai actentso-erecursive multi-step forecastinghyperparameter tuning
Revealed Rationality: Label-Free Evaluation and Regularization from Representation Theorems
The paper proposes a label-free framework for evaluating and regularizing LLMs and AI systems based on representation theorems from decision theory. By leveraging the 'if and only if' structure of these theorems, the method checks axiom compliance using synthetic choice problems without external labels or human feedback. Three instantiations are discussed: probabilistic coherence via de Finetti's theorem, preference rationality via Afriat's theorem, and subjective expected utility via Echenique and Saito's theorem, each yielding continuous penalties that are zero for rationalizable behavior. These penalties complement existing evaluation signals by exhaustively testing rationality standards.
representation theoremslabel-free evaluationaxiom complianceprobabilistic coherencepreference rationality
ORACLE: A Multi-Objective Reinforcement Learning-Based Analog Circuit Design Optimizer with Large Language Models-Guided Exploration
ORACLE introduces a multi-objective reinforcement learning framework for analog circuit design optimization, replacing scalar rewards with vector-valued learning and preference-aware conditioning. The method employs normalized-weight and cosine-aligned guidance for convergence, alongside an LLM-guided action selection mechanism to filter suboptimal designs. Evaluated on 2,000 test cases across multiple circuit topologies, ORACLE reduces runtime by 20.4x–104.4x, meets 99.9% of target specifications, and achieves 5.1x–318.6x better figure of merit compared to state-of-the-art approaches.
multi-objective optimizationreinforcement learninganalog circuit designpreference-aware conditioninglarge language models
Protoreasoning in Tiny Transformers
The paper introduces protoreasoning, a simplified Chain of Thought method enabling step-by-step reasoning in tiny transformers (~1M parameters), facilitating detailed analysis of reasoning generalization. Using Dyck languages (nested bracket sequences) as reasoning-friendly tasks, the study demonstrates that protoreasoning traces significantly improve out-of-distribution generalization. Ablations confirm the improvement stems from trace content rather than additional tokens, addressing questions about general algorithmic learning in small-scale models.
protoreasoningtiny transformerschain of thoughtdyck languagesout-of-distribution generalization
SciCode-Verified: How Benchmark Defects Underestimated the Scientific-Coding Ability of Language Models
This work identifies and corrects benchmark defects in SciCode, the standard evaluation for scientific-coding ability in language models, which had underestimated model performance. A domain-expert audit of all 65 test problems revealed 263 defects, with 192 causing correct solutions to be wrongly rejected due to issues like non-reproducible answers and overly strict tolerances. After correcting these defects to create SciCode-Verified, re-evaluation of twelve frontier models showed substantial improvements: subproblem accuracy increased from 45-60% to 84-98%, and main-problem accuracy from 9-27% to 69-92%. The results demonstrate that state-of-the-art models possess significantly stronger scientific-coding capabilities than previously indicated.
scientific-codingbenchmark defectsdomain-expert auditsubproblem accuracygold answers
WorldCycle: Self-Verifiable Reinforcement Learning for Long-Horizon Video World Models
WorldCycle introduces a self-verifiable reinforcement learning framework for long-horizon video world models, addressing compounding errors through reversible action cycles. The method constructs closed action cycles from ordinary sequences, optimizing spatial closure and temporal consistency rewards to enforce state-operator consistency rather than memorized patterns. Evaluated on CycleBench, WorldCycle reduces state-returning drift by 44% and improves composite-action accuracy by nearly 4x over base models.
reinforcement learningworld modelsaction cyclesself-verificationlong-horizon planning
A General Sufficient Condition for Rewriting Horn-ALCHI Atomic Queries into GQL
The paper establishes a sufficient condition for rewriting Horn-ALCHI ontology-mediated queries (OMQs) into GQL, addressing the challenge of evaluating atomic queries over non-first-order-rewritable ontologies. The authors introduce DL automata, a formalism capturing OMQ semantics via runs over fact sets, and identify a stratified subclass that avoids cyclic dependencies. This stratification enables rewriting into unions of conjunctive two-way regular path queries (UC2RPQs), a core GQL fragment, thus demonstrating GQL-rewritability for a broad class of Horn-ALCHI OMQs.
horn-alchidl automatagql-rewritabilityuc2rpqsontology-mediated queries
CheMLFlow: An Open-Source Platform for Cheminformatics and Materials Informatics Applications
CheMLFlow introduces an open-source platform for constructing reproducible, high-throughput workflows in cheminformatics and materials informatics. The system addresses pipeline fragmentation by providing modular components for data acquisition, model training, validation, and reporting, with deterministic splits and structured artifacts. Benchmarks demonstrate literature-matching performance on quantum mechanical, physicochemical, and bioactivity prediction tasks, including time-series applications beyond molecular datasets. The architecture supports agent-assisted experimentation through configuration-driven workflows and standardized outputs.
cheminformaticsworkflow orchestrationreproducible pipelinesmaterials informaticsagent-assisted experimentation
SVI-DAG: A Structured Variational Inference Approach to Bayesian Causal Discovery
SVI-DAG introduces a structured variational inference method for Bayesian causal discovery, leveraging normalizing flows to model edge dependencies and Stein variational gradient descent to enhance multimodal posterior learning over DAGs. The approach incorporates prior beliefs and mitigates mode-seeking behavior via acyclicity-space kernel updates. Evaluated against 5 state-of-the-art Bayesian DAG learning methods, SVI-DAG achieves superior uncertainty quantification while maintaining competitive structural accuracy.
bayesian causal discoverystructured variational inferencenormalizing flowsstein variational gradient descentdirected acyclic graphs
Consistency-Driven Co-Evolution for Self-Supervised Cross-Representation Learning
Proposes CoCoEvolve, a self-supervised framework for cross-representation learning across charts, tables, and code by optimizing consistency across modalities. The method establishes explicit one-to-one correspondences between representations, using agreement as a training signal without additional annotations. CoCoEvolve@Train performs cyclic co-evolution during training, while CoCoEvolve@Test applies consistency objectives at inference. Evaluated on six cross-representation tasks across four benchmarks, the approach improves performance in both training-time and test-time settings.
cross-representation learningconsistency optimizationself-supervised learningmultimodal alignmenttest-time adaptation
A Chain Is Only as Strong as Its Weakest Link: A Scoping Review of System Integration Audits in AI
This scoping review highlights the critical role of system integration audits in AI risk assessment, analyzing 58 studies from 4,259 documents that treat integration as a core evaluation tenet. Using reflexive thematic analysis, the study categorizes integration across three sites (inter-component, system-environment, multi-system) and identifies four functions (risk exploration, risk determination, coordination, procedural regularity). Findings reveal fragmented practices with limited measures targeting integration-specific risks and gaps in meeting traditional audit expectations. The review emphasizes compatibility, completeness, and oversight as key qualities assessed in integration audits and calls for prioritizing system integration in AI risk mitigation strategies.
system integrationai auditsrisk assessmentreflexive thematic analysissafety-critical domains
When Shared Rollouts Fail in Defensive Driving Evaluation: A NAVSIM Score Basis Audit
The article identifies a critical flaw in defensive driving evaluation metrics where shared rollout instability propagates reference failures into unwarranted compliance credit. The authors audit NAVSIM v2.2's scoring, demonstrating that route-blind and actor-blind probes erroneously outperform human replay and PDM-Closed on 12,146-token navtest splits due to numerical instability in velocity refit. Diagnostic controls (32-token set, 450-token pool) isolate the solver as the instability source, with fixes restoring proper ranking. The work proposes an audit protocol mandating score basis disclosure, blind probes, and stability tests for reliable defensive driving assessments.
defensive drivingrollout instabilityreference-conditioned forgivenessnumerical solveraudit protocol
When Does Latent Communication Pay? A Causal Audit of Relayed KV Caches in Multi-Agent LLMs
The study causally audits the efficacy of latent communication in multi-agent LLM systems by manipulating key--value (KV) caches. It replaces caches with deranged, zeroed, and moment-matched random variants across two regimes: one where the receiver requires sender-specific information, and one where it does not. Results show ceiling performance (100%) when sender information is necessary, contrasting with 23--25% for irrelevant relays. When sender information is unnecessary, equivalence is established within a 2.8-point margin across benchmarks like GSM8K, ARC-Challenge, and MedQA. The findings demonstrate that large cache effects do not necessarily imply pairing effects, and mismatched-cache audits are essential to validate latent-thought transmission.
key-value cachelatent communicationmulti-agent llmcausal auditmismatched-cache
A-SR: Self-Evolving Agentic LLMs for Symbolic Regression via Hierarchical Coordination
A-SR introduces a self-evolving agentic framework for symbolic regression, replacing unified proposal loops with hierarchical role-conditioned agents. The method coordinates formula discovery via routing protocols, evaluator-reward policies, and state-routed memory, adapting search processes without LLM parameter updates (intra-run) and distilling trajectories into role-conditioned priors (inter-run). On LLM-SRBench's LSR-Synth domains, A-SR with Llama3.1-8B improves Acc@0.01 from 25.79% to 48.30%, while A-SR-LoRA boosts Qwen3-4B from 24.58% to 38.29%. It achieves best NMSE on 7/8 metrics across real-world scientific tasks.
symbolic regressionagentic llmshierarchical coordinationrole-conditioned priorsnormalized mean squared error
Towards a satellite image manipulation and deepfake localization benchmark dataset
The authors introduce a prototype benchmark dataset for satellite image manipulation detection and localization, addressing the lack of high-quality datasets with pixel-level ground truth in remote sensing. The dataset comprises 60 images (30 manipulated via copy-paste splicing and diffusion model inpainting, 30 authentic), each with ground-truth masks and acquisition metadata. This enables pixel-level localization metrics and analysis of detection performance relative to image collection parameters. The dataset, available on Hugging Face, supports research in geospatial deepfake detection and image forensics.
satellite imagerydeepfake detectionimage forensicsdiffusion inpaintinglocalization benchmark
ContextWeave: A Real-World Workflow Benchmark
We introduce ContextWeave, a longitudinal benchmark evaluating memory-augmented language agents in realistic office workflows. The benchmark reconstructs privacy-preserved, multi-month workflows from 14 participants into 1,005 executable tasks, including 568 core evaluation tasks, with detailed instructions, containerized environments, trajectories, and task-specific rubrics. It measures workspace quality, preference alignment, and robustness to misleading recall. Experiments across six memory components and five base models show that actionable, experience-rich memory improves workspace scores from 68.08 to 78.20 and preference scores from 41.50 to 70.60, outperforming compact summaries but being more susceptible to misleading recall. These findings highlight the need for memory systems optimizing both retrieval relevance and reliable execution.
longitudinal benchmarkmemory-augmented agentscontainerized environmentstask-specific rubricsmisleading recall
Scrouting: Cost-Aware Routing of Coding Agents by Scouting the Repository First
SuperScout introduces a cost-aware routing system for repository-level code fixing, where a 7B parameter searcher (SuperScout-7B) first explores the repository to produce a sandbox-verified structured handoff, enabling a resume-based router to dispatch tasks to one of four frontier fixers without retraining. On SWE-bench Pro's Python tasks (N=266), SuperScout matches the best single model's solve rate (159 vs. 158) at ≈20% cost per solve, with the handoff—not routing—driving performance. The handoff redistributes solving ability, lifting cheaper fixers while slightly hurting the strongest (N=99, directional effects). Searcher hidden states improve cost routing, adding <$0.005 GPU time per task.
repository-level codingcost-aware routingstructured handoffresume-based routersandbox-verified
Privileged, but Biased: How PI-Conditioned Teachers Break Self-Distillation
The study investigates self-distillation (SD) as a compute-efficient alternative to reinforcement learning, where a privileged-information (PI)-conditioned teacher provides per-token supervision to a student. Across diverse tasks (question answering, mathematics, coding, tool use) and model sizes, SD fails to improve validation accuracy despite reducing loss, due to PI bias: the teacher's targets favor specific reference solutions over general correctness. This bias causes the student to penalize exploratory tokens and prioritize low-information tokens, resulting in flatter, less decisive reasoning. The authors quantify PI bias with a novel score and show SD's objective decouples from task success.
self-distillationprivileged informationpi biasper-token supervisionvalidation accuracy
Agentic Reinforcement Learning with Observation-Calibrated Self-Distillation
The paper introduces Observation-Calibrated Self-Distillation (OCSD), a method to address confounding in token-level supervision for LLM agents trained with reinforcement learning. OCSD contrasts Full and Observation-Ablated replay views to isolate privileged information from scaffold-induced score shifts, applying the resulting residual to modulate GRPO updates at high-uncertainty steps. Evaluated on ALFWorld, WebShop, and Search-QA with Qwen3 models, OCSD outperforms baselines and shows better alignment with local environment feedback.
reinforcement learningself-distillationtoken-level supervisionprivileged informationgrpo updates
RepoProbe: Benchmarking Architecture-Aware Repository Comprehension with Checklists
RepoProbe introduces a benchmark for repository-level code comprehension, addressing limitations in existing benchmarks that rely on GitHub Issues and suffer from edit bias. The benchmark uses GitHub Discussions to focus on architectural inquiries rather than defect reporting. A Checklist-Based Verification Protocol decomposes answers into atomic, verifiable facts, replacing subjective scalar scoring. Evaluation of SOTA LLMs reveals a gap between clarity and technical correctness, confirms edit bias, and demonstrates improved evaluation reliability with the proposed protocol.
repository comprehensionedit biaschecklist-based verificationgithub discussionsarchitectural analysis
IMFACT: Counterfactual Explanations for Time Series via Intrinsic Mode Function Substitution
IMFACT introduces a model-agnostic framework for generating plausible counterfactual explanations for time series classifiers by operating in the Empirical Mode Decomposition space. The method decomposes signals into Intrinsic Mode Functions (IMFs), substitutes selected IMFs with those from a Nearest Unlike Neighbour (NUN) until classification flips, and evaluates six IMF-selection strategies. On UCR benchmarks FaultDetectionA and FruitFlies, the variance-based strategy with three NUNs outperforms baselines in reliability and plausibility, while multi-NUN cycling achieves the best proximity.
counterfactual explanationsempirical mode decompositionintrinsic mode functionstime series classificationnearest unlike neighbour
NSF-HRPT: Neural Semantic Field meets Hierarchical Risk Perception Tree for Safety-Critical Scenario Assessment
The NSF-HRPT framework introduces a novel approach for quantitative risk assessment in safety-critical autonomous driving scenarios by combining learning-based perception with structured reasoning. It employs a Neural Semantic Field (NSF) to model scene semantics, trajectory predictions, and probabilistic Time-to-Collision (TTC) distributions from simulation data, coupled with a Hierarchical Risk Perception Tree (HRPT) for efficient parallel computation and spatial reasoning. Enhanced by a Sim2Real strategy incorporating foundation model priors, the framework achieves state-of-the-art performance on synthetic benchmarks and competitive results on real-world datasets for TTC estimation and risk localization.
neural semantic fieldhierarchical risk perception treetime-to-collisionsim2realrisk localization
Guideline-as-Oracle: Zero-Annotation Training of an Ophthalmic Telephone Triage Agent
The paper introduces Guideline-as-Oracle (GAO), a method for training medical triage agents without human-annotated dialogues by compiling clinical guidelines into a 70-row rule table. Eight construction strategies (e.g., cited-row tier assignment, label repair) generate 3,000 training dialogues, with supervision derived solely from rules. Fine-tuning a 9B model yields GAO-Triage, improving agreement with a 201-case reference from 61.7% to 74.1% (p=0.0046) and emergent-case recall from 9.5% to 69.0%, outperforming seven general-purpose systems. Ablations confirm the signal comes from rule-based assignment, not dialogue surface form.
zero-annotation trainingmedical triage agentrule-based supervisionlabel repairemergent-case recall
Fewer Tokens, Smaller Cache: Reward-Coordinated Efficient Reasoning
ReCo (Reward-Coordinated Compression) introduces a step-wise framework to optimize KV-cache compression and token generation in Large Reasoning Models (LRMs) by leveraging process reward. The method employs a lightweight reward estimator to adaptively compress the KV-cache (prioritizing high-reward steps), penalize redundant reflection tokens, and enable confidence-based early stopping. Evaluated on three reasoning models and six benchmarks, ReCo reduces generated tokens by 37%-65% and latency by 2.08x-2.35x while maintaining accuracy compared to Full Chain-of-Thought reasoning.
kv-cache compressionchain-of-thoughtprocess rewardearly stoppingreasoning models
FUSEP: A Multi-Center Benchmark for Diverse Tasks in Early Pregnancy Fetal Ultrasound Screening
The study introduces FUSEP, the first publicly available benchmark dataset for fetal ultrasound screening in early pregnancy, comprising 4,017 images (45,820 box-level annotations) across CRL and NT views from three hospitals. The dataset includes 14 annotated anatomical structures, diverse acquisition conditions (sonographers, devices, hospitals), and evaluates four learning paradigms: semi-supervised, fully supervised, UDA, and source-free UDA for multi-object detection. Baseline results demonstrate applicability to standard plane recognition, quality control, and automated diagnostics in early fetal screening.
fetal ultrasoundmulti-object detectiondomain adaptationsemi-supervised learningmedical imaging
Explicit Language Memory for Long-Horizon Planning in Vision-Language-Action Models
The authors propose a hierarchical vision-language-action (VLA) architecture with explicit language memory to address long-horizon task challenges in robotic control. The method decouples high-level semantic reasoning (via a vision-language model) from low-level continuous control, using textual memory sequences with temporal logic for persistent tracking and error correction. Evaluations in simulation and sim-to-real experiments demonstrate improved success rates and interpretability on complex tasks.
vision-language-action modelslong-horizon planninglanguage memorytemporal logicsim-to-real transfer
InsightEmb: Learning Action-Intent Embeddings for Agentic Insight Retrieval
The paper introduces InsightEmb, a contrastive embedding framework for agentic insight retrieval that learns progress-oriented retrieval geometry from mathematical reasoning data. The method jointly aligns concrete situations with abstract heuristic rules and clusters reasoning trajectories with similar progress structures, without environment-specific training. Evaluations on dynamic agent tasks and a static skill-retrieval benchmark show that InsightEmb outperforms existing reasoning embedding models, demonstrating transferable state-insight matching across domains using only publicly available reasoning data.
agentic insight retrievalcontrastive embeddingprogress-oriented retrievalreasoning trajectoriesstate-insight matching
PURPOSE: Poisoning Conflict Resolution in RAG via Proxy-Fact-Grounded Updates
PURPOSE introduces a black-box poisoning attack for Retrieval-Augmented Generation (RAG) systems that minimizes conflict resolution detection by framing injections as updates rather than counter-claims. The method extracts query-related facts approximating the resolver's reference, grounds a pivot event in these facts, and ensures consistency with resolver verification while steering the generator toward the target answer. Evaluated across three QA benchmarks, five generators, and three conflict-resolution methods, PURPOSE achieves the highest attack success rate (ASR) in 35 of 45 settings, outperforming prior methods by +9.7 mean ASR points. Results demonstrate the efficacy of non-contradicting injection as a practical poisoning strategy.
retrieval-augmented generationconflict resolutionblack-box poisoningattack success ratequery-related facts
EviGraph: Evidence-Guided Autonomous Research Agents
EviGraph introduces an evidence-graph framework for autonomous research agents, replacing sequential pipelines with a typed graph (Problem, Gap, Hypothesis, Experiment, Finding, Claim nodes) that maintains real-time claim-evidence consistency. The system validates semantic alignment, detects inconsistencies, and regenerates affected subgraphs via graph checkpointing, ensuring manuscripts derive only from validated evidence chains. On ARC-Bench-ML and NanoResearch-20, EviGraph improves Claim Support Rate by 40.19% over baselines and achieves 87.73% Experimental Data Consistency, demonstrating the efficacy of explicit evidence-state maintenance.
autonomous research agentsevidence graphclaim-evidence consistencygraph checkpointingexperimental data consistency
Chain-of-Thought Monitoring Can Be Unreliable in Implicit-Influence Settings
The study introduces a benchmark comparing chain-of-thought (CoT) monitorability in explicit- versus implicit-influence settings, where prompts either directly incentivize concealment or subtly bias behavior without instruction. Evaluating seven frontier reasoning models across four task formats, the authors find CoT monitors detect 60-94% of behavior shifts under explicit influence but show 41-46 percentage-point lower detection in implicit settings, dropping to 5% with system-prompt interventions. Results indicate that explicit-influence evaluations overestimate monitorability and that deployment choices can further reduce detection without altering behavioral influence.
chain-of-thought monitoringimplicit influenceexplicit influencebehavioral biasfrontier reasoning models
Toward Integrating Adaptive Experience Replay and Online Uncertainty Estimation in Safe Actor-Critic Optimal Control
The paper proposes an integrated architecture for safe actor-critic control that jointly optimizes barrier filtering, uncertainty estimation, and experience replay. The method updates obstacle geometry via uncertainty estimates, prioritizes replay based on filter interventions and residuals, and trains the critic using executed actions. Evaluated on a 2D robot-navigation task with corrupted measurements, the integrated configuration achieved zero contacts and 100% goal attainment in extreme-stress tests (cost: 7.63±0.44, belief RMSE: 3.52±0.55 cm), outperforming ablated variants. A finite-training bound formalizes replay exposure, while a robust barrier condition specifies estimation-error requirements.
actor-critic controluncertainty estimationcontrol barrier functionexperience replayrobot navigation
When Prompts Become Pixels: Prompt-Region Grounding for Multimodal Reasoning
The paper introduces Visualized Task Semantics (VTS), a controlled intervention to evaluate multimodal large language models (MLLMs) by embedding task questions within images while keeping the underlying problem unchanged. Across six MLLMs and four benchmarks, VTS reveals a 17.8-point average accuracy drop, indicating a semantic channel gap beyond OCR. To address this, the authors propose prompt-region grounding, which aligns visual question regions with typed semantics and reconstructs their clean representations from masked views. The method improves four-benchmark VTS accuracy from 58.0 to 66.3 without compromising original interface performance or requiring OCR/region metadata at inference.
multimodal reasoningvisualized task semanticsprompt-region groundingsemantic channel gapocr
Diagnosing Tool-Selection Reasoning in LLM Agents with Canary Tools
The study introduces canary tools, a diagnostic framework for probing tool-selection reasoning in LLM agents, categorizing weaknesses into a six-type taxonomy. The method evaluates eight models across three capability tiers using 120 tasks under varying canary-density conditions, totaling 8,640 runs, with task success graded by independent judges. Results show susceptibility to canary tools decreases sharply with model capability, with Claude Opus 4.8 being the least susceptible and Llama 3.1 8B the most. The taxonomy stratifies by capability, with capability mirages affecting frontier models most reliably. The framework, canary schemas, tasks, and logs are publicly released.
canary toolstool-selection reasoningcapability tiercapability miragessusceptibility rate
What We Observe as LLM Behavior Can Be a Side-effect of Inference Backend
The study demonstrates that inference backends (e.g., HuggingFace, vLLM, Ollama) significantly influence LLM benchmark performance, challenging the assumption that benchmarks reflect model properties alone. In a fully-crossed experiment (3 models × 5 frameworks × 6 benchmarks × 4 generation modes), backend choice caused structural, model-dependent variance, contributing ~39% of observed variability under greedy decoding. Divergences were stronger on factual than social-bias benchmarks. The authors recommend disclosing backend details, versions, and generation configurations, and using deterministic decoding for cross-backend comparisons.
inference backendbenchmark variabilitydeterministic decodinggeneration hyperparametersinstruction-tuned models
A 6G Integrated Sensing and Communication Framework for Railway Intrusion Detection and Collision Prediction
The authors propose a 6G Integrated Sensing and Communication (ISAC) framework for railway intrusion detection, combining sensing and communication via Channel State Information (CSI) to detect intruders and predict collisions. Using a 3D-rendered railway environment and Sionna radio simulator, they generate 22,695 CSI matrices and develop a hybrid 3D CNN-BiLSTM model for intruder detection and trajectory estimation. The model achieves 99.57% detection accuracy and a combined MAE of 0.4240 for position, velocity, and time-to-collision prediction on synthetic data, demonstrating ISAC's potential for railway safety.
integrated sensing and communicationchannel state information3d convolutional neural networkbidirectional lstmrailway intrusion detection
Design Choices That Matter: A Functional ANOVA Analysis for Remote Sensing Multi-Label Classification
The study employs functional ANOVA (fANOVA) to quantify contributions of design choices in multi-label classification (MLC) of remote sensing images, moving beyond model rankings. Analyzing 48 and 20 deep learning models across seven datasets, it evaluates architecture, fine-tuning, learning strategy, and initialization. Results show dataset-specific sensitivity: large-scale datasets prioritize fine-tuning and architecture, data-limited regimes rely on initialization, and intermediate regimes depend on architecture-learning strategy interactions. Hierarchical clustering reveals dataset groupings tied to scale, resolution, and label complexity.
functional anovamulti-label classificationremote sensingdeep learningdesign choices
Teaching MLLMs to Say No: Generalized Referring Expression Comprehension via Refusal Calibrated GRPO
The paper introduces Refusal-Calibrated Group Relative Policy Optimization (RC-GRPO), a reinforcement learning method to improve Multimodal Large Language Models' (MLLMs) ability to reject nonexistent objects in Generalized Referring Expression Comprehension (GREC) without degrading localization accuracy. RC-GRPO enforces "None" outputs for negative samples while penalizing over-refusal on positives, supplemented by reasoning reinforcement for interpretability. Evaluations on three GREC benchmarks show RC-GRPO achieves superior localization accuracy (exact metrics unspecified) while maintaining robust refusal capability.
generalized referring expression comprehensionmultimodal large language modelsrefusal calibrationreinforcement learninggroup relative policy optimization
Traceable LLM-Generated Hazard Scenarios for Operational Safety Analysis of Aviation Systems Using ASRS Reports
The paper introduces an AI-assisted method for generating traceable hazard scenarios in aviation safety analysis using NASA's Aviation Safety Reporting System (ASRS). The approach produces structured hypotheses and narrative scenarios with plausibility scores, leveraging large language models (LLMs) with zero-shot or few-shot prompting and optional fine-tuning. Evaluations compare model performance on validity and realism, with a hybrid variant combining evolutionary abduction and narrative generation showing improved correctness and reduced variability. Results demonstrate the impact of prompting strategies and model selection on scenario quality.
hazard scenario generationaviation safetylarge language modelsevolutionary abductionasrs reports
Personalized Federated Sparse Adaptation of Time-Series Foundation Models
The paper proposes a personalized federated sparse adaptation framework for time-series foundation models (TSFMs) in building energy forecasting, addressing data privacy and non-IID challenges. The method employs a heterogeneous temporal mixture-of-experts (MoE) adapter with a sequence-level router that selects top-$k$ experts for specialized temporal patterns (e.g., periodicity, trend-residual). Evaluated across 50 buildings and three TSFM backbones, personalized federated adaptation outperforms global FL-MoE and local MoE, with routing analysis revealing client-level expert specialization and backbone-dependent optimal strategies.
federated learningtime-series foundation modelsmixture-of-expertssparse adaptationnon-iid data
Active-SWE: Benchmarking Coding Agents for Proactive Bug Fixing without Issue Reports
The paper introduces Active-SWE, a benchmark for evaluating coding agents' ability to proactively discover and fix bugs without issue reports, addressing a gap in existing reactive bug-fixing benchmarks. The benchmark comprises 1,663 tasks across six bug categories and eight languages, employing a difficulty-aware task formulation pipeline and dual-track evaluation framework. Experiments show state-of-the-art coding agents struggle with proactive bug-fixing, particularly in locating/resolving recorded bugs, handling multiple bugs, and discovering potential bugs.
coding agentsproactive bug-fixingsoftware engineering benchmarksmulti-language evaluationdifficulty-aware formulation
Easy to Complete, Hard to Choose: Investigating LLM Performance on the ProverbIT Benchmark
The paper introduces ProverbIT, a 100-item Italian multiple-choice benchmark evaluating LLMs' proverb comprehension through completion and selection tasks. Testing 13 models (including Large Reasoning Models) reveals a performance gap: while models successfully complete proverbs (demonstrating memorization), accuracy drops sharply in multiple-choice tasks without correct answers, with LRMs showing literal synonym bias and failure to detect absent correct options in Chain-of-Thought analysis. Results suggest LLMs rely on surface-level pattern matching rather than deep semantic understanding of culturally embedded figurative language.
large language modelscultural linguisticschain-of-thoughtfigurative languagemultiple-choice evaluation
Calibrating Artificial Guilt: Neurally Grounded Reward Shaping for Prosocial Multi-Agent Reinforcement Learning
The study proposes neurally calibrated guilt signals for prosocial multi-agent reinforcement learning, derived from human fMRI data. Using the SoDec responsibility dataset (N=40), the authors fit a subject-fixed-effects regression to quantify guilt ($\hat{w}=1.118$, Cohen's $d=0.214$) and embed this weight in a two-agent Social Lottery environment. Evaluated over 1,000 episodes, neurally calibrated Proximal Policy Optimization agents closely matched human safe-choice rates (0.459 vs. 0.484; KL=0.0012), outperforming uniform, selfish, and oracle baselines by 1-3 orders of magnitude in KL divergence.
prosocial reinforcement learningneural calibrationmulti-agent systemsreward shapingfmri dataset
CSGen: A Multi-Domain Curvilinear Structure Generation Model via Hierarchical Multimodal Diffusion
The paper introduces CSGen, a hierarchical multimodal diffusion model for generating high-fidelity images with precise curvilinear structures. The method employs three innovations: a multi-domain dataset (24K samples, 5 domains, 7 annotation types), hierarchical progressive control to decouple topology from visual context, and sparsity-aware loss re-weighting for thin structures. Experiments show CSGen improves structure accuracy, visual realism, and downstream segmentation performance while maintaining robustness across diverse prompts.
curvilinear structuremultimodal diffusionhierarchical controlsparsity-aware losstopology decoupling
DisMix: Order-Aware Mixup for Medical Imaging via Disentangling Ordinal and Non-Ordinal Features
DisMix introduces an order-aware mixup framework for ordinal classification in medical imaging, addressing the limitation of standard mixup which blends ordinal and non-ordinal features indiscriminately. The method employs a dual-codebook VQ-VAE to disentangle ordinal (disease-severity) and non-ordinal (appearance-level) features, enabling independent interpolation of ordinal codes for meaningful intermediate ranks and variation of non-ordinal codes for appearance diversity. Evaluated on four medical imaging datasets, DisMix outperforms six image mixup baselines across six ordinal classifiers, demonstrating robustness under data scarcity and clinical grading variability.
ordinal classificationimage mixupvq-vaedata augmentationmedical imaging
AI Literacy for Legal Translation: Developing Digital Resilience
The chapter proposes an AI literacy framework for legal translation, addressing linguistic, technical, legal, ethical, and cognitive risks introduced by generative AI. The framework integrates four dimensions—foundational, procedural, critical, and strategic—to extend professional competence and foster digital resilience. It conceptualizes AI literacy as a transversal component of legal translation competence, emphasizing critical and responsible AI integration. Pedagogical implications are explored through classroom activities designed to develop AI literacy in legal translator education, aligning with professional standards.
generative ailegal translationai literacydigital resilienceprofessional competence
A/B Agent: A Self-Evolving Agent for Strategy Iteration in Industrial A/B Testing
The A/B Agent introduces a self-evolving framework for industrial recommendation strategy iteration, addressing limitations in traditional manual tuning and flat RAG approaches. It organizes historical strategies hierarchically via a Tree-RAG, autonomously generates target-aware strategies, and self-evolves through experiment-guided feedback. Evaluations show a 4.829% GMV improvement in a short-video e-commerce system while preserving guardrail metrics.
a/b testingrecommendation systemstree-ragself-evolving agentstrategy iteration
Masked diffusion enables coherent beat tracking
The authors propose a masked diffusion approach to address incoherent beat tracking in neural networks, where current models generate invalid outputs like consecutive downbeats and erratic tempo changes. Their method introduces three modifications: independent masking of beats and downbeats, a balanced masking scheduler, and peak-picking across inference steps, enabling coherent predictions through iterative inference. The approach reduces erratic behaviors and improves performance by modeling multiple plausible output beat grids.
masked diffusionbeat trackingiterative inferencetempo changesdownbeats
Agreement Before Diversity: Verification-First Complementarity for Heterogeneous Language-Model Coordination
The paper introduces Agreement-Before-Diversity (ABD), a verification-first method for heterogeneous language-model ensembles that decouples candidate diversity from replacement authority. ABD uses a frozen, label-free decision rule: an anchor answer is retained if two trusted samples corroborate it under a fixed equivalence relation; otherwise, it is replaced by heterogeneous synthesis. The method provably bounds accuracy gaps relative to unconditional synthesis or no synthesis, without independence or calibration assumptions. Evaluated on LiveCodeBench-v6 (n=175) and GPQA-Diamond (n=180), ABD achieves 59.43% (vs. 52.57% for Single9) and 75.00% (vs. 72.78% for controls), respectively, with error localization to enumerable protected strata.
heterogeneous ensemblesverification-firstequivalence relationlabel-free decisionerror localization
The Order Is the Guarantee: Verifier-Budgeted Code Deletion with Static-First Learned Proposals
DELSCOUT introduces a verifier-budgeted approach to redundant-code reduction via proposal scheduling, where a ranker orders single-statement deletion candidates and an execution suite accepts the first passing candidate within a finite budget. The method prioritizes deterministic shortest-first candidates (3 slots) and learned candidates (2 slots) under domain validation, improving verified-deletion coverage by 9.5% (+6.7 tasks) across nine MBPP replications with 0.5B–8B rankers. Without validation, a static-first schedule ensures non-decreasing coverage at a 4.8–62.5% verifier-call overhead. MBPP+ ablation confirms scheduling governs search while test suites define behavior preservation.
proposal schedulingverified-deletion coverageexecution-verificationstatic-first rankingredundant-code reduction
Rethinking Reservoir Pruning: A Dynamical Perspective for Echo State Networks
Proposes Dynamical Mode Pruning (DMP), a reservoir pruning method for Echo State Networks (ESNs) that ranks neurons by their contribution to dominant transition modes derived from a trajectory-averaged Jacobian Gramian, addressing redundancy in over-parameterized reservoirs. Unlike static connectivity-based approaches, DMP focuses on input-driven state transitions, removing low-impact units and retraining only the readout layer. Experiments on chaotic and real-world time-series benchmarks demonstrate maintained or improved forecasting accuracy with reduced reservoir redundancy, indicating dynamical influence as a superior pruning criterion to static structural metrics.
echo state networksreservoir pruningdynamical mode pruningjacobian gramiantemporal prediction
When Absence Is Evidence: Evaluating Completeness-Sensitive Negative Reasoning in Large Language Models
The study introduces CROWN-QA, a benchmark for evaluating completeness-sensitive negative reasoning in large language models (LLMs), comprising CROWN-Synth (controlled paired core) and CROWN-Real (real-document contrast-set). Across three LLM families, models exhibit unstable closure judgments and significant over-closure, failing to reliably distinguish Certified-Negative from Unknown responses. The primary failure mode is asymmetric: models recognize implicitly complete evidence but treat implicitly partial evidence as query-covering. Prompting redistributes errors without consistent resolution, and structured certificate elicitation traces errors to evidence-coverage mischaracterization. CROWN-Real confirms the persistence of partial-coverage asymmetry in real-document contexts, with variations in strength and error balance across models, prompts, and sources.
completeness-sensitive negative reasoningcrown-qaclosure judgmentsevidence-coverage mischaracterizationstructured certificate elicitation
Joint UAV Flight and Opportunistic Routing under Reinforcement Learning for Delay-Tolerant Networks
The study proposes JUROR, a joint optimization framework for decentralized opportunistic routing and UAV flight control in delay-tolerant networks (DTNs), addressing sparse delivery and congestion issues. The method employs a cooperative factored routing-UAV control approach under centralized training and decentralized execution (CTDE), utilizing the proximal policy optimization (PPO) framework. It formulates the problem as a factored partially observable Markov decision process with sequential motion-routing coupling and per-step team rewards. Simulation results across four traffic modes show significant performance improvements over PRoPHET and MaxProp while maintaining decentralized execution under contact-limited conditions.
delay-tolerant networksopportunistic routinguav flight controlproximal policy optimizationcentralized training decentralized execution
The First EgoCross Challenge at EgoVis 2026: Cross-Domain Egocentric Video Question Answering
The EgoCross Challenge introduces a cross-domain egocentric video question answering benchmark to test multimodal large language models' generalization beyond daily-life scenarios. The benchmark comprises first-person videos from surgery, industrial assembly, extreme sports, and animal perspectives, requiring models to select correct answers from four options. The challenge featured two tracks: Source-Limited (restricted baseline model and support set) and Open-Source (broader model choices), attracting 1,500 submissions from 130 participants. Results and winning solutions are summarized, with all resources made publicly available.
egocentric videomultimodal llmscross-domain generalizationvideo question answeringbenchmark
EASy: Towards Efficient LLM-Based Agentic System
EASy introduces a trainable agentic framework that jointly optimizes task performance and computational efficiency in LLM-based systems. It employs a milestone-plan-act workflow to decompose complex tasks, construct dependency-aware execution graphs, and parallelize independent steps, while adapting to intermediate outcomes. The orchestrator is trained using a tree-structured rollout procedure with multi-component rewards capturing task correctness, execution efficiency, and trajectory completeness. Extensive experiments on mathematical reasoning, embodied decision-making, and deep research benchmarks demonstrate that EASy achieves superior performance-efficiency trade-offs compared to existing agentic baselines.
agentic systemsmilestone-plan-actdependency-aware executiontree-structured rolloutperformance-efficiency trade-offs
Breaking the Curse ofMultilinguality inMany-to-Many Speech-to-Text Translation via a Resource-AwareMixture of Speech Encoders
We propose MSRT, a resource-aware Mixture of Speech Encoders (MoSE) framework to address the curse of multilinguality in many-to-many speech-to-text translation. MoSE employs an explicit language router to assign utterances to expert encoders, combining a frozen expert for high-resource languages with trainable experts for medium- and low-resource languages. A five-stage curriculum learning strategy reduces data dependence to 10 hours per language. Experiments on 45 languages across all 1,980 translation directions show that our 4B-parameter model achieves state-of-the-art performance, with largest gains on low-resource speech while maintaining high-resource capabilities.
mixture of speech encoderscurse of multilingualitylanguage routercurriculum learningspeech-to-text translation
PhysMind: From Video to Executable Worlds for Training-Free Physical Reasoning
PhysMind introduces a training-free framework for physical reasoning from video by constructing reusable executable worlds per video. It recovers dynamic scenes via object segmentation, mesh reconstruction, and 6D pose tracking, then fits analytic continuous-time dynamics and latent physical parameters without time-stepped simulation. The system answers questions by inspecting, continuing, or editing the world and analyzing resulting trajectories. Compared to direct chain-of-thought reasoning with the same vision-language model, PhysMind improves accuracy by 38.23 points on CLEVRER and 8.08 points on Physion++, and outperforms GPT-5.5 by 19.25 points on counterfactual questions.
physical reasoningexecutable worlds6d pose trackingcontinuous-time dynamicscounterfactual questions
Breadcrumbing Search Agents
The paper introduces Authority-Chain Hijack (ACH), a strategy-driven attack on LLM-based search agents that coordinates evidence manipulation across multiple queries to bypass existing defenses. Under a constrained tool-intermediary threat model, ACH leverages mediated search interfaces to steer agent evidence-gathering, achieving 55.9% ASR and 83.3% MaxN ASR on SafeSearch. The authors also propose Trace-Guided Strategy Evolution (TGSE), an automated refinement method that improves attack strategies from execution traces, reaching 71.4% ASR and 95.0% MaxN ASR in held-out evaluation.
llm-based search agentsauthority-chain hijacktrace-guided strategy evolutionprompt injectiongoal hijacking
What Is a Skill Worth? Structure-Aware Shapley Valuation of Agent Skills
The paper introduces SkillSV, a structure-aware Shapley-style framework for valuing internal units (e.g., rules, scripts) within agent skills by accounting for dependencies, hierarchy, and context constraints. The method compiles skills into units with dependencies, uses paired deletion and length-neutral padding to isolate content value from context cost, and employs a rollout-budgeted estimator for noisy evaluations. Evaluations on four agentic benchmarks demonstrate SkillSV's ability to recover unit interactions, preserve aggregate skill lift, and guide pruning and compression.
skill valuationshapley valueagent skillsstructure-awarecounterfactual evaluation
EuroExec: Frontier Language Models Fall Short of Expert Judgment on European Executive Decision Tasks
The study introduces EuroExec, a benchmark of 413 expert-authored European executive decision tasks, to evaluate frontier LLMs on open-ended complex questions. Using 4,000+ expert hours, six models were assessed via multi-attribute rubrics, item-specific checklists, and preference rankings, yielding an aggregate Solve Rate. The top model achieved only 56.9% Solve Rate, while expert references scored near-ceiling and were preferred in 74% of rankings, highlighting LLMs' subpar performance on professional-standard tasks. Human evaluation proved essential, as automatic metrics failed on subjective, real-world problems.
euroexecsolve rateopen-ended taskshuman evaluationexpert benchmark
A Model Merging Approach for Continual MLLM Unlearning
The paper introduces Merging for Continual Unlearning (MCU), a method for continual unlearning in multimodal large language models (MLLMs) that dynamically merges one-shot unlearning adapters to mitigate cumulative utility degradation and unlearning rebound. MCU projects adapters into a shared representation space, preserves dominant directions, suppresses over-concentrated coordinates, and reconfigures cross-task dependencies to balance transferability and interference. Evaluations on ICU-Bench and MLLMU-Bench show MCU achieves superior unlearning effectiveness while preserving retained knowledge and general multimodal utility.
multimodal large language modelcontinual unlearningadapter mergingcross-task dependenciesutility degradation
Leak-Resistant Unlearning: A New Benchmark for Evaluating Multi-Hop Reasoning Consistency and Recovery Robustness
The paper introduces Leak-Resistant Unlearning, a novel benchmark for evaluating robust knowledge removal in large language models (LLMs) across diverse multi-hop reasoning paths and recovery attacks. The benchmark addresses two challenges: knowledge leakage through interconnected reasoning paths and the fragility of unlearning against recovery attacks. Experiments conducted on 3 models, 6 unlearning methods, and 2 datasets reveal that existing methods are vulnerable to multi-hop reasoning and recovery attacks. The study further explores the trade-offs between forget quality, robustness, and model utility in LLM unlearning.
machine unlearningmulti-hop reasoningrecovery attacksknowledge leakagemodel utility
CARVE: Cross-Slice Anisotropic Reallocation of Visual Evidence for Efficient 3D Medical Volume Understanding
CARVE introduces a training-free token compression framework for 3D medical volume understanding, addressing inefficiencies in slice-based MLLMs where redundant visual tokens from adjacent slices burden LLM backbones. The method partitions the depth axis into windows, non-uniformly allocates tokens via cross-slice evidence normalization, and merges eligible tokens into spatial anchors. On Hulu-Med-7B, CARVE removes 80% of visual tokens while outperforming baselines, achieving 6.2 points higher quality retention than the strongest baseline on AMOS-MM report generation and preserving 98.1% of full-token performance across three VQA benchmarks.
token compression3d medical volumecross-slice evidencespatial anchorsvqa benchmarks
GUARD: Grounding Uncertainty and Ablation-Based Risk Detection for Diffusion-Based VLAs
The paper introduces GUARD, a test-time failure detection method for diffusion-based vision-language-action (VLA) policies that assesses grounding uncertainty without modifying pretrained models. GUARD analyzes key-value (KV) cache entries, constructs counterfactual caches via ablation, and compares denoising responses to derive diagnostic metrics (sensitivity, attention entropy, modality bias, grounding efficiency). Evaluated on Pi0, SmolVLA, and Alpamayo-1.5 across LIBERO, SimplerEnv, MetaWorld, and PhysicalAI-AV, GUARD achieves top ROC-AUC in 4/5 unseen-task settings, improving average performance by 5.73 percentage points over baselines while maintaining seen-task accuracy within 0.19 points.
diffusion-based vlakv-cache ablationgrounding uncertaintyruntime monitoringmultimodal failure detection
CARGO-VL: Counterfactual Arbitration with Risk-Constrained Group Optimization for Vision-Language Models
CARGO-VL introduces a group-relative optimization framework for vision-language models, addressing evidence conflicts by jointly optimizing matched variants across aligned, image-correct, text-correct, and both-wrong (A/V/T/N) states. The method combines condition-wise correctness with transition rewards for answer invariance, source equivariance, and abstention switching, using a primal-dual controller to balance unsafe answers against excessive deferral. Evaluated on XMC, CMC-Bench, and Modality-Bias, CARGO-VL improves conflict handling (+12.3% over baselines), unsupported-answer avoidance, and modality balance, with ablations confirming the benefits of relational transition signals and adaptive risk control.
vision-language modelscounterfactual arbitrationgroup-relative optimizationprimal-dual controlmodal conflict
GeoReward: Mitigating Contextual Variable Overestimation in Vision-Language Models for Cross-Market Preference Prediction
The paper introduces GeoReward, a reward model addressing Contextual Variable Overestimation (CVE) in vision-language models (VLMs), where dominant visual-textual cues overshadow sparse but critical contextual variables. The method combines Market-Aware Retrieval Augmentation, Context-Guided Visual Modulation, and Selective Sensitivity Loss to predict advertisement image preferences across geographic markets. Experiments on a new multimodal dataset of ad creatives show GeoReward mitigates CVE and outperforms baselines, enabling market-aware ad generation via VLM fine-tuning.
contextual variable overestimationvision-language modelsretrieval augmentationvisual modulationreward modeling
AFD-Ledger: Deployment Provisioning for Attention--FFN Disaggregation
AFD-Ledger introduces an offline analytical provisioning system for Attention--Feed-Forward Network (FFN) Disaggregation (AFD) architectures, addressing deployment optimization for Mixture-of-Experts (MoE) language models. It employs an analytical execution model and evaluation-bounded hardware search to independently provision AFD and collocated deployments, reducing complete deployment evaluations by 68.8%--83.5% while recovering globally optimal deployments. Validation on three LongCat 2.0 deployments shows AFD-Ledger predicts AFD-to-collocated throughput within 6.6%--9.6% of measured values. Findings indicate homogeneous AFD improves throughput in minority settings, heterogeneous AFD requires deployment-level hardware complementarity, and role-specific hardware improvements primarily impact deployment organization.
attention--ffn disaggregationmixture-of-expertsdeployment provisioninganalytical execution modelhardware complementarity
AudioScape-TTA: A Structured Soundscape Benchmark for Fine-Grained Text-to-Audio Evaluation
The paper introduces AudioScape-TTA, a structured benchmark for fine-grained evaluation of text-to-audio (TTA) systems. The benchmark uses modality-aware semantic structures and complexity metrics (event density, structural complexity) to assess 2,258 audio-text pairs via 25,707 binary QA rubrics, enabling granular analysis of event realization, acoustic attributes, and speech content. Evaluation of 13 TTA models reveals persistent limitations in attribute control and compositional generation, with human validation showing the rubric-based approach outperforms global similarity metrics in semantic alignment.
text-to-audiosemantic structuresevent densityrubric-based evaluationacoustic attributes
EndoVLM: An Endoscopy Vision-Language Pre-training Model via Anatomy-Guided Sparsity and Progressive Alignment
EndoVLM introduces a vision-language foundation model for endoscopic analysis, pre-trained on 348K examination pairs (images + clinical reports). Key innovations include: (1) Anatomy-Guided Sparse Pooling, using text queries to aggregate salient frames via sparse attention; (2) Progressive Semantic-Aware Alignment, modeling clinical taxonomy through structured soft targets; and (3) Semantic-Concentrated Masked Autoencoder for semantic-rich frames. The model outperforms existing foundation models and matches task-specific methods, demonstrating strong zero-shot generalization across downstream tasks.
vision-language pre-trainingsparse attentionclinical taxonomymasked autoencoderzero-shot generalization
Beyond Linear Dynamics: Neural Bilinear Dynamical Models for Time Series Forecasting
The Neural Bilinear Dynamical Model (NBDM) is proposed for nonlinear time series forecasting, addressing limitations of linear and Koopman-based approaches. NBDM employs a bilinear latent dynamical formulation, leveraging Koopman theory to lift nonlinear dynamics into a higher-dimensional latent space and incorporating a parameterized error compensation term. It explicitly integrates control inputs using auxiliary variables or learned feedback signals, with a memory-enhanced controller handling missing inputs through multiplicative interactions. Evaluations on five real-world datasets show NBDM outperforms baselines in both given-control and missing-control settings, particularly for multi-step and long-horizon forecasting.
neural bilinear dynamical modelkoopman theorylatent dynamical formulationerror compensation termmemory-enhanced controller
Tropical Algebraic Geometry for Neuronal Representations: An Arakelov-Green Measure Based Descriptor for Graph Learning
We propose a tropical algebraic geometry-based geometric prior for learning on 3D neuronal morphologies, overcoming the expressivity limitations of message-passing GNNs bounded by the 1-Weisfeiler-Lehman test. Our method applies the tropical Abel-Jacobi transform and polarization distances, augmented by cycle space transformation and quotient space construction, to embed spatial trees into the Tropical Jacobian. We avoid NP-Hard Closest Vector Problem computations via continuous relaxation on the Albanese torus, deriving exact Arakelov-Green measures from graph Laplacian inverses. This yields node-level eigenvector coordinates and graph-level eigenvalue signatures. Evaluations on BREC and 3D morphology datasets (ACT-4, JML-4, BIL-6) demonstrate expressivity beyond 1-WL and improved classification accuracy when integrated into VAEs, GNNs, and Tree-LSTMs.
tropical algebraic geometryarakelov-green measureabel-jacobi transformalbanese toruscycle space augmentation
Architectural Implications of Agentic AI Workflows
The paper presents the first architectural characterization of agentic AI workflows, identifying key mismatches with conventional server designs. Through a production study at Microsoft Azure and controlled experiments with open-source frameworks, the authors demonstrate that agentic execution exhibits fragmented resource usage across CPU-GPU boundaries, bursty demand patterns, and heterogeneous software roles. Their taxonomy explains how workflow structure (orchestration decisions, tool invocations) translates to resource demands. The proposed Agora prototype addresses these inefficiencies via dynamic CPU core harvesting, GPU memory oversubscription with prefetching, and role-aware core pooling, improving utilization by 1.4-2.3× while preserving tail latency.
agentic aigpu memory oversubscriptiontail latencyworkflow fragmentationheterogeneous provisioning
Eigenius: A Typed Knowledge-Graph DBMS with Epistemic Stratification and Institution-Mediated Reasoning
Eigenius introduces a typed knowledge-graph DBMS designed to address the scalability and reliability challenges of AI-driven research. The system integrates a dependent type theory, institution-mediated reasoning, and an immutable storage layer to enforce epistemic stratification and data provenance as structural invariants. It eliminates polystore bottlenecks through shared intermediate representations and supports both empirical justification logic and formal mathematical proofs via Lean 4. In a case study recomputing a Nature publication, Eigenius validated all 52 derived conclusions and identified four discrepancies, demonstrating its robustness for scientific reproducibility.
knowledge-graphepistemic stratificationdependent type theorypolystorejustification logic
TwinIR: Coordinated Invisible Dual-Point Attacks on Online HD Map Construction
The paper introduces TwinIR, a physical attack methodology targeting online HD map construction in autonomous driving. TwinIR addresses the cross-boundary compensation effect by jointly optimizing attack effectiveness and point sparsity, minimizing visible-spectrum changes through near-infrared illumination modeling. Evaluated on nuScenes with state-of-the-art models, TwinIR reduces mAP by 8.18-8.96 percentage points under RSA and 2.84-5.62 under ETA, while increasing unreachable-goal and unsafe-trajectory rates by 25-28 and 19-20 points respectively. Real-world validation confirms its efficacy in inducing road deformations while remaining visually inconspicuous.
online hd map constructionphysical attacknear-infrared illuminationcross-boundary compensationautonomous driving
Q-CueGraph: Query-Conditioned Visual Evidence Graphs for Multimodal Reasoning
Q-CueGraph introduces query-conditioned visual evidence graphs for multimodal reasoning, explicitly determining where to inspect images based on task conditions. It maps questions and image representations to budgeted, coordinate-level observations for a frozen reader, leveraging OCR/layout graphs for text-rich images and query-conditioned visual nodes for natural-image search. Optional utility refinement learns effective candidate crops from training-answer correctness without region-box supervision. With a frozen Qwen2.5-VL-7B reader, Q-CueGraph achieves 0.833 accuracy on V*Bench (vs. 0.696 full-image) using 19% image area and 92% of full-image ANLS on InfographicVQA with half the image area. Explicit observation proves most valuable when evidence is localizable and resolution limits full-image reading.
multimodal reasoningvisual evidence graphsquery-conditionedocr/layout graphutility refinement
When does training on downscaled images yield the same gradients?
The study analyzes gradient preservation in diffusion transformers when training on downscaled images, identifying two key terms governing signal changes: a noise-dependent term decaying at high noise levels and a σ-independent floor tied to token count. Empirical validation reveals a noise window (0.65 < σ < 0.95) for 1024→768 downscaling where gradients remain near-native. Training LoRA adapters within validated noise windows reduces training time by 14.6% while maintaining near-native weight-space performance.
diffusion transformersdownscaled traininggradient preservationlora adaptersnoise-dependent decay
D$^2$F-ReAG: Dynamic Decomposition and Filtering for Multi-Hop Reasoning-Augmented Generation
The authors propose D2F-ReAG, a dynamic decomposition and filtering method for multi-hop reasoning-augmented generation in large language models (LLMs). The approach adaptively controls reasoning depth by assessing root-level reasoning reliability: if reliable, it generates answers directly; otherwise, it decomposes questions into sub-questions and refines root reasoning using verified sub-question derivations. Evaluated on three multi-hop benchmarks, D2F-ReAG outperforms existing methods like graph-structured RAG and question decomposition by improving both efficiency and accuracy in cross-document reasoning tasks.
retrieval-augmented generationmulti-hop reasoningdynamic decompositionroot-level reliabilityreasoning refinement
ExeCRE: Execution-Consistency Guided Reliability Estimation for Self-Correcting Code Generation
ExeCRE introduces an Execution-Consistency guided Reliability Estimation framework to improve self-correcting code generation pipelines. The method statistically analyzes consistency patterns in execution outputs across randomly generated inputs, projects these into consistency signals, and applies the Dawid-Skene model to infer latent code reliability. Integrated into self-correction pipelines, ExeCRE reduces misleading feedback cases from 113.2 to 14.0 on LiveCodeBench under GPT-5.2, enhancing both effectiveness and stability. The framework also demonstrates similar benefits in code-based mathematical reasoning, suggesting broader applicability for reliable execution-based pipelines.
execution-consistencyreliability estimationself-correctiondawid-skene modellivecodebench
MERaLiON-GR: Speech Gender Recognition Model for English and SEA Languages
MERaLiON-GR introduces a speech gender recognition system for English and Southeast Asian languages, achieving state-of-the-art performance through parameter-efficient fine-tuning of a conformer-based transformer. The model combines a pre-trained MERaLiON-SpeechEncoder-2 with LoRA adaptation, a multi-scale ECAPA-TDNN downstream network, and attention pooling. Evaluations across eight languages (English, Chinese, Malay, Tamil, Thai, Vietnamese, Indonesian, Khmer) show superior accuracy over Vox-Profile and Audio-LLM in both full-utterance and segment-level tasks, demonstrating robust cross-lingual generalization.
gender recognitionconformerloraecapa-tdnncross-lingual
Generative Optimization for Incentivized Advertising with Global Level Constraints
The paper introduces GOAL, a generative framework for constraint-aware incentive allocation in advertising, formulated as conditional sequence generation with hierarchical causal state encoding to capture user dynamics. It proposes Safe Constrained Policy Optimization (SCPO) to enforce ROI constraints without retraining. Evaluations on real-world and synthetic fatigue-aware environments demonstrate GOAL's improvements in long-term revenue (15.2%) and user retention (8.7%) while reducing ROI violations by 32.4% compared to baselines.
generative optimizationincentivized advertisinghierarchical causal encoderconstrained policy optimizationnon-markovian dynamics
SPOT: Sparse Probing and Outcome Calibration for On-Policy Distillation
SPOT introduces a novel on-policy distillation (OPD) framework addressing limitations in standard reverse-KL training by jointly optimizing where to probe and what to distill. The method employs a three-phase procedure: acquisition combines teacher entropy, top-k probability mass, and student-teacher mismatch to allocate probing budget; exploration evaluates teacher-proposed candidates via verifier-scored student continuations; exploitation produces KL-regularized targets favoring candidates with better downstream outcomes. Experiments across multiple student models and reasoning benchmarks demonstrate SPOT's effectiveness in improving reasoning performance while balancing solution quality and coverage.
on-policy distillationreverse-kl trainingsparse probingoutcome calibrationkl-regularized targets
Not Every Divergence Should Be Suppressed: Counterfactual Recoverability in On-Policy Distillation
The paper introduces counterfactual recoverability as a decision variable for on-policy distillation (OPD), categorizing erroneous trajectory prefixes into recoverable, irreversible-but-avoidable, or ambiguous states based on teacher-continuation and rollback branches. This recoverability-aware control method achieves a success rate of 0.578 on held-out AIME2025, outperforming baselines (0.517), and improves AIME2024-2025 average@32 from 0.2656 to 0.3125. Diagnostic metrics show a mean continuation-minus-rollback effect of 0.185 for recoverable states and -1.000 for irreversible-but-avoidable states, with a recoverability proxy AUC of 1.000 versus 0.392 for divergence alone.
on-policy distillationcounterfactual recoverabilitytrajectory supervisionbranch diagnosticsselective supervision
MESH: Memory-Efficient Sinkhorn Optimization for Mixture-of-Experts Training
MESH introduces a memory-efficient Sinkhorn optimization method for Mixture-of-Experts (MoE) training, addressing failures in direct Sinkhorn application due to conditional, temporally varying gradients in routed expert matrices. The hybrid approach combines hidden-momentum Sinkhorn updates with optional block-preconditioning, restoring temporal first-moment signals without storing full optimizer state. Evaluated on a 110M-parameter DeepSeek-style MoE, MESH reduces optimizer-state memory by 62.5% and peak CUDA allocation by 12.6% versus AdamW, with modest evaluation-loss gaps (3.8265 vs. AdamW's 3.58–3.64). Ablations confirm temporal smoothing as critical, while block preconditioning offers optional memory-quality trade-offs.
mixture-of-expertssinkhorn optimizationmemory-efficient traininggradient preconditioningtransformer optimization
Training-Free Hashing-Based Attention via Binary Principal Components
BinaryPC introduces a training-free, hashing-based sparse attention mechanism for long-context LLMs that avoids accuracy degradation and computational overhead. The method constructs binary hash codes via binary principal components, preserving data structure without gradient-based training, unlike Locality-Sensitive Hashing or learned hashing. Evaluations across model families and benchmarks show BinaryPC matches full attention accuracy while outperforming sparse and hashing baselines, achieving 3.56× decoding throughput over FlashAttention on GPUs.
sparse attentionbinary principal componentskey-value cachelocality-sensitive hashingdecoding throughput
Approximate Multi-Objective Search Under Rulebooks
We introduce RA*pex, a best-first search algorithm for computing epsilon-approximate rulebook-optimal solutions in multi-objective robotic planning. RA*pex employs epsilon-rule-dominance, a novel approximate dominance concept under rulebooks, and leverages dimensionality reduction while maintaining separate closed sets and performing dominance checks over truncated and residual rule sets to respect rule hierarchies. Formal analysis proves that every rulebook-optimal solution is epsilon-rule-dominated by at least one solution in RA*pex's output. Empirical evaluation shows RA*pex achieves computation times over 100× faster than existing methods.
epsilon-rule-dominancerulebook-optimaldimensionality reductionbest-first searchmulti-objective planning
Improving Auto-Design of Neural PDE Solvers with a Domain-Specific Language
ADSL-PDE introduces a domain-specific language (DSL) to improve neural PDE solver auto-design by structuring the search space between conceptual decisions and executable code. The method uses a deterministic compiler to map valid search states—encoding architecture, constraints, objectives, sampling, and optimization—to executable solvers, eliminating invalid programs and increasing meaningful candidate density. An evolutionary agent refines designs iteratively using empirical feedback. Evaluations show a 52% improvement in search efficiency within ten iterations, demonstrating that structured search representations enhance LLM-driven auto-design more effectively than raw code generation.
neural pde solversdomain-specific languagesearch-space representationevolutionary optimizationauto-design
NodeJEPA: Structure-Conditioned Latent Prediction for Node-Level Graph Self-Supervised Learning
NodeJEPA introduces a joint-embedding predictive architecture for node-level graph self-supervised learning, addressing limitations of contrastive and generative methods that entangle representations with low-level input statistics. The method masks structure-aware k-hop ego-subgraphs, trains a context encoder to predict latent representations of masked nodes using an EMA-updated target encoder, and integrates spectral and centrality descriptors via a structure-conditioned predictor with cross-attention. Regularization techniques stabilize embedding geometry, and an optional curriculum increases masking difficulty. Evaluations on node classification benchmarks under linear probing and fine-tuning protocols demonstrate effectiveness, with ablations clarifying structural conditioning's role in representation learning.
joint-embedding predictive architecturenode-level graph learninglatent predictionstructure-conditioned predictorspectral regularization
Image Classification Using CNN-QNN Hybrid Model with Optimized Correlated Features
A CNN-QNN hybrid model improves image classification by optimizing feature correlations to match quantum entanglement structures, unlike prior orthogonal decomposition approaches. The method introduces a correlation-regularization term on CNN outputs to drive off-diagonal entries of the feature correlation matrix toward a target constant (0.5, per Monte Carlo simulations). Evaluated on CIFAR-10 (automobile vs. truck), Fashion-MNIST (shirt vs. coat), and radar micro-Doppler signatures (robotic dogs vs. non-robots), intermediate correlation consistently enhanced accuracy and reduced variance, demonstrating alignment with QNN's entanglement advantages without circuit modification.
quantum neural networkfeature correlationconvolutional neural networkquantum entanglementcorrelation-regularization
Towards Trustworthy Hypergraph Neural Networks under Label Noise
The paper introduces HyperTrust, a robust framework for hypergraph neural networks (HGNNs) under label noise, addressing the underexplored challenge of noisy-label learning on hypergraphs. HyperTrust employs a pretraining-based entropy-aware strategy to estimate hyperedge trustworthiness, coupled with HyperedgeBoost and HyperedgePrune modules to enhance reliable supervision and suppress noisy propagation by adjusting hypergraph structure. Evaluated on multiple datasets under various noise settings, HyperTrust demonstrates superior robustness compared to adapted LLN and GLN methods, establishing a unified benchmark for future research in noisy hypergraph learning.
hypergraph neural networkslabel noiserobust learninghyperedge trustworthinessnoisy-label learning
FinReportBench: Measuring and Improving Institution-Grade Financial Report Generation
FinReportBench introduces a benchmark for institution-grade financial report generation, addressing gaps in report identity, institutional components, source discipline, and visual delivery. Derived from expert partial orders and multimodal evidence, a 35-item rubric evaluates deliverability, report identity, and institutional completeness. The benchmark includes 244 bilingual tasks across three research objects and two input tiers, validated by three independent judge families. Across nine model families, basic deliverability is nearly saturated, but report identity and institutional completeness remain bottlenecks. Benchmark-guided skill distillation improves mean G1 by 33.85 points and mean G2 by 13.83 points over no-skill runs.
financial report generationexpert partial ordersinstitutional completenessskill distillationmultimodal evidence
Combating Knowledge Corruption in Agent Systems: A Byzantine-Tolerant Secure Collaborative RAG Framework
SecureCollaRAG introduces a Byzantine-tolerant collaborative retrieval-augmented generation (RAG) framework to mitigate knowledge corruption attacks in agent systems. The method employs a multi-source knowledge validation mechanism with dynamic graph neural network (GNN)-based credibility scoring to verify document provenance, preventing adversarial poisoning while preserving domain knowledge. Evaluations demonstrate robustness against non-IID data distributions and stealthy attacks.
retrieval-augmented generationknowledge corruptionbyzantine-tolerantgraph neural networknon-iid
NeuMoSync: End-to-End Neuromodulatory Control for Plasticity and Adaptability in Continual Learning
NeuMoSync introduces a novel architecture for continual learning (CL) that enhances plasticity and adaptability by integrating neuron-specific neuromodulation inspired by biological mechanisms. The method extends standard neural networks with learnable feature vectors per neuron and a higher-level module that synthesizes neuron-specific signals based on current inputs and network state, dynamically regulating activation and synaptic plasticity. Evaluated on diverse CL benchmarks—including Random Label CIFAR-10, Shuffle Mini-ImageNet, and Class Split ImageNet—NeuMoSync outperforms existing methods in forward and backward adaptation, retaining plasticity effectively. Ablation studies confirm the necessity of its components, and learned modulatory signals exhibit interpretable task coordination patterns.
neuromodulationplasticitycontinual learningsynaptic plasticityactivation dynamics
HyPASE: Hyperbolic Geometry for Parameter-Efficient Speech Emotion Fine-Tuning Framework for Large Audio-Language Models
HyPASE introduces a hyperbolic Parameter-Efficient Fine-Tuning (PEFT) framework for adapting Large Audio-Language Models (LALMs) to Speech Emotion Recognition (SER). The method employs a Poincare ball model with a Hyperbolic Geometric Adapter (HGA) for layer-adaptive weight modulation and an Emotion-aware Multi-capacity Cross-modal Aggregator (EMCA) for multi-scale feature compression. Evaluations on MELD and IEMOCAP show HyPASE outperforms Euclidean PEFT baselines, with notable gains in Unweighted Accuracy for class-imbalanced recognition and robust zero-shot generalization, while maintaining parameter efficiency.
hyperbolic geometryparameter-efficient fine-tuningspeech emotion recognitionpoincare ball modelcross-modal aggregation
iStructTab: Structured Feature Sequencing for Multimodal Learning of Image and Tabular Data
The paper introduces Graph-Enhanced Descriptor Sequencing (GEDS), a structured feature sequencing algorithm for multimodal learning of image and tabular data. GEDS refines statistical descriptors via similarity graph computations, solving a Column Permutation Problem (CPP) to determine optimal feature ordering, integrated with order-aware memory tokens in a transformer framework using a dedicated loss. Evaluations on multimodal benchmarks show iStructTab reduces feature dispersion, enhancing predictive performance and robustness, demonstrating the value of structured sequencing in multimodal learning.
graph-enhanced descriptor sequencingcolumn permutation problemmultimodal learningfeature sequencingorder-aware transformer
Equitable System-Prompt Selection via Constrained Mixed-Strategy GroupDRO
The paper proposes a constrained mixed-strategy GroupDRO framework for equitable system-prompt selection in large language models (LLMs), minimizing worst-case information-quality loss across metrics and groups while preserving average performance. The method assigns weights to prompts in a pre-existing pool, enabling ensemble-based mitigation without direct prompt optimization. Evaluated on five LLMs across bilingual medical and consumer-finance benchmarks, it reduces Overall Mean, Worst 25% Mean, and Worst losses by 13.1%, 13.2%, and 13.7% on average compared to no mitigation, maintaining overall quality near average-based selection. Multi-prompt weighting reveals complementary metric-group interactions.
group distributionally robust optimizationsystem promptslarge language modelsensemble weightingworst-case optimization
COMPAS: Difficulty-Aware Joint Search for Optimizing Code Generation
COMPAS introduces a difficulty-aware joint optimization method for code generation, addressing limitations of existing approaches that tune only subsets of model, prompt, and decoding parameters. The method learns group-specific quality-cost fronts via low-cost model selection and joint prompt-decoding search, then routes test tasks to matching fronts without additional search. On LiveCodeBench, COMPAS improves pass@1 from 45.9% to 52.8% while reducing cost from $36.57 to $4.92, and resolves 76.0% of SWE-bench tasks versus 70.0% for the best baseline.
code generationjoint optimizationdifficulty-awaredecoding settingsmodel selection
ATLAS: Adaptive Topological Learning with Abstract Successors for Continual Learning
ATLAS (Adaptive Topological Learning with Abstract Successors) improves continual learning in non-stationary environments by combining a Grow When Required network with Successor Features, decoupling transition dynamics from reward signals. This hybrid approach enhances sample efficiency and mitigates catastrophic forgetting. Evaluated on spatial navigation tasks, ATLAS demonstrates near-instantaneous adaptation to new goals and positive backward transfer, outperforming standard on-policy and off-policy baselines in dynamic settings.
continual learningsuccessor featuresgrow when required networknon-stationary environmentsbackward transfer
Efficient Online Lexicographic Generalized Low-Rank Matrix Bandits
The paper introduces Lexi-LowGLM, an efficient online algorithm for generalized low-rank matrix bandits with multiple prioritized objectives. The method estimates objective-specific low-rank subspaces and performs lexicographic learning in reduced feature spaces, updating each objective-specific estimator via an online Newton step. This reduces estimator-update complexity from O(T^2) to O(T) over T rounds. The algorithm achieves a regret bound of Õ(W_i^lex√m(d1+d2)r√T) for each objective i, where r is the rank bound of parameter matrices and W_i^lex captures lexicographic trade-offs. Numerical experiments confirm the method's effectiveness and computational efficiency.
lexicographic learninglow-rank matrixonline newton stepregret boundgeneralized linear estimator
Trident : How to Break Deep Reinforcement Learning Cyber Defenses (Agentic)
Trident introduces an agentic LLM red teaming framework to evaluate DRL-based cyber defenses against adaptive threats, addressing a critical gap in robustness testing. The framework comprises a dynamic benchmark with CybORG CAGE 4 and CyberWheel sandbox servers, a dataset of 13,000 red-blue interaction trajectories, and a 'Code-as-Policy' RLVR architecture. The architecture employs a tripartite Log Summarizer--Planner--Coder design, where a trainable Planner generates attack strategies from compressed logs, and a frozen Coder translates them into executable Python policies. Empirical results show Trident reduces blue agent defensive performance by 522% compared to static baselines, uncovering emergent behaviors like decoy avoidance and adaptive state prioritization.
deep reinforcement learningred teamingrlvrcybersecuritycontextual bandit
MIDAS: Multi-LLM Iterative Data-Adaptive Summarization
MIDAS introduces a Multi-LLM Iterative Data-Adaptive Summarization framework that automates prompt adaptation for enterprise summarization tasks, eliminating manual prompt engineering. The method leverages data-driven pattern learning and use-case-specific personalization across multiple LLMs, enabling dynamic adjustment to diverse summarization requirements. Evaluated on enterprise customer ticket summarization across five output formats, MIDAS outperforms state-of-the-art frameworks like CriSPO and ZERA, achieving improvements of up to 11.0% in ROUGE-1, 18.2% in ROUGE-2, and 8.0% in ROUGE-L, while consistently enhancing BERTScore F1. Cross-model and cross-domain generalization is demonstrated through multi-LLM configurations and finance-domain benchmarks.
multi-llmprompt adaptationrougebertscoreenterprise summarization
SafeCommit: Certifying When Memory-Grounded Agents May Safely Act
SafeCommit introduces a risk-controlled layer for memory-grounded agents to prevent premature commitment to side-effectful actions under memory uncertainty. The method constructs a calibrated set of plausible latent worlds from memory, observations, tool outputs, provenance, and policy constraints, permitting an action only if a conformal action certificate ensures safety across all retained worlds. Otherwise, it selects low-side-effect probes or conservative fallbacks. Theoretical guarantees show that the probability of unsafe certified commits is bounded by the target level α, separating calibration and representation errors. A dependency-free simulator validates the safety-utility tradeoff and reproduces results.
memory-grounded agentsconformal action certificatelatent worldsside-effectful actionscalibration error
The RAIL Principles for Neurosymbolic AI: Reasoning, Assurances, Interfacing and Learning
The article proposes the RAIL principles (Reasoning, Assurances, Interfacing, Learning) as a unified framework for designing neurosymbolic AI systems, arguing that such integration is fundamental rather than niche. It demonstrates RAIL's applicability across diverse AI domains, including physics-aware ML, neuro-guided search (e.g., Alpha-*), causal learning, and tool-augmented LLMs. The framework provides practitioners with structured design guidelines for developing reliable, efficient, and trustworthy production systems by combining symbolic reasoning with data-driven learning.
neurosymbolic aisymbolic reasoningmachine learningtrustworthy systemsdesign principles
EA-Graph: Artifact-Anchored Verification Memory for Coding Agents under Upstream Drift
EA-Graph introduces an artifact-anchored memory system for coding agents to maintain verifiable claims across sessions, addressing upstream drift by anchoring claims to sub-path artifact content and distinguishing evidence strength from freshness. The method evaluates on synthetic repositories with known ground truth, testing claim classification under value/logic drift and withheld content across 42 sessions. Results show artifact-anchored memory outperformed prose notes and no memory in smaller models (p=0.0156), with perfect performance in one tier, though cross-model equivalence remains unproven.
artifact-anchored memoryupstream driftverification claimssub-path granularityevidence strength
Compass: Continuously Aligning Social Media Feeds via In-Situ Reflections
Compass introduces a system for continuously aligning social media feeds with users' reflective preferences through in-situ reflections, addressing the divergence between stated preferences and behavioral signals. The system employs lightweight notifications to prompt reflection during browsing and periodically simulates behavioral signals to manipulate feed content. Embedded within YouTube Shorts, Compass was evaluated in a 10-day field study (N=15) against a baseline lacking continuous support. Results indicate that Compass enhanced reflective and purposeful feed consumption, enabled iterative preference adjustment, and improved feed alignment while maintaining the casual browsing experience.
social mediafeed alignmentbehavioral signalsin-situ reflectionspreference adjustment
Strategic Evaluation of Planning Strategies for LLM Agents in Cyber-Physical Systems
The paper introduces a physics-grounded benchmark for evaluating LLM planning agents in cyber-physical systems, focusing on planning-induced control trajectories in a smart-grid demand-response system with 40 heterogeneous prosumers. The benchmark employs predefined, sequential, hierarchical, and search executors, bounded LLM policy declarations, and explicit code for schedule construction and power flow. Results show architecture significantly impacts outcomes, with forced search as the oracle in baseline seeds, objective substitution increasing voltage shortfall by 2.68x, and a 144-scenario bank yielding feasible oracles from three architectures. Mean regret is reduced from 90.7 to 29.0 by applying deadline feasibility, highlighting challenges in within-feasible quality selection.
planning-induced control trajectoriessmart-grid demand-responseheterogeneous prosumersvoltage shortfalldeadline feasibility
Hallucinations on the Board: Tool-Augmented Evaluation of LLM Chess Commentary
The paper introduces ACT-Eval, a tool-augmented framework for evaluating LLM-generated chess commentary by decomposing it into atomic claims verified via engine-supported tools and expert-annotated references. The benchmark includes 325 position–move pairs with 125 expert-verified gold atoms and a five-class error taxonomy. Evaluations show pervasive factual hallucinations: GPT-5.4 produces incorrect sub-claims 22.0% of the time, while smaller open-weight models exceed 40%. Tool augmentation improves factual correctness and move-quality assessment but fails to fully cover expert strategic concepts. ACT-Eval's judgments align with human inter-annotator agreement and strategic completeness assessments.
llm evaluationchess commentaryfactual hallucinationtool augmentationerror taxonomy
Interoceptive Attention as Dynamic Homeostatic Prioritization in a Foraging Agent
The paper introduces interoceptive attention as a dynamic precision allocation mechanism for homeostatic regulation in foraging agents. Using active inference, the agent reallocates a fixed perceptual budget toward its most urgent bodily need, shaping both belief updates and planning. In the AffectWorld gridworld (4-channel, 11 layouts), this method more than doubles survival rates (0.414 vs 0.199, p≤10^-4) compared to uniform allocation. Key findings show the benefit is planning-dependent (halved when denying shaped likelihood to planner) and need-aligned (worse than uniform when targeting least-needed channel). Precision routing also accelerates learning in attended channels by 2×.
interoceptive attentionactive inferencehomeostatic regulationprecision allocationforaging agent
A Unified Model for Cross-Domain Clone Detection via Model Merging
The paper introduces a unified approach for cross-domain code clone detection via model merging, addressing fragmentation in current specialized detectors that suffer >70% F1 drops across domains. It evaluates parameter merging (TIES, five task-vector methods), architecture merging (greedy layer stitching), and cross-tokenizer alignment across four code models and three benchmarks. TIES merging achieves 0.865 combined F1 on UniXcoder (93% of multi-task performance without training data) and generalizes best to AI-generated clones, while cross-base merging shows marginal gains. Merged detectors outperform zero-shot code LLMs on GPTCloneBench with lower inference cost and 4× better OOD robustness to unseen AI clones.
model mergingcross-domain clone detectiontask vectorsout-of-distribution robustnessparameter stitching
MatrAIx: Simulating the World with 8.3 Billion Persona Agents
MatrAIx introduces a population-scale simulated-user evaluation infrastructure for testing AI systems with heterogeneous users, comprising three components: Persona 8B (8.3 billion persona records with 1,290 categorical dimensions), the MatrAIx Playground (four interactive environments), and 1,010 application tasks across 25 domains. The system employs LLMs (Claude Opus 4.8, GPT 5.5, Claude Haiku 4.5) to power persona agents, validated through 18,189 trials. Controlled studies showed 91.5% adherence to declared behaviors (366/400 trials) and high extraction quality for human-grounded personas.
simulated-user evaluationpersona recordsdependency graphllm-powered agentsbehavioral adherence
Patients-like-me: A Variational LM--GNN Framework for Explainable Clinical Prediction
Proposes Patients-like-me (PLM), a variational LM--GNN framework combining local EHR semantics with global cohort structure via inter-patient relationships for clinical prediction. Introduces a Variational EM algorithm alternating LM and GNN updates under supervised variational objectives. On MIMIC-III/IV, PLM outperforms SOTA across encoder/decoder-only backbones (+1.2-3.8% accuracy) with minimal overhead, while providing explainability via reference-patient retrieval (validated by edge-masking).
electronic health recordsgraph neural networksvariational emexplainable aiclinical prediction
Behavioral Skill Reconstruction: Reconstructing Hidden Functionality from LLM Agent Skills
This work introduces behavioral skill reconstruction (BSR), a method to clone hidden functionality in closed-source LLM agent skills through ordinary interactions. The authors propose SkillClone, a black-box attack that forms an interface hypothesis, issues structured benign probes, synthesizes executable replicas, and iteratively repairs them via differential validation. Evaluated on 30 skills spanning rules, tables, procedures, and algorithms, SkillClone achieves exact or partial recovery on held-out inputs, with iterative requerying improving reconstruction gaps. Results demonstrate that file secrecy alone does not ensure functional secrecy, as cumulative information leakage from legitimate interactions enables skill cloning despite disclosure-focused defenses.
behavioral skill reconstructionskillcloneblack-box attackdifferential validationfunctional secrecy
Adversarially Robust Abductive Fusion of Pre-trained Transformer-based Perception Models
The paper introduces a domain-knowledge-free method for robustly fusing pre-trained Vision Transformer (ViT) detectors under distribution shift and adversarial attacks. By constructing Label Vector Pools (LVP) from each model's training embeddings, the approach derives geometric error-detection rules that match domain-knowledge-based metacognitive methods (within 0.002 F1). A neurosymbolic framework combines these rules via consistency-based abduction, solved either exactly (Integer Program) or heuristically (polynomial-time). Evaluated on 15 weather-shifted aerial-imagery test sets with six ViT detectors, the method matches majority voting on clean data (within 0.005 F1) and outperforms all baselines under label-flipping attacks (0.42 vs. 0.35 F1 at 90% flip rate, 22% relative gain).
vision transformerlabel vector poolsmetacognitive learningadversarial robustnessabductive reasoning
Visualizing Graph-to-Answer Mechanism Recovery in Materials-Science Hypothesis Generation
The study introduces a visual diagnostic workflow for tracing mechanism preservation in Graph-PRefLexOR-8B, a Qwen3-8B variant adapted for materials-science hypothesis generation. The method combines semantic backtracking, graph corruption, activation-based recovery measurements, and layer-by-token-region grids to inspect the model's graph-to-answer pathway (brainstorming, graph construction, pattern extraction, synthesis). Results from 100 materials-science questions show final answers align closest to synthesis stages, with mechanism recovery concentrated in late layers (30, 36) under graph corruption, while early layers (7–10) exhibit minimal recovery. The workflow aids in identifying mechanism loss or recovery before experimental planning.
graph-to-answer mechanismactivation-based recoveryresidual-stream checkpointsmaterials-science hypothesissemantic backtracking
BrainBench: Benchmarking Large Language Models for Comprehensive EEG Understanding
The authors introduce BrainBench, a unified benchmark for evaluating large language models (LLMs) on comprehensive EEG understanding through instruction-conditioned tasks. The benchmark spans four subsets (Foundational Analysis, Sleep Assessment, Neurocognitive Assessment, Physiological Integration) across 17 datasets, assessing models via numerical, categorical, and semantic validation under two paradigms: autonomous code execution (CodeAct) and structured agentic analysis (BrainAgent). Evaluations of multiple LLMs over 100K executions reveal significant performance variations across models, subsets, and difficulty levels, demonstrating EEG competence depends on model choice and operationalization. BrainBench provides a reproducible testbed for advancing LLM-based EEG analysis.
eeg understandinginstruction-conditioned tasksagentic analysisneurocognitive assessmentcodeact
TRNet: Topography-Guided Frequency Rectification and Structure-Aware Decoding for Multimodal Paddy Rice Segmentation
TRNet improves paddy rice segmentation in mountainous terrain by jointly processing 0.5-m RGB imagery and 5-m TanDEM-X DEM. The method introduces Topographic Energy-Spectral Rectification for terrain-conditioned frequency modulation and a Topography-guided Paddy Structure Decoder combining semantic, boundary, and interior cues. Evaluated on two test areas, TRNet achieved 85.10% and 80.68% IoU, outperforming Dual-Encoder U-Net by 9.15 and 18.83 percentage points, with gains attributed to frequency rectification and reduced steep-terrain false positives.
multimodal segmentationfrequency rectificationterrain-conditioned modulationstructure-aware decodingpaddy rice mapping
AgentForge: An Immersive Role-Playing Platform for Learning Agentic Software Engineering
AgentForge introduces an immersive role-playing platform where novice developers learn agentic software engineering by assuming one of four roles (Task Planner, Patch Author, Code Reviewer, or Test Runner) in a multi-agent code-repair workflow, with AI agents handling the remaining roles. The system provides role-based scaffolding and metacognitive support to clarify responsibilities, visualize agent coordination, and encourage decision evaluation. In a study with 37 novices, participants achieved high task-completion rates but faced varying interaction demands, with Code Reviewer requiring significantly more turns and time (p_adj = .004). Despite challenges, participants reported improved understanding of software repair and agent collaboration (p_adj < .001).
agentic aisoftware engineeringmulti-agent workflowmetacognitive supportcode-repair
LiNC: Lightweight Noise Correction via Per-Sample Trust and Gaussian Mixture Modeling
Lightweight Noise Correction (LiNC) introduces a per-sample trainable trust parameter to mitigate label noise in medical imaging datasets. The method combines observed labels and model predictions via convex weighting, with trust parameters learned during training. A 3-component Gaussian Mixture Model separates trust values into clean, ambiguous, and noisy cases, followed by soft and hard correction phases. Evaluated on ten 2D datasets from MedMNISTv2 with up to 50% label noise, LiNC improves accuracy and mislabel detection while maintaining negligible asymptotic overhead, with memory scaling linearly with dataset size.
label noisegaussian mixture modelmedical imagingtrust parametermislabel detection
InvFlowFD: Reference-Free and Background-Set-Free Perceptual Music Quality Metric with Flow Matching Inversion
The authors propose InvFlowFD, a reference-free and background-set-free metric for perceptual music quality assessment, eliminating the need for paired noisy-clean data or aggregated clean audio statistics. The method leverages unconditional Flow Matching inversion via Euler integration to detect artificial distortions and rank music generation models against human judgments. Evaluations against prior work, including quantitative metrics and human studies, show InvFlowFD correlates strongly with human perception of sound distortions and generative model quality while offering greater flexibility than existing metrics.
flow matchingperceptual qualitymusic generationreference-freeeuler integration
Perception Before Reasoning: Dynamic Latent Reasoning for Video Understanding and Question Answering
Dynamic Latent Reasoning (DyLaR) improves video question answering by first grounding queries in perception latents encoding visual evidence, then adaptively deciding whether to append reasoning latents for further inference. DyLaR trains perception latents on verified visual evidence, distills rationales into reasoning latents, and refines reasoning decisions via reinforcement learning. Evaluated across nine video benchmarks with four multimodal language model backbones, DyLaR increases average accuracy while reducing response length, e.g., improving Qwen3-VL-4B accuracy from 54.0 to 58.2 and decreasing tokens per query from 1,220.7 to 18.5. Ablations confirm the contributions of grounded perception, rationale-supervised reasoning, and adaptive routing.
perception latentsreasoning latentsvideo question answeringmultimodal language modelsadaptive routing
Interpretable Fuzzy Inference for UAV Target Tracking Using Bounding-Box Geometry
The paper presents an interpretable fuzzy-inference framework for UAV yaw control in target tracking, using low-dimensional YOLO bounding-box features (centroid, area, aspect ratio) without explicit geometric modeling. A Mamdani fuzzy system and a first-order Takagi–Sugeno model (27 rules derived from training-set quantiles) achieve mean absolute error of 0.140°±0.003° and 99.676%±0.270% accuracy within ±1° on 6,169 VICON motion-capture samples. The method demonstrates transparency, data efficiency, and real-time suitability for resource-constrained UAVs.
fuzzy inferenceuav guidanceyolo featurestakagi–sugenointerpretable control
Out-Of-The-Loop Multi-Fidelity Bayesian Optimization
The paper addresses suboptimality in multi-fidelity Bayesian optimization (MF-BO) when the highest-fidelity function is too expensive for active optimization loops. It proposes incorporating historical high-fidelity data with task descriptors (explicit or metadata-derived) to improve optimization. The method demonstrates effectiveness on synthetic functions and real-world chemistry and hyperparameter optimization tasks, outperforming standard MF-BO approaches under practical constraints.
multi-fidelity optimizationbayesian optimizationtask descriptorsblack-box optimizationmetadata extraction
FinPerMA: A Theory-Informed, Event-Grounded Personalized-Memory Benchmark for LLM Agents
FinPerMA introduces a theory-informed, event-grounded benchmark for evaluating personalized memory in LLM agents, focusing on longitudinal investor trajectories. The benchmark combines deterministic impact rules, controlled LLM narration, and automated quality screening, with a Post-Shock checkpoint to assess event integration. Testing 2,994 questions across 276 personas, seven frontier LLMs achieve ≤0.47 overall accuracy, with summary-based memory preserving facts but losing preference signals, leading retrieval to outperform purpose-built systems post-shock.
personalized-memoryllm agentsevent-groundedlongitudinal trajectoriespreference adaptation
FinProBench: Evaluating Financial AI Agents with Role-Grounded Rubrics Derived from Professional Deliverables
We introduce FinProBench, a benchmark for evaluating financial AI agents using Role-Grounded Rubric Construction (RGRC), a pipeline that derives rubrics from practitioner deliverables. RGRC comprises Deliverable Collection, Competency Extraction, Rubric Synthesis, and Validation, capturing tacit standards and transferring across tasks within a role. Evaluations on 57 occupations show RGRC outperforms prompt-only methods for role-specialized roles (99.1% vs. 78.0%) but matches for conventional roles (89.2% vs. 90.7%). FinProBench includes 1,723 deliverables across 57 occupations, 8 sub-industries, and 161 deliverable types, with human deliverables ranking highest (73.7 vs. 70.3, 70.2, and 69.6). Role-level rubric reuse reduces per-task effort by 6.7x.
role-grounded rubric constructionfinprobenchtacit standardsdeliverable collectionrubric synthesis
OneDayAgent: Towards a Long-Horizon Harness for Autonomous Agents
OneDayAgent introduces a long-horizon harness for autonomous agents to manage open-ended, cross-environment, and multimodal tasks by decomposing them into bounded subtasks, maintaining execution memory under context pressure, and verifying deliverables. The method addresses goal drift, state loss, and context overflow jointly, generalizing across five backend LLMs from three families without tuning. Evaluated on AgentIF-OneDay (104 tasks), OneDayAgent achieves a state-of-the-art score of 0.821 with GLM-5.2, demonstrating consistent performance despite varying execution styles induced by different models.
long-horizonautonomous agentscontext overflowgoal driftmultimodal tasks
Spatiotemporal Graph Transformer for Traffic Intelligence in Edge Computing
The authors propose a spatiotemporal graph Transformer framework for traffic forecasting in edge computing, addressing limitations of recurrent approaches in modeling long-horizon traffic evolution under non-stationary conditions. The method combines graph neural networks to capture spatial correlations among service regions with Transformer-based self-attention for learning long-range temporal dependencies. Evaluated on a real-world cellular network dataset, the framework outperforms recurrent graph-based baselines (GCN-RNN, GCN-LSTM, GCN-GRU) across multiple forecasting horizons, enabling more effective proactive resource provisioning and reduced overload risk compared to reactive strategies.
spatiotemporalgraph transformeredge computingself-attentiontraffic forecasting
Spend Bits Where Queries Look: KV Cache Vector Quantization with Attention-Preserving Transforms
NOVA-KV introduces an attention-preserving transform coding method for KV cache quantization, optimizing both keys and values via closed-form transforms derived from calibration statistics. The approach formulates distortion as attention product error, enabling MSE-optimal vector quantizers in the transform domain and equal-volume coefficient partitioning for fixed-width layouts. At 2 bits per element, NOVA-KV recovers most long-context retrieval accuracy lost by scalar quantization methods while maintaining comparable throughput.
kv cachevector quantizationattention productstransform codinglong-context retrieval
FBID: Adaptive Personalized Federated Learning for Robust Out-of-Distribution Attack Detection in IoT Networks
FBID introduces an adaptive personalized federated learning framework for robust out-of-distribution attack detection in IoT networks, addressing over-personalization via server-side control. The method employs a contextual multi-armed bandit to dynamically regulate local training intensity and a trust-based blending mechanism to balance global-local model interpolation. Evaluated on CICIoT2023 under heterogeneous and OOD settings, FBID improves client OOD detection rate by up to 7.66% and F1-score by 5.08% over baselines while enhancing robustness to unseen attacks.
personalized federated learningout-of-distribution detectioncontextual banditiot securitynon-iid data
Monte Carlo Tree Search for Table-to-Multimodal Report Generation
The paper introduces MCTS-Report, a Monte Carlo Tree Search framework for generating multimodal reports from tabular data by decomposing the task into atomic actions (chapter planning, visualization, chart generation, insight organization, narrative refinement) executed via LLM-based dynamic reasoning. The method employs a multi-dimensional reward function evaluating fact consistency (SQL-verified), chart quality, chart-text alignment, and structural completeness, with diversity penalties and precondition checks. On the MMRBench benchmark (6 domains), MCTS-Report achieves a 77.9 overall score, outperforming baselines in structural completeness, numerical accuracy, alignment, and insight novelty.
monte carlo tree searchmultimodal report generationtabular datallm reasoningreward function
The LLM Proposes, the Executive Disposes: A Self-Verifying Agent Instrument that Dissociates Commitment Drift from Binding Drift in Long-Horizon Agents
The paper introduces a self-verifying agent instrument that structurally dissociates commitment drift from binding drift in long-horizon agents. The architecture features a deterministic Executive that owns all belief states, while a language model submits typed proposals; claims are admitted only when pre-registered predictions match observations. The instrument invalidates runs upon detecting write-errors or canary breaches, with shadow references enabling drift metrics in ablation studies. Results show that ablating the commitment mechanism increases goal-abandonment from 0.00 to 1.00 while binding error remains at 0.00, demonstrating clean separation of failure modes. Task efficacy was null (zero completions on ARC-AGI-3), pre-registered as a structural limitation.
commitment driftbinding driftself-verifying agentpre-registered predictionablation cell
An Inline Control Architecture for Language Models in Intelligent Transportation Systems
Guarded-V2X introduces an inline semantic guardrail architecture for securing large language model (LLM)-enabled vehicle-to-everything (V2X) services, addressing prompt-level attack surfaces not covered by traditional V2X security mechanisms. The architecture integrates rule-based ingress filtering, a lightweight safety classifier, policy-constrained structured generation, trusted-only retrieval, and post-decision adjudication to enforce safety boundaries. Evaluated via a four-stage pipeline including intrusion vulnerability analysis and adversarial stress testing, Guarded-V2X reduces intrusion acceptance rates and eliminates unsafe completions in two-turn settings while maintaining V2X latency constraints.
semantic guardrailv2xllmadversarial stresslatency benchmarking
Toward Skill-Native LLMs: Skill Entropy for Benchmarking and Training Long-Horizon Reasoning
The authors introduce Skill Entropy, a metric quantifying the difficulty of skill transitions in long-horizon reasoning tasks, and Skill^2-Bench, a benchmark comprising 558 skills across 9 domains with task-level skill-entropy scores. They evaluate 12 models, revealing a skill-switching gap where accuracy declines on higher-entropy tasks. To address this, they propose Skill-Entropy RL, a reinforcement learning framework combining step-level correctness with skill-entropy rewards. This method improves Skill^2-Bench scores from 34.4% to 68.4% on Qwen3-4B-Instruct and from 14.6% to 40.1% on Qwen3-1.7B, demonstrating its effectiveness as a reusable training signal.
skill entropylong-horizon reasoningskill-switching gapskill-entropy rlskill^2-bench
The Loss Does Not See the Basis, but Adam Does
This work identifies gauge symmetry as a key factor differentiating optimization methods in low-rank matrix recovery. The authors prove that gauge-equivariant optimizers (e.g., gradient descent) inherit gradient flow's low-rank bias, while coordinate-wise methods (e.g., Adam) do not. Through theoretical analysis and experiments on matrix sensing and transformer models, they demonstrate that Adam's anisotropy breaks gauge symmetry, leading to distinct optimization paths even for gauge-equivalent initializations. Empirical results show gradient descent achieves 43-44% lower held-out error than Adam on hyperspectral datasets, with lower effective rank solutions. A spectral schedule is proposed to reconcile conflicting reports about Muon's performance.
gauge symmetrylow-rank recoverycoordinate-wise optimizationmatrix sensingspectral schedule
Predicting Brain Morphometry with MT-GNN: Mesh Evolution in Continuous Time with Graph-Based Metric Tensor Embeddings
MT-GNN introduces a graph-based approach for predicting subcortical brain morphometry evolution in continuous time, conditioned on a Fourier-encoded lead time and arbitrary causal multiple-visit histories. The model predicts per-vertex first fundamental forms (metric tensors) via a single per-structure graph network, decoded into surfaces using a differentiable As-Rigid-As-Possible solver trained end-to-end on rigid-aligned vertex error. Evaluated on 14 subcortical structures from the ADNI dataset, MT-GNN achieves a -2.29% mean vertex error reduction compared to the temporal mean, outperforming geodesic shape regression (DCM, -0.19%) and a mesh transformer (TransforMesh, -0.45%) across all structures, with performance gains increasing over longer prediction horizons.
metric tensorgraph networkas-rigid-as-possiblesubcortical morphometrycontinuous-time prediction
DASyR-LLM: Domain-Aware Symbolic Regression with LLMs for Kinetic Model Discovery
We introduce DASyR-LLM, a domain-aware symbolic regression framework that integrates large language models (LLMs) for kinetic model discovery in chemical engineering. The framework embeds an LLM module within an iterative SR algorithm, where the LLM critiques SR candidates based on physicochemical plausibility and proposes new rate expressions guided by chemical knowledge. Evaluated on four in silico case studies, DASyR-LLM reduces iterations to identify ground-truth models by 41.7-79.3% compared to state-of-the-art SR, with the LLM directly proposing correct structures in over 50% of runs. Validation performance remains equivalent (R²>0.98), demonstrating LLMs' efficacy in injecting domain knowledge into scientific model discovery.
symbolic regressionkinetic model discoverylarge language modelsdomain-awarechemical engineering
Stable Density Ridges: Consistency and Convergence of Subspace Constrained Mean Shift
The paper introduces the 'stable ridge', a novel geometric structure that corrects the misconception that Subspace Constrained Mean Shift (SCMS) trajectories converge to the classical 'static ridge'. By analyzing SCMS through dynamical systems and the Jacobian of the projected density gradient, the authors prove that the stable ridge is the true theoretical target of SCMS. They develop a generalized SCMS framework with constant step size, demonstrating its uniform R-linear convergence and topological surjectivity onto the stable ridge. The framework also improves computational efficiency, addressing the polynomial-time complexity of the original SCMS algorithm.
density ridgesmean shiftdynamical systemshausdorff distancetopological surjectivity
Reward Structure Shapes the Interaction Between Episodic Exploration and Neural Memory in Reinforcement Learning
This work investigates the interaction between episodic exploration bonuses and neural memory architectures in partially observable reinforcement learning, demonstrating that reward structure shapes their complementarity. Through controlled experiments across three environments with varying memory acquisition dynamics, the authors show that exploration bonuses amplify architectural capacity differences when memory content requires unsupervised discovery, equalize architectures when memory relies on reward-supervised cues, and become null when observations are scheduled. Reward manipulations reveal that dense rewards neutralize bonuses only when directly supervising latent memory, while avoidable penalties induce suboptimal convergence. The study formalizes reward sparsity via observation-anchored reward machines, distinguishing structural and potential sparsity to organize task regimes by retention burden.
exploration bonusesneural memoryreward sparsitypartially observable reinforcement learningreward machines
BnBERT-iPET: Sparse Few-Shot Language Modeling for Bengali via Lottery Ticket Pruning
BnBERT-iPET introduces a sparse few-shot language modeling approach for Bengali, achieving 90% sparsity via Lottery Ticket Hypothesis pruning and iterative pattern exploiting training (iPET). The method retains only 10% of BERT's edges, yielding a lightweight model competitive with Bangla Electra, Indic-BERT, and XLM-RoBERTa on Bengali benchmarks. Results demonstrate comparable performance to larger models despite few-shot learning and extreme sparsity, addressing computational constraints for resource-scarce languages.
few-shot learninglottery ticket hypothesisiterative pattern exploiting trainingsparse language modelingbengali nlp
Multimodal Spatiotemporal Atmospheric Data Assimilation with Latent Flow-matching
The study introduces a unified atmospheric data assimilation approach using latent video flow-matching to sample temporally consistent trajectories from an ERA5 reanalysis prior (69 variables over 8 days). Posterior sampling assimilates real observations from NOAA sources, enabling continuous trajectory generation that propagates information between observed and unobserved frames. This method supports filtering and smoothing by varying observed frames and generates full-state ensemble forecasts from sparse observations, matching state-of-the-art observation-to-forecast models.
data assimilationlatent flow-matchingera5 reanalysisposterior samplingensemble forecasts
MALT: Lightweight Curvature-Aware Muon via Diagonal Preconditioning
MALT introduces lightweight diagonal preconditioning to Muon, a momentum-based optimizer for language model pretraining, addressing curvature anisotropy while maintaining low memory and computational overhead. The method employs two-sided diagonal preconditioners, orthogonalizes momentum via Newton-Schulz iterations, and uses norm grafting for update magnitude control. MALTER extends MALT with adaptive stepsize rescaling for robustness to stochastic gradient noise. Theoretical convergence guarantees are provided for non-convex settings. Experiments on GPT-2 Small, Medium, and Large demonstrate superior performance over Muon with comparable resource usage.
momentum orthogonalizationdiagonal preconditioningnewton-schulz iterationsnorm graftingcurvature anisotropy
Learning When to Stop: Prefix-Optimal Dynamic Diffusion Policies for Continuous Control
The paper introduces Prefix-Optimal Generative Policies (POGP), a framework for reducing computational costs in diffusion-based continuous control policies by dynamically adapting denoising steps. POGP learns a prefix value function through Bellman-style recursion over the denoising chain, serving as both an auxiliary training objective and a test-time stopping criterion. Evaluated across four MuJoCo environments against 12 baselines, POGP achieves a 2.7× reduction in denoising iterations while maintaining near-full task performance and improves final task performance by 3.5% over dynamic diffusion baselines.
diffusion policiescontinuous controlprefix value functiondenoising stepsbellman recursion
Optimizing What Policies Learn From: Recoverability-aware Rollout Intervention Learning
The paper introduces Recoverability-Aware Intervention Learning (RAIL), a framework for optimizing rollout generation in critic-free group-based reinforcement learning for post-training large language models. RAIL addresses limitations of fixed heuristics by modeling intervention selection as an online contextual-bandit problem, using a shadow-to-live procedure to train a recoverability controller that adapts to policy changes. Evaluations demonstrate RAIL's effectiveness, adaptivity, and efficiency, showing consistent performance improvements under constrained rollout budgets by generating more informative and less redundant learning signals.
recoverability-aware interventioncontextual-bandit problemrollout generationshadow-to-livepost-training
SparseDitto: Customizing GPU Kernels for Different Sparsity Patterns with LLM-Based Agentic System
SparseDitto introduces an LLM-based system for generating GPU kernels tailored to specific sparsity patterns, operators, and target GPUs. It employs a lightweight additive model to rank strategies based on matrix structural features, followed by an architecture-aware planner and coding agents to implement and refine designs. Evaluated on SpMV, SpMM, and SpGEMM across diverse matrices, SparseDitto achieves geometric-mean speedups of 2.68x (max 146.61x) on an RTX PRO 6000 and 2.79x (max 78.5x) on an H200 over cuSPARSE, with GCN training acceleration up to 3.39x.
sparse matrixgpu kernelsllm-basedspmmspgemm
Canonical Joint Energy-Based Model on CIFAR-10: failure modes and practical indistinguishability of Predictor-Corrector and SGLD samplers
This work systematically evaluates the Predictor-Corrector (PC) sampler against stochastic gradient Langevin dynamics (SGLD) in canonical Joint Energy-Based Models (JEM) on CIFAR-10. Using WideResNet-28-10 without normalization layers, the study compares PC and SGLD across three protocols: full training replacement, cold-start generation (FID), and multi-OOD detection (AUROC). Results show practical indistinguishability between PC and SGLD, with AUROC differences below 0.007 and FID differences below 0.5. Two failure modes are documented: catastrophic divergence via outlier-buffer mechanism and run-dependent SVHN OOD dynamics. Theoretical analysis confirms PC's predictor step degenerates under fixed noise, explaining the observed equivalence.
joint energy-based modelspredictor-corrector samplerstochastic gradient langevin dynamicsout-of-distribution detectionwide-resnet
Link prediction on multi-relational graphs from an influence propagation perspective
We propose Influential Graph Neural Predictor (IGNP), a novel link prediction framework for multi-relational graphs that models relationships through node influence propagation. The method extends the Susceptible-Infectious-Recovered (SIR) epidemic model to capture influence propagation at scale via sub-graph structures, which are compressed using virtual edges to reduce computational complexity. IGNP integrates both local and global graph information through this influence propagation perspective. Experimental results demonstrate IGNP's superior performance, outperforming strong baselines by significant margins on widely-used real-world datasets.
link predictionmulti-relational graphsinfluence propagationgraph neural networksepidemic model
Stochastic Emulation using Generalized Stratified Sampling for Performance-Based Risk Optimization of Structures
The study introduces a GSS-SPCE framework combining Generalized Stratified Sampling (GSS) with Stochastic Polynomial Chaos Expansion (SPCE) to improve tail accuracy in Performance-Based Risk Optimization (PBRO) of structures. GSS partitions the input space by hazard intensity, enabling stratum-specific SPCE emulators, while total probability theorem recombines conditional exceedance probabilities. Applied to buckling-restrained brace design in a steel building, the method accurately estimates tail responses and reduces nonlinear evaluations by 50% compared to conventional SPCE.
stochastic emulationgeneralized stratified samplingperformance-based risk optimizationpolynomial chaos expansionprobabilistic constraints
Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes
The study systematically investigates multimodal pretraining dynamics through controlled experiments on synthetic and real-world datasets, revealing four key mechanisms: (i) asymmetric knowledge flow between language, visual understanding, and generation modalities; (ii) data complexity-dependent modality synergy, enabled by shared attention with modality-specific feed-forward layers; (iii) superior performance of early modality unification over late alignment, exposing vision laziness in delayed integration; (iv) efficient pretraining recipes achieving strong generative performance with 5% compute. Findings are validated at scale using 13.5B MoE models trained on 2T tokens.
multimodal pretrainingknowledge flowmodality synergyearly unificationvision laziness
EvolveNet: Collaborative Harness Evolution for Agent Self-Improvement
EvolveNet introduces collaborative harness evolution for LLM agents, enabling decentralized improvement of execution programs without weight updates. The method broadcasts a shared harness to local deployments, where each evolves it independently on isolated workloads; adaptations are then composed into an updated shared harness via scope-typed, evidence-guided aggregation. Evaluations across text-to-SQL, data-science coding, competitive programming, software engineering, and agentic workflows show universal improvement, with maximal gains under heterogeneous workloads, attributed to adaptation composition rather than selection.
harness evolutionllm agentsdecentralized adaptationscope-typed aggregationheterogeneous workloads
SpecRoll: Fast-Slow Verifier-Feedback Adaptation for Speculative Reinforcement Learning Rollouts
SpecRoll introduces a speculative rollout engine that accelerates reinforcement learning (RL) post-training for large language models while preserving the target model's sampling distribution. It employs lightweight future-token heads for parallel proposals and a Reflex module for trajectory-local hidden-state corrections using delayed verifier feedback, alongside a slow path for head parameter updates. The method integrates concurrency-aware sparse-tree verification and exact target verification, maintaining the GRPO objective. Evaluated across five models (1.5B to 14B) and three mathematical reasoning datasets, SpecRoll achieves 1.26-2.15x generation speedup and 1.21-2.04x end-to-end speedup over vanilla GRPO, outperforming FastGRPO in all 15 matched settings.
speculative decodingreinforcement learninghidden-state correctionssparse-tree verificationgrpo objective
A geometry-based deep equilibrium model for image restoration under multiplicative Gamma noise
The authors propose a geometry-based deep equilibrium (DEQ) model for image restoration under multiplicative Gamma noise and blur, learning an explicit regularizer parameterized by geometric priors (surface area, mean curvature) rather than implicit neural regularization. They develop a mirror descent algorithm for Gamma-noise fidelity terms, proving global convergence via Kurdyka-Lojasiewicz properties in $o$-minimal structures. Experiments on grayscale and color images show superior performance to model-based approaches and parity with implicit DEQ models, despite fewer parameters.
deep equilibrium modelmultiplicative gamma noisekurdyka-lojasiewicz propertymirror descentimage restoration
State2State: Environment-Derived Mid-Training for LLM Agents
State2State introduces an environment-derived mid-training method for LLM agents that autonomously generates training objectives from explored environment states, eliminating the need for external task specification or expert supervision. The approach converts environment states into training targets, verified through rule-based state matching, enabling scalable and diverse agent training. Evaluations on ALFWorld and ScienceWorld demonstrate improved agent performance as a standalone training stage and enhanced downstream RL efficiency, with indications of cross-environment generalization.
llm agentsenvironment interactionstate matchingmid-trainingcross-environment generalization
Optimal Training-Time Scaling in Gradual Adaptation
The work derives optimal per-task training time scaling for gradual adaptation in overparameterized linear regression, where tasks change smoothly and share a zero-loss solution. Analyzing the continuum limit as task count $N$ grows, the authors prove final learning progress scales as $Θ(τ)$ for small $τ=Ns_N$ and $Θ(τ^{-1})$ for large $τ$, implying minimal progress at extreme training durations. Optimal per-task training time follows $s_N^\star=Θ(N^{-1})$, confirmed empirically on gradually rotated MNIST and Yearbook time-shift tasks.
gradual adaptationoverparameterized regressiontraining-time scalingcontinuum limitzero-loss solution
Visual Representation Matters: Exploiting Temporal Differences in Video-to-Audio Generation
TD-V2A introduces temporal differences (TD) as a key representation for video-to-audio (V2A) generation, requiring minimal architectural modification compared to existing diffusion-based methods. The method investigates TD at frame and feature levels, proposing a hierarchically continual learning strategy and annealed TD guidance to enhance visual conditioning during diffusion training and sampling. Experiments on benchmark datasets show TD-V2A outperforms dedicated V2A representations like contrastive audio-visual pretraining in generation quality.
temporal differencesvideo-to-audio generationdiffusion modelsvisual representation learninghierarchically continual learning
Variational Bounds for Perceptron Learning from Structured Data
The paper introduces a variational approach for analyzing a finite-temperature continuous-spin perceptron trained on Gaussian mixture data, accommodating concave utilities and log-concave separable priors. Combining interpolation with log-concavity and concentration estimates, the authors derive minimax variational bounds for the quenched pressure, which differ only in the optimization order of two variational parameters. When these optimizations commute, the bounds coincide, yielding exact solutions for ground-state energy, training loss, and generalization error. The variational potential also provides fixed-point equations for these quantities.
variational boundsperceptrongaussian mixturelog-concavequenched pressure
Training Crossroads for Recurrent Vision Transformers: Recurrence, Neural ODEs, and Deep Supervision
The paper empirically evaluates three training regimes for recurrent Vision Transformers (bViT) under controlled CIFAR-100 experiments, comparing them to standard ViTs. Key findings include: (1) recurrent ViTs outperform standard ViTs in accuracy-parameter trade-offs under memory constraints, though standard ViTs remain superior under FLOP constraints; (2) higher-order ODE solvers for residual recurrent blocks act as architectural biases rather than numerical refinements, with non-uniform gains; (3) deep supervision improves robustness beyond the training horizon but not nominal accuracy. The study clarifies design choices for recurrent ViTs without proposing new architectures.
recurrent vision transformersneural odesdeep supervisionparameter efficiencyarchitectural bias
The Neural Echo: A Signal Processing Perspective for Understanding Neural Networks
The paper introduces the neural echo, a novel tool for analyzing neural networks by generalizing classical signal processing concepts like impulse responses and filter echoes to learning-based methods. Neural echoes provide input-dependent, local impulse responses and filter kernels, applicable to various architectures including CNNs, transformers, and recurrent networks without requiring differentiability. For differentiable networks, they subsume Jacobian-based methods like saliency maps. The authors demonstrate the framework on DnCNN, revealing its pixel-weighting behavior aligns with classical denoisers such as bilateral filtering, bridging model-based and learning-based approaches.
neural echoimpulse responseexplainable aisignal processingdenoising
Nonparametric Goodness-of-fit Testing under Covariate Shift
The paper proposes nonparametric goodness-of-fit testing procedures for covariate shift scenarios, where labeled data originates from a source population but evaluation targets a different distribution. The method combines truncated importance-weighted kernel ridge regression with multiplier bootstrap to construct confidence sets for the regression function, addressing instability from heavy-tailed density ratios. Theoretical guarantees include nonasymptotic validity, sharpness under operator compatibility conditions, and explicit coverage error rates dependent on spectral decay of the kernel operator. Experiments validate the approach under bounded moment or sub-exponential tail conditions on the target-to-source density ratio.
covariate shiftkernel ridge regressiongoodness-of-fit testingimportance weightingmultiplier bootstrap
Robust Control under Stationary Ambiguity
The paper introduces stationary ambiguity, a method for training robust control policies in simulators where parameter uncertainty does not systematically decay over time. By ensuring the simulator induces a stationary filter process over latent states, policies maintain robustness to shifting latent factors, unlike traditional approaches where ambiguity vanishes as the policy infers parameters. The authors demonstrate this on financial hedging problems, showing improved performance on real market data. The method also informs simulator design decisions, including parameter randomization and initialization. While tested on hedging, stationary ambiguity may generalize to other sequential control problems with exogenous stochastic processes.
stationary ambiguityrobust controllatent stateparameter uncertaintysequential control
Intrinsic-Hybrid Latent Diffusion Models for Generative Modeling on Unknown Manifolds
The Intrinsic Hybrid Latent Diffusion Model (ILDM) introduces a generative framework combining probabilistic dimensionality reduction with geometry-aware diffusion on unknown manifolds. ILDM interprets the latent space as a chart of an unknown Riemannian manifold, employing a hybrid diffusion process that switches between Riemannian and Euclidean dynamics based on local uncertainty. The Riemannian component is governed by a probabilistic metric tensor derived from a probabilistic decoder. An approximate denoising score matching method enables hybrid Langevin dynamics for the backward process. Experiments on COIL-100, MNIST, and cardiac MRI datasets show ILDM achieves superior generation quality, with lower FID and LPIPS scores compared to standard diffusion and latent diffusion models.
latent diffusion modelriemannian manifolddenoising score matchingprobabilistic decoderhybrid langevin dynamics
MGSB: Manifold Gated Signature Branch Pressure-Domain Baseline Architecture for Two-Phase Pipeline Flows Under Distributional Shift
The Manifold Gated Signature Bias (MGSB) architecture improves leak detection robustness in multiphase pipelines under distributional shift by explicitly modeling flow regime transitions. MGSB integrates regime-conditioned feature fusion, a TT-RoughPath encoder, and Mean-Teacher consistency regularization to enhance out-of-distribution (OOD) performance. Evaluated via leave-one-group-out testing, MGSB achieves a detection F1 of 0.930 and an OOD F1 of 0.783, significantly outperforming CNN-LSTM and fully connected baselines under severe feature corruption. Ablation studies confirm the architecture's design, not the training procedure, drives OOD robustness, while Mahalanobis-distance analysis validates the OOD nature of held-out conditions. This demonstrates regime-aware modeling's efficacy for robust, sensor-agnostic pipeline leak detection.
manifold gated signature biastt-roughpath encodermean-teacher consistencydistributional shiftmahalanobis-distance analysis
Above-ground Biomass Estimation with Geospatial Foundation Models
The study benchmarks Geospatial Foundation Models (GFMs) for global Above-Ground Biomass (AGB) estimation using the AGBD dataset, comparing 11 GFMs as frozen encoders and pre-computed embeddings (AlphaEarth Foundations, TESSERA) against a supervised SOTA model. Frozen GFMs underperform (R²=0.62 vs. SOTA's 0.72), while MLPs on AEF embeddings surpass SOTA (R²=0.75) and exhibit superior spatiotemporal generalization. Augmenting SOTA with AEF embeddings yields the best performance (R²=0.78), demonstrating the utility of embedding products for quantitative regression tasks in remote sensing.
geospatial foundation modelsabove-ground biomassremote sensingembedding productsregression benchmarking
Continual-Learning Physics-Informed Neural Networks for Parameterized Partial Differential Equations
The paper introduces continual-learning physics-informed neural networks (CL-PINNs) to improve parameterized PDE solutions by treating instances at different parameter values as sequential tasks. The method combines Bayesian-optimization-based active parameter selection, task-wise dynamic loss weighting, sparse physics-constrained replay, and an optional parameter subnetwork to enhance task allocation and knowledge retention. Evaluations on five benchmarks demonstrate that CL-PINN achieves higher and more balanced accuracy than fixed-sampling and grid-greedy baselines, with Bayesian selection reducing objective-loss queries and sparse replay mitigating forgetting.
physics-informed neural networksparameterized pdescontinual learningbayesian optimizationsparse replay
Attention, Anomalies! Handling Attention Layers in Unsupervised Federated Outlier Detection
The paper introduces novel aggregation techniques for Memory Augmented Autoencoders (MemAE) in federated learning, addressing challenges in unsupervised outlier detection with non-IID data. The method focuses on specialized attention-layer aggregation to enhance robustness in unbalanced, multi-node environments. Experiments demonstrate improved performance for shallow autoencoders in resource-constrained settings, validating the approach's efficacy for federated anomaly detection.
memory augmented autoencodersfederated learningattention layersunsupervised anomaly detectionnon-iid data
Benchmarking Deep Learning Models for Dense Event Classification of Offshore Wind Infrastructure in Sentinel-1 Time Series
The study benchmarks deep learning models for dense event classification in Sentinel-1 SAR time series to monitor offshore wind infrastructure deployment. It compares ten variants of LSTM, Transformer, and fully connected models with monotemporal, unidirectional, and bidirectional context awareness, with and without self-supervised pretraining. The supervised BiLSTM achieves the best performance (AUC: 0.8509, perfect match rate: 0.5063), outperforming the rule-based baseline (AUC: 0.7853). An ensemble combining BiLSTM predictions with baseline labels further improves agreement, enabling global analysis of turbine deployment durations (median: 84 d in China, 242 d in EU, 258 d in UK) and revealing regulatory and environmental drivers.
sentinel-1bilstmdense classificationtime seriesoffshore wind infrastructure
Suppression Sticks, Locality Is Fragile: A Closed-Loop Target-and-Control Audit of Task-Vector Negation in VLA Policies
The study investigates the behavioral locality of task-vector subtraction in multitask vision-language-action (VLA) policies through a closed-loop target-and-control audit. Using LIBERO-Goal skills, the authors identify three regimes: target-control separation (5 skills), resistance (3 skills), and global collapse (2 skills). While target suppression maintains 0% success on held-out initial states, control retention averages 52%, with edits adversely affecting unrelated controls. Experiments across policies with continuous-regression, discrete-token, and flow-matching action heads reveal separation in Goal panels but collapse in Object and Long-horizon panels. Task-vector cosine fails to explain variation, and retain-aware gradient baselines require optimization, highlighting subtraction's brittleness despite its efficiency.
task-vector subtractionvision-language-actionclosed-loop auditbehavioral localitycontrol retention
The Sample Complexity of Distributionally Robust PAC Learning under Cressie--Read Divergences
The paper establishes tight sample-complexity bounds for distributionally robust PAC learning under Cressie--Read divergences, closing prior gaps between upper and lower bounds. Using VC theory, the authors analyze both realizable and agnostic cases for hypothesis classes with dimension $d$, showing rates depend on divergence order $k>1$ and radius $ρ\geq 0$. Key results reveal a phase transition: robustness changes the realizable $\varepsilon$-dependence from $\varepsilon^{-1}$ to $\varepsilon^{-k_\star}$ as $\varepsilon\downarrow0$, where $k_\star=k/(k-1)$. Bounds correctly interpolate to standard PAC learning when $ρ\to0$.
distributionally robust learningcressie--read divergencepac learningsample complexityvc dimension
Kathleen Writes: Autoregressive Generation and Data Scaling Without Attention
The Kathleen series demonstrates that an attention-free, byte-level autoregressive model with wavetable encoding and multi-scale reverberant state achieves competitive performance in generation tasks. On WikiText-103 (raw UTF-8), it outperforms parameter-matched transformers across dataset scales (2-512 MB), e.g., 1.84 vs. 2.04 bits/byte at 512 MB (~0.5M params). The study introduces FORM DISTANCE, a non-parametric metric for text realism, and shows decoding policy significantly impacts generation quality—retrieval-augmented decoding reduces distance from 1.52 to 1.14. Key findings include corpus-specific retrieval efficacy and four ineffective architectural modifications.
attention-freebyte-levelreverberant stateform distanceretrieval-augmented decoding
Diverse and Plausible Algorithmic Recourse via Tractable Recourse Distributions
The paper introduces Tractable Recourse Distributions (TRD), a probabilistic framework modeling the space of feasible algorithmic recourse alternatives as a probability distribution over favorable outcomes. For proximity- and sparsity-based cost functions, TRD represents this distribution exactly as a probabilistically tilted circuit, enabling closed-form sampling of diverse, plausible recourses without model retraining. Experiments on benchmark datasets show TRD achieves simultaneous diversity, plausibility, and feasibility, with practical rejection sampling via controlled tilting parameters. MNIST visualizations demonstrate the trade-off between proximity and validity.
algorithmic recourseprobabilistic circuitscounterfactualsrejection samplingfeature sparsity
Differentiating Through Dual Prices: End-to-End Policy Learning Under Capacity Constraints
The paper introduces an end-to-end learning approach for resource allocation policies under capacity constraints, differentiating through dual prices to optimize off-policy value estimates. Two formulations are proposed: an exact nonconvex method and a convex relaxation guaranteeing expected capacity satisfaction with suboptimality bounded by a term linear in smoothing temperature and logarithmic in arm count. Evaluated on six datasets including a 70k-patient hospital cohort, the end-to-end methods outperform decision-blind baselines in policy value and constraint adherence, particularly when resources are scarce.
resource allocationdual pricescapacity constraintsend-to-end learningoff-policy estimation
Automatic Statistical Test for Rationally Expressible Algorithms by Selective Inference, with Applications to Feature Selection
AutoSI introduces an automated framework for selective inference (SI) that generates statistically valid p-values without manual derivation of selection events. The method supports algorithms expressible through rational functions of data, extending beyond existing linear/quadratic inequality constraints. By automatically constructing selection events from NumPy-like code, AutoSI enables exact SI for feature-selection methods like lasso with cross-validated tuning, previously intractable. Empirical results on synthetic and real datasets confirm type I error control at nominal levels while maintaining power.
selective inferencerational functionsfeature selectionp-valuestype i error
Active Learning Guided Design Space Refinement for Scalable Multi-Objective Bayesian Optimization in Materials Discovery
The authors propose an active-learning-guided adaptive search-space refinement framework integrated with multi-objective Bayesian optimization to enhance materials discovery efficiency. The method combines active learning for design space reduction with Bayesian optimization to focus on Pareto-relevant regions, avoiding low-value evaluations. Evaluated on CH4/N2 separation in covalent-organic frameworks and pressure-vessel design, the approach reduces candidate spaces by approximately 50% while preserving over 99% of the original hypervolume. Results demonstrate improved early convergence and cumulative Pareto-front discovery, enabling scalable optimization in constrained materials discovery settings.
active learningbayesian optimizationmaterials discoverypareto frontdesign space refinement
An entropic explanation of insistence on sameness in autism
The article proposes an information-theoretic framework to explain insistence on sameness in autism as a strategy to minimize surprise and uncertainty. The framework defines autism as a cognitive impairment limited to discrimination, memorization, and prediction of tangible environmental properties, formalized via the entropy metric $D_H(R, M) = H(R|M) + H(M|R)$, where $R$ is random stimuli and $M$ is memory. Results suggest insistence on sameness arises from restricting $R$ to known $M$, enabling quantification of behavioral concepts (e.g., sensory overload, anxiety) and guiding therapeutic algorithms for robotic caregivers. Validation via Turing test-like methods is proposed.
information theoryentropy minimizationcognitive impairmentbehavioral quantificationtherapeutic algorithms
Why Ranking Anomaly Detection Algorithms Isn't as Reliable as You May Think
This work investigates the reliability of anomaly detection algorithm rankings by analyzing the impact of benchmarking choices. Using seven anomaly detection algorithms and 690 datasets from OddBench, the study evaluates ranking instability across varying dataset selections, evaluation metrics, hyperparameter configurations, and random seeds. A rank instability metric quantifies variability, revealing that rankings are highly unstable, with competitive algorithms often appearing as top performers under specific configurations. Dataset selection and hyperparameter choice are the primary contributors to uncertainty, while random seeds and metrics have limited impact. The findings emphasize the need for larger, more diverse dataset collections for reliable benchmarking.
anomaly detectionbenchmarkingrank instabilityhyperparameter configurationdataset selection
On MUON optimization: From non-convergence to an error analysis with Polar Express and the Newton-Schulz polynomial from implementations
The authors propose a generalized variant of the MUON optimizer, incorporating arbitrary Newton-Schulz (NS) polynomial steps, extending beyond the original five-step implementation. This framework encompasses both the standard MUON and its combination with the Polar Express method. For a specific class of stochastic optimization problems (SOPs), they demonstrate non-convergence of MUON for almost all mini-batch sizes as gradient steps increase. Additionally, they provide an error analysis, deriving convergence rates based on gradient steps and mini-batch size, validated on quadratic SOPs and ℓ2-regularized logistic regression.
muon optimizernewton-schulz polynomialstochastic optimizationpolar expresserror analysis
Relevant but Incomplete: Referential Dangling as a Paradigm-Level Failure Mode in Hard Prompt Compression
The paper identifies referential dangling as a critical failure mode in hard prompt compression, where independent selection of text units splits evidentiary pairs, retaining answers while deleting required entity definitions. Evaluating six compressors on multi-hop QA datasets (HotpotQA, LongBench-v2, MuSiQue), the authors find dangling rates up to 60%, causing accuracy drops of 8.8-34 points across models (Qwen3-8B, GPT-5.5). They propose a compact classifier to reinsert critical omitted sentences, improving Qwen3-8B accuracy by 4.7 points on HotpotQA with minimal compression ratio change (0.30→0.31).
hard prompt compressionreferential danglingmulti-hop qacontext pruningevidence pairs
Relational Response Fields: A General Theory of Black-Box LLM Response Consistency and Recovery
The paper introduces relational response fields (RRFs), a theoretical framework for analyzing black-box LLM response recoverability under query transformations. It defines a recoverability metric γ_k(D,A) based on relation operator D and anchor operator A, proving it determines identifiability bounds and showing no estimator can surpass its stability guarantees. The work separates information-theoretic identifiability from optimization requirements, develops sparse repair algorithms, and validates four theoretical predictions through theorem tests and black-box experiments.
relational response fieldsresponse recoverabilityidentifiability boundssparse repairblack-box consistency
Learning Compression Rules for Network Traffic
The authors present RECAP, a method for learning compact rule-based compressors for structured network traffic by optimizing header field redundancy. RECAP employs a two-stage approach: (i) unsupervised structure discovery via recursive partitioning using a normalized entropy-ratio criterion, and (ii) constrained dynamic programming to select rules maximizing compression gain under a fixed budget. Evaluated on Static Context Header Compression (SCHC) across IoT and 5G core-network datasets, RECAP outperforms manual rule sets with fewer learned rules, eliminating expert design. Results demonstrate robust compression efficiency with small sample robustness.
rule-based compressionentropy-ratio criteriondynamic programmingstatic context header compressioniot networks
Discretization and Statistical Consistency of Functional Flow Matching
(No summary returned.)
ODRA: Synthesizing Cognitive Behavioral Therapy Sessions with Structured Chain-Of-Thought and Dynamic Patient Resistance
ODRA introduces a novel framework for synthesizing Cognitive Behavioral Therapy (CBT) sessions by combining structured Chain-of-Thought (CoT) reasoning with dynamic patient resistance modeling. The method addresses sycophancy in synthetic dialogues through a resistance orchestrator that steers patient behavior to match resistance levels, ensuring adherence to CBT guidelines while capturing realistic patient unpredictability. Evaluations show ODRA outperforms existing methods on therapeutic skills, CBT alignment, and behavioral fidelity, with licensed psychologists preferring it on 12 of 13 clinical metrics. Fine-tuned models trained on ODRA-generated data exhibit improved robustness against both cooperative and resistant patients.
cognitive behavioral therapychain-of-thoughtresistance orchestratorsycophancytherapeutic robustness
DIVE: Dynamic Iterative Visual Evidence Construction for Efficient Vision-Language Models
DIVE introduces a training-free framework for efficient vision-language model inference by dynamically constructing visual evidence through iterative token selection. The method repeatedly selects tokens with highest residual-conditioned scores, updates visual and prompt residuals, and re-evaluates remaining tokens to build complementary, prompt-relevant evidence. Experiments on eight image-understanding benchmarks show DIVE retains 98.2% of uncompressed performance with an 88.9% reduction in visual tokens.
vision-language modelstoken pruningdynamic evidence constructionresidual-conditioned scoringefficient inference
DeepInvert: Semi-Supervised Embedding Inversion Against Obfuscated Language Models
DeepInvert introduces a semi-supervised embedding inversion attack that exposes vulnerabilities in obfuscation-based defenses for language models (e.g., ObfusLM, SentinelLMs). The method combines supervised training on labeled shadow data with an unsupervised consistency objective, leveraging residual semantic structure in perturbed embeddings. Evaluated across nine defenses, five tasks, and four architectures, DeepInvert achieves 73.5% top-1 token recovery against ObfusLM, outperforming prior attacks by 47.3 percentage points. Results reveal a trade-off: obfuscations preserving utility remain invertible, while robust perturbations degrade task performance.
embedding inversionobfuscation defensessemi-supervised learninglanguage model privacytoken recovery
Local Violation Certification for Linear Predict-Then-Optimize Pipelines
The paper introduces a local violation certification framework for linear predict-then-optimize pipelines under input uncertainty, addressing computational inefficiencies in traditional scenario generation for rare failures. By analyzing fixed decision boundaries, the method computes local failure risk via a single optimization solve and provides feature-level attributions through exact sampling and closed-form statistics. Evaluated on an economic power dispatch system subject to emissions regulations, the approach yields precise risk assessments with significantly reduced computational overhead compared to random sampling.
violation certificationpredict-then-optimizeinput uncertaintydecision boundaryrisk attribution
Multi-Objective Ranking for Live-Streaming: Balancing Fresh and Delayed Signals with Segment-Aware Targeting
The paper proposes a multi-objective ranking system for live-streaming recommendation, addressing sparse/delayed user behaviors and segment biases via three contributions: delayed feedback windows, a multi-model architecture combining fresh/delayed signals with segment-aware targeting, and Multi-gate Mixture-of-Experts (MMoE) integration reducing parameters by 41.9%. The system achieves +0.09% Daily Active Viewers (DAV), +0.56% capped ARPU, and +1.12% positive interactions on Twitch, with segment-specific gains (e.g., +0.15% DAV for newer users) while maintaining low-latency processing.
multi-objective rankingdelayed feedbacksegment-aware targetingmulti-gate mixture-of-expertslive-streaming recommendation
Beyond Global Routing Aggregation: Phase-Aware Expert Merging for MoE Vision-Language Models
RoleMerge introduces a phase-aware expert merging method for Mixture-of-Experts Vision-Language Models (MoE-VLMs), addressing limitations of global routing aggregation. By constructing Routing Role Profiles (RRP) from phase-normalized routing statistics, it captures experts' relative phase preferences and merges compatible experts while preserving answer-decoding distinctions. Experiments across three models and multiple benchmarks demonstrate that RoleMerge outperforms alternative methods, achieving up to 9.6% relative improvement in six-task macro-average performance at matched expert-retention ratios. This validates phase-conditioned expert roles as a superior basis for MoE-VLM expert merging.
mixture-of-expertsvision-language modelsrouting role profilephase-normalized statisticsexpert merging
A Counterexample to Fourier Alignment in Single-Neuron Modular Addition
The work presents a counterexample disproving the general validity of Fourier alignment in single-neuron modular addition (MAIS-O60). Through construction of an initially active ReLU neuron that becomes permanently inactive, the authors demonstrate equal Fourier energy distribution across nonzero frequency classes, occurring with positive probability under Gaussian initialization. GPT-5.6 Sol extends this result to Clarke trajectories, smooth ReLU approximations, and full-batch gradient descent, showing single-frequency alignment is not universally achievable in this setting.
fourier alignmentmodular additionrelu neurongradient descentclarke trajectory
Robustness Emerges Early in Training Dynamics, but Is Not Preserved
The paper identifies a robustness fading phenomenon where shallow layers in deep neural networks spontaneously develop robust representations and flat loss landscapes early in training, but lose these properties during standard convergence. To address this, the authors propose Early-Phase Stabilization (EPS) and Asymmetric Weight Reversion (AWR), two parameter-free strategies that stabilize or recover robust shallow configurations without architectural modifications or additional learnable parameters. Extensive experiments demonstrate significant improvements in downstream transfer, dynamic adaptation, and diverse computer vision applications across various benchmarks and architectures.
robustness fadingearly-phase stabilizationasymmetric weight reversionshallow layersloss landscapes
The Price of Isolation: Estimating the Ecosystem Cost of Symmetric Two-Sided A/B Testing
The article quantifies the engagement cost of symmetric two-sided isolation in A/B testing on content platforms, showing it persists under heavy-tailed match quality distributions. Using extreme-value theory and order-statistics modeling, the authors demonstrate that catalog expansion fails to eliminate isolation-induced loss when match quality has heavy tails, contrasting with light-tailed cases. Empirical validation from two large-scale platform experiments (millions of creators) confirms the model's predictions, with catalog ablation revealing depth-graded per-viewer costs. A preflight estimation procedure is proposed to budget isolation costs and recommend fallback designs when thresholds are exceeded.
two-sided isolationextreme-value theoryorder-statistics modelheavy-tailed distributioncatalog ablation
Deltoris: Enabling Real-time VLA Inference in Embodied AI via Bit-level Sparsity and Speculative Inference
Deltoris introduces an algorithm-hardware co-design framework for efficient diffusion-based vision-language-action (VLA) model inference in embodied AI. The framework employs temporal-aware bit-sparsity to compute differences between consecutive inputs, reducing redundant bit-level operations, and speculative inference to amortize data loading across control steps. A dedicated accelerator with customized 1D systolic bit-serial PE arrays is co-designed to support these techniques. Evaluations demonstrate up to 34.2× speedup over mobile GPUs and 6.1× over prior accelerators while maintaining comparable accuracy.
diffusion-based vlatemporal-aware bit-sparsityspeculative inference1d systolic bit-serial pe arraysalgorithm-hardware co-design
Elbow-Based MoE Routing: A Training-Free Inference Time Plugin for Expert Selection
The paper introduces elbow-based routing, a training-free inference-time modification for Mixture-of-Experts (MoE) models that dynamically adjusts the number of activated experts per token by identifying an elbow point in the sorted router probability distribution. This method replaces fixed top-k selection, reducing unnecessary computation while preserving expert load balance. Experiments on a state-of-the-art MoE model show a 5.3% average latency reduction without accuracy loss across six benchmarks.
mixture-of-expertsrouter probabilityelbow-based routinginference-time optimizationload balance
When Proxy Prediction Becomes Equation Reconstruction: Diagnostics and Residual Learning for Factor-Derived Proxy Supervision
The study identifies and addresses a critical pitfall in scientific machine learning where high predictive accuracy may stem from reconstructing proxy-generating equations rather than learning robust representations, demonstrated via RUSLE-derived soil-loss prediction under degraded soil-erodibility factor $K$. A diagnostic framework combining degraded-formula references, tree baselines, contextual ablations, and degradation scoring is introduced, alongside RASPL—a formula-preserving residual learning method that anchors predictions to degraded estimates while learning gated contextual corrections. RASPL outperforms direct prediction, with statistical and convolutional encoders excelling in macro-averaged $R^2$ and Tail95 MAE, respectively, establishing formula preservation as key for robust proxy supervision.
proxy supervisiondegradation robustnessresidual learningfactor-derived targetsscientific machine learning
NeuroPB: Scaling Neural Decoding with Pretrained Behavioral Representations
NeuroPB introduces a framework for scaling neural decoding by leveraging pretrained behavioral representations to improve brain-computer interfaces (BCIs). The method pretrains a motor encoder on large-scale behavioral data, aligns neural activity with this representation using paired neural-behavioral recordings, and optimizes a neural encoder and lightweight motor decoder for trajectory reconstruction. Results show an 11% $R^2$ increase on center-out tasks and 8% on random-target tasks compared to training from scratch. Pretraining on robotic trajectories matches macaque trajectory performance, demonstrating transferable kinematic structure. Scaling robotic pretraining data enhances decoding with fixed neural data, and pretraining improves generalization across sessions, subjects, and tasks, reducing calibration needs to 10%.
neural decodingbehavioral representationsmotor trajectoriesbrain-computer interfaceskinematic structure
Non-asymptotic implicit bias of logistic regression at early-stage gradient descent dynamics
The work analyzes early-stage gradient descent dynamics in logistic regression, revealing weak alignment with the max-margin direction before asymptotic convergence. By examining radial and tangential flows, the authors avoid asymptotic expansion and directly model alignment dynamics using dataset geometry. Theoretical results show the parameter vector aligns weakly with the max-margin direction within $O(\exp(\exp(-δ)))$ iterations, where $δ>0$ is the permissible alignment error, and prove this bound is tight.
gradient descentimplicit biasmax-marginlogistic regressionalignment dynamics
Helping Music Co-Creation Agents 'Listen' Well: Hierarchical Self-Supervised World Models for Understanding and Generation
A hierarchical self-supervised world model for symbolic music co-creation is proposed, featuring a 2.55M-parameter Swin V2 encoder trained on MIDI piano-roll images using JEPA-style objectives without labels or music-theory vocabulary. The model's embeddings capture musical properties at varying time scales, with phrase boundaries decodable at coarse levels and harmonic detail at fine levels. A conditional flow-matching decoder achieves pixel F1 $0.996$ in reproducing target windows, enabling graphical prompting for masked inpainting. The pipeline runs in $2.8$ s on CPU and $0.6$ s on Apple MPS, demonstrating potential for interactive music creation agents that enhance human agency.
self-supervised learningsymbolic musicflow-matchingmidi piano-rollhierarchical embeddings
EvtGraph: Event-Adaptive Compression for Sparse Temporal Graph Learning in Multimodal Time Series
The paper introduces EvtGraph, a framework for event-adaptive compression in sparse temporal graph learning to address inefficiencies in uniform discretization of multimodal time series. The method employs event-adaptive compression (EAMC) to reparameterize sequences into event-level tokens, selects a compact subset via node budget constraints (NBC), and performs temporally constrained sparse graph reasoning (T2SG). Evaluated on multimodal clinical (MIMIC-IV + CXR) and cross-domain benchmarks, EvtGraph outperforms Transformer-based and recurrent baselines while improving computational efficiency, demonstrating that budget-constrained event-centric representation effectively handles high-redundancy temporal data.
event-adaptive compressionsparse temporal graphmultimodal time seriesnode budget constraintstemporally constrained reasoning
Manipulation-Proof Oblivious Audits against Deceptive Model Providers
We propose a manipulation-proof audit protocol for fairness evaluations in machine learning models, leveraging Private Information Retrieval to enable oblivious queries. The method requires model providers to label a large dataset without revealing the audit subset, preventing strategic manipulation of fairness metrics. The protocol operates efficiently without modifying the model, training procedure, or inference pipeline. Theoretical analysis shows that providers attempting to hide unfairness must falsify significantly more responses, increasing detectability. Empirical results demonstrate the approach's effectiveness across diverse audit scenarios.
fairness evaluationsprivate information retrievalaudit protocolmanipulation-proofalgorithmic governance
Looking in the Mirror: Introspecting Side-Effect Misalignments Induced by Fine-Tuning
The paper introduces side-effect introspection, a novel problem setting for detecting unintended alignment degradation in fine-tuned models, distinct from explicitly implanted behaviors. It proposes Delta-Aware Introspection Adapter (DAIA), which processes both base-model activations and fine-tuning-induced differences to improve sensitivity to misalignment. Experiments demonstrate DAIA's superior performance over existing introspection adapters, with generalization to unseen models and safety categories.
fine-tuningalignment degradationintrospection adapterside-effect misalignmentdelta-aware mechanism
Cost-Aware Multi-Objective Bandits: Theory and Application to Budgeted LLM Configuration Evaluation
The paper introduces a cost-aware multi-objective bandit framework for large language model (LLM) configuration evaluation, addressing online selection and Pareto identification under budget constraints. For online selection, a hypervolume-based UCB algorithm achieves a budgeted regret bound of $O(\sum_{i\ne i^\star}\frac{\log B}{\Delta_i})$, while a cost-aware gap elimination algorithm attains an error probability of $O(\exp(-B/H_{\mu,c}))$ for Pareto identification. Experiments confirm efficient decision-making and accurate Pareto identification with limited budgets.
multi-objective banditsllm configurationhypervolume efficiencybudgeted regretpareto identification
Right Reset: Chunking by Prefix Removal
The paper introduces Right Reset (RR), a method for text chunking by measuring hidden-state trajectory preservation when removing left context in causal language models. Using prefix-removal probing, RR identifies boundaries where right-hand token processing remains stable, converting edge scores into variable-length chunks via dynamic programming. On flattened text with deleted separators, RR recovers 47.7% of original records versus 25.9% for BGE embeddings, maintaining gains post-rendering and OCR. RR outperforms passive Qwen3-4B layer scores and same-scale instruction-model prompting, with consistent output stability across six models. Likelihood-ratio readouts show context dependence alone can signal boundaries without surface structure.
chunkingprefix-removal probinghidden-state trajectorydynamic programmingcontext dependence
Real-time probabilistic tsunami forecasting via generative AI
The study introduces a probabilistic tsunami forecasting framework using generative AI to address uncertainty in inundation predictions. The method employs a conditional diffusion model to generate ensemble forecasts, reconciling accuracy with calibration while tracking uncertainty evolution post-earthquake. Validated on 2011 Tohoku-oki earthquake data, the model accurately predicts inundation depth and extent, demonstrating improved uncertainty quantification over deterministic approaches. This advances tsunami early warning systems by enabling real-time probabilistic forecasts.
conditional diffusion modelprobabilistic forecastingtsunami inundationuncertainty quantificationgenerative ai
ArborEnum: Decision Tree Rashomon Sets over Continuous Features
ArborEnum introduces the first exact algorithm for enumerating decision-tree Rashomon sets while preserving continuous feature structure, alongside approximate and anytime variants. The method exploits feature ordering to avoid binarization, with a relaxation for efficient approximation and progressive threshold refinement. Experiments demonstrate that prior binarization-based approaches miss up to 99% of valid trees and key features, while ArborEnum achieves 100-1000× speedups over existing methods, with approximations maintaining >95% recall at further 10× acceleration.
rashomon setsdecision treescontinuous featuresapproximate enumerationpredictive multiplicity
Adaptive Finite-Budget Training for CVaR Risk-Aware Q-Learning
The paper introduces an adaptive training controller for Conditional Value-at-Risk (CVaR) risk-aware Q-learning (RaQL) to address finite-budget instability issues. The method employs six coordinated mechanisms, including per-cell inner-step sizing and coverage-first sample allocation, without altering the CVaR estimator or Bellman fixed point. Evaluated on Bitcoin trading, the controller reduces CVaR Bellman residuals by 85% (MeanBEQ: 1.2202 to 0.1854) and achieves a Sharpe ratio of 0.9281 with 6.46% max drawdown, outperforming buy-and-hold in volatility and risk metrics.
risk-aware q-learningconditional value-at-riskbellman residualadaptive trainingfinite-budget optimization
Sample Complexity of Multicalibration for Multilevel Properties
The paper establishes tight sample complexity bounds for multicalibration across sequential properties, where each property depends on preceding ones. Using a framework that generalizes Bayes pairs, the authors prove matching upper and lower bounds of Θ̃(ε^−(k+2)) for k properties under regularity conditions. Their randomized learner achieves this with O(ε^−(k+2) + ε^−2 log|G|) samples for any finite group family G, demonstrating polynomial complexity for polynomial-size groups. The theory is instantiated for three canonical examples.
multicalibrationsample complexitysequential propertiesbayes pairsgroup families
Eliciting Intrinsic Hallucinations in LLMs via Semantically Equivalent Adversarial Attacks
The paper introduces a framework to elicit intrinsic hallucinations in LLMs by generating semantically equivalent adversarial queries, revealing vulnerabilities in retrieval-augmented generation systems. The method employs adversarial optimization under strict semantic equivalence constraints across white-box, gray-box, and black-box settings, evaluating 10 models (5 open/closed-source) on 3 datasets. Results show meaning-preserving perturbations degrade contextual faithfulness by up to 50% (e.g., GPT-5-mini), highlighting fragility in evidence grounding despite state-of-the-art performance.
intrinsic hallucinationsretrieval-augmented generationsemantic equivalencecontextual faithfulnessadversarial optimization
Geometry-Informed Parameter-Efficient Fine-Tuning of Pre-trained Molecular GNNs for Blood-Brain Barrier Permeability Prediction
BBBP-GeoPEFT introduces a geometry-informed parameter-efficient fine-tuning framework for pre-trained molecular graph neural networks (GNNs) to predict blood-brain barrier permeability. The method constructs distance-based graphs and line graphs from molecular conformers to capture spatial atom and second-order edge interactions, integrating these via lightweight auxiliary geometric graph encoders with node-wise cutoff attention and gated residual connections. This approach preserves pre-trained knowledge while incorporating geometric information with minimal parameter updates. Experiments on a BBBP dataset demonstrate competitive performance, achieving improved ROC-AUC and accuracy under random and scaffold splitting while updating only 10.1% of model parameters.
graph neural networksparameter-efficient fine-tuningblood-brain barrier permeabilitymolecular conformersline graphs
PriDyG: Privacy-preserving Dynamic Graph Inference with LLM-GNN Collaboration
The paper introduces PriDyG, a privacy-preserving framework for dynamic graph inference that combines GNN-based structural learning with LLM-based semantic reasoning under edge-level differential privacy. The method employs incremental private multi-hop aggregation to process each edge exactly once, maintaining constant privacy cost regardless of update frequency, and supplements GNN outputs with LLM predictions from node text without additional edge privacy loss. Experiments on four benchmarks demonstrate PriDyG's superior performance over geometrically decaying baselines, matching naive retraining utility while reducing cumulative privacy cost by up to 1000x.
differential privacydynamic graph inferencegnnllmmulti-hop aggregation
Dynamical Lie Algebras Cannot Describe Shallow QAOA: Cragged Terrains, Barren Plateaus, and Empirical Hardness Models
The study demonstrates that dynamical Lie algebraic (DLA) predictions of exponentially vanishing gradients fail for shallow Quantum Approximate Optimization Algorithm (QAOA) circuits applied to maximum independent set (MIS). Analyzing ∼23,000 problem instances, the authors find barren plateaus rare and instead identify 'cragged terrains'—landscapes with polynomially increasing variances—across graph families, including random and vertex-transitive graphs. Empirical hardness models, though poorly generalizing, correctly classify landscape scaling. The results challenge unitary-design-centric theories for shallow variational quantum algorithms, advocating empirically-informed models.
quantum approximate optimization algorithmdynamical lie algebrabarren plateausvariational quantum algorithmsmaximum independent set
Attention-based representations for multi-task computation
The paper establishes theoretical bounds on the number of attention heads required for multi-task computation in transformer architectures. Through two concrete scenarios—computing extremal values from a list and computing n-bit XOR—the authors prove that single-head attention requires exponentially higher embedding dimension or precision compared to multi-head variants. For XOR computation, they demonstrate a trade-off between head count and polynomial degree, showing the product must be ≥n and providing matching constructions. Results extend to symmetric Boolean functions via threshold degree analysis.
multi-head attentionthreshold degreeboolean functionsembedding dimensionxor computation
Physics-informed reduced-order modelling with equivariant spectral submanifolds
The authors introduce equivariant spectral submanifold (eSSM) reduction, an extension of spectral submanifold (SSM) reduction that incorporates symmetries of full-order models to accelerate computations and enhance robustness. By proving SSMs are naturally equivariant submanifolds and that their charts and reduced dynamics inherit group actions, they develop a novel eSSM reduction algorithm. This approach explicitly leverages model symmetries, enabling faster computation compared to traditional SSM methods while maintaining accuracy. Empirical validation on benchmark problems, including a Common Task Framework for Science test case, demonstrates the method's effectiveness in high-dimensional systems where linear techniques like Dynamic Mode Decomposition (DMD) fail.
spectral submanifoldequivariant reductiondynamic mode decompositiongroup actionsnonlinear reduced-order models
Multimodal Alignment Through Joint Kernel Entropic Gromov--Wasserstein Optimal Transport
The authors propose joint kernel entropic Gromov--Wasserstein Optimal Transport (JK-EGW), a structure-preserving alignment framework for multimodal data that leverages fine-grained similarity relationships via a global affinity kernel. JK-EGW minimizes a quadratic optimal transport objective, provides explicit control over latent space geometry, and achieves $n^{-1/2}$ parametric sample complexity. A scalable alternating procedure with low-rank kernel approximation and variational lifting enables efficient optimization. Empirical results demonstrate improved multimodal retrieval performance for post-hoc alignment of pretrained encoders in data-scarce regimes compared to baselines.
optimal transportmultimodal alignmentgromov-wassersteinentropic regularizationkernel approximation
Transferable Dual-Stream Representations for Mesoscale-Preserving Sea Surface Temperature Downscaling
EddyFlow introduces a physics-informed representation learning framework for kilometer-scale sea surface temperature downscaling, addressing limitations of existing methods that fail to preserve mesoscale variability. The approach employs dual-stream representations to balance predictive accuracy, scale-dependent structure, and regional generalization. Trained on the Gulf of St. Lawrence, EddyFlow is evaluated in zero-shot and few-shot settings on the Bay of Fundy and the Gulf of Mexico. Results show a 21% reduction in zero-shot RMSE, up to 85.6% skill relative to persistence on unseen domains, and near-ideal spectral fidelity with a PSD ratio of ≈1.00.
downscalingmesoscalerepresentation learningspectral fidelityzero-shot
Random features for Grassmannian kernel approximation with bounded rank-one projections
The authors propose a family of random feature maps for scalable kernel approximation on Grassmannian manifolds, addressing computational bottlenecks of classical Grassmannian kernels like projection and Binet-Cauchy kernels. Their method uses rank-one projections of subspace matrices with bounded non-linear transforms (periodic or binary) to approximate rotation-invariant kernels dependent on principal angles between subspaces. Theoretical guarantees show uniform approximation for sufficiently large feature dimensions, with closed-form kernels for periodic transforms and compact binary features. Structured rank-one projections via randomized FFTs reduce computation. Experiments on synthetic data and ETH-80 demonstrate preserved geometry with reduced computational and memory costs.
grassmannian manifoldrandom featuresbinet-cauchy kernelrank-one projectionssubspace classification
TIDE: A Physically Diverse 3D Turbulence Benchmark Dataset for Advancing Scientific Machine Learning
TIDE introduces a 256^3 DNS benchmark dataset for 3D incompressible turbulence, addressing limitations of prior 2D and single-realization studies with 15 configurations across eight controlled axes, ensemble variants, pressure fields, and equation-level verification. The benchmark includes five tasks, standardized baselines, generalization splits, and physical-fidelity metrics, revealing that current learned models (1) barely outperform persistence, (2) exhibit ~2× higher error than spectral solvers given true equations, and (3) show discordance between pointwise accuracy and small-scale dynamics fidelity. Forced-to-decay transfer experiments identify missing conditioning variables as a key failure mode, establishing measurable gaps in accuracy, physics fidelity, and conditioning.
turbulence3d simulationbenchmark datasetphysical dynamicsgeneralization splits
Attention-Only White-Box Transformer via LeJEPA-Based Self-Supervised Pretraining
The authors propose an attention-only white-box Transformer optimized via LeJEPA-based self-supervised pretraining, eliminating MLP layers and ISTA structure while maintaining competitive performance. The method jointly optimizes sparse rate reduction terms using ADMM, aligning with an isotropic Gaussian embedding assumption. Results show classification accuracies of 88.88% on CIFAR-10 and 63.54% on CIFAR-100, comparable to CRATE but with a 31% parameter reduction. Additionally, replacing MLP blocks with ReLU activations in standard ViTs reduces parameters by 66% without significant accuracy loss, suggesting MLP redundancy.
attention-only transformerlejepasparse rate reductionadmmwhite-box optimization
From Non-Convex Self-Concordant Regularization to Scalable Quasi-Newton Training of PINNs
We propose SCORE, a self-concordance-inspired quasi-Newton method with decrement-coupled shifted secant geometry for training physics-informed neural networks (PINNs). SCORE jointly determines a strong-Wolfe-tested candidate step and an adaptive shift for secant geometry using a single quasi-Newton decrement, avoiding Hessian construction or Hessian-vector products. Under local spectral-equivalence, it recovers normalized self-concordant rules and maintains quasi-Newton decrement comparability. Experiments on viscous Burgers, Kuramoto--Sivashinsky, Korteweg--de Vries, and complex Ginzburg--Landau equations demonstrate SCORE's superiority over BFGS and self-scaled Broyden baselines, with shifted curvature stabilization and decrement-based step selection contributing to high-accuracy refinement.
quasi-newtonself-concordancephysics-informed neural networkssecant geometrystrong-wolfe
Unscented KalmanNet: a hybrid deep learning filter with calibrated posterior covariance for nonlinear state estimation
The paper introduces Unscented KalmanNet (UKN), a hybrid deep learning filter combining the Unscented Kalman Filter (UKF) with two learned components—NoiseNet for time-varying noise covariance prediction and GainNet for analytical gain correction—while preserving sigma-point covariance recursion. UKN employs a calibration-aware training objective optimizing both state accuracy and covariance calibration. Evaluated on three synthetic systems and UZH-FPV flight data, UKN achieves 26.4-49.7% lower RMSE than UKF on synthetic tasks and reduces mean position/velocity RMSE by 22.4%/34.3% in cross-validation on real-flight sequences, with superior covariance calibration.
unscented kalman filternonlinear state estimationcovariance calibrationdeep learninghybrid filtering
From Financial Sentiment Classification to Return Predictability: A QLoRA Benchmark of Large Language Models
This study evaluates the effectiveness of financial sentiment classifiers in predicting stock returns through two experiments. First, it benchmarks TF-IDF Naive Bayes, FinBERT, Financial-RoBERTa, zero-shot Qwen2.5-7B, and QLoRA-adapted Qwen2.5-7B, LLaMA3-8B, and Mistral-7B on a unified three-class financial text dataset, with Mistral-7B achieving the highest test accuracy (0.8840) and macro-F1 (0.8771). Second, it assesses economic validity using a temporally separate Benzinga dataset, converting model probabilities into sentiment scores aligned with next-session returns. Results show small mean rank information coefficients, with FinBERT achieving the largest (0.0143), but no significant predictive advantage after statistical corrections. QLoRA improves Qwen2.5's macro-F1 from 0.7274 to 0.8615, yet classification accuracy does not translate into robust tradable signals.
financial sentiment classificationreturn predictabilityqloramacro-f1mean rank information coefficient
SiMDex: Mining Similar Egocentric Videos for Cross-Embodiment Dexterous Manipulation
SiMDex introduces a similarity-based data mining framework for selecting egocentric human videos to improve vision-language-action (VLA) models in dexterous manipulation. The method employs a three-layer recall-ranking-re-ranking pipeline to extract task-relevant subsets from ~32M human samples, operating in a morphology-agnostic action space without VLA architecture modifications. Using only ~1.49M mined samples (<5% of the pool), SiMDex boosts success rates from 47.7% to 61.1% compared to random sampling, demonstrating the efficacy of selective curation over indiscriminate data mixing.
egocentric videosdexterous manipulationdata miningvla modelsmorphology-agnostic
A Comparative Study of Feature Selection Methods for EHR Diagnosis Codes in Opioid Use Disorder Prediction
This study compares five feature selection methods for opioid use disorder (OUD) prediction from EHR diagnosis codes: recurrence enrichment, NTK-motivated early gradient sensitivity, LightGBM-SHAP, Elastic Net, and LLM-guided semantic selection. Evaluated via a unified framework, methods are assessed on predictive performance, resampling stability, and representation of infrequent codes. Results show diminishing returns beyond moderate feature budgets, with NTK sensitivity offering the best accuracy-stability tradeoff, while LLM-guided selection provides clinically meaningful but lower-performing signals.
feature selectionelectronic health recordsopioid use disorderneural tangent kernelshapley values
TS2TabPFN: Time Series Classification and Extrinsic Regression through Feature Extraction and a Tabular Foundation Model
TS2TabPFN introduces a novel framework for time series classification (TSC) and extrinsic regression (TSER) by combining explicit feature extraction with TabPFN 2.5, a state-of-the-art foundation model for tabular data. The method bridges feature engineering and end-to-end learning by leveraging TabPFN's predictive capabilities on extracted temporal features. Experimental results show statistically significant improvements over current TSER benchmarks and competitive TSC performance, establishing a new state-of-the-art for time series analysis.
time series classificationextrinsic regressionfeature extractiontabular foundation modeltemporal sequences
Understanding Fault Tolerance of Adversarially Robust Pruned Models
The paper empirically investigates interactions between pruning, adversarial training, and hardware faults in convolutional neural networks. Using a three-layer CNN on MNIST, experiments evaluate fault tolerance under adversarial attacks, pruning's effect on robustness, and joint accuracy across fault rates, perturbation magnitudes, and pruning levels. Results show adversarial training increases sensitivity to stuck-at-zero faults but improves input perturbation robustness, while pruning minimally affects fault sensitivity. The study underscores the need for co-optimizing adversarial robustness and hardware reliability.
adversarial robustnessmodel pruningfault toleranceneuromorphic hardwarestuck-at-zero faults
Mind the Cap: Output-Budget Regimes Change the Measured Multilingual Reasoning Gap
The study demonstrates that output-token budgets significantly affect measured multilingual reasoning gaps, challenging standard evaluation practices. By varying token caps (up to 2048 tokens) and analyzing Qwen3-8B and Llama-3.1-8B-Instruct on MGSM (German, Thai, Swahili), the authors show that gap magnitudes swing by 57 accuracy points across budgets, with length normalization shifting results by 38.9 points. A vocabulary extension for Thai closed 4.9 points of the gap where truncation occurred (19% of traces). Accuracy also varied by 5.1 points when announced budgets differed despite fixed enforced caps. A timing identity model predicted MGSM accuracy peaks within 0.65 points, generalizing to other benchmarks (0.92 error).
output-token budgetmultilingual reasoning gaplength normalizationvocabulary extensionaccuracy peaks
MINT: Tensor Decomposition on Stacked Recurrence Matrices for Time Series Data Mining
The paper introduces MINT, a tensor decomposition method for mining patterns in univariate and multivariate time series data via stacked recurrence matrices. The approach constructs self-similarity tensors of size $N \times (n-m+1) \times (n-m+1)$ from $N$ time series of length $n$ using subsequence windows of length $m$, then applies tensor decomposition to identify co-clustered motifs. Evaluated on mass transit, electricity demand, wind turbine, and traffic datasets, MINT successfully detects cross-sensor patterns in data with regular intervals.
tensor decompositiontime series miningrecurrence plotsself-similarity matricesco-clustering
Sublogarithmic Swap Regret in Multiplayer General-Sum Games via Hybrid Regularization
The paper presents uncoupled learning dynamics achieving $O(nm^2\sqrt{\log m\log T})$ swap regret per player in multiplayer general-sum games, the first sublogarithmic guarantee in this setting. The method combines the Blum-Mansour reduction with optimistic follow-the-regularized-leader (FTRL) using a hybrid regularizer weighting negative Shannon entropy and log-barrier, controlling prediction error and transition-matrix movement via Bregman divergence. Results show the time-averaged play distribution is an $O(nm^2\sqrt{\log m\log T}/T)$-approximate correlated equilibrium, with extensions to adversarial robustness and horizon-free settings.
swap regretcorrelated equilibriumfollow-the-regularized-leaderbregman divergencemarkov chains
Neighborhood-Aware Dual Biomedical Entity Linking
PILOT introduces a three-stage biomedical entity linking framework combining neighborhood-aware retrieval, dual reranking, and score fusion to address KB scale, mention ambiguity, and corpus-specific annotation. The method reformulates mentions and pools entity embeddings to inject ontological structure, then scores retrieved entities from surface-form and contextual views before fusion. It achieves state-of-the-art performance on five benchmarks while maintaining inference efficiency.
biomedical entity linkingknowledge baseontological structuremention ambiguityscore fusion
SpecDrop: Parameter-Free Category-Conditioned Routing for Modular Specialization
SpecDrop introduces a parameter-free category-conditioned routing mechanism for mixture-of-experts networks, addressing the bottleneck in learned routers by aligning training-signal granularity with target categories. The method assigns fixed weights to each branch based on its assigned category and a small leakage for other categories, using a category-independent denominator without learned parameters or auxiliary losses. On CIFAR-100 and ImageNet-1K, SpecDrop achieves 79.23% and 79.89% accuracy, outperforming parameter-matched baselines by +4.75 and +6.53 points respectively. Results show that category supervision enhances modular structure, with 58%/100% branch-category alignment and masking gains of 0.00/+1.06, indicating internalized output-space restriction during training. On fuzzy partitions, SpecDrop reduces to matched No-Routing controls, confirming granularity alignment as the key factor.
mixture-of-expertsparameter-free routingcategory-conditionedgranularity alignmentmodular specialization
Latent Reward Registers for Diffusion Preference Alignment
Latent Reward Registers introduce a mechanism for aligning diffusion models with human preferences by estimating terminal rewards from intermediate noisy latents. The method prepends learnable register tokens to a frozen Diffusion Transformer (DiT) input sequence, enabling dense reward signals without altering the generator's hidden states. Two alignment strategies are proposed: Reward-Gradient On-Policy Distillation (RG-OPD) for training and Reward-Guided Sampling (RGS) for inference. Empirical results show RG-OPD reduces GPU hours by up to 33x compared to online reinforcement learning baselines, while RGS achieves state-of-the-art performance among training-free methods, enhancing both alignment and perceptual metrics.
diffusion modelslatent reward registersreward-gradient on-policy distillationreward-guided samplingdiffusion transformer
Robust Low-Tubal-Rank Tensor Completion under Cross-Concentrated Sampling
The authors propose Robust Iterative t-CUR (R-ItCUR), a tensor completion method for recovering low-tubal-rank tensors from cross-concentrated sampling (t-CCS) with sparse gross corruptions. The algorithm partitions sampled tensor crosses into blocks, applies adaptive Welsch correction for outlier suppression, and updates the low-rank component via projected blockwise gradient descent, avoiding full tensor reconstruction. Experiments on synthetic data, cardiac MRI, and seismic data demonstrate robust recovery with computational efficiency. The method explicitly leverages t-CCS structure for improved performance under corruption.
tensor completionlow-tubal-rankcross-concentrated samplingrobust recoverywelsch correction
Quantization Effects on Biomedical LLM Reliability
The study quantifies how implementation choices affect biomedical LLM reliability by evaluating Mistral-7B variants (Base, BioMistral, Instruct) on PubMed RCT sentence classification (n=2000) under FP16/INT8/INT4 precision. Using four prompt templates and two scoring rules (summed vs. mean token log-likelihood), results show calibration rankings reverse between models based on scoring: BioMistral's expected calibration error increases from 0.097 to 0.289 under mean scoring, while Instruct's decreases from 0.237 to 0.096. Prompt templates cause 7-24pp accuracy variations, exceeding model differences (1.3pp between BioMistral/Instruct). INT8 quantization impacts accuracy/F1 by ≤2pp for specialized models but up to +4.2pp for Base. Temperature scaling improves calibration only under summed scoring.
quantizationcalibration errorprompt templatetoken log-likelihoodbiomedical llm
SJEPA: Learning Elegant Latent Dynamics with Hybrid Symbolic-Neural Predictors
SJEPA introduces a hybrid symbolic-neural predictive architecture for learning interpretable latent dynamics by combining symbolic laws with regularized neural corrections. The framework enforces simplicity through representation constraints and operator compression, avoiding collapse while maintaining predictive adequacy. Theoretical analysis addresses non-identifiability and collapse risks from unconstrained compression. Experiments on pendulum dynamics show joint learning yields simpler symbolic transitions with lower rollout error (37% reduction) versus post-hoc fitting, validating the trade-off between fidelity, parsimony, and neural-symbolic allocation under grammar misspecification.
joint-embeddingsymbolic-neurallatent dynamicsoperator compressiongrammar misspecification
📰 Industry Media (9)
Cloudflare Introduces Kitesurf: An Agent-First Web Browser That Runs Entirely in V8 Isolates on Cloudflare Workers
Cloudflare introduces Kitesurf, an agent-first web browser optimized for AI workloads, running entirely in V8 isolates on Cloudflare Workers without Chromium. It eliminates human-centric features like tabs and extensions, focusing on machine-readable content, scalability, and isolation against threats like prompt injection. Kitesurf splits the browser into stateless components, leveraging Blitz for HTML/CSS parsing, Boa JS for eval, and Parley for text shaping. Benchmarks show 3.1–3.8× lower CPU usage and 4.7–7.0× lower memory consumption compared to Chromium, albeit with 1.7–1.8× slower wall time. It supports existing Puppeteer and Playwright clients via a single parameter and is available as a free beta.
v8 isolatescloudflare workersblitzboa jsstateless
Adaptive Experimentation with Meta’s Ax: A Practical Coding Guide
The tutorial demonstrates adaptive experimentation using Meta's Ax platform for hyperparameter optimization of a RandomForestClassifier on synthetic data. It implements a mixed search space with integer, float, log-scaled, and categorical parameters, employing constrained Bayesian optimization and multi-objective optimization to balance accuracy against model size. Results include Pareto frontier visualization, experiment persistence via JSON, and built-in analysis tools for interpreting optimization trajectories.
bayesian optimizationpareto frontierhyperparameter tuningmulti-objective optimizationexperiment persistence
Prime Intellect Releases Prime Agent: An Open-Source RLM Harness Where Sub-Agents Are Function Calls Inside Persistent IPython Kernel
Prime Intellect introduces Prime Agent, an open-source Recursive Language Model (RLM) harness leveraging a persistent IPython kernel for sub-agent delegation via function calls. The system employs Continual Harness to dynamically manage prompts, skills, and memory as mutable state, enabling self-improvement through trajectory-based refinement. With Opus 5, it achieves 95.5% accuracy on ARC-AGI-3, surpassing the human expert baseline (95.4%), and demonstrates efficiency gains in token usage and long-context tasks. Deployment supports API-based and self-hosted models (e.g., GLM-5.2, vLLM), targeting engineering orgs and AI labs.
recursive language modelpersistent ipython kernelcontinual harnessarc-agi-3in-context learning
Microsoft’s SkillOpt Shows Optimized Agent Skill Artifacts Transfer Across Model Scales and Between Codex and Claude Code Harnesses
Microsoft's SkillOpt introduces a text-space optimizer for transferring natural-language skill artifacts across model scales and execution harnesses without modifying target models. The method employs an optimizer model that proposes bounded edits to a skill document (best_skill.md) based on scored rollouts, with edits accepted only when strictly improving performance. Results show strong cross-harness transfer (e.g., 102% retention from Codex to Claude Code on SpreadsheetBench) and variable cross-model retention (16-140% within GPT-5.4 family), with procedural skills demonstrating greater portability than reasoning-heavy ones.
text-space optimizerskill artifactscross-harness transferprocedural skillsin-context learning
End-to-End Bayesian Marketing Mix Modeling with Google Meridian: Media Measurement, ROI Analysis, and Budget Optimization
The tutorial presents an end-to-end Bayesian marketing mix modeling workflow using Google Meridian, enabling media performance measurement, ROI analysis, and budget optimization. The method involves GPU-accelerated NUTS sampling with interpretable ROI priors, adstock/saturation modeling, and posterior analysis via TensorFlow Probability. Results include channel-level ROI distributions (mean ± 90% CI), probabilistic comparisons (e.g., P(ROI_Channel1 > ROI_Channel0)), and optimized budget allocations under fixed/target-ROI constraints, validated through R-hat diagnostics and predictive accuracy metrics.
bayesian marketing mix modelingnuts samplingadstock decayhill saturation curvesroi optimization
Meta AI Releases Muse Code (Beta): A Terminal Coding Agent Powered by the New Muse Spark 1.2 Model
Meta AI introduces Muse Code (beta), a terminal coding agent powered by the Muse Spark 1.2 model, designed for complex software engineering tasks across large repositories. The system employs persistent async background agents to reduce redundant information gathering and features a replay-exact, restart-safe local event log for crash recovery. Muse Spark 1.2 was co-trained with the harness, emphasizing long-horizon tasks like whole-repository generation and GPU kernel optimization, demonstrated in a 24-hour case study with 1,000+ tool calls on NVIDIA Hopper GPUs. Evaluations on Terminal-Bench 2.1 (89 tasks) and DeepSWE v1.1 (113 tasks) show competitive performance against models like GPT-5.6 Terra and Claude Opus 5.
terminal coding agentasync background agentsreplay-exactgpu kernel optimizationco-training
NVIDIA Releases Alpamayo 2 Super: A 34B Open Vision-Language-Action Model for Robotaxis and Autonomous Driving Under OpenMDW-1.1
NVIDIA introduces Alpamayo 2 Super, a 34B-parameter vision-language-action (VLA) model for autonomous driving, designed to address long-tail multi-agent scenarios. The architecture combines a 32B Cosmos 3 Super Reasoner backbone with a 2.3B diffusion-based action decoder, processing multi-camera RGB video, egomotion history, and text to output trajectories, causal explanations, and meta-actions. Trained on 115,000 hours of driving video and 1B+ images, it achieves a Lingo-Judge score of 79.2 on LingoQA, outperforming Qwen2.5-VL 72B, Gemini 2.5 Pro, and GPT-4o. It also attains an AlpaSim score of 1.50 ± 0.13 and minADE₆ of 0.911m at 6.4s. Released under OpenMDW-1.1, it supports commercial use and fine-tuning.
vision-language-actiondiffusion decoderegomotion historylingo-judgeminade₆
CopilotKit Open Sources Channels SDK: An MIT Licensed Library That Runs Any AG-UI Agent Inside Slack And Microsoft Teams
CopilotKit released Channels SDK, an MIT-licensed library enabling AG-UI-compatible agents to operate within Slack and Microsoft Teams without platform-specific rewrites. The SDK employs a transport-layer architecture, converting agent outputs (expressed as JSX) into native UI formats (e.g., Slack Block Kit, Teams Adaptive Cards) while preserving tools, memory, and human-in-the-loop workflows. Early benchmarks support Node.js 22+ deployments with managed platform ingress via CopilotKit Intelligence, achieving interoperability with LangGraph, CrewAI, and custom HTTP agents.
ag-ui protocolgenerative uiadaptive cardstransport-layer architecturehuman-in-the-loop
Pixel-Native RAG: A Practical Guide to Visual Document Indexing
Pixel-Native RAG introduces a novel retrieval-augmented generation pipeline for visual document indexing, bypassing traditional text extraction methods. The system renders web pages and PDFs as images, divides them into overlapping tiles, and generates multimodal embeddings using SigLIP, CLIP, or Qwen3-VL. These embeddings are indexed in FAISS for efficient similarity search, enhanced by OCR-based BM25 scoring and reciprocal rank fusion. The pipeline aggregates tile-level evidence into document-level results, achieving robust retrieval quality measured by Recall@k and mean reciprocal rank. Optional vision-language model integration enables grounded answer generation from retrieved evidence tiles.
retrieval-augmented generationmultimodal embeddingsfaiss indexreciprocal rank fusionvision-language model
Generated automatically at 2026-08-07 01:00 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.