Daily Digest — 2026-08-08
287 items · 3 research labs, 273 arxiv papers, 11 industry media
🏛️ Research Labs (3)
Responding to the next frontier of critical cyber capabilities
OpenAI reports preliminary evidence that its upcoming Astra model may achieve Critical cybersecurity capabilities under its Preparedness Framework, defined as autonomously developing zero-day exploits or novel attack strategies against hardened systems. Internal evaluations show significant advancements in agentic coding and cybersecurity, prompting enhanced safeguards including isolated testing environments, restricted access, and universal monitoring of model reasoning. The company is implementing stricter security controls and collaborating with external partners to assess risks, following protocols established during previous capability transitions like biological threat modeling in 2025.
agentic codingzero-day exploitspreparedness frameworkmodel monitoringcybersecurity capabilities
How HSP GRUPPE builds AI capabilities for tax advisory
HSP GRUPPE implemented ChatGPT Enterprise across 81 organizational groups to transform tax advisory workflows, achieving 84% weekly active usage and 500,000+ conversations in six months. The organization focused on organizational transformation rather than tool deployment, establishing AI forums, governance protocols, and custom Agents for tasks like client communication and booking assistance. Results include 98.6% reported productivity gains, 40,000+ estimated annual hours of additional capacity, and improved client service. The firm plans to advance toward agentic AI workflows using ChatGPT Work, aiming to automate complex processes while maintaining professional judgment and enhancing client relationships.
chatgpt enterpriseagentic aiorganizational transformationcustom agentsworkflow automation
TutorMoments: Do AI tutors know when to help and when to hold back?
TutorMoments introduces a framework to evaluate LLMs' ability to balance pedagogical trade-offs in tutoring, specifically when to scaffold versus push for rigor. Built on 462 real math tutoring transcripts annotated by 27 experienced teachers, the method replays key decision points using LLMs as tutors and simulated students. Results show LLMs tend to over-help, with evaluation-aware prompts improving performance but not closing the gap to human tutors. The dataset, replay pipeline, and model replays are released for reproducibility. Limitations include reliance on simulated students and narrow subject focus.
llmsscaffoldingrigorpedagogical trade-offssimulated students
📜 arXiv Papers (273)
Learning When to Trust via Selective Context Preference Optimization
The paper introduces MIST, a human-annotated benchmark for evaluating selective trust in language models, and SC2W, a paired metric measuring susceptibility to misleading context. The authors propose SCOPE, a method using Direct Preference Optimization (DPO) over balanced preference pairs (clean-correct, misleading-wrong, correct-context, irrelevant-context) to reduce SC2W while preserving accuracy. Experiments show SCOPE significantly reduces susceptibility to misleading signals without compromising performance on trustworthy contexts, advocating for selective trust evaluation over mere resistance.
selective trustdirect preference optimizationcontext susceptibilitybenchmark evaluationlanguage model robustness
Tracing the Heart: An Evidence-Linked Pipeline for Heart-Failure Feature Engineering
The Nimblemind Multi-Agent System (nMAS) introduces an evidence-linked, rubric-grounded pipeline for automated heart-failure feature engineering from electronic health records (EHRs), addressing a major bottleneck (39-45% of data scientists' workload). The system processes 500 dummy patient records from nine EHR source tables, generating 132 structured and 70 rubric-scored aggregated features with verified structural integrity, rubric compliance, and provenance. Results show AUROC improvements from 0.895 to 0.963 for HFrEF and 0.870 to 0.910 for HFpEF phenotyping, with independent LLM-based rubric assessment scoring 81.5% on evidence support and methodological soundness.
electronic health recordsfeature engineeringheart failuremulti-agent systemclinical reasoning
Investigating Artificial Intelligence Digital Sovereignty in Mobile Shopping Apps: A Case Study of Nigeria
This study contributes empirical evidence on AI transparency and platform practices in Nigerian mobile shopping apps, advancing understanding of digital sovereignty in AI-driven environments. Using an interpretive approach, the research combines forensic analysis of selected Android applications with contextual document analysis to identify AI features and evaluate disclosure practices. Findings reveal widespread AI implementation but limited transparency about its use, alongside Nigeria's increasing dependence on consumer digital platforms, moderate AI awareness, and uneven interaction patterns. The analysis highlights challenges in protecting user control over digital technologies.
digital sovereigntyforensic analysisplatform transparencyai-driven environmentsdisclosure practices
An Optimal Agnostic PAC Algorithm
The paper presents an optimal agnostic PAC learning algorithm for hypothesis classes with finite VC dimension d. The learner achieves a statistically optimal risk bound, matching known lower bounds up to universal constants. For an i.i.d. sample of size n, the algorithm guarantees with probability ≥1-δ that the excess risk L(ĥ) - L* is bounded by O(√(L*(d + log(1/δ))/n) + (d + log(1/δ))/n), where L* is the minimal risk in the class. This result settles the sample complexity of agnostic PAC learning for all fixed L*.
agnostic pac learningvc dimensionrisk boundsample complexitystatistical learning theory
AV-AIVAT: 74x Cheaper Agent Evaluation with Certified Anytime-Valid Stopping in Imperfect-Information Games
AV-AIVAT introduces an anytime-valid stopping method for agent evaluation in imperfect-information games, combining AIVAT's variance reduction with Confidence Sequences (CSs) to enable early stopping while maintaining statistical guarantees. The approach uses conditional mean-zero corrections and online value models trained only on past games, avoiding bias from current game data. Empirical results show median improvements of 54× variance reduction across 71,439 HUNL hands and 74× fewer hands needed for stopping at 95% confidence. The method separates asymptotic screening (AsympCS) from exact certification (EB-CS), with a structural payoff bound demonstrated for Leduc hold'em.
anytime-valid stoppingvariance reductionconfidence sequencesimperfect-information gamesonline value model
The Low Frequency Trap: Video Language Models Fail at Simple Event Bookkeeping
We introduce trace-grounded parametric profiling to diagnose temporal reasoning failures in video language models, addressing limitations of existing benchmarks that lack executable ground truth. The method evaluates event counting across 2,190 videos with controlled event counts (N) and frequencies (F), using executable event traces for timestamp-level capability assessment. Results reveal staged temporal failures: Gemini 3.6 Flash reliably counts persistent state transitions up to 12 events at 0.5-1.0 Hz but fails completely on transient blinking events. High-count, high-frequency scenarios yield only 0.2% correct final counts and 18.1% true event recovery. Increased sampling rates improve aggregate accuracy (19.6% to 29.3%) but yield only 3.7% sequence agreement, indicating inflated scores without faithful event recovery.
trace-grounded profilingtemporal reasoningevent countingexecutable tracesparametric evaluation
Resourced Authority A Mechanism-Design Model for Participatory Governance of Deployed AI Agents
The paper proposes a formal mechanism-design model for participatory governance of deployed AI agents through resource allocation, establishing compute budgets as a governance lever. The mechanism operates via an extensive-form game where verified human stakeholders sequentially contribute governance currency in provision/rejection markets, with a funding aggregator converting contributions into breadth-weighted supports. A two-threshold gate with hysteresis authorizes compute budgets via signed licenses, bounded by a certified safety ceiling. Results characterize governable agent classes and identify agent manipulation of the electorate as a key open problem.
mechanism designparticipatory governancecompute budgetsextensive-form gamesafety ceiling
Challenges in Evaluating Explanation Methods for Static and Evolving Data
The paper identifies limitations in Explainable Artificial Intelligence (XAI) evaluation methodologies, demonstrated through DetoxAI, an image recognition system for bias detection and concept unlearning. It presents a human-grounded evaluation framework for image classification explanation methods and explores adaptation techniques for evolving data streams with concept drift. The study discusses experiences in modifying counterfactual explanations for dynamic environments and highlights challenges in tracking the co-evolution of data, models, and explanations.
explainable artificial intelligenceconcept unlearningcounterfactual explanationsconcept drifthuman-grounded evaluation
TRAJDEBUG: Tracing Error Lifecycle to Identify Critical Failures in Long-Horizon Agent Trajectories
TrajDebug introduces a framework for tracing error lifecycles in long-horizon agent trajectories, addressing challenges in critical error detection through multi-granularity history compression and evidence-based error identification. It traces each error's resolution status and terminal impact to attribute critical failures accurately. The authors construct TrajErrBench, a benchmark of 486 annotated failed trajectories from Tau2Bench and SWE-Bench Pro, covering tool-use and coding scenarios. Experiments show TrajDebug outperforms existing baselines, providing actionable feedback for improving agent success. Code and data will be released to support further research.
error lifecyclehistory compressioncritical attributiontool-use scenariosagent trajectories
Tytan: Interactive Neurosymbolic Construction of Analytic Semantic Schemas from Relational Data
TYTAN introduces a neurosymbolic system for automating the construction of analytic semantic schemas from relational databases, addressing the scalability bottleneck of manual schema creation. Combining symbolic database analysis with LLM-based semantic inference, TYTAN proposes entities, assigns roles, and names elements, engaging users with targeted natural-language questions when ambiguity arises. Evaluated on eight databases, TYTAN achieves 100% coverage of entities and attributes, 100% correct retrieval instructions (1,678/1,678), and 92-100% accuracy in semantic role assignment. In a blind test on a ten-table database, TYTAN successfully recovers entity structures and meets 100% of annotators' expectations.
neurosymbolicsemantic schemarelational databasellm-based inferenceentity proposal
Benchmarking the Benchmarks: Evaluating Benchmarks for Conversational Agents
The authors propose a reference-free framework for evaluating benchmarks of task-oriented conversational agents, addressing issues like inconsistency, simplicity, and limited policy coverage. The method employs LLM judges to assess benchmark quality along dimensions of consistency, complexity, and policy coverage, providing diagnostic insights. Validation shows agreement with human annotations and distinguishes quality levels across domains, synthetic benchmarks, and manually curated benchmarks, demonstrating robustness across varying LLM capabilities and controlled perturbations.
benchmark evaluationconversational agentsllm judgespolicy coveragetask-oriented dialogue
Does FLAIR super-resolution erase or hallucinate small white-matter lesions?
This study evaluates whether FLAIR super-resolution (SR) preserves small white-matter hyperintensities (WMH) or introduces artifacts, using 1-mm isotropic high-resolution FLAIR scans from 29 ADNI cohort participants. Simulated 3mm and 5mm thick-slice acquisitions were reconstructed using multi-contrast implicit neural representation (INR), single-contrast self-supervised ECLARE, and cubic interpolation. WMH segmentation was performed using MARS-WMH, the most sensitive method for small lesions. Results show SR primarily erases small real lesions rather than hallucinating absent ones, with erasure increasing with slice thickness. ECLARE outperformed INR and cubic interpolation in recovering small lesion signal, though all reconstructions improved detection over raw thick slices.
super-resolutionwhite-matter hyperintensitiesimplicit neural representationeclaremars-wmh
Beyond Top-K: Replacing Black-Box Retrieval with Interpretable Agentic Operations
We propose READ (Reliable Embedding-free Agentic Document-search), an interpretable alternative to black-box retrieval methods for long, structured documents like financial reports. READ replaces embedding-based top-k retrieval with three deterministic operations: normalized lexical search, structural navigation, and bounded span reads, exposed via the Model Context Protocol for auditability. Evaluated on a 780-page government financial report with 51 verified questions, READ achieves 58.8% accuracy, outperforming dense retrieval (15.7%, p_Holm = 2 x 10^-5) and a top-k-equipped agent (27.5%). Results demonstrate that the performance gain stems from READ's interface design rather than iterative refinement, while BM25 remains statistically indistinguishable from READ.
retrieval-augmented generationmodel context protocolembedding-free retrievalstructural navigationbounded span reads
HarnessOpt-Bench: Evaluating LLMs at Harness Optimization
The authors introduce HarnessOpt-Bench, a benchmark for evaluating LLMs' capability in harness optimization—iterative improvement of prompts, tools, and control flow surrounding LLMs under constrained evaluation budgets. The method provides optimizers (LLMs + coding harness) with seed harnesses and graded feedback, enforcing evaluation boundaries via trusted execution. Testing 5 frontier LLMs across 4 tasks (111 runs), results show optimizer models differentiate more than coding harnesses, native harnesses lack consistent superiority, and gains vary by task and seed regime. This establishes harness optimization as a measurable, discriminative capability with improvement potential.
harness optimizationtrusted execution environmentevaluation boundarygraded feedbackseed harness
Bias Analysis of L2 Speaking Assessment Systems Using Concept Activation Vectors
This work extends Concept Activation Vectors (CAVs) to analyze bias in neural L2 speaking assessment systems, distinguishing between concept encoding and score influence. The authors apply CAVs to a BERT-based text grader and a Whisper-based multimodal grader, investigating whether sparse autoencoders (SAEs) improve concept recoverability in complex embedding spaces. Results show that concept recoverability and sensitivity vary by architecture, with SAEs enhancing linear separability but reducing activation-space sensitivity, particularly in low-dimensional layers. The study emphasizes the importance of separating concept recoverability from concept influence when auditing bias in automated speaking assessment systems.
concept activation vectorssparse autoencodersspeaking assessmentmultimodal graderbias analysis
QuanTiMedAI: Quantum-Enhanced Time-Series Model guided by Agentic AI for Cardiac Arrest Mortality Prediction
QuanTiMedAI introduces a quantum-agentic framework for cardiac arrest mortality prediction, combining an agentic LLM for feature discovery with a compact quantum recurrent network for temporal modeling. The method leverages clinically informed feature selection and quantum-enhanced nonlinear processing, achieving an AUROC of 0.852 (2.9% improvement over SOTA) using only 605 parameters. Experiments on MIMIC-IV validate the architecture's efficiency and performance gains through systematic ablation.
quantum-agentictemporal modelingfeature selectionquantum recurrent networkmortality prediction
BaKron: Efficient Quantization with Kronecker-Factored Hessians
BaKron introduces an efficient solver for neural network quantization leveraging Kronecker-factored Hessian approximations, reducing computational complexity while exploiting richer curvature information. The method combines anti-diagonal parallelism with a recursive divide-and-conquer construction, achieving $O(m+n)$ sequential steps and $O(mn(m+n))$ total work for an $m\times n$ weight matrix, matching GPTQ's cubic scaling. BaKron is modular with respect to both the base quantizer and Hessian estimator, and practical benchmarks demonstrate its effectiveness across various Hessians. The paper also presents an efficient technique for computing these Hessians, validated through experimental evaluation.
kronecker-factored hessianadaptive roundingneural network quantizationanti-diagonal parallelismrecursive divide-and-conquer
The Illusion of Visual Tool-Use: A Causal Audit of Thinking with Images
The study investigates the causal effectiveness of visual tool-use in multimodal LLMs, revealing policy miscalibration despite marginal accuracy gains. Using a causal graph framework, the authors audit visual tool-use through interventions at policy, trajectory, and step levels, introducing the Visual Evidence Gain estimand to isolate individual observation contributions. Across six models and five perception benchmarks, they identify two failure modes: Calling Without Looking (observations lack causal effect) and Looking Without Planning (informative but incoherent call schedules). Trajectory-level diagnostics show accuracy gains concentrated in a Calibrated minority, exposing the illusion of visual tool-use. Code is available at OpenCausaLab/CauAudit.
multimodal llmscausal graphvisual evidence gainpolicy miscalibrationtrajectory-level diagnostics
Improving the Realism of Synthetic Clinical Benchmarks Under Utility Constraints
The paper proposes utility-constrained realism improvement for synthetic clinical benchmarks, addressing structural unrealism while preserving operational utility. The method formulates benchmark revision as an optimization problem with realism metrics (missingness structure, simplicity, plausibility, population alignment) constrained by utility floors. Evaluated on a Synthea-derived care-gap benchmark, deterministic revisions improved realism (reducing sampled-pair missingness from 79.44%, increasing actionable rows from 12.75%) while maintaining utility, unlike naive densification. Results distinguish between internal realism and source fidelity, advocating explicit realism optimization alongside utility constraints.
synthetic benchmarksutility constraintsmissingness structureclinical airealism metrics
Toward Deployable Bangla Sign Language Recognition with Expert-Validated Data and a Lightweight Attention-Based Model
We introduce RSBdSL38, a deployable Bangla Sign Language (BdSL) recognition system comprising 10,874 expert-validated images spanning all 38 BdSL hand signs and a lightweight attention-based convolutional network of 298,470 parameters. The model employs grouped bottleneck residual blocks, channel and spatial attention, a multi-scale depthwise hand-feature block, dual pooling, and Swish activations, achieving 96.37% accuracy on RSBdSL38 and 97.04% on a merged corpus. It outperforms nine ImageNet-pretrained architectures with 8.5 to 68x fewer parameters and 1.3 to 21.7x fewer MACs. Quantized to 0.48 MB, it runs at 3.98 ms per image on a smartphone. The dataset, code, and models are publicly released.
bangla sign languageattention-based convolutional networkgrouped bottleneck residual blocksmulti-scale depthwise hand-feature blockdual pooling
DASH: Divergence-Adaptive Supervision Horizons for On-Policy Self-Distillation of Reasoning Models
The paper introduces Divergence-Adaptive Supervision Horizons (DASH), a method to improve on-policy self-distillation (OPSD) in reinforcement learning with verifiable rewards (RLVR) for reasoning models. DASH addresses the limitation of standard OPSD, which uniformly weights local divergences, by dynamically adjusting token-level supervision weights based on the temporal evolution of discrepancies between teacher and student distributions. This is achieved through adaptive propagation gates that aggregate multi-step divergence sequences. Experiments on three mathematical reasoning benchmarks across three model scales demonstrate consistent improvements over vanilla OPSD, with no additional computational overhead. Code is available at https://github.com/DBtxy/DASH-OPSD.
reinforcement learningself-distillationdivergence-adaptivetoken-level supervisionmathematical reasoning
PRISM: Distribution-Gated Flow Matching for Controllable Unpaired Image Translation
PRISM introduces a GAN-free flow-matching framework for unpaired image-to-image translation, replacing global noise control with a learned per-feature gate derived from standardized distance to target feature distributions. The gate controls both initialization, mixing real source latent with task-matched corruption, and transport timing during ODE integration, enabling local override at inference via text or detectors without retraining. Evaluated on five benchmarks (AFHQ cat->dog, CelebA-HQ appearance translation, day->night relighting, virtual staining, breast frozen->permanent histopathology), PRISM achieves the best Inception FID and KID on four benchmarks and competitive results on the fifth, with histopathology yielding the closest nuclei-count ratio to ideal, balancing target realism and structural preservation.
flow-matchingunpaired translationfeature gateode integrationhistopathology
Depth-Guided Video Object Counting in Crowded Scenes
The paper introduces Depth-Guided Detector (DG-Det), a method for robust video object counting in crowded scenes by integrating depth cues with multi-scale RGB-D cross-attention and explicit occlusion prediction. A unified de-duplication framework addresses cross-frame redundancy, and a new RGB-D Video Object Counting dataset is released. Experiments show a 62.01% reduction in MAE and consistent RMSE improvements over baselines.
depth-guided detectionrgb-d cross-attentionocclusion predictionvideo object countingde-duplication framework
From Passive Mirrors to Active Agents: Holonic Digital Twins for Physical AI over Networks
The paper introduces Holonic Digital Twins Networks (HDT-Nets), a framework enabling real-time physical AI inference through holonic agents that actively reason about their environment rather than passively mirroring physical assets. HDT-Nets employ hierarchical structures spanning physical agents and network edges, facilitating local autonomous reasoning and cooperative intelligence among neighboring HDTs. The framework integrates causal Markov blankets for spatiotemporal coordination, active inference for unified perception-action-learning cycles, category theory for semantic preservation across heterogeneous agents, and integrated information theory to quantify collective intelligence gains. This approach addresses the limitations of current AI tools in maintaining reliable world models and generalizing to unseen scenarios in physical systems.
holonic digital twinsactive inferencecausal markov blanketsintegrated information theorycategory theory
TS-RAG: Retrieval Augmented Generation for Time Series Forecasting
TS-RAG introduces retrieval-augmented generation (RAG) for time series forecasting, addressing limitations of existing models through specialized reference tokens that fuse input sequences with retrieved similar sequences. The method enhances transformer-based architectures by capturing complex temporal dynamics more robustly than simple concatenation approaches. Experiments show TS-RAG achieves state-of-the-art performance across multiple real-world forecasting benchmarks.
retrieval-augmented generationtime series forecastingtransformer-based architecturesreference tokenstemporal dynamics
Continual Learning in Transition
The paper characterizes a paradigm shift in continual learning (CL) from parameter-centric adaptation to system-level evolution through a tri-axial framework analyzing When, How, and Where learning occurs. It examines CL across temporal stages (pre-training, post-training, inference-time), optimization mechanics (off-policy, on-policy, beyond-gradient), and spatial domains (internal parameters, external structural constraints). The framework systematically surveys emerging CL methods, including on-policy learning, test-time training, and external harness components like memory and skill libraries. Results highlight the transition from static parameter adaptation to dynamic system-level capabilities, identifying key challenges and future directions for CL research.
continual learningsystem-level adaptationon-policy learningtest-time trainingexternal harness
What Current AI Benchmarks Leave Unmeasured: Modality, Search, Citations, and Implications (for Safety Evaluations)
This study critiques current LLM benchmark practices by demonstrating how modality, search, and multi-run consistency affect model behavior in safety evaluations. Using ChatGPT's chat UI and OpenAI API, the authors evaluated 401 prompts from BBQ and SafetyBench across 4,812 responses, analyzing accuracy, response consistency, citation grounding, and abstention behavior. Results show chat UI was less accurate than API without search, search reduced accuracy by up to 8%, and 21% of prompts yielded inconsistent responses across runs. Modalities also differed in citation grounding and abstention patterns, highlighting the need for more comprehensive safety evaluations.
llm benchmarksmodality effectsresponse consistencycitation groundingabstention behavior
EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning
EnvACE proposes world rehearsal as a method for training LLM agents without external environment interaction, alternating between action generation and environment response simulation within the policy. The jointly optimized policy internalizes environment dynamics, forming an agent world model that supports decision-making. Evaluated on BFCL-v4, tau^2-Bench, VitaBench, and FinMCP-Bench, EnvACE outperforms environment-scaling baselines, with controlled studies confirming scalability and test-time private rehearsal yielding additional gains.
world rehearsalagentic reinforcement learningllm agentsenvironment dynamicstask-success rewards
Comparative Approaches to Agent Retrieval over Large Skill Libraries
The study compares two approaches for agent skill retrieval from large libraries (690 skills): a hybrid ranker (lexical + dense embeddings) and a typed knowledge graph encoding workflow relations. On 117 realistic queries, the hybrid ranker achieves 73.5% hit@5 accuracy (±8.0%), while the graph-based approach underperforms by 11.2 points (p=0.0007) due to redundant edge connections (98.6% from embedding neighborhoods) and limited reach (73% of ranker misses unreachable). The graph adds no value over local embeddings, and author-written queries overestimate hit@5 by up to 44 points. The work identifies conditions where structural interdependence fails to enhance retrieval over strong rankers.
skill retrievalhybrid rankerknowledge graphembedding neighborhoodshit@5 accuracy
MicroEvo: Knowledge-Guided LLM Sampling for Efficient Microarchitecture Design Space Exploration
MicroEvo introduces a knowledge-guided framework combining off-the-shelf LLMs with Monte Carlo Tree Search (MCTS) for efficient microarchitecture design space exploration. The method integrates LLM-driven evolutionary operators, a Pareto-aware tree policy, active knowledge accumulation, and state-aware directives to optimize multi-objective search. Experiments demonstrate a 36.2% improvement in Pareto-front quality over NSGA-II and 10.6× higher search efficiency, with scalability to industrial-scale cores.
microarchitecturemonte carlo tree searchpareto optimizationknowledge-guided samplingdesign space exploration
Schema-Guided Hierarchical Information Extraction and Semantic Evaluation Using Generative AI
A schema-guided framework for hierarchical information extraction and semantic evaluation using generative AI is proposed, enabling structured data extraction from unstructured text in zero-shot mode. The method employs a domain-specific schema as an information model, coupled with a path-based semantic matching algorithm for aligning nested, variable-cardinality attributes with a gold standard. Semantic comparison leverages generative AI, classifying results as exact, semantic, useful, or non-matches. Experiments on NICE health technology assessment documents achieved >90% F1 score for 12 out of 14 attributes, reducing extraction time by ~30x compared to human experts. The framework demonstrates generalizability across generative models and transferability across HTA organizations and languages.
schema-guidedzero-shotsemantic matchingvariable-cardinalitygenerative ai
Audio-to-Score Transcription using Pre-trained Features, Data Augmentation, and the New SheetSage-A2S Dataset
The paper introduces SheetSage-A2S, a novel dataset comprising 61 hours of audio with **kern score encodings for 9,468 clips from 6,066 popular music songs, addressing the gap in audio-to-score (A2S) research for popular music. The proposed A2S model leverages MuQ, a pretrained feature-extraction model, and data augmentation to enhance generalization. It achieves a 4.98% symbol error rate (SER) on classical music (Quartets collection), outperforming the prior state-of-the-art (15.3% SER), and 20.92% SER on SheetSage-A2S, establishing a benchmark for popular music A2S.
audio-to-scorepretrained featuresdata augmentationsymbol error ratemusic transcription
iARCS: Iterative Agentic RL for Controllable 3D Scene Generation
The paper introduces iARCS, an iterative agentic reinforcement learning framework for controllable 3D scene generation that addresses functional constraints critical for downstream tasks. The method employs a two-stage approach: universal-reward pretraining for physical plausibility and layout quality, followed by task-specific fine-tuning with LLM-generated reward programs refined via training feedback. Results demonstrate improved constraint fidelity for walkability, reachability, and clearance tasks, effective task-specific optimization, and competitive scene diversity. Additionally, iARCS-generated data enhances base generator performance, validating its utility as a synthetic data generation tool.
3d scene generationreinforcement learningconstraint optimizationllm-generated rewardssynthetic data
Visual Grounding in Zero-Shot Vision-Language Control
The study evaluates visual grounding in zero-shot vision-language control by systematically ablating visual inputs across 16 models in three simulators (32,874 trials). Through blind controls, reflection tests, and non-visual baselines, it reveals most direct-action VLMs fail basic grounding criteria, with constant policies often outperforming them. However, a deterministic visual-only positive control demonstrates sufficient visual information exists (0.090m MAE for gap estimation). A post-hoc symmetry-consensus guardian achieves 0.954 balanced accuracy on held-out data, suggesting VLMs function as bounded hazard assistants rather than reliable monolithic controllers.
visual groundingzero-shot controlinput ablationsymmetry-consensushazard detection
Learning Globally Reusable Skills for Coding Agents
The paper introduces GSE, a globalized skill evolution framework for LLM-based coding agents that jointly optimizes skill compatibility and generalization. GSE maintains a Skill Relation Graph (SRG) to model inter-skill relationships and employs cluster-based consolidation with replay-driven verification to prevent overfitting. Evaluated on bug-revealing test generation and false-positive bug report filtering, GSE improves precision by 6.1%-96.4% and recall by 13.1%-180.0% over baselines across OpenHands and mini-SWE-agent, with a 61.4% F1-score gain in industrial deployment.
skill evolutionskill relation graphreplay-driven verificationcoding agentsgeneralization
Reducing belief in conspiracy theories as they unfold using large language models
This work demonstrates that multi-turn conversational interventions with large language models (LLMs) can reduce belief in emerging conspiracy theories following high-profile events. Across two experiments (N=472 and N=1035) conducted after the 2024 Trump assassination attempt and 2025 Kirk assassination, conspiratorial U.S. adults engaged with an LLM optimized for belief reduction showed significantly decreased conspiracy beliefs compared to control conditions (irrelevant LLM dialogue or static fact sheets). Longitudinal follow-ups revealed persistent effects, with reduced susceptibility to unrelated conspiracies 1-2 months post-intervention. The findings suggest LLMs may offer scalable cognitive interventions against real-time misinformation.
large language modelsconspiracy theoriesmisinformation interventioncognitive debiasinglongitudinal effects
CogVis: Must Open-Vocabulary Change Detection Perceive the Scene Anew for Every Query?
CogVis introduces a cognitive memory-guided framework for Open-Vocabulary Change Detection (OVCD), reformulating it as a perception-memory-verification paradigm. The method employs a Scene Change Perceptron (SCP) to extract reusable, category-agnostic change priors from frozen bi-temporal features, decoupling temporal evidence from semantic decisions. A Semantic Memory Calibrator (SMC) dynamically estimates image-query-specific decision thresholds to compensate for category-dependent score shifts. An Adaptive Region Filter (ARF) filters candidates using semantic, temporal, and structural reliability. Evaluated across seven benchmarks, CogVis achieves state-of-the-art performance and improves inference throughput by 28.50% by avoiding redundant temporal perception across queries.
open-vocabulary change detectionscene change perceptronsemantic memory calibratoradaptive region filterbi-temporal features
PaDoc: Layout-Grounded Parallel Decoding for Document Parsing
PaDoc introduces a layout-grounded parallel decoding approach for document parsing that eliminates serial dependencies while maintaining full-page context. The method employs a prefix-conditioned factorization with concurrent layout and content streams, implemented via packed variable-length ancestor attention and masked parallel decoding in a single multimodal LLM. On OmniDocBench Full, PaDoc achieves 91.1 layout F1 and 94.24 Overall score, with 67.4-118% higher throughput and 39.2-54.9% lower P95 latency versus sequential baselines at five concurrency levels.
document parsingparallel decodinglayout-groundedmultimodal llmancestor attention
FinEvo-Bench: A Longitudinal Benchmark for Self-Evolving Agents in Professional Financial Workflows
FinEvo-Bench introduces a longitudinal benchmark for evaluating self-evolving agents in professional financial workflows, addressing gaps in existing benchmarks by covering professional workflows, open-ended deliverables, and multi-aspect evaluation. The benchmark comprises 120 real-case-grounded tasks across 20 business scenes in six financial domains, with institution-provided procedures and manually reviewed rubrics. Four self-evolving agent scaffolds were compared using Qwen3.7-Max, with outputs evaluated by Claude Code backed by Claude Opus 4.6. Results show evolving conditions improve scores by 9.33-19.37 points and reduce compliance issues by 0.12-0.44 per task, with Letta achieving the highest evolved score (91.65) and Codex the largest self-evolution gain (+19.37).
self-evolving agentslongitudinal benchmarkfinancial workflowsrubric feedbackcompliance issues
Hardware Keystores for AI Agent Signing Workflows: A Zero-Trust MCP Enforcement Architecture
The paper introduces a hardware-based keystore architecture for securing AI agent cryptographic workflows, enforcing both key confidentiality and content-aware authorization. The method replaces software-resident keys with hardware-confined keys accessed via a vendor-neutral PKCS#11 interface, supported by a five-layer Zero-Trust enforcement stack comprising session identity, scope bounds, semantic validation, taint tracking, and hardware execution boundaries. Evaluation against 12 injection scenarios derived from AgentDojo's ImportantInstructionsAttack template shows a reduction in Attack Success Rate (ASR) from 19.3% in baseline mode to 0% in the protected setup, with zero false positives across benign task scenarios.
hardware keystorezero-trustpkcs#11semantic validationtaint tracking
Contextual Information Policy Optimization for Search Agents
The paper introduces Contextual Information Policy Optimization (CIPO), a reinforcement learning framework that aligns policy optimization with evidence use in search agents. CIPO assigns turn-level credit to reasoning actions influenced by retrieved information, combining this with global outcome rewards to discourage prior-driven reasoning while preserving answer correctness. Evaluated on seven benchmarks, CIPO reduces confirmation bias and improves performance without requiring human annotations or additional reward models.
contextual information policy optimizationsearch agentsreinforcement learningprior-driven reasoningevidence-use signal
Poli-Bias: Understanding and Measuring Large Language Model Biases in International Political Conflicts
We introduce Poli-Bias, a counterfactual framework for measuring political bias in large language models (LLMs) through systematic country-swapping across geopolitical relationships, legal violations, and reasoning tasks. The framework decomposes response disparities into five interpretable dimensions, revealing how unequal treatment manifests in descriptions, evaluations, and legal defenses. Evaluations across 13 LLMs from diverse families and sizes demonstrate that country identities and user affiliations systematically influence responses to legally equivalent conflict scenarios. Poli-Bias thus provides a fine-grained method for auditing political even-handedness and sycophancy in LLMs.
counterfactual frameworkpolitical biaslarge language modelsgeopolitical relationshipslegal violations
Is Self-Pretraining really useful to improve diagnosis in medical Time Series?
This study demonstrates that Self-PreTraining (SPT) consistently improves transformer-based models' performance on medical time-series tasks, even with univariate inputs. The authors evaluate SPT using four masking-based objectives across three datasets: rehabilitation robotics (Camargo), stress detection (Non-EEG Stress), and Parkinson's disease detection (Gait Parkinson's Disease). Results show classification accuracy improvements of 0-6 percentage points, with deeper models benefiting more from enriched temporal representations learned during pre-training. SPT proves effective without task-specific architectural modifications, suggesting its utility for enhancing robustness in data-limited clinical settings.
self-pretrainingtransformer architecturesmasking-based objectivestemporal representationsmedical time-series
Mind the Gaps: Mixture-of-Minds for Human Simulation
The paper introduces Anacreon, an audience simulation model that improves individual-level prediction accuracy in narrow domains. The method combines authorship embeddings, cluster-specific adapters (mixture of minds) on a Gemma-4 12B base, and augmentation with demographics, psychological traits, and chain-of-emotion from public text. It addresses prompt brittleness via response option shuffling and reduces positive bias through balanced training. On an external survey, Anacreon achieves state-of-the-art ordinal alignment of 0.775 with minimal residual bias, advancing individual-level simulation for aggregate insights.
audience simulationmixture of mindsauthorship embeddingordinal alignmentchain-of-emotion
Beyond Sequence Order: Syntax-Informed Positional Embeddings for Transformers
The paper introduces Syntax-informed Positional Embeddings (SiPE), a method that injects syntactic structure from dependency parses into Transformer positional embeddings across three dominant PE families (absolute, relative, rotary) without modifying self-attention. SiPE learns a lightweight syntactic prior during pretraining, with optimal injection points varying by architecture: multiplicative coupling with relative-position terms for autoregressive decoders, and direct addition to input embeddings for encoders. Experiments show SiPE improves SyntaxGym performance by 10.3%, reduces perplexity by 9.0%, and boosts GLUE scores by 8.2%, while maintaining inference efficiency by conditioning on a single parse.
positional embeddingssyntactic priordependency parsesautoregressive decoderssyntaxgym
From Siloed Algorithms to Compliance-First Agentic Platforms: A Multi-Layered Architecture for Hospital AI Systems
This research proposes a multi-layered Agentic AI architecture for hospitals, addressing governance gaps and fragmented data in AI deployments. The architecture introduces three key layers: an Agent Orchestration Layer for multi-agent workflows, a Compliance and Policy Layer centralizing policy-as-code for regulatory standards, and a Privacy-Preserving Data Fabric integrating federated learning and differential privacy. Using a synthetic hospital dataset and a prototype implementation, the study demonstrates end-to-end orchestration of triage risk prediction and workflow optimization, achieving reductions in task turnaround times and manual documentation effort while maintaining policy-guarded data access.
agentic aifederated learningdifferential privacypolicy-as-codehospital information management system
ECHO: A Locally-Deployable Agentic Health Assistant with Temporal Memory, Safety Guardrails, and Speech Assessment
ECHO introduces a locally-deployable health assistant integrating three modules: (1) an agentic chatbot with ReAct loop, 17 clinical tools, and temporal knowledge graph (94.9% tool-execution pass rate); (2) a hybrid safety layer combining rule-based filtering (1ms latency) and signed GNN classification (88.8% accuracy, 90.6% unsafe recall); (3) a multimodal speech module with Whisper-BERT fusion (0.652 macro F1). The system operates on consumer hardware with GDPR/KVKK compliance.
agentic chatbottemporal knowledge graphsigned gnnreact loopmultimodal fusion
Evaluating Investment Logic in Large Language Models: A Real-World Benchmark Towards Personalzied Financial Agents
The authors introduce extsc{InvestLogicBench}, a benchmark for evaluating financial large language models (LLMs) through personalized investment logic. The benchmark comprises 201,247 documented decisions from 151 real-world investors, structured as P$ ightarrow$E$ ightarrow$R$ ightarrow$D$ ightarrow$O traces: Profile, Events, Reasoning, Decision, and Outcome. It supports profile-conditioned generation, comprehension, and end-to-end replay. Evaluation of four leading LLMs reveals logical plausibility scores near 4/5, while event grounding scores range from 0.8 to 2.8/5, indicating polished but weakly grounded reasoning. The authors advocate for P$ ightarrow$E$ ightarrow$R$ ightarrow$D$ ightarrow$O as a data-system interface, emphasizing versioned profiles, temporal provenance, and inspectable retrieval.
large language modelsinvestment logicbenchmarktemporal provenanceprofile-conditioned generation
Does Latent Context Help? A Controlled Evaluation of Inverse Reinforcement Learning in Arctic Shipping
This work evaluates the utility of latent context variables in inverse reinforcement learning (IRL) for Arctic shipping navigation, comparing linear, nonlinear, and latent-context reward models on 3,186 AIS-derived voyages across nine seasons. The nonlinear reward model improves held-out likelihood by 50.9% over the linear baseline, but adding vessel-specific latent context reduces performance by 16.5%. Behavioral analysis and feature-hiding ablation reveal that observable route and environmental conditions explain most behavioral variation, not hidden vessel-specific factors. The study demonstrates that predictive accuracy, route fidelity, and reward transfer yield divergent model rankings, highlighting the need for multi-metric evaluation in safety-critical AI deployment.
inverse reinforcement learninglatent contextarctic shippingais-derived voyagesbehavioral analysis
Signal or Spurious Cue? A Randomized Audit of Survey-Country Metadata in LLM Social Inference
This study investigates the impact of survey-country metadata on LLM social inference, distinguishing between informative and spurious cues. A randomized audit evaluates whether disclosing a label's random origin reduces its country-directed influence and whether verified metadata improves forecast accuracy. The experiment spans five fixed API models, six countries, and seven development-selected targets, using a 72-record panel. Results show that both opaque and disclosed-random labels induce country-direction shifts of 0.214, with minimal attenuation (0.0003). Verified metadata reduces Brier loss by 0.040, while random-label regret remains negligible. The findings suggest verified metadata is useful, but disclosure does not reliably mitigate random-label uptake.
survey-country metadatallm social inferencebrier lossrandom-label uptakeverified metadata
Domain-Grounded Candidate Selection for Agentic Image Editing: A Shadow Removal Case
The study introduces a physics-grounded candidate-selection pipeline for agentic image editing, specifically addressing shadow removal. Leveraging a commercial generative editor, the pipeline generates multiple candidates, filters failures, and selects results balancing shadow removal with scene preservation. By grounding the process in shadow-formation physics—treating shadows as illumination effects rather than material or object structure—the method improves reliability and consistency. Evaluated on the ShadowRemovalRefine benchmark, the pipeline achieves a CDD of 0.0075, reducing CDD by at least 47% over prior methods. This demonstrates that commercial vision-language models benefit from classical low-level vision priors to constrain physically underconstrained generation.
shadow removalvision-language modelscandidate-selection pipelinephysics-informed priorscdd metric
When History Lies: Evaluating and Improving Tool Use under Misleading Multi-Turn Histories
The paper introduces a benchmark (bench) and method (ours) to address tool-use failures in LLMs caused by misleading multi-turn histories. Bench provides synchronized Original, Polluted, and Oracle State views to evaluate failures in decision state, entity binding, and interface execution. The proposed method transfers Oracle-conditioned teacher policies to students observing polluted history via soft supervision on student-generated prefixes, achieving 87.0% Balanced Tool-Use Accuracy on Qwen3-1.7B, outperforming baselines. Scaling experiments show consistent improvements, with an 8B teacher raising a 1.7B student to 91.9%. The method generalizes to clean histories, unseen functions, and external benchmarks.
tool-usemulti-turn historypolicy transfersoft supervisionbenchmark
Integrating Implicit and Explicit Relational Biases through Graph-Based Multiple Instance Learning: A Case Study in Skin Lesion Diagnosis
This work proposes a dual-level relational framework combining implicit and explicit structural modeling for skin lesion diagnosis. The method first learns implicit inter-patch relationships via a convolutional masked autoencoder on EfficientNetB3 features, then applies explicit graph-based message passing with various topologies (grid, random, k-NN). Evaluated on ISIC-2018 and ISIC-2019, the integrated approach improves upon baseline EfficientNetB3 (76.17% balanced accuracy) to 79.27% (ISIC-2018) and 60.67% (ISIC-2019), demonstrating gains from combining self-supervised patch relations with graph attention networks.
relational inductive biasesconvolutional masked autoencodergraph attention networkpatch-based learningskin lesion diagnosis
FormBharo: Designing and Evaluating a Voice Agent for Conversational Form Filling in Rural India
FormBharo introduces a voice agent for conversational form filling in rural India, combining Large Language Models (LLMs) with rule-based validation and flow control to address literacy barriers. The system, piloted with ARMMAN, targets low-income Hindi-speaking mothers for antenatal and postnatal care enrollment. FormVoiceAgentBench, a benchmark with 3,760 multi-turn conversation tests, evaluates transcription, extraction, and reply generation components. Results show form completion drops by ~41 points with error-prone transcripts, but rule-based controls mitigate extraction errors, enabling smaller models to match frontier models. End-to-end evaluation reveals error propagation and cancellation, necessitating Pareto-based weighted-sum scalarization for optimal model selection across accuracy, cost, and latency.
voice agentlarge language modelsrule-based validationmulti-turn conversationpareto-based scalarization
From Economic Agents to Agentic Economies: A Systems Blueprint for Economic World Models
The paper proposes a six-level capability ladder for Economic World Models (EWMs), generative systems simulating economies via heterogeneous agents, market mechanisms, and institutional evolution. It organizes EWMs from rule-based agents to LLM-based adaptive agents, self-evolving systems, and empirically aligned sim-to-real twins. A literature survey reveals current work predominantly addresses lower-level agent simulations, with limited progress in endogenous institution modeling and empirical validation. The authors provide an implementation blueprint and release curated resources to advance high-fidelity economic sandboxes for human and AI decision-making.
economic world modelsheterogeneous agentssim-to-real alignmentendogenous institutionscapability ladder
ProDVI: Programmatic Dynamics Priors for Value Network Initialization
ProDVI introduces a framework leveraging large language models to initialize reinforcement learning agents without requiring pre-collected datasets or simulators. It prompts code-generating language models to produce executable Python functions encoding coarse hypotheses about environment dynamics, generating synthetic transitions for pretraining the state-action encoder in an actor-critic framework. This pretraining provides dynamics-aware inductive biases before online RL begins. Experiments on OpenAI Gym and DeepMind Control Suite demonstrate improved sample efficiency for model-free RL algorithms.
reinforcement learninglarge language modelsactor-critic frameworksample efficiencydynamics prediction
HERALD: Counterfactual Audits and Minimal Repairs for Proof-of-Retrieval Rewards
HERALD introduces an offline audit framework for proof-of-retrieval rewards, employing counterfactual interventions and detector contract enumeration to separate candidate-visible from oracle information. The method identifies citation-laundering vulnerabilities in Qwen3-8B pools across HotpotQA, 2WikiMultiHopQA, and MuSiQue, with a minimal repair strategy ($R[L]$) reducing empirical attack success rate to zero while maintaining answer quality. Results show improved citation precision (2.02 points) and support recall (1.46 points), with unsupported citations decreasing by 1.69 points, though MuSiQue fails non-inferiority. The approach isolates robust scoring from sparse learning signals.
proof-of-retrievalcounterfactual auditscitation-launderingqwen3-8bnon-inferiority
Hybrid Machine Learning Framework for Herd-Level Cattle Growth Pattern and Weight Gain Forecasting in Grazing-Based Production Systems
A hybrid machine learning framework was developed for herd-level cattle weight forecasting in grazing systems, addressing irregular livestock observations. The framework integrated weekly live weight data, demographic variables, and lagged environmental predictors, employing four hybrid architectures (residual, stacked, cascade, ensemble) alongside ARIMA, LSTM, and GRU baselines. The cascade GB to RF to NN architecture achieved optimal performance with a test R² of 0.889, RMSE of 21.319 kg, and MAE of 15.462 kg. Hybrid models demonstrated robustness under sparse observation conditions, though forecasting error increased with longer horizons. Feature importance analysis highlighted animal age, rainfall, and temperature as key predictors. The framework supports feed allocation, grazing management, and livestock marketing decisions.
hybrid architecturetemporal aggregationfeature importancesparse observationsgrazing systems
OPERA: Operator-residual feedback for reliable autonomous optical experiments with language-model agents
The OPERA framework improves autonomous optical experiments by combining operator-residual feedback with language-model agents, addressing the discrepancy between action scores and physical outcomes. OPERA represents experimental actions as optical operators and evaluates outcomes via physically interpretable residuals, which guide operator selection and generation. In three optical tasks, operator-residual feedback reduced score-physical mismatch to 0.9–1.9% (vs. 23.6–39.0% for score-only feedback), increased target achievement probability, and lowered experimental budgets. Validated on digital twins and transferred to optical instruments, OPERA reduced projection budgets in structured-light reconstruction.
autonomous agentsoptical operatorsresidual feedbackstructured-light reconstructiondigital twins
AgentOPSD: Recursive Self-Distillation for Agentic Reinforcement Learning
AgentOPSD introduces a critic-free, recursive method for turn-level credit assignment in agentic reinforcement learning, addressing sparse supervision in long-horizon tasks. The method aggregates token-level teacher-student log-probability gaps into turn-level evidence, recursively updating a Bayesian belief state in log-odds space to identify pivotal turns. Evaluated on ALFWorld, WebShop, and Search-QA using Qwen2.5 models (3B and 7B), AgentOPSD outperforms GRPO and self-distillation baselines, achieving 89.1% success on ALFWorld with Qwen2.5-7B. Gains are attributed to turn-level aggregation and history-dependent recursive belief updates.
credit assignmentrecursive belieflog-odds spaceself-distillationagentic reinforcement learning
Temporal Bridges for Spatial Resolution: Enhancing Climate Data Super-Resolution with Bidirectional Alignment
The authors propose Temporal-Enhanced, a novel framework for climate data super-resolution (SR) that leverages bidirectional temporal alignment to enhance spatial resolution. The method employs Paired Latent Mapping for spatial alignment and noise reduction, followed by Bidirectional Temporal Alignment to capture temporal correlations via forward and backward networks on consecutive latent frames. Temporal Enhanced Super-resolution then optimizes the framework end-to-end. Experiments on large-scale real-world datasets demonstrate superior performance compared to existing approaches that neglect temporal correlations.
climate data super-resolutionbidirectional temporal alignmentpaired latent mappingtemporal correlationsspatial resolution
TRACE: Learned Proprioceptive Odometry for Legged Robots under Unreliable Contact Conditions
TRACE (Tokenized Robust Attention for Contact-Aware Estimation) is a learned proprioceptive odometry estimator for legged robots that predicts relative displacement, rotation, and body-frame velocity from inertial and joint measurements. It employs a foot-aware cross-attention module to adaptively weight IMU and kinematic tokens without manual contact thresholds, trained with direct supervision and physics-inspired auxiliary losses. Policy randomization and partial real-world fine-tuning enhance sim-to-real transfer. Experiments show reduced position drift across diverse terrains compared to classical, hybrid, and learning-based baselines, with ablations validating training objectives and robustness under unreliable contacts.
proprioceptive odometrylegged robotscross-attention modulesim-to-real transferkinematic consistency
SkillMemo: Expert-guided Skill Memory Framework for Compositional Embodied Manipulation
The SkillMemo framework enhances embodied manipulation by decomposing trajectories into latent atomic skills via an expert-guided Mixture-of-Experts segmentation module and storing them in a skill-level episodic memory bank. This architecture retrieves and fuses relevant skill primitives during inference, refining action predictions with a contextual prior. Evaluations on simulation and real-world tasks show SkillMemo improves Diffusion Policy and Vision-Language-Action models, achieving state-of-the-art performance (outperforming π₀.₅) and strong compositional generalization to unseen configurations.
embodied manipulationmixture-of-expertsepisodic memorycompositional generalizationlatent skills
Big, Bright, or Invisible: A Frozen-Feature Benchmark of 3D CT Foundation Models
The study benchmarks ten frozen 3D CT foundation models across thoracic CT cohorts using k-nearest neighbors, zero-shot prompting, and linear probing, revealing no universal SOTA due to performance variability across contexts. Models combining fine-grained tokenization with vision-language alignment generally excel, though a lightweight supervised encoder remains competitive, indicating label utility over scale. Key findings highlight a physical bottleneck: detectability correlates with lesion contrast and spatial extent, with high-contrast/widespread abnormalities (e.g., devices, effusions) reliably detected but small, low-contrast lesions remaining challenging due to global embedding limitations.
3d ct foundation modelsfrozen encodersvision-language alignmentglobal embeddingslesion detectability
Stability of Ranking-dependent Pair-wise Comparison Patterns in the Analytic Hierarchy Process
The study identifies the most stable incomplete ranking-dependent pair-wise comparison pattern for the Analytic Hierarchy Process (AHP), reducing expert workload while maintaining credibility. It compares three patterns—Best-Worst Method (BWM), Best-Second Best (Top 2), and the original maximum difference method—analyzing their stability under expert errors via simulation. Results show BWM and Top 2, though incomplete, outperform the complete maximum difference method in error resilience, enabling fewer comparisons without compromising decision reliability. This advances algorithmic and cognitive decision support in uncertain environments.
analytic hierarchy processpair-wise comparisonbest-worst methoddecision supportexpert errors
Training a Conditioned Video Game Agent on a VLM Annotated Dataset
The paper proposes using Vision Language Models (VLMs) to annotate video game datasets with human-defined rewards, enabling offline Reinforcement Learning (RL) to train conditioned agents. This approach addresses challenges in RL, such as sparse rewards and the trial-and-error process of reward identification and weighting. The method leverages VLMs to extract rewards directly from the dataset, bypassing the need for access to the game engine. Early experiments demonstrate the feasibility of training agents that respond to desired returns, though limitations and difficulties are acknowledged.
vision language modelsoffline reinforcement learningconditioned agenthuman-defined rewardssparse rewards
VLMs for Videogame Data Annotation
This work investigates Vision Language Models (VLMs) for annotating video game frame sequences with reward signals, addressing challenges in conditioned training and offline reinforcement learning. The study evaluates VLMs on racing video games, revealing their limitations in basic question answering due to synthetic scenarios and non-physical environments. Techniques such as VLM output mixing and prompt optimization are proposed as countermeasures. Empirical analysis demonstrates the impact of input sequence length, resolution, and question batching on annotation quality and token consumption. Results highlight VLMs' struggles in gaming contexts despite their broader AI advancements.
vision language modelsreward signalsoffline reinforcement learningprompt optimizationtoken consumption
GAUGE: A Measurement-Grounded Benchmark for Physical Fidelity in Simulation Engines and Video World Models
The paper introduces GAUGE, a measurement-grounded benchmark for evaluating physical fidelity in simulation engines and video world models. It comprises 22 controlled task families covering rigid bodies, flexible materials, and deformable objects, with real-world trajectories, physical metadata, and uncertainty annotations. Benchmarking Isaac Sim, Genesis, Newton, and 6 image-to-video models reveals no uniformly faithful physics engine, with significant discrepancies in impulsive contact, textile motion, and deformation. Video world models exhibit correct trajectory forms but incorrect accelerations and momentum transfer. GAUGE provides a foundation for improving physical fidelity in simulators and world models.
physics enginesvideo world modelsphysical fidelitydeformable objectstrajectory errors
BALANCE: Hybrid Autoregressive-Speculative LLM Inference in Wireless Edge Networks
The paper proposes BALANCE, a hybrid autoregressive-speculative framework for efficient LLM inference in wireless edge networks. It combines autoregressive decoding (AD) and speculative decoding (SD) modes, where an edge server jointly schedules users and allocates resources under latency and memory constraints. A polynomial-time algorithm solves the NP-hard throughput maximization problem with constant approximation guarantees. Experiments show BALANCE outperforms standalone AD/SD, improving task throughput by 1.4-2.1× under heterogeneous user demands.
edge inferenceautoregressive decodingspeculative decodingresource allocationthroughput maximization
CourseGraph: Finding overlaps and differences in Computer Science courses across universities
CourseGraph automates the evaluation of external course overlaps for student mobility programs by semantically representing course information using BERT embeddings and computing pairwise similarities. A Random Forest classifier then determines overlap with a student's home curriculum. Evaluated on the Computer Science programs at Eindhoven University of Technology and Lund University, CourseGraph effectively identifies overlapping courses, supporting curriculum alignment across institutions.
bert embeddingsrandom forest classifiercourse overlapstudent mobilitycurriculum alignment
GSBF: Gaussian Splatting for Environment-Aware Beamforming
The paper introduces GSBF (Gaussian Splatting for Environment-Aware Beamforming), a method for MIMO beamforming that bypasses instantaneous CSI requirements by modeling radio propagation via 3D Gaussian representations. GSBF uses bidirectional spherical Gaussian (Bi-SG) kernels to capture environmental scattering and performs electromagnetic rasterization to render angular propagator maps, which are then projected to constant-modulus beamformers via an array-manifold dictionary. Evaluations show GSBF outperforms exhaustive beam alignment (EBA) in latency while maintaining performance.
beamforminggaussian splattingmimochannel state informationelectromagnetic rasterization
ECG-LENS: Lead-Aware Clinical Context Enriched ECG Report Generation and Evaluation
ECG-LENS introduces an end-to-end framework for automated ECG report generation by integrating multi-lead signal modeling, diagnosis-aware representations, and clinically grounded text generation. The method combines lead-wise encoders for localized waveform morphology with a global encoder for inter-lead dependencies, fusing signal representations with clinical prompts to condition a GPT-2 decoder. Evaluations on PTB-XL and MIMIC-IV-ECG demonstrate state-of-the-art performance, with absolute gains of 4.0% (METEOR), 6.3% (ROUGE-L), and 11.5% (F1-ECGBERT) over baselines.
electrocardiographymulti-lead encodersclinical text generationdiagnosis-aware representationsf1-ecgbert
AppDeltaWorld: Transition-Grounded Delta Code World Model for Mobile GUI Agents
AppDeltaWorld introduces a transition-grounded delta code world model for mobile GUI agents, addressing scalability and fidelity limitations in existing simulated environments. The model predicts next-GUI states as executable HTML updates, leveraging Level-1 HTML references, action-conditioned Level-2 HTML generation, and visual asset insertion. Evaluated on CMGUIBench-500, it achieves superior fidelity in structural layout and UI element reconstruction compared to image-only and code-only baselines. When used for training, AppDeltaWorld enables closed-loop SFT data construction, yielding state-of-the-art performance on AndroidLens (56.7% improvement) and consistent gains on MobileGym and MobileWorld. Test-time RL further enhances policy adaptation without real-app interaction.
gui world modeldelta code generationhtml reconstructionclosed-loop sfttest-time rl
The em-dash em-beds in Congress: A population-level rise in em-dash frequency in U.S. congressional press releases at the dawn of the large-language-model era, 2021-2025
The study detects a population-level rise in unspaced em-dash usage (word---word) in U.S. congressional press releases (2021-2025), potentially signaling LLM-assisted writing. Analyzing 146,239 releases via preregistered Poisson/negative-binomial models (OSF: 10.17605/OSF.IO/U5NEY), em-dash density doubled from 0.10-0.12 to 0.217 per 1,000 characters in 2025, with 24.8% of releases containing at least one (vs. ~13% baseline). The frequency ratio (2023-2025 vs. 2021-2022) was 1.55 (95% CI 1.28-1.93), exceeding the prespecified 1.5x threshold. Results held within offices (75.6% increased; p ~ 1e-16) and passed falsification tests, though the preregistered validation gate was not fully met. The rise correlates with LLM maturation but lacks causal evidence.
em-dashlarge language modelspreregistered designpoisson regressionstylometric analysis
CodeGrep: An RL-Trained Retrieval Agent for LLM Coding Agents
CodeGrep is a 14B retrieval agent trained with GRPO to optimize file-finding efficiency for LLM coding agents, reducing token usage in repository exploration. It issues parallel grep, glob, and read calls, preserving resolve rates (27.0% vs. 25.8% baseline) while cutting rounds by 15% and tokens by 19% on SWE-Bench Verified. Precision analysis shows downstream utility improves above a threshold (CodeGrep: 0.677 vs. BM25: 0.375). The method leverages 67K agent trajectories mined via CATM and a Git-worktree RL environment, with advantage-layer training minimizing KL drift.
retrieval agentllm coding agentsgrposwe-benchkl drift
Beyond Feature Importance: A Comparative Analysis of Pattern Detection Methods in Cluster Interpretation
The study conducts a comparative evaluation of post-hoc analysis methods for pattern detection in clustering results, addressing a critical gap in cluster interpretation. Using a suite of synthetic datasets with systematically injected predefined patterns, the authors assess three techniques: Random Forest surrogate models with permutation feature importance, LIME (Local Interpretable Model-agnostic Explanations), and principal component analysis. Results reveal that while each method successfully recovers relevant features, none consistently detects all injected pattern types. This highlights the limitations of existing explainability tools for pattern-level cluster interpretation and underscores the need for dedicated methodologies.
pattern detectioncluster interpretationpost-hoc analysispermutation feature importancelocal interpretable model-agnostic explanations
D-CLOT: Double Closed Loop Optimal Transport for Unsupervised Action Segmentation
D-CLOT introduces a double closed-loop optimal transport framework for unsupervised action segmentation, addressing representation–prototype inconsistency by refining both frame embeddings and action prototypes. The method employs a graph-constrained module to preserve local neighborhood geometry and periodically re-anchors prototypes using either $k$-means (D-CLOT) or OT barycenters (D-CLOT$_{B}$). Evaluated on five benchmarks, D-CLOT variants outperform CLOT, achieving up to +12.7 F1 and +10.2 mIoU gains per video on YTI and +8.9 F1 on FS-Eval. The approach also establishes the first unsupervised baseline on the fine-grained Assembly101 benchmark, demonstrating robustness and complementarity of the refinement mechanisms.
optimal transportaction segmentationgraph-constrained moduleprototype refinementunsupervised learning
Personalized Deep Research Query Refinement with Graph-Scaffolded Evidence Grounding
G-STEER introduces a method for refining user research queries into personalized specifications by resolving three coupled decisions: relevance of framing factors, sufficiency of user context, and evidence acquisition strategy. The approach uses an Intent Elicitation Graph to organize framing factors as elicitation targets and trains a clarification policy via graph-scaffolded trajectories. Experiments demonstrate G-STEER achieves 1.3× higher weighted target coverage and 1.5× greater downstream report personalization versus baselines, while reducing user questions by 66%.
intent elicitation graphclarification policyframing factorsevidence acquisitionquery refinement
MACRO: Markov Chain Routing of Transformer Layers
MACRO introduces a framework for Markov Chain Routing of Transformer Layers, enabling task-specific dynamic routing through LLM architectures without weight updates or ground-truth labels. It models layer routing as a context-dependent Markov policy, incorporating layer indices, computation budget phases, directional displacements, and operator context, supporting skip, repeat, and residual hidden-state addition operations. The route distribution is updated via training feedback and decoded using a top-k Viterbi algorithm. Evaluated across diverse reasoning and knowledge benchmarks, MACRO achieves a +5.0% average accuracy improvement over baselines, outperforming Dr. LLM by +7.2% while reducing route-search time 9.4x (from 14.8 to 1.6 hours).
markov chain routingdynamic routingtransformer layersviterbi algorithmcomputation budget
Improving Interoperability among Defence and National Security Ontologies: Analysis and Evaluation Tasks
The authors address interoperability challenges in defence and national security ontologies by analyzing over 60 publicly available ontologies and introducing a new evaluation track for the Ontology Alignment Evaluation Initiative (OAEI). The track features eight matching tasks with consensus alignments derived from multiple state-of-the-art ontology alignment systems, supplemented by manually-curated silver-standard mappings. Results include validated mappings from both consensus alignments and unique system suggestions, providing a benchmark for future ontology alignment research in this domain.
ontology alignmentknowledge graphsdefence ontologiessilver-standard mappingsconsensus alignments
Seeing Is Not Deciding: Can Multimodal LLMs Act as Effective CEOs?
This work introduces C-SUITEBENCH, a multimodal benchmark assessing executive decision-making in LLMs, addressing limitations of text-only evaluations. The study evaluates nine frontier models across five decision tasks under paired text-only and multimodal conditions in 50 scenarios. Results show multimodal inputs improve evidence-centric reasoning, particularly in risk forecasting and board-facing justification, but reveal a multimodal integration paradox: visual information degrades constrained resource allocation despite improving visual grounding. Ablation experiments attribute this failure to signal crowding, demonstrating separable bottlenecks in visual perception and constrained action, motivating selective grounding strategies for executive AI systems.
multimodal benchmarksignal crowdingconstrained resource allocationevidence-centric reasoningvisual grounding
Runtime Observability for Heterogeneous Attention Memory
The paper introduces a runtime observability contract for heterogeneous attention memory systems, addressing modern models that employ latent caches, learned sparse selectors, and recurrent states instead of plain KV-caches. The method defines three operators covering four memory classes, instantiated across six model configurations from five architecture families, and composes per-stage bounds into an executable request-level risk ledger. Results show zero violations over 12.4M entry reads under eight-way concurrency, with a fused probe observing a one-layer subset within CUDA graphs. Applied to DeepSeek-V4, the system localizes silent corruptions to structural boundaries, validated by machine-adjudicated discrimination.
runtime observabilitykv-cacheattention memoryrisk ledgercuda graphs
Evidential Rule Learning for Interpretable Classification with Abstention
We introduce Fast Evidential Rule Learning (FERL), an interpretable classification method that learns fuzzy rule models with evidential outputs for abstention and transparency. FERL generates belief, plausibility, and abstention directly from fuzzy memberships in a single deterministic pass, ensuring Lipschitz stability for smooth evidential outputs. On a 30-dataset benchmark, FERL achieves statistically significant accuracy improvements (+2.6% over the second best) and superior utility-discounted accuracy (u65/u80=0.80/0.83) at higher set coverage (0.92). It matches dedicated out-of-distribution detectors in tabular near-OOD detection (77.7 AUROC) and performs competitively on concept-bottleneck tasks, achieving the best AUPR-Out (68.3) and novel-class rejection (57.2) on AwA2 while identifying anomalous attributes.
evidential rule learningfuzzy rule modelslipschitz stabilityout-of-distribution detectioninterpretable classification
MameLoshnLM: Yiddish Language Model and Evaluation Benchmark
MameLoshnLM introduces the first open-source 8B-parameter language model specifically for Yiddish, addressing gaps in Yiddish NLP due to limited digital resources and noisy multilingual corpora. The model is developed by pretraining Llama 3.1 8B on Oytser, a high-quality Yiddish corpus combining web-native and literary sources, and evaluated using Kashes, a multi-task benchmark for translation, linguistic analysis, information extraction, and language understanding. MameLoshnLM outperforms open baselines of similar scale, demonstrating superior capture of Yiddish lexical and morphological patterns. The results highlight the limitations of noisy web-scale multilingual data for low-resource languages and provide a foundation for Yiddish NLP development.
language modelyiddishmultilingual corporalinguistic analysisinformation extraction
ViSR-KGC: Visual Subgraph Reasoning with Vision-Language Models for Multimodal Knowledge Graph Completion
ViSR-KGC introduces a visual subgraph reasoning framework for multimodal knowledge graph completion (MMKGC), addressing limitations in embedding-based and LLM-based methods by leveraging vision-language models (VLMs). The approach integrates global topology dependencies, local multimodal evidence, and pre-trained commonsense knowledge. It extracts query-aware subgraphs from MMKGs, transforms them into visually interpretable images via empirical layout strategies, and combines these with entity images, textual descriptions, and candidate answers into unified prompts for VLM-based inference. This enables effective multimodal reasoning while preserving graph topology and visual semantics.
multimodal knowledge graph completionvisual subgraph reasoningvision-language modelsquery-aware subgraphempirical layout strategy
Cautious Context Steering for Language Model Personalization
Cautious Context Steering (CCS) introduces a lightweight adapter for frozen backbone language models (LMs) to dynamically control user context influence during generation, addressing limitations of in-context learning (ICL) and Context Steering (CoS). CCS learns from an oracle context-conditioned LM to decide token-level context steering, preserving base LM behavior when context is unhelpful. Trained on a single dataset, CCS improves generation quality in-domain and across four out-of-distribution personalization benchmarks, demonstrating robust generalization to new users and domains. CCS eliminates per-user fine-tuning and reduces inference cost by avoiding CoS's additional forward pass.
cautious context steeringin-context learningcontext steeringlightweight adapteroracle context-conditioned lm
When Self-Evolution Backfires: Pre-Commit Gating against Skill Contamination in LLM Agents
The paper introduces Verifier-as-Gatekeeper (VaG), a method to prevent skill contamination in self-evolving LLM agents by filtering skills through a progressive trust hierarchy. VaG employs three heterogeneous critics—structural validity, behavioral harmlessness, and semantic consistency—to individually assess skills, combined with marginal-gain subset selection to eliminate combinatorial contamination. Empirical results on Terminal-Bench 2 show that VaG achieves 72% pass@1 with a skill pool 5x smaller than unconditional accumulation, which degrades past a critical pool size due to irreversible contamination. VaG's frozen skill pool also transfers positively to other backbones and benchmarks without re-evolution.
skill contaminationverifier-as-gatekeepermarginal-gain subset selectionstructural validitybehavioral harmlessness
Hierarchical Latent Prediction for Language Models
Hierarchical Latent Prediction (HiLP) improves long-horizon reasoning in language models by introducing an auxiliary higher-level abstract latent to mitigate error accumulation in latent-space rollouts. Unlike Next-Token Prediction (NTP) or Multi-Token Prediction (MTP), HiLP reduces compounding errors through hierarchical abstraction, enabling longer-horizon coherent belief state representation. Experiments demonstrate HiLP's effectiveness in coding and multi-step reasoning benchmarks, along with improved speculative decoding efficiency.
hierarchical latent predictionnext-token predictionmulti-token predictionlatent-space rolloutsspeculative decoding
When Agentic AI Meets Integrated Sensing and Communication
The paper introduces Agentic Integrated Sensing and Communication (AISAC), a paradigm shift from function-oriented to goal-driven intelligent systems in ISAC. It proposes a six-stage closed-loop framework (observation, contextualization, reasoning/prediction, planning/orchestration, execution/collaboration, feedback/resilience) and five levels of agentic maturity to unify disparate research areas like multimodal intelligence, large language models, and reinforcement learning. An audit of representative studies reveals gaps in agentic-specific evaluation criteria, with no system reporting more than one or two metrics. Key challenges include physical-to-semantic grounding, predictive world models, real-time agent-PHY interaction, and resource-efficient autonomy.
agentic aisacclosed-loop frameworkagentic maturityphysical-to-semantic groundingpredictive world models
A Two-Tier Perspective on Inference-Time Parallelism in Multi-Agent LLM Systems
The paper proposes TIPEX, a framework unifying two levels of inference-time parallelism in multi-agent LLM systems: Replica Parallelism (exploring multiple solution paths) and Structural Parallelism (concurrent execution within a path). It introduces a unified execution semantics for coordinating these strategies, enabling systematic analysis of their interactions. Experiments on GAIA show parallelism improves accuracy by 12% and reduces latency by 30%, though with increased token costs. Complementary effects emerge, with intermediate-complexity tasks benefiting most from coordinated parallelism.
multi-agent systemsinference-time parallelismreplica parallelismstructural parallelismllm coordination
ChainClaw: A Layered Agent Framework for Reliable On-Chain Execution
ChainClaw introduces a blockchain-native agent framework addressing three fundamental gaps in on-chain execution: Reactivity, Irreversibility, and Observability. The framework employs a layered architecture with an event-driven orchestration layer, a simulation-based safety intelligence layer, and an on-chain monitoring runtime layer, unified by a cross-layer memory subsystem. It enhances reactivity through event ingestion and simulation feedback, mitigates irreversibility via a pre-execution safety pipeline, and improves observability using an on-chain read adapter and transaction monitor. Evaluated on a purpose-built benchmark covering seven tasks across four categories and five dimensions, ChainClaw consistently outperforms baselines in both safety and task completion.
blockchain-nativeevent-drivensimulation-basedpre-executionon-chain monitoring
Task-Conditional Flow Matching for Balanced Multilingual Text Embedding Adaptation
We introduce Task-Conditional Flow Matching (TCFM), a multilingual embedding adaptation framework that selectively applies Flow Matching to translation tasks while optimizing retrieval, classification, and pair-classification tasks with task-specific objectives. TCFM combines teacher-guided representation preservation with a three-stage curriculum for stable adaptation. Evaluated on the Indic Massive Text Embedding Benchmark, TCFM establishes a new state-of-the-art, consistently improving embedding quality across diverse multilingual tasks and generalizing across embedding model families. Codebase and datasets will be released upon paper acceptance.
task-conditional flow matchingmultilingual embeddingflow matchingrepresentation preservationthree-stage curriculum
Activity Frames: Deterministic Screen-Activity Compilation for Agent Memory and Replay
The paper introduces Activity Frames, a deterministic pipeline for compiling passive screen-capture data into agent-memory representations without model inference. The method segments raw screen activity into typed frames (application, site, timing, input volume) with evidence pointers, enabling byte-identical, cacheable outputs. Evaluated on a 128,756-frame corpus (51 days), the compiler reduces raw data by 86x (68 ms processing) and achieves 98.4% accuracy (Wilson 95% CI 91.7-99.7%) for day-recall tasks, outperforming LLM summaries (66-80%). It also measures Routine Overhead Ratio (R=60-343x) and delegable recurrence (7.7-9.0%), enabling zero-token deterministic replay. Open-source schema and tools are provided.
activity framesdeterministic compilationroutine overhead ratioagent memoryscreen-capture
GROM: Gradient-Free Rapid One-Shot Machine Unlearning
GROM introduces a gradient-free, one-shot machine unlearning method for LLMs, replacing iterative fine-tuning with an exact analytical solution. The approach formulates unlearning as a ridge-regularized least-squares problem, deriving closed-form additive updates to weight matrices that suppress unwanted content while preserving retained behavior. Evaluated on TOFU-5%, TOFU-10%, MUSE-Books, MUSE-News, and WMDP, GROM achieves state-of-the-art forgetting-utility trade-offs, resists quantization attacks, and applies updates in seconds without backpropagation or iteration.
machine unlearningridge regularizationclosed-form solutionquantization attackgradient-free optimization
When Do Prompt-Side Agent Playbooks Transfer? Accuracy, Cost, and Runtime Shift in Agent Deployment
The study investigates the transferability of prompt-side agent playbooks across different settings without retraining, employing a shared distill--validate--transfer protocol. Evaluations on ALFWorld, TAU2-Bench, and XBench-DeepSearch reveal that transfer benefits are context-dependent: distilled guidance outperforms fixed demonstrations in ALFWorld under controlled decoding, while TAU2-Bench shows modest matched-domain advantages with heterogeneous compatibility. XBench-DeepSearch highlights runtime-cost tradeoffs and heuristic preservation. Target-side validation is essential for success, termination, and cost compatibility. Frozen transfer is a conditional cold-start option, not universally preferable to redistillation.
prompt-side playbooksdistill--validate--transfergreedy decodingruntime-cost tradeoffscold-start option
HyTBE: Hyperbolic Target-Background Expert Model for Cross-Domain Infrared Small Target Detection
HyTBE introduces a Hyperbolic Target-Background Expert model for cross-domain infrared small target detection (IRSTD), addressing target-background relation shift via hyperbolic relation modeling. The method employs Target-Background Relation Intervention to diversify training patterns and Hyperbolic Relation Modeling in a Poincaré ball to encode multi-scale target-background relations. A Hyperbolic-guided MoE Adapter then calibrates features based on these relations. Evaluated on NUAA-SIRST, NUDT-SIRST, and IRSTD-1K in leave-one-domain-out experiments, HyTBE outperforms baselines in cross-domain generalization.
infrared small target detectioncross-domain generalizationhyperbolic relation modelingtarget-background relation shiftpoincaré ball
UniVVT: A Unified End-to-End Framework for High-Fidelity Video Virtual Try-on
UniVVT proposes a unified end-to-end framework for Video Virtual Try-On (VVT) by reframing it as semantically conditioned video generation, eliminating explicit geometric preprocessing. The method employs a scene-task perceiver based on a Multimodal Large Language Model to encode source video, target garment, and task instruction into task-aware latent tokens, coupled with a semantic bridge to align these tokens with a diffusion-based video generator. A three-stage progressive training strategy ensures robust component coupling. Experiments show state-of-the-art performance across benchmarks, validating implicit semantic guidance over geometric preprocessing.
video virtual try-onmultimodal large language modeldiffusion-based video generatorsemantic bridgeprogressive training
Multivariate Time Series Forecasting needs Cross Variable Loss
We propose Cross-Variable Loss (CvLoss), a structural regularizer for multivariate time series forecasting that addresses the objective gap in Direct Forecasting (DF) paradigms. CvLoss constrains forecast residuals on a cross-variable graph, penalizing inconsistent edge-wise residual differences across synchronous and asynchronous interactions. Experiments demonstrate that CvLoss consistently enhances competitive forecasting models, outperforms representative learning objectives, and maintains compatibility with diverse forecasting backbones.
multivariate time seriescross-variable lossdirect forecastingstructural regularizerforecast residuals
Once a Response, Always a Response: Detecting LLM-generated Text via Latent Prompt Restoration
The paper introduces EchoPrompt, a training-free detector for LLM-generated text based on latent prompt restoration, addressing limitations of probability-based zero-shot methods that ignore generation mechanisms. The method prepends a generic prefix to restore assistant-response context, measures likelihood gain using an instruction-tuned model, and calibrates against a base model to quantify latent prompt dependency. Experiments demonstrate state-of-the-art performance in zero-shot detection, with robust generalization across challenging settings.
llm-generated textzero-shot detectionlatent prompt restorationlikelihood gaininstruction-tuned model
ABC: Numerical Data Collection under Local Differential Privacy without Prior Knowledge
The paper proposes Adaptive Bounding of Clipping regions (ABC), a local differential privacy (LDP) framework for numerical data collection without prior domain knowledge. ABC dynamically estimates the data domain by having users submit perturbed values alongside privatized clipping signals, iteratively adjusting the domain bounds. Theoretical analysis confirms domain convergence, while empirical evaluation shows improved data quality across datasets and LDP mechanisms, with robustness to hyperparameters.
local differential privacynumerical data collectiondomain estimationadaptive clippingprivacy-preserving
Subliminal Learning is Non-Semantic Distillation
The paper investigates Subliminal Learning (SL), a phenomenon where language models transfer biases or behaviors from teacher to student models via non-semantic distillation from synthetic data. Through experiments with Gemma and Llama models, the authors demonstrate that Gaussian noise augmentation of model weights increases subliminal transfer by factors of 1.9 and 1.3 respectively, revealing the importance of non-semantic weight structures. They show that steering vectors, prompting, and fine-tuning can induce subliminal data, with students inheriting both semantic biases and intervention types. Gradient analysis reveals linear correlations between steered data and teacher vectors, suggesting potential for data auditing.
subliminal learningnon-semantic distillationsteering vectorsgaussian noisesynthetic data
Unified Agent: Managing Interactions across Devices
Unified Agent introduces a stateful AI agent architecture designed to manage cross-device, cross-time interactions effectively by maintaining a compact, action-ready state. The agent organizes engagement evidence, stated facts, and standing requests to inform decisions based on current observations. A benchmark for user-agent interactions across devices and time was constructed to evaluate state designs. Unified Agent significantly outperforms adaptations of four published designs in default settings and maintains superiority across variations in multimodal large language model (MLLM) families, capabilities, and reasoning efforts. Code and data will be publicly available on GitHub.
stateful agentcross-device interactionmultimodal large language modelaction-ready statebenchmark
BlockPython: A Process-Aware Agent-Supported Platform for the Transition from Block-Based to Python Programming
BlockPython introduces a process-aware platform supporting learners' transition from block-based to text-based programming via bidirectional block-Python translation. The system employs a four-stage workflow (Task Decomposition, Block-Based Practice, Code Challenge, Extended Interaction) with deterministic diagnosis and rule-based stage control. Process evidence (block artifacts, code versions, run outcomes) informs program visualization and a learning assistant providing explanations and prompts. The design facilitates cognitive mapping between program structure, runtime behavior, and textual syntax, offering a reference for transition-support systems.
block-based programmingbidirectional translationprocess-aware learningdeterministic diagnosisprogram visualization
Hijacking Robots with a Piece of Paper: A Systematic Study of Physical Prompt Injection in VLM-Controlled Robots
The study investigates physical prompt injection attacks on Vision-Language Model (VLM)-controlled robotic systems, introducing a taxonomy of four attack types (indirect signage, task redefinition, authority impersonation, conflict injection) and evaluating them across 5,670 trials with GPT-4o, Gemini 2.5 Flash, and Qwen3-VL-32B. Attacks succeeded at rates of 27.0%, 29.4%, and 5.0%, respectively, with authority impersonation and negation attacks being transferable. Mitigations like prompt-based defense (75-100% effective), two-stage verification (85-100%), and text masking (100%) reduced vulnerability, though trade-offs exist for tasks requiring in-scene text reading.
vision-language modelsphysical prompt injectionrobotic systemsadversarial attacksmitigation strategies
RA-CAD: Learning Post-Execution Critique for State-Aware Text-to-CAD Generation
RA-CAD introduces a state-aware agent for text-to-CAD generation that optimizes feedback utilization through a Generate--Execute--Critique--Rewrite loop. The agent generates parametric CAD codes, executes them, critiques outcomes, and rewrites based on feedback, making critique a learnable policy decision. CAD Code Bootstrapping establishes foundational coding capabilities via supervised fine-tuning, while Feedback-Driven Agent Optimization applies Group Relative Policy Optimization to trajectory-level code and critique sequences, rewarding execution validity and geometric quality. RA-CAD achieves state-of-the-art performance on CADFusion and Text2CAD benchmarks, surpassing existing methods and proprietary language models in execution validity and geometric precision.
text-to-cadparametric cadgroup relative policy optimizationexecution validitychamfer distance
Shaping Human-AI Interactions to Provide Improvement Pathways and Balance Competing Objectives
This thesis proposes design principles for human-AI interaction loops to achieve three objectives: fostering accurate user beliefs about AI systems, incentivizing improvement over gaming, and maintaining system performance. Methodologically, it combines theoretical analysis, data-driven modeling, human-subject experiments, and empirical evaluations on real-world and semi-synthetic datasets. The work contributes to human-centered machine learning by developing approaches that align AI systems with human needs while balancing competing objectives like accuracy and strategic behavior.
human-ai interactionstrategic behaviorhuman-centered machine learningempirical evaluationtheoretical analysis
Spectral Aliasing Pretext: A novel task for Self-Supervised fault diagnosis in rotating machinery
We introduce Spectral Aliasing Pretext (SAP), a self-supervised learning method for machinery fault diagnosis that pretrains models on unlabeled vibration data by exploiting spectral aliasing. SAP deliberately undersamples signals to create folded spectra, then trains a Transformer to reconstruct the original unfolded spectrum, enabling the model to learn frequency-domain invariants characteristic of mechanical faults without destructive augmentations. Experiments on the CWRU dataset demonstrate that SAP learns stable, discriminative representations, achieving high classification performance in linear probing with minimal labeled data and low variance. Results indicate that SAP with linear probing outperforms fully supervised training in effectiveness and reliability for fault diagnosis with limited labeled data.
spectral aliasingself-supervised learningtransformerfault diagnosislinear probing
DreamGuard: Efficient Runtime Guardrail for LLM Agents via Risk-Aware World Model
DreamGuard introduces a proactive runtime guardrail for LLM agents that mitigates long-horizon risks via a risk-aware world model. The method maintains a recurrent latent state over the trajectory, predicts future latent states, and fuses immediate-hazard and prefix-risk evidence into intervention decisions prior to execution. Evaluated across four benchmarks and an online guardrail assessment, DreamGuard outperforms generic, reactive, and proactive baselines, achieves optimal safety-utility trade-offs, and operates with an average latency of 25 ms per call.
runtime guardrailrisk-aware world modelrecurrent latent stateprefix-risk evidencelong-horizon risks
Answer First, Reason Later: Commitment Order in Diffusion LLMs
The study identifies a reasoning pathology in masked diffusion language models (dLLMs) where unconstrained token commitment during decoding leads to premature answer generation and reasoning collapse. Analyzing LLaDA-8B on GSM8K, the authors find that final answers are committed at 15-24% of the trajectory while half the reasoning region remains masked, with answer-only outputs occurring in up to 90% of problems. Through a 2x2 prompt-decoder design, they demonstrate that chain-of-thought reasoning improves performance only under ordered commitment (+34.8 percentage points). A frontier-gated commitment intervention recovers the performance gap (0.528 to 0.852) while maintaining up to 4x parallel decoding, with optimal window size transitioning from w=1 to unconstrained at 8 tokens/step.
diffusion language modelstoken commitmentchain-of-thoughtfrontier-gated commitmentparallel decoding
Grounded Well-Condition Anomaly Detection on the Volve Field: Constructed Labels, a Baseline, and a Dual-Head Model
The authors introduce a grounded anomaly-detection framework for the Volve field dataset, addressing the lack of fault logs in real-world production fields. They construct anomaly labels validated against engineering documents and release detailed provenance. A dual-head model, adapted from defect detection in metal parts, predicts event presence and type, while an unsupervised baseline identifies flagged regions. Results show the unsupervised detector aligns with rule-based labels, and the supervised model generalizes well across unseen wells, though temporal localization remains approximate. The dataset, labels, baseline scores, and code are publicly released under CC-BY-NC-SA 4.0.
anomaly detectionvolve fielddual-head modelunsupervised baselinelabel provenance
Nonvisual Classification of Ground-Condition by Artificial Proprioception in an Amoeba-Inspired Autonomous Walking Robot
The study presents a nonvisual ground-condition classification system for an amoeba-inspired four-legged robot, achieving high accuracy without image processing. The method integrates artificial proprioception via a three-axis accelerometer, eight foot pressure sensors, and reservoir computing (RC) to handle sensor-output fluctuations during dynamic walking. Results demonstrate successful classification of flat versus rough terrain and on-site gait switching. Sensor contributions to classification are analyzed.
artificial proprioceptionreservoir computingground-condition classificationmultimodal sensingwalking robot
DistMedVL: Distributional Vision-Language Alignment for Uncertainty-Aware Medical Image Segmentation
DistMedVL introduces a probabilistic vision-language framework for uncertainty-aware medical image segmentation, addressing aleatoric and epistemic uncertainties in cross-modal alignment. The method employs a lightweight Probabilistic Cross-Modal Adapter (PCM-Adapter) atop frozen encoders, featuring a Mahalanobis Alignment Module (MAM) for variance-conditioned patch-text matching and a Distribution Flow Module (DFM) for vision-guided refinement of textual distributions. Evaluated across eight medical segmentation benchmarks, DistMedVL achieves state-of-the-art performance with only 6.3M trainable parameters, demonstrating enhanced data efficiency, robustness to perturbations, and cross-dataset generalization.
probabilistic alignmentmahalanobis distancecross-modal adaptermedical segmentationuncertainty modeling
A Unified Framework for Trajectory Prediction with Explicit Planning and Reaction Decomposition
INTraJ proposes a unified trajectory prediction framework that decomposes social influence into explicit planning and reaction stages, where planning constructs reference trajectories using future social information and reaction recovers local adjustments from residuals. The method supports both multi-target and single-target paradigms, demonstrating state-of-the-art performance on Argoverse 2, Argoverse 2-ped, ETH/UCY, and SDD benchmarks, with notable improvements in FDE and long-horizon consistency. Results validate that staged social modeling enhances prediction stability.
trajectory predictionsocial modelingplanning-reaction decompositionmulti-target paradigmlong-horizon consistency
SafeDivertor: Faithful Divertor Heat Flux Reconstruction from Macroscopic Plasma State Signals via Time-Frequency Prior Exploitation
The paper introduces SafeDivertor, a framework for online reconstruction of time-resolved radial heat-flux profiles from multi-source plasma-state signals in magnetic-confinement fusion devices, bypassing conventional post-discharge infrared-based inversion. The method employs physical prior-aware initialization, input perturbation, spectral-aware reconstruction optimization, and progressive training to address challenges like signal heterogeneity and transient dynamics. Evaluated on the DivMPS2HF dataset, SafeDivertor outperforms baselines across five metrics, setting a new benchmark for signal-based divertor heat-flux reconstruction.
divertor heat-fluxplasma-state signalstime-frequency priorsspectral-aware optimizationmagnetic-confinement fusion
F$^2$Agent: Financial Fusion of Agentic Intelligence for Multimodal Trading
F$^2$Agent introduces a novel multimodal agentic paradigm for financial trading, addressing limitations in cross-modal dependency capture and noise robustness. The method employs a hierarchy of specialized agents for modality-specific signal extraction, coupled with a modality-aware adaptive fusion mechanism and noise-robust consistency regularization. Extensive experiments on six stocks and cryptocurrency assets demonstrate F$^2$Agent's superiority over 16 baselines, achieving over 20% relative improvement in annualized return on average. Notably, it delivers returns of 120.48% on GOOG and 148.41% on TSLA, showcasing robustness across varying market dynamics.
multimodal fusionagentic intelligenceconsistency regularizationnoise robustnesstrading signals
Studying People to Study AI: Expert Perspectives on the Epistemic Fit and Barriers of Human Research in AI Safety & Ethics
The study investigates barriers to human research adoption in AI Safety & Ethics (AISE) through an expert survey (n=93) and interviews (n=17) across Technical, Sociotechnical, Governance, and Normative backgrounds. Findings reveal consensus on human research's value but identify constraints: perceived validity issues, resource barriers, methodological preferences, and infrastructural limitations. Technical researchers exhibit lower valuation of human research and interdisciplinary collaboration, indicating epistemic tension. Recommendations address establishing epistemic fit and overcoming adoption barriers without performative 'human-washing'.
ai safetyhuman researchepistemic fitinterdisciplinary collaborationmethodological preferences
Relay, Don't Route: Adaptive Population Handoff for Cost-Efficient LLM-Driven Evolution
The authors propose \model, a training-free framework for cost-efficient LLM-driven evolution that adaptively allocates inference budgets at the population level rather than individual queries. \model employs a bandit scheduler to explore multiple trajectories with a cheap model, using Relay Gain—the marginal improvement of a quality-diverse candidate bank—to determine when to hand off to a strong model for refinement. Evaluated across four benchmarks and three budgets, \model achieves the highest mean score in 11 of 12 settings, demonstrating that stateful search benefits from population-level budget allocation.
llm-driven evolutionbandit schedulerrelay gainpopulation handoffstateful search
Refining Over Resampling: Test-Time Self-Correction for LLM Reasoning
The paper introduces a verifier-free breadth-depth refinement framework for improving large language model (LLM) reasoning at test time. The method samples multiple independent reasoning rollouts, refines each through iterative self-critique and self-correction, and aggregates answers via majority voting. This approach preserves diverse initial attempts while repairing local reasoning errors before aggregation. Evaluated on benchmarks including AIME24, AIME25, AMC, OlympiadBench, and MATH500, the method outperforms greedy decoding, majority voting, verifier-based best-of-N, beam search, and lookahead decoding. For example, accuracy on MATH500 with Qwen2.5-1.5B increased from the strongest verifier-based baseline to 58.0%, demonstrating the effectiveness of refinement over resampling.
test-time refinementself-correctionmajority votingreasoning rolloutsiterative self-critique
Bayesian Expected Uncertainty Reduction (B-EUR) Model: A Computational Account of What Makes Design Options Worth Trying
The Bayesian Expected Uncertainty Reduction (B-EUR) model formalizes the epistemic value of design actions as their expected reduction in uncertainty about action-outcome relations, addressing a gap in the Uncertainty Driven Action (UDA) framework. Through simulations and human experiments involving a graph-shape guessing task, the study examines how environmental properties—generalizability (knowledge transfer across trials) and outcome discriminability (distinctness of outcomes)—affect action selection. Results show an inverted-U relationship between epistemic value and generalizability, a positive correlation with outcome discriminability, and similar patterns in subjective value and enjoyment. The model informs prototype selection, problem framing, and feedback design.
bayesian expected uncertainty reductionuncertainty driven actiongeneralizabilityoutcome discriminabilityepistemic value
SkillHEX: Improving Agent Skills via Hypothesis-Driven Autonomous Exploration and Exploitation
SkillHEX introduces a hypothesis-driven framework for autonomous skill evolution in LLMs, addressing sparse rewards and exploitation traps in on-demand refinement. The method couples self-verification with evidence-guided tree search, translating failure hypotheses into executable tests to generate dense rewards and balance skill-revision exploration-exploitation. Evaluated on SkillsBench (87 tasks), SkillHEX achieves 55.9% (GPT-5.3-Codex) and 57.9% (Claude Opus 4.7) pass rates under 5-iteration budgets, outperforming prior self-evolving methods.
autonomous skill evolutionhypothesis-driven verificationsparse rewardtree searchllm refinement
Measuring and Detecting Harmful AI Sycophancy
The paper introduces CAP (Contrastive Anchor Probing) to measure and detect harmful preference-induced stance reversal sycophancy (PSRS) in LLMs, where models reverse stances to align with user preferences. Using CAP, the authors collect 290,460 labeled responses from 17 LLMs across 12 domains, revealing PSRS rates of 5-56% (lower in more capable models). They demonstrate PSRS detection feasibility from response text, highlight cross-model generalization challenges, and propose initial solutions. Dataset and code will be released.
preference-induced stance reversal sycophancycontrastive anchor probinglarge language modelscross-model generalizationharmful ai behavior
GAUGE: Granularity-Adaptive Counterfactual Gating of Evidence for Incomplete Multimodal Classification
GAUGE introduces a granularity-adaptive counterfactual gating framework for incomplete multimodal classification, addressing coarse modality-level fusion by operating on fine-grained evidence units. The method imputes missing modalities, encodes inputs uniformly, and scores counterfactual effects via prediction-aware Taylor evidence scores in a single forward-backward pass, mapping scores to continuous gates for unit-wise modulation. Experiments on six benchmarks show GAUGE outperforms baselines in incomplete-input settings, supported by a theoretical analysis of first-order approximation error. The framework provides principled, scalable fine-grained evidence control under modality incompleteness.
multimodal classificationcounterfactual gatingtaylor evidence scoresmodality incompletenessgranularity-adaptive
SkillZip: Contract-Preserving Graph Compression for Scalable Agent Skill Libraries
SkillZip introduces a contract-preserving graph compression framework for scalable agent skill libraries, addressing unit mismatch in procedural knowledge reuse. The method performs execution-aware procedural abstraction by rewriting recurring contract-valid motifs into reversible ported macros, preserving boundary signatures, dependency closure, verifier reachability, and source-level expansion. It hydrates compact contexts at inference and integrates new skills via execution evidence. Experiments on technical and embodied agent benchmarks demonstrate SkillZip outperforms baselines by up to 12.2 points, achieves a 3.46x compression ratio with 99.2% dependency preservation and 98.7% verifier reachability, and scales robustly across libraries from 200 to 100K skills.
contract-preservingprocedural abstractiondependency closureverifier reachabilityreversible macros
Epistemic Trustworthiness in Generative AI: A Normative Framework for Warranted Reliance in High-Stakes Workflows
The paper proposes a normative framework for assessing epistemic trustworthiness in generative AI systems deployed in high-stakes workflows, arguing that current evaluations focusing on accuracy, fairness, and explainability fail to address warranted reliance. Drawing on philosophical accounts of trustworthiness, the authors identify three constitutive conditions: epistemic humility (communicating competence limits), epistemic access (enabling output inspection), and resistance to epistemic injustice (recognizing users as legitimate epistemic agents). Through case studies in legal, medical, and hiring contexts, they demonstrate how failures in these dimensions cause harms unaddressed by standard metrics, concluding with design implications for systems prioritizing epistemically warranted reliance.
epistemic trustworthinessgenerative ainormative frameworkhigh-stakes workflowswarranted reliance
LC-GRPO: Bridging Train-Inference Gap for Flow-Based GRPO with Langevin Correction
LC-GRPO introduces a Langevin correction framework for flow-based GRPO to address the train-inference gap between deterministic ODE sampling and stochastic SDE rollouts in reinforcement learning. The method combines an ODE Euler step with a stochastic Langevin correction, recovering the required score from the flow velocity without additional models, while maintaining tractable likelihoods for policy optimization. Theoretical analysis shows reduced Wasserstein error compared to imperfect ODE steps and standard SDE discretization. Experiments on SD3.5-Medium, FLUX.1-Dev, and HunyuanVideo demonstrate improved reward optimization, preserved generation quality, and reduced train-inference discrepancy in text-to-image and text-to-video tasks.
flow-based modelslangevin correctiontrain-inference gapwasserstein errorpolicy optimization
Search-Aided Joint Agent-Environment Reinforcement Learning for Robust Lifelong Multi-Agent Path Finding with Rotations
The paper introduces Search-Aided Joint Reinforcement Learning (SJRL) for Lifelong Multi-Agent Path Finding with Rotations (LMAPF-R2), addressing motion constraints in automated warehouses. SJRL combines neural policies with Causal PIBT for collision resolution and jointly optimizes agent-environment policies via learned graph edge costs guided by backward Dijkstra search. Experiments show SJRL outperforms Causal-PIBT in high-density scenarios and validates performance with 8 physical and 248 virtual robots in a mixed-reality warehouse.
lifelong multi-agent path findingreinforcement learningcausal pibtdijkstra searchmotion constraints
StepReflect: Structured UI Transition Reflection for Mobile GUI Agents
StepReflect introduces a structured prediction approach for per-step GUI state transition reflection in mobile GUI agents, addressing the inefficiency of open-ended multimodal reasoning. The method employs a staged training pipeline combining supervised fine-tuning, teacher-student distillation, and preference- and reward-based refinement. The resulting 8B parameter model achieves 82.16% transition-level accuracy on AndroidWorld, outperforming zero-shot GPT-5.2 by 11.83 percentage points. In online evaluations across four agent configurations, StepReflect achieves higher task success in three configurations and remains competitive in the fourth, while reducing API costs compared to GPT-based reflection.
structured predictiongui state transitionteacher-student distillationpreference-based refinementtransition-level accuracy
The Judgment-Consequence Gap: LLM Moral Reasoning in Healthcare Decisions
This study investigates the moral reasoning of large language models (LLMs) in healthcare decisions, focusing on responsibility judgments and resource allocation. Using clinical vignettes adapted from prior studies, the authors evaluate a range of LLMs across different model families and capability levels. Results reveal a judgment-consequence gap: LLMs agree with humans on patient responsibility for health-harming behaviors but default to random resource allocation, diverging from human preference for less-culpable patients. LLMs also emphasize access to information, reducing responsibility judgments when health-risk knowledge is absent. Notably, higher-capability LLMs amplify normative disagreement with humans, suggesting a systematically different moral framework.
large language modelsmoral reasoningresource allocationclinical vignettesjudgment-consequence gap
SkillTV-Bench: Benchmarking How Well Judges Perform on Skill-Augmented Agentic Execution
The paper introduces SkillTV-Bench, a 681-case benchmark for evaluating skill-aware trajectory verification in LLM and agent judges, covering 50 tasks across 11 domains. It proposes SkillTV-Evolve, which externalizes verification knowledge as a reusable JudgeSkill to guide targeted inspections and evidence-grounded verdicts, refined via an automated evolution loop. Results show a 14.8 percentage point accuracy improvement for agent judges and increased success rates from 22.9% to 45.5% in rollout-pool selection with ten rollouts.
skill-augmented agentstrajectory verificationllm-as-a-judgeagent-as-a-judgejudgeskill
When Experience Becomes Instruction: Trajectory Poisoning in Self-Evolving Agent Skill Systems
The paper introduces PoisonedEvolution, a trajectory-poisoning attack on self-evolving skill (SES) systems where untrusted experience becomes trusted instruction. The attack operates under a skill-visible black-box setting, requiring Inclusion, Evolution Attribution, and Realization, with Attribution as the bottleneck. Evaluated on six LLM evolvers in SkillClaw and Trace2Skill, it achieves 91.0% and 61.5% success rates (SER) at 10% attacker support, demonstrating cross-architecture transfer. Key success factors include recurring support, causal framing, and domain-aligned encoding, revealing evidence promotion as a critical security boundary.
trajectory poisoningself-evolving skill systemsblack-box attackevolution attributionskill bank
Turing's Frist Imitation Game: Design Concepts and a Human-Approximates-Machine Reading
The paper analyzes Turing's 1948 'Intelligent Machinery' report as a foundational source for imitation games, identifying key design concepts: machine fallibility, exclusion of physical attributes, human judge's role, and intellectual activity as search. It proposes that limiting human chess players to suboptimal performance enhances search-driven behavior, making human cognition more machine-like. This reframes the 1948 game as a human-approximates-machine paradigm, extending imitation games to study machine-like cognition in humans under constrained tasks.
imitation gameintelligent machineryintellectual searchtask constraintshuman-machine comparison
Vibe Compiler: A Research-Logic Synthesis Tool That Runs without Prompt Engineering -Toward Enhancing Metacognition for Sustaining Agency in the Age of Generative AI-
The Vibe Compiler introduces a research-logic synthesis tool that enhances metacognition in AI-assisted intellectual work by preserving human epistemic agency. Grounded in the Synthesis-Analysis Reciprocity Model, it transforms vague ideas into coherent research logic using a 16-parameter ontology, identifying structural gaps through cognitive function and executing agent dimensions. The system prompts researchers with reflective questions rather than autonomously filling gaps, encouraging critical engagement with AI-generated reasoning. A prototype built on NotebookLM and Gemini demonstrates that effective AI-assisted reasoning relies more on structured knowledge than sophisticated prompting. The paper itself was developed using the proposed Vibe Compiler.
metacognitionepistemic agencysynthesis-analysis reciprocityresearch-logic compilerknowledge structure
Hyper-ES: Effective Evolution Strategies for LLM Reasoning via Descent Direction Merging
Hyper-ES introduces a subspace-based Evolution Strategy (ES) framework for efficient LLM reasoning optimization, addressing the ineffectiveness of direct ES application in high-dimensional parameter spaces. The method first extracts descent directions via gradient-based fine-tuning, constructs an adaptation subspace, and applies CMA-ES to optimize layer-wise DARE-TIES merging coefficients within this subspace. Evaluated on Qwen2.5-Instruct and DeepSeek-R1-Distill across six mathematical reasoning datasets, Hyper-ES outperforms GRPO-LoRA by 1% accuracy while reducing gradient updates by 10%.
evolution strategyllm reasoningdescent directioncma-esdare-ties
EcoAgent-Bench: Evaluating Economic Decision-Making in Budget-Constrained LLM Agents
The paper introduces EcoAgent-Bench, a benchmark evaluating LLM agents' economic decision-making under budget constraints across 304 real-derived tasks from GAIA, HotpotQA, and MuSiQue. It tests four key decisions: avoiding unnecessary escalation, escalating when needed, model tier selection, and stopping on unsupported premises. Seven LLM agents were evaluated in tool-API and workspace-CLI settings, revealing poor performance (3.9-24.0% micro strict success, ≤7.3% economic consistency) and frequent overspending or premature stopping. The benchmark includes task bundles, transformation pipelines, and evaluation environments for reproducibility.
llm agentseconomic decision-makingbudget constraintstask completionevaluation benchmark
APQF: Agentic Profiling-Guided Structured Pruning and Mixed-Precision Quantization with Adaptive Fine-Tuning
APQF introduces an automated framework for neural network compression via agentic profiling-guided structured pruning and mixed-precision quantization with adaptive fine-tuning. The method employs an LLM-driven planner to propose per-layer pruning ratios and bit-widths based on profiling data, validated before execution, and integrates quantization-aware training for accuracy recovery. Evaluated on ResNet, VGG7, ViT, DeiT, and Swin using ImageNet-1k and CIFAR-10, APQF reduces compute to 5.6-7.7% of original bit-operations (13-18x) while maintaining baseline accuracy, outperforming existing joint methods by +17 Top-1 points under a 200K-image budget. On VGG7, it achieves 93.15% accuracy using 0.41% of baseline bit-operations, surpassing full-precision performance.
structured pruningmixed-precision quantizationllm-guided planningquantization-aware trainingagentic profiling
Learning Context-Free Grammars for Grammar-Constrained Decoding via Declarative Agentic Programming with Guarantees
Autogrammar, a declarative agentic programming framework, automatically learns context-free grammars (CFGs) from documentation and execution data for grammar-constrained decoding in language models (LMs). Formalized as a Kripke structure with nondeterministic choices resolved by an LM, Autogrammar employs linear temporal logic constraints for declarative control. Evaluated on three domain-specific languages (Amazon CloudWatch Logs Insights, Dynatrace Query Language, Datadog Search Syntax), Autogrammar achieves near-perfect precision on unseen data, reduces execution time by 3.8x via temporal restrictions, and improves end-to-end LM performance on 8/10 real tasks, matching or exceeding professionally-maintained grammars. Existing LM baselines and state-of-the-art formal techniques perform significantly worse.
context-free grammardeclarative programmingkripke structurelinear temporal logicgrammar-constrained decoding
Innovation-Residual Auditing of Autonomous Analysis Agents: Localization, Detection Limits, Error Control, and Identifiability
This paper provides a formal analysis of innovation-residual auditing for autonomous analysis agents, addressing localization capabilities, detection limits, and error control. The authors demonstrate that error localization depends on scoring methodology: operations scored against immediate predecessors produce single flags per mistake, while longer reconstruction-based scores spread errors across multiple operations. They quantify error spread patterns and establish procedures for controlling false flag rates under exchangeability assumptions, even with imperfect models or content-dependent analysis selection. Results show a fundamental detection limit where small errors become indistinguishable from normal variation, with representation dimension rather than training data volume being the primary constraint.
autonomous analysis agentsinnovation-residual auditingerror localizationfalse flag controldetection limits
Matrix Zonotopic Attention: A Context-Adaptive Value Projection for Set Transformers
The paper introduces Matrix Zonotopic Attention (MZAttn), a context-adaptive value projection for set transformers that replaces fixed linear projections with input-dependent matrix zonotopes (center matrix plus gated generator matrices). Theoretical analysis defines Transformation Degrees of Freedom (TDOF) to quantify target complexity, showing MZAttn avoids depth separation issues of standard attention for high-TDOF targets. Experiments confirm MZAttn outperforms standard attention on set-prediction tasks requiring high-rank, combinatorial input dependence, while remaining competitive on low-TDOF aggregate-statistic tasks.
matrix zonotopic attentiontransformation degrees of freedomset transformerscontext-adaptive projectionpermutation equivariance
Recursive Synthesis for Long-Horizon Terminal Tasks
(No summary returned.)
Stochasticity Is Not the Hard Part: Reduction and Complexity in Instructional Sequencing over Prerequisite DAGs
The paper demonstrates that instructional sequencing with stochastic concept mastery can be reduced to a deterministic shortest-path problem on prerequisite order ideals, preserving optimality but retaining NP-hardness via reduction from feedback arc set in tournaments. The analysis identifies tractable cases: polynomial-time exact dynamic programming for fixed prerequisite width and optimal topological ordering when transfer preferences are jointly acyclic with prerequisites. A diagnostic metric $mΔ$ bounds sequencing value, validated on 70,893 CS course interactions showing doubly easy regimes, while constructed instances exhibit high myopic regret but efficient A* search with linear state expansion.
instructional sequencingstochastic shortest-pathprerequisite dagsdynamic programmingfeedback arc set
SCP-NL2TL: Selective Conformal Prediction with Semantic Verification for Natural Language to Temporal Logic Specifications
The paper introduces SCP-NL2TL, a selective conformal prediction framework for translating natural language instructions into formal temporal logic specifications with semantic verification. The method combines two reliability signals—fidelity of back-translated specifications and dispersion of repeated translations—to assess correctness, calibrated via conformal risk control for abstention decisions. It also employs a conformal anomaly detector on instruction embeddings to filter out-of-distribution inputs. Experiments on Signal Temporal Logic (STL), Linear Temporal Logic (LTL), and Spatio-Temporal Logic (SpaTiaL) demonstrate improved reliability, robustness to cross-tier shifts, and effective uncertainty-aware abstention. This work advances trustworthy natural language interfaces by enabling AI systems to recognize unreliable specifications.
conformal predictiontemporal logicsemantic verificationformal specificationsabstention
The ethics of artificial intelligence in the life sciences: Universality, cultural diversity and an architecture of care
The article argues that ethical governance of artificial intelligence in life sciences should derive from universal human cognitive architecture rather than AI-specific principles. It contrasts the brain's efficient computational model, characterized by a global neuronal workspace and cyclical reward processes, with AI's costly reward maximization paradigm. The authors propose that AI systems built on neurocognitive principles would shift governance from restraint to upbringing, requiring new institutional frameworks. They highlight the tension between universal ethical judgment and cultural diversity, emphasizing epigenetic influences on moral development. Open questions remain regarding the implementation of such governance structures.
global neuronal workspaceepigenetic appropriationreward maximizationethical governancecomputational architecture
Why the Third Axis Is Freedom
The paper identifies freedom—defined as the weakness of a model's behavioral constraints—as the key factor in Explorative Modeling (XM), challenging the prior framing of exploration as a 'third pretraining axis'. The author proves that XM's average loss depends on the probability of missing acceptable regions, with exploration (K outputs) raising this probability to power K, and shows freedom increases match probability. Empirical results demonstrate that XM optimizes for freedom: larger K increased measured freedom, and a freedom-based selector outperformed validation selection in 29/30 cases. The work formalizes freedom's role in generalization, linking it to generative expressivity.
explorative modelingfreedom selectionbehavioral constraintgeneralizationgenerative expressivity
Perturbation Sensitivity at Convergence: A Simple Signal for Identifying Spuriously Correlated Samples
The paper introduces a post-convergence perturbation sensitivity method to identify spurious correlations in trained models without group annotations. By applying fixed input perturbations to a converged model, it detects samples that rely on spurious features through their higher prediction fragility compared to robustly classified samples. This approach requires only two forward passes per sample and no validation labels. Evaluated on Waterbirds, it improves worst-group accuracy from 57.3% to 80.8%, approaching the 85.8% achieved with ground-truth labels.
spurious correlationsperturbation sensitivityworst-group accuracyempirical risk minimizationinput fragility
Negotiating Risk Boundaries in AI for Policing Through Mixed-Stakeholder Deliberation
The study examines risk assessment of AI in policing through mixed-stakeholder deliberation, explicitly addressing racial bias. A workshop with 30 participants (community representatives, police, academics) evaluated 13 AI use cases, finding broad openness to adoption but rejecting three, particularly recidivism risk assessment due to foundational objections. Analysis revealed that foregrounding racial equity expanded deliberation to core questions of efficacy, benefit distribution, and universal applicability, demonstrating the value of integrating bias considerations early in risk-benefit analysis.
racial biasrisk assessmentmixed-stakeholder deliberationai in policingrecidivism prediction
Evaluating and Improving Pedagogical Fit in LLM-Based AI Tutors with the Pedagogical Suitability Index
The study introduces the Pedagogical Suitability Index (PSI), a composite metric evaluating the instructional fit of LLM-generated tutoring responses based on learner readiness and curricular progression. PSI comprises six theory-informed sub-scores and serves as structured feedback for response improvement. Four LLMs (ChatGPT, Gemini, Gemma4, Qwen3) were evaluated across 240 scenario-based assessments using standard and defective prompts, with PSI-guided regeneration applied to 62 weak cases. Baseline PSI scores ranged from 0.557 to 0.638, showing modest differences across models. PSI-guided feedback improved 82.3% of weak cases, demonstrating instructional meaningfulness and alignment with human judgment.
pedagogical suitability indexlarge language modelsinstructional fitcurricular progressionfeedback signal
Adaptive Arena-based Contestable Argumentative Network-of-Experts for Open-Ended Care Plan Coordination
We introduce CANOE (Contestable Argumentative Network-of-Experts), a multi-agent neuro-symbolic framework for transparent and safe care plan coordination. CANOE integrates five modules: complexity assessment, adaptive team recruitment, role-based argumentative computation via Arena-based Quantitative Bipolar Argumentation Framework (A-QBAF), human-in-the-loop contestation, and care-plan synthesis. Role-specialized agents generate arguments for interventions, resolved through arena-based clash resolution, with deterministic recomputation of final plans based on human input. Evaluations on Discharge Me! and MedicalRAG benchmarks demonstrate that medically fine-tuned models achieve optimal clinical correctness and safety, while CANOE provides faithful explanation and contestability.
neuro-symbolicargumentation frameworkmulti-agentcontestabilitycare-plan synthesis
C$^3$PO: Evaluating Cross-Modal Composition and Counterfactual Performance in Omnimodal Models
The C$^3$PO benchmark introduces 3,404 multimodal samples to evaluate cross-modal composition (IC) and counterfactual conflict (CC) in Multimodal Large Language Models (MLLMs). Using 25 logically grounded templates and an automatic pipeline, it reveals a 56-point accuracy gap between humans (88.64%) and top models (Gemini-3.1-Pro at 73.17%), with open-source models failing under conflict. Attention probes show 86-95% of errors stem from modality dominance, where models fixate on text (87-95% attention) and ignore contradictory evidence. Mid-layer attention entropy predicts success, highlighting the need for sustained cross-modal exploration in architectures.
multimodal large language modelscross-modal reasoningattention probesmodality dominancecounterfactual conflict
DoctorAgents: an agentic framework to iteratively refine AutoML pipeline for small clinical temporal data
DoctorAgents introduces an agentic framework for optimizing AutoML pipelines in small clinical temporal data settings, replacing brute-force search with reasoning-driven refinement. The system employs specialized LLM agents for pipeline generation, validation, and refinement, utilizing natural-language feedback via textual gradient descent to iteratively improve models. Experimental results demonstrate consistent outperformance over traditional AutoML baselines while yielding more interpretable task-specific representations across diverse clinical tasks.
automlclinical machine learningllm agentstextual gradient descenttemporal data
Counterfactual Analysis via Large Language Models
The paper demonstrates GPT-3.5's capability for counterfactual analysis in online lending, achieving an R-squared of 2.84% with prompt engineering, nearing gradient-boosted regression's 3.48%. The method involves evaluating predictive performance and generating counterfactual return on investment (ROI) under alternative interest rates, showing logical coherence in causal reasoning. Results suggest LLMs like GPT-3.5 can effectively support decision-making in financial scenarios through counterfactual prediction.
counterfactual analysislarge language modelsprompt engineeringgradient-boosted regressioncausal reasoning
CASCADE: An Agentic Regulatory Network Framework for Patient-Data-Validated Downstream Perturbation Prediction
CASCADE introduces an agentic framework for predicting downstream transcriptional effects of gene perturbations using precomputed ARACNe regulatory networks. It validates predictions by comparing the direction of gene expression changes against real TCGA patient tumor data, using focal-gene copy-number amplification as a proxy. For MYC, CASCADE achieves high concordance rates (BRCA: 90.0%, COAD: 72.0%, STAD: 85.7%; p<0.0013) and replicates in an independent cohort (METABRIC, 87.2%). Validation shows gene-specific accuracy, with proliferation-machinery regulators replicating well but lineage-identity transcription factors failing. An LLM-based agent achieves 71.4% exact match in grounding natural-language requests to CASCADE's tool calls.
regulatory networksgene perturbationcopy-number amplificationtranscriptional effectsllm-based agent
Multi-Agent Reinforcement Learning for Online Traffic Scheduling in Time-Sensitive Application
The paper proposes a multi-agent reinforcement learning (MARL) framework for dynamic traffic scheduling in time-sensitive networking (TSN) for mobile edge computing (MEC) applications like extended reality (XR). Using Heterogeneous-Agent Proximal Policy Optimization (HAPPO), the method models each TSN queue as an autonomous agent to capture inter-queue dependencies and jointly optimize service delivery. Simulations show reductions of 26.8% in average frame waiting times and 16.8% in worst-case delays compared to existing approaches.
multi-agent reinforcement learningtime-sensitive networkingmobile edge computingheterogeneous-agent proximal policy optimizationextended reality
Multi-Agent Transformer for Queue-Level XR Traffic Scheduling in TSN Networks
A multi-agent transformer (MAT) is proposed for queue-level scheduling of eXtended Reality (XR) traffic in Time-Sensitive Networking (TSN) environments. The method leverages multi-agent reinforcement learning with attention mechanisms to model inter-queue dependencies and enable implicit coordination across heterogeneous XR applications. Evaluations demonstrate significant improvements over baselines, achieving up to 71.42% latency reduction, 83.2% reduction in failure rate, and consistent reliability across all queues.
multi-agent transformertime-sensitive networkingextended realityqueue-level schedulingattention mechanisms
Hierarchical Server Architecture for Agentic Science
The paper introduces a hierarchical, dynamic architecture for resource discovery across cloud, edge, and HPC systems, tailored for agentic science workloads. The design employs secretary agents to concurrently and asynchronously negotiate, select, and dispatch work requests, probing 51 real and simulated providers across 7 categories. Through 19,973 negotiation and 6,952 selection simulations, the system achieves 87.71% negotiation accuracy and selection costs comparable to traditional strategies. The architecture, extensible and currently supporting the Genesis Mission, highlights the critical coordination between agents, discovery tools, and infrastructure in agentic science.
agentic sciencehierarchical architectureresource discoverysecretary agentsnegotiation accuracy
Failing Gracefully: Mitigating Impact of Inevitable Robot Failures
The paper introduces a safety formulation for service robots that quantifies both the probability and severity of impactful interactions during failures, enabling informed planning decisions balancing safety and efficiency. The method evaluates failure impacts on household entities (humans, pets, objects) and presents FailBench, a MuJoCo-based simulation framework for studying diverse failure modes (sensing issues, actuator malfunctions). This combined approach provides a foundation for safer motion planning and learned policies in real-world household environments.
service robotsfailure mitigationsafety formulationmujoco simulationmotion planning
Agentic self-driving microscopy benchmarks support qualification but do not necessarily generalize to unseen tasks
The study introduces a benchmark and trace-logging framework to evaluate agentic microscopy controllers, assessing 105 configurations across 53 tasks (1,949 test runs, 49,109 RAG retrievals). It compares one-, two-, and three-agent graph topologies, five LLMs, and RAG parameters, quantifying differences in latency, token use, cost, and failure modes. Results indicate benchmarks aid qualification and direct comparison but fail to predict performance on unseen tasks, as surrogate models trained on architecture-test correlations lacked generalizability.
agentic microscopyretrieval-augmented generationllm agentsbenchmark frameworktask generalization
OPD-V: Visual On-Policy Self-Distillation with Modality Balance
The paper introduces OPD-V, a visual on-policy self-distillation (OPSD) paradigm addressing modality imbalance in multimodal large language models (MLLMs). By constructing Positive and Negative Teachers with zoom-in and masked images respectively, the method identifies modality-balance trust regions through logit margins, selecting optimal tokens for self-distillation. Experiments across 6 benchmarks, 4 MLLM architectures, and 5 post-training methods demonstrate consistent reasoning improvements with reduced training costs.
on-policy self-distillationmodality imbalancemultimodal llmstrust regionlogit margins
OrchestraBench: Evaluating Multi-Agent Orchestration Failure Modes, Recovery, and Decomposition Quality
OrchestraBench introduces a diagnostic framework for evaluating multi-agent orchestration failures, recovery, and decomposition quality through controlled failure-injection harnesses over templated workflows. It proposes cascade radius and per-failure-mode recovery metrics, comparing routing policies with bootstrap confidence intervals and paired tests. Key results include: intent-reasoning routers outperformed keyword-based ones (100% vs. 0% on adversarial cases), three distinct failure-handling tiers emerged (full recovery for tool faults, partial for ambiguous delegation, none for latent/semantic modes), and cascade radius scaled with pipeline depth (0.9 to 4.7 across depths 3-7). Results were consistent across Claude variants (Sonnet, Opus, Haiku) and workflow domains.
multi-agent orchestrationfailure-injectioncascade radiusintent-reasoning routertrusted-state repair
IMMENSE: Inductive Multi-perspective User Classification in Social Networks
IMMENSE introduces a hybrid inductive learning method for malicious user detection in social networks, combining semantic content analysis, social relationships, and spatial information without requiring retraining for new users or networks. The approach outperforms five state-of-the-art baselines on a real-world Twitter/X dataset, demonstrating the efficacy of multi-perspective contextual integration over text-only analysis for LEAs.
inductive learningsocial network analysismulti-perspective classificationmalicious user detectioncontextual integration
Posture and Sustainment Optimization Under Adversarial Uncertainty
The paper introduces a scenario-weighted adversarially robust posture optimization engine for the Posture and Sustainment Allocation (PSA) problem, modeled as a finite-horizon Markov Decision Process. The Composite Expected Value (CEV) optimizer maximizes scenario-weighted expected posture efficiency over threat distributions, while RobustCEV extends this by iterating against a Bayesian adversary. Experiments in an Indo-Pacific basing environment show greedy baselines suffer 25.1% efficiency loss and 57.3% readiness collapse under adversarial threat; CEV recovers 19.8% efficiency with geographic signals, and RobustCEV achieves 158% efficiency gains against deceptive adversaries, validated via statistical testing.
posture optimizationmarkov decision processadversarial robustnessscenario-weighted efficiencybayesian adversary
An Emerging Retail Portfolio Management Application: Personalized, Tax-Aware Reinforcement Learning with Natural Language Goals
We introduce a personalized, tax-aware reinforcement learning system for retail portfolio management, enabling users to specify investment goals in natural language. The system comprises a self-supervised cross-asset encoder, a Mixture-of-Experts allocation policy with a learned intent router, and a LoRA adapter for personalization without retraining the shared model. Integration-tested end-to-end with a live brokerage API (Alpaca, paper-trading mode), it includes multi-user authentication, a preview-before-apply confirmation flow, daily email digests, and an auditable action-integrity chain. Preliminary validation via 14-day walk-forward backtests shows promise, though the system remains pre-deployment. Practical engineering lessons highlight challenges with integration paths and third-party API calls.
reinforcement learningmixture-of-expertslora adapterself-supervised encoderbrokerage api
PRISM: Priority-aware Rubric Internalization via Structured Multimodal Data Synthesis
PRISM introduces a four-stage data synthesis framework for rubric comprehension in multimodal instruction following, addressing the gap in handling prioritized, multi-rule instructions. The framework generates persona-task pairs, prefix-guided rule sets, quality-filtered rubrics, and structured verification traces. PRISM-Eval, its evaluation metric, employs deterministic matching without requiring inference-time judge models. With 10K synthesized samples, PRISM improves Qwen3-VL-4B's Strict accuracy on PRISM-Eval from 9.5% to 30.1%, while maintaining general benchmark performance. Gains transfer to four additional open-source multimodal language models, demonstrating scalability of structured rubric supervision.
rubric comprehensionmultimodal instructiondata synthesisdeterministic matchingstructured verification
WorldClaw: Agentic 3D Open-World Generation at Scale
WorldClaw introduces an agentic, coarse-to-fine framework for generating large-scale 3D open worlds from text prompts. The system employs planning agents to decompose prompts into structured specifications (regions, terrain, assets) and constructs globally coherent scenes via semantic layouts, procedural materials, and region-aware height fields. Detail-rich regions are enhanced through terrain-conditioned compositions, mesh reconstruction, and render-based refinement. Evaluations demonstrate coherent spatial organization, visually compelling local content, and editable instance-level assets across diverse prompts while maintaining global terrain consistency.
3d scene generationagentic frameworkprocedural materialssemantic layoutsmesh reconstruction
LUNAR: Benchmarking Personalized Large Language Models on UNiversal User BehAvioR Logs
LUNAR introduces the first benchmark for evaluating personalized LLM responses grounded in longitudinal, cross-domain user behavior logs (clothing, food, housing, mobility). It employs a multi-stage coarse-to-fine synthesis pipeline to generate realistic behavioral distributions while addressing data sparsity and privacy. Experiments with 19 LLMs reveal that behavioral log access is necessary but insufficient for deep personalization, with performance depending on evidence selection and cross-domain integration; fine-grained retrieval outperforms compressed memory, though stronger personalization compromises privacy.
personalized llmsbehavioral logscross-domain integrationsynthetic benchmarkprivacy-utility tradeoff
Search2Skill: Skill Distillation Beyond Knowledge Boundaries Via Rubric-Based Reinforcement Learning
Search2Skill introduces a framework for skill distillation in LLM-based agents that transcends parametric knowledge boundaries by identifying capability gaps, searching external sources, and distilling retrieved evidence into reusable skills. The method employs rubric-based reinforcement learning to jointly optimize search timing, search strategy, and skill generation. Evaluated across eight expert-level domains from three benchmarks, Search2Skill outperforms search-augmented and trajectory-based baselines in both streaming and held-out protocols. Gains are attributed to skill abstraction rather than raw evidence, and acquired skills demonstrate cross-model transferability.
skill distillationrubric-based reinforcement learningcapability gapssearch-augmentedcross-model transferability
Marginal Matching Does Not License Factorized Sampling: Auditing Conditional Style Leakage in Factorized Generative Models
The paper demonstrates that marginal distribution matching in factorized generative models does not ensure independence between latent style variables and class labels, contrary to common assumptions. Through an exact decomposition, the authors identify four necessary conditions for factorized sampling and show that eliminating marginal mismatch is insufficient for proper factorization. Empirical evaluations reveal that models achieve near-zero global MMD while still allowing linear probes to recover class labels with 74%-100% accuracy. Mitigation strategies reduce probe accuracy to 21%-46%, but within-class dependence persists. Conditional priors improve class generation to 0.97 on MNIST but only 0.41 on CIFAR-10, highlighting the limitations of marginal statistics in verifying independence.
factorized generative modelsmarginal distribution matchinglatent style variablesclass-conditional distributionsmmd
CalibForge: Adversarial Solver Calibration for Scaling Learnable Terminal Tasks
CalibForge introduces adversarial solver calibration for synthesizing learnable terminal tasks by revising candidates based on verified solver behavior. The method employs multi-solver calibration (targeting solver disagreement) and contrastive calibration (enforcing strong-pass/weak-fail relations) to create tasks within a solver-relative learnable zone. Evaluated on 5,431 calibrated tasks, models trained with CalibForge achieve 32.58% and 47.57% on Terminal-Bench 2.0, with improvements up to 30.04 points on Doc2Repo, demonstrating the efficacy of solver-relative learnability for agent training data.
adversarial calibrationterminal taskssolver disagreementcontrastive calibrationlearnable zone
Scalable estimation of VARMA models
The authors introduce a scalable framework for estimating vector autoregressive moving-average (VARMA) models, overcoming traditional computational barriers. Their method employs a partial-autocorrelation reparametrization ensuring stationarity, Gaussian priors with distinct scales, and Parseval-based evaluation for efficient computation. Two estimators—regularized least-squares and covariance-marginalized MAP—are proposed, achieving near-parametric rates in recovering infinite-autoregressive representations. Empirical results show competitive performance against VAR, Bayesian-VAR, and sparse-VARMA baselines, scaling to dimensions (d=10 to d=40) where classical methods fail. The framework extends to seasonal dynamics, VARMAX, and rolling-window refits with minimal additional cost.
varmareparametrizationparsevalstationarityvarmax
Optimal Rates for Learning with Monotone Adversaries
The work establishes minimax optimal error rates for learning with monotone adversaries, showing that the logarithmic overhead in Larsen et al.'s bound is inherent for VC dimension d ≥ 2. Using elementary constructions and a leave-one-out analysis, the authors prove worst-case expected errors of Θ(1/n) for d=1 and Θ((d/n)log(n/d)) for d≥2, matching upper and lower bounds. The same rates hold for Littlestone dimension, demonstrating that correctly labeled adversarial insertions can degrade learning performance logarithmically even for online-learnable classes.
monotone adversaryvc dimensionminimax ratesonline-to-batchimproper learning
RRC: Unlocking Generative Reward Models in LLM Reinforcement Learning via Ranking-Based Reward Construction
The paper introduces Ranking-based Reward Construction (RRC), a method to bridge the gap between generative reward models and reinforcement learning (RL) by deriving rewards from relative preference rankings. RRC employs two strategies: self-competitive ranking, which compares sampled responses, and anchor-guided ranking, which uses a small set of reference responses for scalable reward construction. Experiments on open-ended chat and reasoning benchmarks show that RRC significantly enhances RL training with generative reward models, outperforming existing reward construction approaches.
generative reward modelsreinforcement learningranking-based rewardself-competitive rankinganchor-guided ranking
On-Policy Self-Distillation without Any Supervision
The study introduces Unsupervised On-Policy Self-Distillation (U-OPSD), a method for improving large language models (LLMs) without external supervision. U-OPSD constructs pseudo-solutions via majority vote from model rollouts, conditions a teacher distribution on the shortest correct solution, and distills it into incorrect completions. Evaluated on benchmarks (AIME24, AIME25, HMMT25, MATH500, AMC23), U-OPSD improves base Qwen3 models by 8.5-10.7% and outperforms supervised methods like OPSD by 0.9-3.2%, while matching or surpassing GRPO.
self-distillationmajority votepseudo-solutionrolloutsunsupervised learning
Surv-IPTB: An Attention-Based Model for Estimating Individual Probability of Treatment Benefit with Survival Data
Surv-IPTB introduces an attention-based model for estimating Individual Probability of Treatment Benefit (IPTB) in survival analysis, reformulating it as a binary classification problem via pairwise patient comparisons. The framework handles right-censored data through imprecise probability representations and employs learnable query-key attention mechanisms to aggregate comparisons while inferring soft probabilities for censored cases. Experiments on synthetic datasets with nonlinear structures demonstrate superior performance over meta-learner baselines (T-learner, S-learner) using random survival forests, Cox models, and Beran estimators, particularly in complex feature spaces with varying censoring rates.
attention mechanismsurvival analysisimprecise probabilitytreatment effectpairwise comparison
The Tamed Subgradient Unadjusted Langevin Algorithm beyond Convexity
The paper introduces the Subgradient Tamed Unadjusted Langevin Algorithm (SG-TULA), a novel method for sampling from non-smooth, non-convex target distributions with superlinear gradient growth. SG-TULA operates directly on subgradients, avoiding computationally intensive smoothing procedures, and employs taming techniques for stability. The authors derive non-asymptotic convergence bounds in Wasserstein-2 distance, explicitly tracking constants in terms of dimension and inverse temperature, improving upon existing subgradient-based Langevin algorithms. Additionally, excess risk estimates for the associated optimization problem are provided. The method is validated on the regularized pretraining potential of a GPT-2 lineage LLM, demonstrating competitive performance against finetuned AdamW and Muon.
subgradientlangevin algorithmwasserstein-2non-convexsuperlinear
Stochastic Dynamics on Persistence Diagram Space via Reinforcement Learning
The paper introduces a reinforcement learning framework for stochastic dynamics on persistence diagram (PD) space, enabling probabilistic modeling and topological evolution through local edit operations. The method defines controlled Markov processes on finite PDs with variable cardinality, establishing conditions for irreducibility, aperiodicity, and geometric ergodicity, which guarantee unique stationary probability laws. Objectives include distribution matching, task-specific topological statistics, and structure-preserving compression, balancing fidelity, complexity reduction, and distributional targets. Experiments on synthetic and neuroimaging PDs demonstrate the framework's ability to preserve dominant topological structure while reducing diagram complexity.
persistence diagramsreinforcement learningmarkov processestopological simplificationgeometric ergodicity
OTLesMix: Wasserstein Barycenter and Optimal Transport Map for Synthetic Lesion Generation with Diverse Shapes and Locations
OTLesMix introduces a synthetic lesion generation method using Wasserstein barycenter and optimal transport to enhance shape and location diversity in medical imaging. The approach addresses limitations of existing mix-based augmentation techniques by producing more varied synthetic samples. Evaluated on three brain lesion segmentation tasks, OTLesMix improves Dice scores by 2.9 to 6.6 points over baseline models without synthetic data and outperforms state-of-the-art mixing methods.
wasserstein barycenteroptimal transportlesion segmentationdata augmentationsynthetic generation
Hypothesis Testing with Conditional Queries: Learnability and the Value of Interaction
The paper characterizes the learnability of hypothesis testing under conditional queries, showing that distinguishability requires positive separation in pairwise conditional probabilities between distribution classes. For cases with zero separation, the worst-case error remains 1/2 regardless of query budget. The authors prove that adaptive testing can reduce query complexity quadratically compared to non-adaptive methods: they construct a randomized non-adaptive procedure using O(N²(T + log(1/ρ))) pair queries to simulate any T-query adaptive policy within ρ total variation distance, and demonstrate a matching Ω_ε(N²) lower bound for non-adaptive queries. The adaptivity gap is thus Θ_ε(N²).
hypothesis testingconditional queriesadaptive testingnon-adaptive testingquery complexity
RxnCLF: Contrastive Transformation-Aware Reaction Foundation Model for Improved Reactivity Prediction
RxnCLF introduces a self-supervised contrastive learning framework for reaction representation using condensed reaction graphs (CRGs), unifying reactant and product information to capture explicit transformation structure. Pretrained on 1.7M Pistachio reactions, RxnCLF learns a transformation-aware latent space encoding reaction-center features and side-chain contexts. Fine-tuned on Buchwald-Hartwig, Pd-catalyzed BH coupling, and proprietary datasets, it outperforms graph/sequence baselines in yield prediction (R2 improvement), demonstrating potential for broader reaction informatics tasks like regioselectivity prediction and condition optimization.
condensed reaction graphcontrastive learningreaction yield predictionself-supervised pretrainingtransformation-aware representation
MetaboLLM: a metabolomics-specialized large language model for biochemical knowledge integration and predictive metabolite graph construction
MetaboLLM introduces a metabolomics-specialized LLM via continual pretraining, supervised fine-tuning, and structured retrieval, paired with MetaboLLM-GIN to construct predictive metabolite graphs using a graph isomorphism network. The method outperformed base and medically adapted models across four backbone families on metabolomics tasks (knowledge, relational, description) and transferred to external benchmarks. MetaboLLM-GIN achieved top AUCs for stress hyperglycemia prediction (0.8616) and hormone-regimen classification (0.8123), surpassing conventional models and alternative graph constructions. Interpretability revealed biologically meaningful patterns, demonstrating domain-specialized LLMs can structure biochemical knowledge into actionable representations.
metabolomicsgraph isomorphism networkcontinual pretrainingstructured retrievalauc
Minimax Optimal Early-Stopped Gradient Descent for Gaussian Mixture Classification
The work demonstrates that early-stopped gradient descent (GD) achieves minimax-optimal excess zero-one risk for Gaussian mixture classification with label-flipping noise, particularly for covariance spectra exhibiting fast or continuous decay (e.g., polynomial, exponential). By combining a sharp upper bound for the early-stopped iterate with a matching statistical lower bound, the analysis proves optimal rates, validated empirically. A key technical contribution is a calibration result converting excess logistic risk to zero-one risk, addressing model misspecification and eliminating square-root rate bounds. The paper also shows linear interpolators may require exponentially more samples than early stopping for equivalent excess risk.
gradient descentminimax optimalitygaussian mixture modellabel-flipping noiseearly stopping
A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance
This survey introduces a six-dimensional taxonomy for classifying post-training adaptation techniques, addressing fragmentation in the literature across method families, model classes, and deployment contexts. The taxonomy organizes techniques by mechanism, goal, data requirement, persistence, structural scope, and model type, clarifying distinctions between fine-tuning, retrieval augmentation, and prompting. It traces the evolution of adaptation strategies from traditional machine learning to multimodal large language models, mapping relationships through inheritance, supersession, hybridization, and layered deployment stacks. The taxonomy supports technical documentation, model-change tracking, and governance analysis, while identifying open challenges in evaluation, reproducibility, persistent inference-time adaptation, unlearning, multimodal adaptation, and governance-aware workflows.
taxonomypost-training adaptationmultimodal instruction tuningretrieval augmentationgovernance analysis
Timestep-Conditioned Transformers for Global Weather Forecasting
GEM-3 introduces a timestep-conditioned transformer for global weather forecasting, enabling multi-timestep inference with a single set of trained weights to balance predictability and usability across forecast horizons. The model employs a lightweight neighborhood-attention transformer with ~134M parameters on an equirectangular grid, incorporating architectural advancements over GEM-2. Mixed-timestep training enhances rollout stability compared to timestep-specialist models. GEM-3 achieves near-state-of-the-art medium-range probabilistic skill, stable extended-range rollouts, and efficient training and inference, while providing decision-relevant diagnostics.
timestep-conditionedneighborhood-attention transformermulti-timestep inferenceequirectangular gridrollout stability
Robot Learning from Human Demonstrations: Handwritten Alphabet Trajectories and Human-Likeness Evaluation
The paper presents a framework for learning human-like robot motion from demonstrations, extending Gaussian Mixture Model/Regression to incorporate force and normalized time dimensions while handling multi-segment trajectories. A dataset of 3,142 handwriting demonstrations (22 participants, 52 Latin character-case combinations) was collected via touchscreen teleoperation, capturing position, force, and timing. In a user study (N=21), generated trajectories achieved a human-likeness score of 71.50 (SD=22.56) on a 0-100 scale, with geometric positioning and trajectory sequence identified as key perceptual factors.
learning from demonstrationgaussian mixture modelhuman-robot interactiontrajectory generationteleoperation
Muon on the Stiefel Manifold Admits an Exact Closed-Form Update
The authors demonstrate that Muon, a matrix-aware optimization method, admits an exact closed-form solution on the Stiefel manifold, which consists of matrices with orthonormal columns. They introduce Skewon, a practical algorithm leveraging this result for orthogonality-constrained optimization, featuring an efficient implementation. Theoretical analysis establishes first-order convergence guarantees for Skewon in the smooth non-convex setting, addressing limitations of prior heuristic or iterative approaches.
stiefel manifoldmuonorthogonality-constrained optimizationclosed-form solutionnon-convex optimization
Beyond Marginal Validity: Finite-Sample Guarantees for Localized Conformal Prediction
The paper provides finite-sample guarantees for randomly localized conformal prediction (RLCP), addressing the gap between marginal validity and covariate-specific miscalibration. RLCP calibrates near test points while preserving marginal coverage, with new theoretical bounds on conditional validity and oracle efficiency. Under Hölder regularity and standard assumptions, the analysis yields high-probability bounds for conditional-coverage gaps and length errors, decomposing into localization bias and calibration terms. The work also extends to data-split learned scores, showing improved learning sharpens localized guarantees when targeting pivotal scores.
conformal predictionfinite-sample guaranteesconditional coveragelocalized calibrationoracle efficiency
Handling Missing Data in Probabilistic Regression Trees
The paper extends Probabilistic Regression Trees (PRTrees) to handle missing predictor values directly during tree construction, eliminating pre-processing imputation. Three strategies are proposed: uniform-probability, partial-observation, and dimension-reduced smoothing approaches, all preserving PRTree's probabilistic properties (probability conservation, marginal compatibility) under arbitrary missingness patterns. Evaluated on real-world datasets with varying missingness levels, the methods often outperformed CART while maintaining interpretability, with the fill strategy being the dominant performance factor in high-missingness scenarios.
probabilistic regression treesmissing dataimputationmarginal compatibilitypredictive performance
On Same-Sample and Independent-Sample Stochastic Extragradient for Monotone Variational Inequalities
This work advances the theoretical understanding of stochastic extragradient (SEG) methods for monotone variational inequality problems (VIPs) by contrasting same-sample SEG (S-SEG) and independent-sample SEG (I-SEG). The authors demonstrate that S-SEG's convergence depends critically on samplewise Lipschitz parameters, unlike I-SEG, and establish high-probability restricted-gap convergence for both variants under relaxed assumptions. They prove that fundamental improvements to these results are generally impossible and show that a known double step-size selection guaranteeing almost sure last-iterate convergence for I-SEG can fail for S-SEG, which may diverge almost surely under the same conditions.
stochastic extragradientmonotone variational inequalitiessamplewise lipschitzrestricted-gap convergencestep-size selection
SAGA: Score-Weighted Adaptive Generation Alignment for Low-Resource Nordic Language Models
SAGA (Score-weighted Adaptive Generation Alignment) introduces a parser-guided preference optimisation framework for low-resource Nordic languages, eliminating the need for costly human preference annotations. The method converts dependency parser judgements into preference pairs for delta-DPO, integrates parser quality and lexical diversity into a composite reward, filters low-information pairs via reward-gap criterion, and monitors reward hacking. Evaluated on Danish, Icelandic, and Norwegian Bokmål using GPT-SW3-1.3B, SAGA improves grammatical quality: Danish parse success increases from 69.0% to 93.8%, Icelandic achieves a +4.5 percentage-point improvement on Stanza evaluation, and Norwegian Bokmål improves by +28 percentage points. Native speakers prefer SAGA outputs in 80% of pairwise comparisons.
dependency parserdelta-dporeward hackinglow-resource languagesgrammatical alignment
Threshold-Based Early Stopping of Accumulations in Neural Networks with Binary Activation
The paper introduces a threshold-based early stopping mechanism for binary neural networks, reducing computational waste in accumulations without retraining. By analyzing running partial sums during inference, the method predicts final activation signs early, halting unnecessary computations. Evaluated on VGG11 with CIFAR-10, it eliminates 86.6% of accumulation terms in the deepest convolution with a 0.37-point accuracy drop and reduces 25% of full-network arithmetic across three deepest convolutions with a 1.36-point drop.
binary neural networksearly stoppingpartial sumsaccumulation termscomputational efficiency
Verifiable Regularity Criterion for Conditional Expectation Operators and Conditional Mean Embeddings with Applications to Nonparametric Regression, Bayesian Inverse Problems, and Koopman Operators
The paper establishes a verifiable regularity criterion characterizing when conditional expectation operators (CEOs) map between function spaces, particularly reproducing kernel Hilbert spaces (RKHSs). The key insight is that this property depends on the Radon-Nikodym density's regularity, with Sobolev regularity sufficing for Sobolev-equivalent RKHSs. The authors prove boundedness and Hilbert-Schmidt properties under this condition, enabling rigorous error analysis for Galerkin and conditional mean embedding estimators. They validate the framework in nonparametric regression, Bayesian inverse problems, and Koopman operator theory, showing classical probabilistic regularity implies the required operator properties.
conditional expectation operatorsreproducing kernel hilbert spaceradon-nikodym densitysobolev regularitykoopman operators
SkillTFM: Gated Skill Evolution for Training-Free Adaptation of Tabular Foundation Models
SkillTFM introduces a training-free adaptation system for tabular foundation models (TFMs) that addresses distribution shifts and task-specific patterns without fine-tuning. The method employs a verifiable skill bank combining boundary evidence identification with gated skill evolution, enabling reusable skill retrieval and extension. Evaluations on simulated boundary settings and electricity-price forecasting demonstrate AUC improvements of 0.128--0.142 and nonlinear-boundary AUC increases from 0.699 to 0.898. The system's effectiveness is validated across multiple TFM backbones, showcasing its generality.
tabular foundation modelsgated skill evolutionboundary evidence identificationskill banktraining-free adaptation
LLM Inference Under Bursty Workload Distribution: Modifying the WAIT Algorithm
The paper proposes a lightweight extension to the WAIT algorithm for LLM inference scheduling, adapting to bursty workloads without prior traffic knowledge by performing online estimation of request intensity from interarrival times. Using Markov Modulated Poisson Process (MMPP)-based synthetic workloads, simulations show the method outperforms Sarathi-Serve, ORCA, and vLLM in throughput under low arrival-rate shift scenarios while maintaining comparable latency, addressing the limitation of constant-rate Poisson assumptions in prior work.
llm inferencescheduling algorithmmarkov modulated poisson processthroughput optimizationbursty workload
Kastor: An efficient fine-tuning strategy for generative emulation of PDE simulations
The paper proposes Kastor, a fine-tuning strategy for generative PDE emulation that combines a two-stage inference scheme (large-stride auto-regression with temporal super-resolution) and Mean Prediction Regularization (MPR) to reduce error accumulation and improve stability. The method incorporates spatial gradient matching for physical fidelity, achieving a 42.9% average forecasting improvement over Walrus-based baselines and superior VRMSE on 8/10 datasets in The Well benchmark.
partial differential equationsgenerative emulationerror accumulationtemporal super-resolutionmean prediction regularization
ML-for-ML
ML-for-ML proposes a cross-layer optimization framework that jointly selects network-side and ML-side parameters under a shared time-to-target-loss objective, addressing the suboptimal separation between networking controls and ML training choices in cloud clusters. The method co-optimizes network mechanisms (e.g., bandwidth allocation) and ML training parameters (e.g., communication frequency) to reduce contention in shared environments. Preliminary results demonstrate a 42% reduction in time-to-target-loss compared to decoupled optimization approaches.
cross-layer optimizationcloud clusterstime-to-target-lossnetwork mechanismsml training
Dynamic Graph Prompting via Topology-Routed Mixed-Curvature Experts
The paper introduces CurvPrompt, a dynamic graph prompting framework that addresses geometry under-adaptation in temporal graphs by routing nodes to curvature-diverse Riemannian experts. The method employs a bank of experts with learnable prompts, dynamically assigning nodes via a topology-aware gate to construct mixed-curvature representations. Soft routing during pre-training ensures stability, transitioning to hard Top-K routing for downstream tasks. Experiments on four benchmarks demonstrate significant improvements in few-shot link prediction and consistent node classification performance, validating geometry-adaptive prompting.
dynamic graph promptingriemannian expertsmixed-curvature representationgeometry under-adaptationtopology-aware routing
Hybrid-Adaptive Thread Tuning to Mitigate Simulation Execution Bottlenecks in High-Performance Reinforcement Learning Inference
AutoThread, a hybrid adaptive thread-tuning method, mitigates simulation bottlenecks in reinforcement learning (RL) inference by dynamically optimizing thread allocation. It combines a Physics-Informed Neural Operator (PINO) for thread-count prediction with an M/M/1 queueing model to guide estimation under dynamic workloads, supplemented by load-aware online fine-tuning. Experiments demonstrate 18.4% average speedup over static strategies, 1.7x and 1.8x higher throughput than XGBoost and Reinforcer, respectively, and up to 83.8% reduction in execution time versus state-of-the-art methods.
reinforcement learningthread tuningsimulation bottlenecksphysics-informed neural operatorqueueing model
BioKD: Selective Physiology-to-Video Knowledge Distillation via Reliability Gate for Emotion Recognition
BioKD introduces a reliability-aware physiology-to-video knowledge distillation framework for emotion recognition, addressing noisy physiological supervision via a sample-wise gating mechanism and progressive distillation. The method leverages physiological signals as privileged training information to guide a video-based student model, while requiring only video inputs at inference. On DEAP and AMIGOS, BioKD achieves 68.01% trial-wise arousal accuracy (65.29% subject-wise), outperforming baselines by mitigating overconfident teacher errors and negative transfer. The approach maintains inference efficiency by eliminating physiological sensing post-training.
knowledge distillationemotion recognitionprivileged informationreliability gatingmultimodal learning
Do Tabular Foundation Models Agree with Themselves?
The paper evaluates the internal consistency of Tabular Foundation Models (TFMs) by proposing two requirements: marginalization consistency (equality between marginalized conditionals and directly predicted marginals) and factorization consistency (invariance of joint distributions to factorization order). TFMs, which approximate Bayesian posterior predictive distributions via transformer architectures, are tested on classification and regression tasks. Results show all evaluated TFMs violate both consistency requirements across all datasets, questioning their faithfulness in modeling joint distributions when targets are sampled autoregressively.
tabular foundation modelsbayesian posterior predictivemarginalization consistencyfactorization consistencyautoregressive sampling
A Unified Risk View of Uncertainty: Posterior Risk for Disentanglement and Evaluation Beyond Proxies
The paper introduces a unified definition of uncertainty as pointwise posterior risk, combining Bayesian uncertainty over functions with estimator-dependent deviations from the posterior mean. This formulation enables direct computation of oracle epistemic and aleatoric uncertainty using semi-synthetic datasets with real covariates and known generative processes, avoiding proxy evaluations. Empirical results demonstrate that accurate prediction does not ensure reliable uncertainty disentanglement, revealing practical differences between methods and their alignment to oracle uncertainty targets.
posterior riskepistemic uncertaintyaleatoric uncertaintysemi-synthetic datasetsoracle uncertainty
Observation-Grounded Self-Predictive Reinforcement Learning for Visual Continuous Control
The paper introduces Observation-Grounded Self-Predictive Representations (OG-SPR), a model-free visual RL algorithm for continuous control that combines latent self-prediction and observation prediction to improve sample efficiency. OG-SPR employs two auxiliary objectives—multi-step latent self-prediction and next-observation prediction—while avoiding over-constraining the shared representation via lightweight adapters. Evaluated on 28 visual control tasks from the DeepMind Control Suite, OG-SPR outperforms state-of-the-art self-predictive and observation-predictive methods, with significant gains in complex domains like dog and humanoid locomotion.
reinforcement learningcontinuous controlrepresentation learningsample efficiencyvisual control
THBKG: A Temporal Biomedical Knowledge Graph for Decision-Aligned Clinical Advancement Prediction
The Temporal Heterogeneous Biomedical Knowledge Graph (THBKG) introduces a novel framework for predicting therapeutic target–disease linkage advancement by reconstructing evidence profiles as they existed at past decision points. THBKG comprises 110,396 entities and 11.1M edges across 19 relation types, with each edge timestamped by the year its evidence changed. Graph propagation on THBKG outperforms direct-evidence baselines in predicting Phase II to Phase III advancement, achieving a relative success of 4.3–4.5 for top ten pairs per therapeutic area. Notably, it excels for 72.8% of pairs lacking direct evidence, ranking 5-6x above chance by propagating over intervening biological pathways. A path-based explainer decomposes predictions into decision-time evidence landscapes for interpretability.
temporal knowledge graphtherapeutic targetgraph propagationdecision-aligned predictionbiomedical evidence
How Far Do Simple Transformations Translate Across Text Embedding Models?
The study evaluates whether simple linear transformations can translate semantic representations across heterogeneous text embedding models, testing the hypothesis of latent universality. Using centered kernel alignment (CKA), downstream transfer, fidelity, and retrieval metrics, the authors analyze nine embedding models varying in architecture, pooling, and training objectives. Results indicate that lightweight translators recover shared structure for compatible model pairs but fail for others, demonstrating that embedding spaces are not universally related by simple mappings as previously suggested.
text embeddingslatent universalitylinear mappingsckasemantic transfer
Operating Multi-Node Full Fine-Tuning on NVIDIA B300: A Field Report on Telemetry-Based Triage, Negative Results, and Operational Hardening
The study presents operational insights from full fine-tuning a 32.76B-parameter Qwen3-32B model on 16 NVIDIA B300 GPUs (two nodes, FSDP/ZeRO-3), focusing on telemetry-based triage and negative results. Key contributions include: (1) a B300 power-draw triage table distinguishing compute/communication/data-starvation states; (2) empirical refutations of optimization folklore (e.g., NFS reading matches local cache at ~53k tok/s); (3) near-linear strong-scaling metrics for 4/8/16-GPU configurations; and (4) a deadlock case study (NCCL epoch-end imbalance) mitigated via invariant gating (2.7s pre-run check). Findings emphasize monitoring power over utilization and pre-launch invariant verification.
full fine-tuningnvidia b300fsdp/zero-3strong-scalingnfs contention
MirrorNet: Can Medical Image Anonymization Really Protect Patient Identity?
The study demonstrates that de-identified medical images retain patient-identifying information through pixel content, challenging the assumption of anonymity. Using coupled cycle-consistent variational autoencoders (MirrorNet), the authors establish a bidirectional mapping between medical scans and non-medical patient images. Evaluations show the model reconstructs recognizable patient likenesses from scans (identity-region MAE = 0.163) and synthesizes scans from external images, suggesting medical imaging data should be treated as biometric identifiers. Code and models are publicly released.
medical image anonymizationvariational autoencoderscycle-consistent learningbiometric identificationpatient privacy
Deep Generalised Mixed Models: a Novel Neural Network Structure for Analysing Hierarchical Data
The authors propose Deep Generalised Mixed Models (DGMM), a novel neural architecture combining mixed effects models with deep learning to analyze hierarchical Experience Sampling Method (ESM) data. DGMM uses variational auto-encoders and Bayesian data augmentation to model mean and correlation structures via fixed/random effects, handling missing-at-random data and high-dimensional settings. Applied to the GrowIt! adolescent emotion dataset and simulations, DGMM shows promise but exhibits suboptimal performance due to instability.
deep generalised mixed modelsexperience sampling methodmissing-at-randomvariational auto-encodersbayesian data augmentation
BioM-JEPA: joint-embedding prediction of graph-connected gene blocks in single cells
BioM-JEPA introduces a joint-embedding predictive architecture for single-cell transcriptomics that predicts aggregate representations of graph-connected gene blocks, defined by protein-association and coexpression evidence. The method employs a student-teacher framework where the student infers target-block representations from partial gene sets, while the teacher provides targets from full observations. Evaluations on CellBench tasks showed higher embedding effective rank, reduced gene-depth bias, and superior perturbation-response error (5.75× fine-tuning throughput, 3.76× embedding throughput vs scFoundation), while preserving pathway and neighborhood information.
joint-embedding predictionsingle-cell transcriptomicsgraph-connected gene blocksstudent-teacher frameworklinear attention
CohortHijack: Robustness of Single Cell Annotation to Companion Cell Removal
The paper introduces CohortHijack, a robustness audit for single-cell annotation tools that manipulate label refinement by removing non-target companion cells while preserving target expression profiles. The method evaluates random and structured removal strategies (greedy, multi-start, beam search) on PBMC3K and Paul15 datasets using logistic regression and linear SVM classifiers. Results show structured removal altered 19.67-24.33% of target labels with minimal collateral changes (<0.4%), and CellTypist's majority voting exhibited vulnerability to refined-label shifts. Ablations confirmed neighborhood refinement as the attack surface.
single-cell annotationrobustness auditlabel refinementcompanion-cell removalneighborhood voting
Alternating Levenberg-Marquardt Training of Physics-Informed Neural Networks with Fourier-Enhanced Features
The paper introduces Fourier-enhanced alternating Levenberg-Marquardt PINN (FALM-PINN), a framework addressing spectral bias and representation-coefficient coupling in physics-informed neural networks (PINNs) for high-frequency/multi-scale PDEs. The method decouples representation learning (via Fourier-enhanced basis construction) from coefficient fitting (via Levenberg-Marquardt optimization), with global convergence guarantees for both nonlinear and linear PDEs. Experiments demonstrate up to 100× lower relative L² errors compared to state-of-the-art baselines on challenging PDE benchmarks.
physics-informed neural networksspectral biaslevenberg-marquardtfourier featuresnonlinear pdes
A neural operator view on U-Nets for inverse imaging problems
The paper analyzes neural operator variants of U-Nets for solving ill-posed inverse imaging problems, focusing on their behavior across increasing discretization resolutions. It compares operator learning approaches in U-Net architectures, using a 1D toy model for interpretability and extensive numerical experiments on limited-angle CT reconstruction. Results show that while U-shaped neural operators are resolution-invariant by design, classical U-Nets exhibit unexpected robustness to resolution changes.
neural operatoru-netinverse problemsresolution-invariantct reconstruction
Learning to Rank Tensor Network Contraction Plans for GPU-Accelerated Quantum Circuit Simulation
The authors propose a learning-to-rank framework to select efficient tensor-network contraction plans for GPU-accelerated quantum circuit simulation, addressing variability in execution performance due to parallelism, memory traffic, and contraction geometry. Plans are represented by structural features from pairwise contraction sequences, with gradient-boosted rankers trained using listwise and pairwise objectives on GPU measurements. Evaluated on diverse circuit families, the listwise model outperformed MinFill and random baselines, maintaining decision quality under circuit-family shift and partial backend shift across GPU architectures, demonstrating practical utility in reducing search costs.
tensor-network contractionlearning-to-rankquantum circuit simulationgradient-boosted rankersgpu acceleration
On-Policy Delta Distillation for Multilingual Math Reasoning
The paper introduces On-Policy Delta Distillation (OPD$^2$), an enhanced variant of On-Policy Distillation (OPD), for multilingual math reasoning in English, Korean, and Japanese. OPD$^2$ leverages the probability gap between a post-trained teacher and its base model as the learning signal, outperforming standard OPD across languages, with notable gains in Korean and Japanese. Experiments with Qwen3 demonstrate that OPD$^2$ reduces the English-Korean performance gap, while English-only OPD improves non-English performance but biases responses toward English, underscoring the need for multilingual data.
on-policy distillationmultilingual reasoningprobability gapqwen3post-training
KVAE: Family of Tokenizers for Multimodal Generative Models
The paper introduces KVAE, a family of multimodal tokenizers for latent diffusion models, including KVAE-Audio (48 kHz audio), KVAE-3D (video), and KVAE-2D (image), designed for text-conditioned generation. These tokenizers compress inputs into latent representations with specific channel counts and compression ratios (e.g., 8x for images, 4x16x16 for video). Evaluated on reconstruction (PSNR, LPIPS, PESQ) and generation metrics (Frechet Distance, CLIP score, CLAP score), KVAE matches or outperforms state-of-the-art open-source tokenizers like Wan-2.2 and StableAudio. Training details, model selection, and ablation studies are shared, with code publicly available.
latent diffusion modelingtokenizersmultimodal generationcompression ratiostext-conditioned generation
Predicting Task Difficulty Without Rollouts
The paper introduces a method for predicting task difficulty in agent-based environments without requiring costly rollouts, focusing on long-horizon domains where empirical trial-and-error is computationally prohibitive. The authors analyze 17 benchmarks across coding, mathematics, web navigation, and other domains, demonstrating that token-level entropy serves as a reliable predictive signal for difficulty. They critique AUC as a potentially misleading metric and propose using residuals between expected and observed difficulty to detect environment flaws like contamination. The approach aims to aid in benchmark calibration and curriculum design.
task difficulty predictiontoken-level entropylong-horizon domainsbenchmark calibrationcurriculum design
VSMP-IMU: Video-Grounded Semantic Motion Programs for Sensor-Aware Synthetic IMU Generation
VSMP-IMU introduces a video-grounded framework for generating synthetic IMU data via Semantic Motion Programs (SMPs), decoupling activity semantics from label-preserving variations. The method extracts SMPs from input videos, augments them, synthesizes motion, and converts it to virtual IMU signals while maintaining wearable-domain relevance. Evaluated on five IMU-HAR datasets under leave-one-person-out, low-resource, and long-tail settings, VSMP-IMU achieves a 78.33% average Macro-F1, outperforming real-only training by 9.77% and prior synthetic baselines by 4.04%. In low-resource scenarios, it improves by 18.54% over real-only and 6% over baselines, with a 19.86% tail-class Macro-F1 gain in imbalanced settings.
semantic motion programssynthetic imu generationwearable harlow-resource learninglong-tail evaluation
SR-JEPA: Learning Predictive Latent State in 3D Scenes
SR-JEPA introduces a joint-embedding predictive architecture for 3D scene-scale point clouds, focusing on inferring latent representations when entire objects are absent. The method employs a frozen predictive pathway queried at object centroids using shape-free 32-point queries, trained solely with self-contained 3D EMA targets without reconstruction or semantic labels. Evaluation on 5,953 ARKitScenes objects shows 43.13% semantic-identity macro accuracy, 22.18 points above baseline, with randomization and context substitution degrading performance by 9.78 and 21.98 points respectively. On 8,570 Sr3D support pairs, the model achieves 41.15 AP, with identity decoding from missing-object latent reaching 39.37 AP, leaving a 1.78-point residual.
joint-embeddinglatent representationspoint cloudssemantic-identitypredictive pathway
Neuro-Symbolic Closed-Loop Control of Laser Powder Bed Fusion with an In-Loop Ontology
A neuro-symbolic closed-loop control architecture is proposed for laser powder bed fusion, integrating an in-loop ontology to couple symbolic reasoning with statistical learning for constraint-aware predictive control. The architecture employs a geometry-conditioned ontology to map process objectives and constraints onto observable signals, using a description-logic reasoner to generate scan-specific references and bounds. A Gaussian process provides calibrated uncertainty for depth-to-width ratios, enabling feature classification and constraint selection. Evaluated on an Eagar-Tsai surrogate calibrated to the NIST AM-Bench benchmark for IN625, the system eliminates overhang dross, maintains zero dross with minimal lack-of-fusion, and gracefully handles plant mismatch. The architecture supports retargeting to new alloys and constraints via ontology edits, demonstrating architectural feasibility.
neuro-symbolicclosed-loop controlontologygaussian processlaser powder bed fusion
Accelerating nanodrug development in continuous flow systems using informed prediction models based on low-cost surrogate nanoparticles
A shape-constrained predictive modeling framework is introduced to accelerate nanodrug development by estimating nanoparticle properties under varying process conditions. The method employs controlled microfluidic preparation of liposomes and lipid nanoparticles, systematically varying lipid concentrations, flow rates, and aqueous-to-organic mixing ratios. The model integrates experimental data with expert knowledge to predict nanoparticle size and polydispersity index (PDI) accurately. Validation on pharmaceutical applications demonstrates that the approach reduces the need for extensive empirical screening, enabling more efficient nanomedicine process development.
nanoparticlesmicrofluidicpolydispersityliposomesshape-constrained
Equipment-centric workpiece localization in near real-time using deep learning-based vision and event-driven finite state machines
The study introduces an equipment-centric framework for near real-time workpiece localization in hot forging, addressing challenges of extreme temperatures and irregular routing. The method leverages multiple static 2D cameras to infer 3D equipment coordinates and recognize grasp/release activities, validated by event-driven finite state machines. A keypoint-guided attention mechanism within a 3D convolutional neural network enhances activity recognition by focusing on functionally relevant regions. Evaluated in an operational factory, the framework achieved 100% event detection accuracy within a 33-second tolerance, a mean localization error of 317.8 mm, and a mean latency of 21 seconds, enabling interpretable reasoning and visualization of workpiece transfers.
workpiece localizationevent-driven finite state machineskeypoint-guided attention3d convolutional neural networkhot forging
CircuitSteer: Geometrically Aligned Multi-Layer Steering via Sparse Autoencoder Circuits
CircuitSteer introduces a novel framework for multi-layer steering in large language models (LLMs) using Sparse Autoencoders (SAEs) to identify and manipulate coherent semantic circuits. The method constructs feature flow circuits based on feature co-activation and geometric alignment of decoder directions, isolating multi-layer subcircuits responsible for target behaviors. Dense steering vectors are synthesized from sparse features and applied via multi-point interventions. Evaluated on tasks including toxicity, emotion-intensity, sycophancy, and refusal across two model families, CircuitSteer consistently produces fluency-preserving interventions, outperforming static single-point methods in robustness and effectiveness.
sparse autoencodersmulti-layer steeringfeature co-activationgeometric alignmentfluency-preserving interventions
Engram-E2VID: Reference-Based Event-to-Video Reconstruction via Generative Activation of Appearance Engrams
Engram-E2VID introduces a structure-guided framework for reference-based event-to-video reconstruction, leveraging generative activation of appearance engrams. The method encodes reference frames into token-space appearance engrams and transforms event streams into target-time motion-structure scaffolds, enabling cross-layer interaction between structural tokens and appearance engrams via a diffusion backbone. This approach achieves up to 3.29 dB PSNR improvement and 0.08 LPIPS reduction over baselines, with robust performance over long reconstruction intervals.
event-to-video reconstructionappearance engramsdiffusion backbonetoken-space associationmotion-structure scaffold
LILAC: An Idempotent Neural Speech Codec
LILAC introduces an idempotent neural speech codec addressing a key limitation in existing systems: non-idempotency, where baseline codecs rewrite ≥15% of tokens during decode-re-encode passes. The proposed fully convolutional 24 kHz codec operates at 9.375 Hz and 0.75 kbit/s, guaranteeing token-stream identity when re-encoding decoded outputs. LILAC maintains competitive quality, achieving UTMOS scores of 4.14 (LibriSpeech) and 4.24 (LibriTTS-R), comparable to state-of-the-art sub-1 kbit/s neural audio codecs.
neural audio codecidempotencyconvolutionaltoken streamutmos
Sparse Mutual Information Graph Averaging for Improving Random Indexing Embeddings
The paper proposes sparse mutual information graph averaging to refine Random Indexing (RI) word embeddings without dense operations. The method applies weighted averaging on a sparse Positive Pointwise Mutual Information (PPMI) graph with top-K pruning, improving RI's semantic analogy accuracy from 19.4% to 30.7% on a fairytales corpus (272 Google family-category questions). While ineffective for PPMI+SVD, Binary+SVD, CBOW, and Skip-gram baselines, the technique provides a non-gradient repair for weak RI embeddings, with optimal performance at top-K=50 neighborhood pruning.
random indexingpointwise mutual informationsparse embeddingsgraph averagingsemantic analogy
SEAM: Global consistency beyond local accuracy in scientific machine learning
The paper introduces Scientific Explanation-Admissibility Machines (SEAM), a framework ensuring global consistency of local explanations in scientific machine learning. SEAM-$Ω$ represents regions via structured explanations with state, closure, and observation channels, detects inconsistencies through obstruction analysis, and attributes failures to specific channels. Theoretical results include minimum-cost intervention and detectability guarantees. In 19 experiments involving synthetic PDE systems and out-of-distribution Fourier neural operator monitoring, SEAM identified incompatible explanations despite accurate local predictions. The framework extends existing solvers with global consistency auditing.
scientific machine learningexplanation admissibilityglobal consistencyobstruction analysisfourier neural operator
A Low-Power Wearable Respiratory Sensor for Non-Invasive Stress Monitoring
The authors present a low-power wearable respiratory sensor using a force-sensitive resistor (FSR) integrated into an abdominal belt with custom Bluetooth Low Energy acquisition. The system captures abdominal expansion via piezoresistive readout without analog amplification, maintaining robustness across postures and light motion. Evaluation on 12 participants under a stress-induction protocol shows 88.0% test accuracy in distinguishing stress from relaxation phases using time-domain respiratory features, demonstrating feasibility for affective computing.
wearable sensorforce-sensitive resistorpiezoresistive readoutbluetooth low energyaffective computing
Provably Efficient Self-Calibrating Quantum Fault Tolerance
We propose a theoretical framework for self-calibrating quantum fault tolerance that enables continuous hardware stabilization without interrupting computation. The method repurposes syndrome measurements as calibration signals, leveraging the detection rate as a locally strongly convex surrogate objective for analog calibration. We prove convergence to an ε detection rate within O(1/ε²) epochs for time-independent drifts and establish guarantees for time-dependent drifts, with convergence rate independent of code distance for quantum LDPC codes. Pulse-level simulations of neutral-atom arrays and large-scale Clifford simulations validate the theoretical predictions, demonstrating efficient simultaneous logical information protection and hardware stabilization.
quantum fault tolerancesyndrome measurementsanalog calibrationldpc codesconvergence rate
Consistency Has a Computable Blind Spot: A Commutation Theory of Label-Free Reliability for Vision-Language Figure Reading
The paper introduces a computable theory of label-free reliability for vision-language models, demonstrating that systematic misreadings form a joint centralizer set that shrinks with added perturbations. By analyzing equivariance relations where correct answers must change predictably under data edits, the authors propose the Equivariance-Consistency Score (ECS) and release REND-EQUIV, a dataset pairing matched invariance and equivariance sets. Experiments across three models and human-labeled data confirm the predicted error detectability ordering, with cyclic relabeling closing gaps in permutation errors. The work shows detectability depends jointly on the relation and fault class, resolving prior contradictions in metamorphic testing literature.
equivariancecentralizervision-language modelsmetamorphic testinglabel-free reliability
When Does Consensus Mean Correctness? Measuring the Agreement-Accuracy Coupling with Semantics-Preserving Re-Rendering
The study introduces RENDEQ, a generator of semantically equivalent figure re-renderings, to quantify the coupling between model consensus and correctness in vision-language models (VLMs). Using programmatically redrawn scientific figures with exact answers, it evaluates three open-weight VLMs, finding that re-rendering outperforms resampling in accuracy and reliability. Agreement across renders surpasses mean token log-probability as a correctness proxy in two models, with dispersion linked to plotting-library choices. Fine-tuning on cross-render consensus reduces accuracy, contrary to natural-image results, revealing that agreement certifies correctness only above an error-diffuseness threshold.
vision-language modelssemantic equivalencere-renderingagreement-accuracy couplingerror diffuseness
Potential Matching Optimal Transport: Continuous Normalizing Flows for Exact $p$-Wasserstein Dynamics
The paper introduces Potential Matching Optimal Transport (PMOT), a continuous normalizing flow framework for exact $p$-Wasserstein dynamics. PMOT parameterizes velocity fields via scalar potentials in the generalized Benamou-Brenier form, training potential gradients with self-induced matching loss along straight bridges while enabling flexible terminal distribution matching. Theoretical analysis shows zero-loss solutions recover $p$-optimal transport maps under regularity conditions. Experiments demonstrate PMOT's agreement with $p$-OT references on synthetic benchmarks, competitive density modeling on tabular data, and effective color transformation via MMD-based terminal matching.
optimal transportcontinuous normalizing flowswasserstein distancebenamou-brenier formpotential matching
Reasoning Errors Have a Region and a Direction in the Residual-Stream Trajectory of LLMs
The paper introduces a three-stream detector for identifying reasoning errors in LLMs by combining motion with restricted views of location, addressing the trade-off between displacement-based methods and full-state probing. The method employs a coarse region reader using vector quantization and a fine direction reader over normalized multi-layer states, restoring state context without reintroducing shortcut-prone information. Evaluated on reasoning benchmarks unseen during training, it improves selection accuracy by up to 12% over displacement-only baselines and 21% over single-layer probing. The detector also generalizes to factual completion and verification tasks, outperforming all compared methods. Ablations confirm that motion, region, and direction provide complementary signals.
residual-streamvector quantizationmulti-layer statesreasoning benchmarksstate-conditioned motion
RASP-QAOA: Resource-Aware Per-Instance Selection for Exact QAOA Simulation
RASP-QAOA introduces a per-instance selector for exact QAOA simulation, addressing variability in computational representations across graph structure, circuit depth, precision, and memory. The method first filters inadmissible actions based on QAOA semantics and execution requirements, then ranks remaining actions using instance features and analytical work estimates. Evaluated on 60 requests, RASP-QAOA achieves 27/31 top-1 and 31/31 top-2 selections with 1.051 geometric-mean regret, outperforming CUAOA with a PAR10 score of 0.0396 (95% CI: 0.0085-0.1644). Results demonstrate efficacy for n ≤ 35, p ≤ 5, driven by representation features rather than classifier complexity.
qaoaresource-awareper-instance selectionexact simulationrepresentation features
FOCUS: Decoupling Expert Personas in LLMs to Enhance Domain Expert Capabilities
FOCUS proposes a method to decouple expert personas in LLMs via orthogonal decomposition and gated activation, addressing cross-domain coupling issues in persona control. The approach extracts persona vectors, applies orthogonal decomposition to isolate domain-specific expertise, and employs an expert gating module for context-aware activation. A two-stage training strategy and gated selection regularizer optimize persona selection for single- and cross-domain tasks. Evaluations on financial, legal, medical, and cross-domain benchmarks demonstrate improved accuracy over existing persona control methods.
large language modelspersona controlorthogonal decompositionexpert gatingcross-domain tasks
LC-Implicit-QAOA: Active-Workspace-Capped Exact Objective-and-Gradient Evaluation for Training over Bounded QUBO Light Cones
LC-Implicit-QAOA introduces a memory-efficient method for exact objective-and-gradient evaluation in QAOA training, specifically targeting bounded QUBO light cones. The approach profiles causal-cone structures, allocates local-amplitude workspaces, and selects microbatches under a strict active-evaluator memory budget, omitting global state and cost tables. Validation on complex128/float64 dense adjoint implementations shows worst-case relative gradient error of 1.56e-13, with memory usage staying below 80% of budget. On a 3-regular graph (n=512, p=2), LC-Implicit-QAOA achieves equivalent optimization in 101 calls and 189s, compared to 909 calls and 1,565s for central differences.
quboqaoaadjoint differentiationcausal conemicrobatches
Enhancing Anomaly Resilience in Research Networks: A Large-Scale Forecasting Benchmark for Dynamic Security Baselining
The paper introduces a traffic forecasting framework for Research and Education Networks (RENs) to distinguish legitimate bursty traffic from volumetric attacks. Using a 57-day Internet2 dataset (13.7 billion packets from 10 backbone routers), it benchmarks six model families (including SARIMA, TiDE, PatchTST) across 960 configurations. Results show TiDE reduces baseline prediction error by 30-42% (p < 0.001) versus traditional methods, with a novel anomaly-integration strategy improving robustness by 3.3%. This provides the first statistically validated framework for anomaly detection in RENs.
research and education networkstraffic forecastinganomaly detectiontime-series modelinginternet2 dataset
How Much Reconstruction Does Quantum Machine Learning Need? Late Fusion of Independently Trained Quantum Subcircuits
The paper proposes late fusion as an efficient alternative to full reconstruction in circuit-cutting quantum machine learning (QML), where independently trained quantum subcircuits are combined via a classical head instead of costly quantum state reconstruction. Introducing a quantumness dial $Q$ and cut-entanglement diagnostic (Spearman $ρ=0.59$), the method interpolates between pure fusion and full reconstruction. Experiments on synthetic and standard datasets show late fusion matches full reconstruction accuracy within $0.04$ at linear cost, with superior noise robustness. The work identifies conditions where fusion fails, though no quantum advantage over classical ML is claimed.
quantum machine learningcircuit cuttinglate fusionquantum neural networkreconstruction overhead
Align-RAG: Alignment Is All You Need for TSFM In-Context Learning
Align-RAG introduces a training-free method for retrieval-augmented forecasting with frozen Time Series Foundation Models (TSFMs), eliminating the need for learned fusion modules. The approach applies closed-form per-pair amplitude rescaling and integer-lag phase shifts to retrieved past-future windows before input to the frozen backbone. On the Chronos-Bolt benchmark, it reduces MSE by 3.75% on average across seven datasets and improves zero-shot performance on four additional TSFMs by 2.5-13.7%. Analysis shows frozen TSFMs inherently support dynamic in-context retrieval use, with alignment inducing prediction shifts akin to a closed-form ridge predictor.
retrieval-augmented forecastingtime series foundation modelsin-context learningamplitude rescalingphase shift
Behavioral Residualization for Unsupervised Intrusion Detection in Automotive CAN Networks
The paper introduces per-ID behavioral residualization, a CAN-specific representation for unsupervised intrusion detection in automotive Controller Area Network (CAN) systems. The method extracts fourteen temporal, protocol, and payload features from sliding windows and residualizes them against each arbitration ID's normal baseline. Evaluated across six unsupervised detectors and two datasets (HCRL and ROAD), residualization improves mean F1 in 21/24 and 30/36 evaluations, respectively. On the ROAD dataset, it achieves recall ≥ 0.99 with high ROC-AUC for targeted signal-manipulation attacks. Limitations include novel-ID flooding (HCRL DoS, F1 = 0.02) and cross-ID fuzzing (ROAD, F1 = 0.27), defining the coverage boundary of the representation.
behavioral residualizationcontroller area networkunsupervised intrusion detectionarbitration idroc-auc
Equation-Free Period-Aware Forecast-Error Contraction for Estimating Negative Largest Lyapunov Exponents from Short Trajectory Ensembles
The authors propose a period-aware forecast-error contraction method for estimating dominant negative Lyapunov exponents from short scalar trajectory ensembles without requiring governing equations or Jacobians. The approach trains a k-nearest-neighbor predictor on trajectory histories, evaluates geometric-mean absolute forecast errors at phase-consistent horizons, and derives the exponent from logarithmic error profiles. Key innovations include period-synchronized forecasting steps and consensus-based slope validation. On the logistic map, the method achieves MAE=0.0253 and R²=0.886; on a 2D map, three scalar pipelines yield MAE=0.00879–0.01145 and R²=0.983–0.986.
lyapunov exponentsforecast-error contractionk-nearest-neighbortrajectory ensemblesphase-consistent horizons
An Inertial Block Proximal Linearized Method with Adaptive Momentum for Nonconvex and Nonsmooth Optimization
(No summary returned.)
GenGA: Editable and Data-Grounded Graphical Abstract Generation for Academic Papers
GenGA introduces a novel framework for generating editable graphical abstracts (GAs) from academic paper content, addressing the limitations of raster-based outputs in conventional methods. By producing vector-format figures with hierarchical structures, GenGA enables seamless element-level editing in existing drawing tools. The framework incorporates the Structural Independence Coefficient (SIC), a metric quantifying editing simplicity based on local modification propagation. Experiments demonstrate GenGA's superior editing simplicity, conciseness, and semantic alignment compared to conventional methods and human-authored GAs, validating SIC's correlation with manual editing costs. This work redefines GA generation as an editable vector graphic problem, enhancing scientific communication workflows.
graphical abstractsvector graphicsstructural independence coefficienteditable generationscientific communication
KV-Skill: Forging Expertise in the Model's Native Language
KV-Skill introduces external factorized operators for frozen language models, enabling modular task knowledge storage via text-derived or reward-learned operators without prompt expansion. The method supports two pathways: registration (converting text skills to operators) and reward learning (developing latent operators from outcomes), both interfacing through a lightweight layer. Evaluated across ten benchmarks and four backbones, KV-Skill outperforms baselines (e.g., 77.2 vs. 23.4 accuracy on Qwen3.5-4B LiveMath) and retains benefits with minimal task-aligned directions (one per layer). A shared interface allows independent loading of multiple skills without forgetting.
factorized operatorsfrozen language modelstask knowledgelightweight interfacereward learning
A Foundational EDM2-Based Generative Model for High-Resolution Synthetic Fetal Ultrasound Imaging from Open Datasets
The authors propose a high-resolution fetal ultrasound synthesis framework using the EDM2 diffusion model, trained on multiple public datasets to generate 512x512 images across six anatomical classes. The method improves image quality (lower FID scores) and enhances downstream fetal plane classification, achieving 93.36% ensemble accuracy after fine-tuning, outperforming real-data-only training. Clinical evaluation by an expert (10+ years experience) on 100 images yielded a mean realism score of 2.67/5 for synthetic images, with noted artefacts including smoothing, speckle irregularities, and anatomical inconsistencies.
edm2diffusion modelfetal ultrasoundfid scoresensemble accuracy
Effective pruning of task-trained recurrent neural networks using noisy fluctuations and connection rescaling
The study validates noise-prune, a biologically-plausible unsupervised pruning rule for task-trained recurrent neural networks (RNNs), demonstrating its effectiveness in preserving task performance. Noise-prune employs noisy fluctuations to assess connection importance, samples connections probabilistically based on their importance, and rescales retained connections to maintain average synaptic strength. This approach outperforms magnitude-based pruning and matches or exceeds non-local second-order strategies. Empirical results indicate that optimal rescaling parameters differ from theoretical predictions. The findings confirm noise-prune as a viable pruning rule for functional RNN architectures and characterize its optimal settings.
noise-prunerecurrent neural networkssynaptic strengthsecond-order informationtask performance
Hybrid Probabilistic Zonotopes for Identifiable and Refinable Predictive Uncertainty
The paper introduces Hybrid Probabilistic Zonotopes (HProbZ), a neural network output head that disentangles discrete mode selection, bounded systematic drift, and stochastic noise via binary, bounded, and stochastic zonotope generators. HProbZ enables closed-form likelihood computation through convolution, algebraic coupling of future predictions for refinement, and identifiability of generators up to permutation. It outperforms Gaussian mixture baselines on prediction benchmarks while offering unique structural benefits like per-mode risk analysis and multi-modal conformal sets, unattainable by mixture or conformal predictors alone.
hybrid probabilistic zonotopesidentifiable uncertaintyclosed-form likelihoodmulti-modal conformal setszonotope generators
EvoHarness-RL: Learning Self-Evolving Runtime Harness for Long-Horizon LLM Agents
The paper introduces EvoHarness-RL, a framework for learning runtime harness policies that enable long-horizon LLM agents to manage external execution support autonomously. The method exposes Belief, Progress, and Experience (BPE) as harness state, combining supervised fine-tuning for harness construction with cost-aware GRPO for runtime coordination. Evaluated on ALFWorld with Qwen3-8B, the system achieves 96.9% success, demonstrating harness annealing (internalizing harness patterns) and harness evolution (refining state substrates), showing benefits beyond tool augmentation or memory scaling.
long-horizon agentsharness policy learningexternal-state accesscost-aware grporuntime coordination
Discrete energy as an exact label-free training objective for finite-element surrogates
The paper introduces a label-free training objective for finite-element (FE) surrogate models using discrete potential energy, eliminating the need for reference solutions. It proves that minimizing discrete energy is equivalent to supervised regression in the stiffness norm for linear elastostatics, with identical gradients and a unique minimiser. The work includes a conditioning lemma bounding displacement error, a modewise contraction identity, and a Chebyshev bound for conjugate-gradient post-processing. Numerical validation on synthetic and pre-registered datasets confirms all inequalities hold. The method does not extend to elastodynamics via direct action functional minimization but suggests a time-discrete formulation for exactness.
finite-elementdiscrete energystiffness normlinear elastostaticsconjugate-gradient
Robust Context-Aware Detection of Malicious Instructions in Text
We propose a robust context-aware detector for malicious instructions in text, addressing indirect prompt injection (IPI) vulnerabilities in LLM-based agents. Our method combines query-relative segment-level detection with adversarial training (AT) techniques: feature-space AT using projected-gradient optimization and LLM-based paraphrasing for realizable evasion attacks. Experiments on IPI benchmarks demonstrate superior performance over state-of-the-art baselines under static attacks, with AT variants offering higher utility and lower attack success rates under adaptive attacks. Domain-dependent tuning of detectors is shown to be essential for optimal performance. Code is publicly available.
indirect prompt injectionadversarial trainingcontext-aware detectionllm-based paraphrasingevasion attacks
Invisible Shortcuts: Why Vision Encoders Know Your Camera
The paper identifies invisible metadata traces (e.g., image processing artifacts) as a previously overlooked source of shortcut learning in vision encoders, showing that large-scale pretraining (ImageNet, LAION) induces metadata-semantics correlations. Through controlled experiments, the authors demonstrate that stronger correlations increase metadata sensitivity and degrade performance under distribution shifts. They propose mitigation strategies during and after pretraining that reduce sensitivity to both targeted and unseen metadata without compromising downstream task performance, while also improving out-of-distribution generalization and explaining generated-image detection capabilities.
shortcut learningmetadata tracesvision encodersdistribution shiftpretraining
IFlowNets: Extending Generative Samplers to Learn Strategies in Incomplete Information Games
The paper introduces Information Flow Networks (IFlowNets), a generative flow network framework extending Adversarial Flow Networks (AFlowNets) to incomplete information games. The authors prove that prior constraints for generative flow networks in complete information games are inadmissible for valid strategy densities and training objectives, and demonstrate that IFlowNets strictly generalize AFlowNets. Preliminary experiments in three standard game environments show IFlowNets match or outperform Outcome Sampling Monte Carlo Counterfactual Regret (OSMCCFR) and RL-based methods in both performance and speed.
generative flow networksincomplete information gamescounterfactual regretadversarial flow networksreinforcement learning
Can Open-Weight LLMs Produce Kernel-Verified Coq Proofs? A Pilot Study
This pilot study evaluates six open-weight LLMs on generating kernel-verified Coq proofs using 100 theorems from CoqStoq, with correctness verified by the Coq kernel. Models were tested with temperature 0, and proofs were checked in original project environments. Gemma 4 achieved 12/100 verified proofs, Llama 3.3 8/100, and DeepSeek Coder V2 Lite 1/100, while Qwen 3.5, Mistral Small 3.1, and GPT-OSS produced none. Successful proofs covered 15 distinct theorems, 11 unsolved by standard Coq tactics, all with short/medium reference proofs. The overall success rate was 3.5% (21/600 attempts).
coq kernelopen-weight llmsformal verificationcalculus of inductive constructionscoqstoq benchmark
Spectral Distillation: From Nonlinear Dynamics to Linear State-Space Models
The paper presents Spectral Distillation, a provable pipeline for learning compact linear state-space (LDS) representations of nonlinear dynamical systems without solving non-convex system identification. The method first learns an implicit spectral predictor via convex Observation Spectral Filtering (OSF), then distills it into an explicit recurrent LDS. Theoretical guarantees show the distilled LDS's error decomposes into exponentially small distillation loss and OSF learning error bounded by Luenberger observer complexity. Experiments on LDS benchmarks and MuJoCo behavior cloning demonstrate the pipeline matches or outperforms directly trained baselines.
spectral distillationlinear state-space modelsobservation spectral filteringluenberger complexitynonlinear dynamics
Velocity- and Regime-Aware Detection of Intraday Options Market Manipulation, with Explainable Attribution
The paper introduces a velocity- and regime-aware method for detecting intraday options market manipulation, leveraging pump-and-crash patterns in market state velocity rather than price levels. The pipeline combines smoothed state velocity (option-Delta velocity for index options, price velocity for equities) with SHAP attribution for explainability, evaluated under strict out-of-sample conditions. On the BANKNIFTY index-options test, the autoencoder achieved 100% recall (10/10 regulator-identified days), with precision near 25% under a closed-world assumption. The method transferred to U.S. equities (SEC v. Patel), showing AUCs of 0.91 (ARQQ) and 0.81 (ACY) for shape scores, with SHAP attribution confirming unlabeled alerts' similarity to known cases.
market manipulationstate velocityshap attributionhidden markov modelautoencoder
Quantum-Structured World Models (QSWMs) for Predictive Latent Dynamics
We propose Quantum-Structured World Models (QSWMs), a quantum-inspired framework for predictive world modeling using structured latent states, transition operators, and measurement-inspired decoding maps. QSWMs leverage complex-valued representations and density-matrix-like latents to explore inductive biases for world modeling, establishing classical inclusion, predictive sufficiency, and structured compactness as foundational properties. We instantiate complex-valued and density-matrix-like QSWM variants and evaluate them on elementary cellular automata, demonstrating promising local predictive potential for complex-valued QSWMs but revealing limitations in long-horizon rollout for density-matrix variants.
quantum-structured world modelslatent dynamicscomplex-valued representationsdensity-matrixcellular automata
DG-FedReuse: Proxy-Gradient-Gated Cached-Update Reuse with Matched Sparse Uplink Accounting
DG-FedReuse introduces a federated learning mechanism enabling selective clients to reuse age-decayed cached updates based on a stochastic head-gradient discrepancy proxy, constrained by a hard cache-age limit and fresh-client quota. Fresh updates employ adaptive per-tensor Top-K numerical-field representation. Evaluated on six image-classification datasets with 50 virtual clients and Dirichlet label heterogeneity (α=0.5), DG-FedReuse achieves 83.36-85.42% uplink savings versus 76.88% for Top-K FedAvg, with accuracy differences ranging from -5.29 to -0.14 percentage points. Sensitivity analysis shows dense-model downlink reduces savings to 41.68-42.71%, highlighting accounting boundary dependencies.
federated learningcached updatestop-khead-gradient discrepancydirichlet heterogeneity
Positive-Unlabeled Preference Optimization For Chest X-ray Report Generation
We propose PU-DPO, a positive-unlabeled preference optimization framework for chest X-ray report generation that addresses omission noise in clinical reports. The method reformulates the objective under PU learning, treating absent mentions as unlabeled rather than negative, and provides preference supervision through constructed contrastive pairs generated via edits to model responses. Evaluated on semi-synthetic experiments and real-world chest radiograph benchmarks with adjudicated labels, PU-DPO demonstrates consistent improvements in detection rates and recovery of hidden positives across multiple pathologies, outperforming prior approaches in robustness to omission noise.
positive-unlabeled learningpreference optimizationomission noisecontrastive pairsreport generation
Physics-Based Molecular Fingerprints from Spectral Graph Theory Provide Efficient Geometry-Aware Measures of Chemical Similarity
The authors introduce spectral fingerprints, a novel 3D molecular representation based on spectral graph theory, addressing limitations of 2D fingerprints and pairwise 3D methods. Molecules are modeled as complete 3D graphs with physics-based edge weights, with eigenvalue decomposition of the graph Laplacian yielding fixed-length, symmetry-invariant fingerprints. These fingerprints distinguish stereoisomers/conformers while remaining computationally efficient for large-scale screening. Evaluation across organic, inorganic, and biological chemistry datasets shows superior performance in community detection, nearest-neighbor property estimation, and applicability domain analysis compared to baselines, demonstrating utility for cheminformatics and ML tasks.
spectral graph theorymolecular fingerprintsgraph laplaciancheminformaticse(3) invariance
Computationally Efficient Collaborative Communication Via Regularity-Based Coarsening
The work presents a computationally efficient protocol design algorithm for collaborative communication games, leveraging a novel strengthening of the Frieze-Kannan weak regularity lemma. For games with n observations and m actions, the algorithm achieves utility α-ε using 2^O(CC_α(G))/ε² bits of communication, where CC_α(G) is the minimum bits required by any protocol to achieve α. The exponential dependence on CC_α(G) is proven tight unless P=NP. The method constructs a coarsened game Ĝ with constant-size partitions, indistinguishable from G under short protocols, relaxing prior structural assumptions like informational substitutes.
communication complexityregularity lemmaprotocol designcomputational efficiencymulti-agent games
QEvict: Recoverable Quantized KV Eviction for Attention-Drift-Robust Long-Context Decoding
QEvict introduces a recoverable quantized KV-cache eviction scheme to address attention drift in long-context autoregressive decoding. The method maintains three tiers: full-precision high-confidence windows, quantized intermediate windows (recoverable via dequantization), and deleted low-confidence windows, dynamically updating importance scores during decoding. Evaluations on long-context understanding, retrieval, and reasoning benchmarks show QEvict reduces missed attention (measured by Future Missed Mass and Global LIR diagnostics) and outperforms baseline eviction and quantization approaches under fixed memory constraints.
kv-cacheattention driftquantized evictionautoregressive decodinglong-context
Rectifying Geometric Misalignment: Online Source-Free Adaptation for Class-Imbalanced EEG
OSPDIM introduces a source-free online unsupervised domain adaptation framework for EEG-based BCIs, addressing geometric misalignment caused by dynamic label shifts on Riemannian manifolds. The method incorporates a manifold-constrained bias parameter into tangent space mapping, optimized via information maximization to correct skew from imbalanced data streams. Unlike offline approaches relying on batch statistics, OSPDIM estimates and rectifies geometric bias in real-time. Experiments on 2D SPD matrices demonstrate successful misalignment correction, and evaluations on motor imagery datasets show significant performance improvements over standard Riemannian baselines, particularly in severe class imbalance scenarios.
riemannian manifoldunsupervised domain adaptationtangent space mappinginformation maximizationelectroencephalography
EdgeXpert: An Edge Device for Memory-Efficient LLM Inference with Mixture-of-Experts and Speculative Decoding
EdgeXpert introduces a software-hardware co-designed accelerator for efficient LLM inference on edge devices, addressing incompatibility between mixture-of-experts (MoE) and speculative decoding. It employs prompt-wise expert reuse during prefill, routing tokens based on importance via a lightweight encoder and shared expert sets, and depth-aware expert coalescing during decoding, leveraging contextual similarity to minimize memory access. Synthesized in Samsung 28nm technology at 800 MHz, EdgeXpert achieves 56.3% latency reduction and 44.1% energy reduction compared to prior works, while maintaining near-baseline accuracy.
mixture-of-expertsspeculative decodingmemory accessedge devicelatency reduction
Evaluating Machine Learning Models for Post-Wildfire Debris-Flow Prediction
This paper presents a systematic evaluation of machine learning models for predicting post-wildfire debris flows, addressing challenges in overlapping event classes, interpretability, and data scarcity. Using basin-scale observations from the western United States, 15 models were compared, including Tabular Prior-Data Fitted Network (TabPFN), with repeated stratified cross-validation. TabPFN achieved the highest unaugmented threat score (0.637), closely followed by tree-based models. SHapley Additive exPlanations (SHAP) identified short-duration rainfall intensity and storm accumulation as top predictive features. Synthetic data augmentation improved performance for most models, with deep learning models showing the largest mean threat score increase (+0.041).
tabular prior-data fitted networkshapley additive explanationssynthetic data augmentationstratified cross-validationthreat score
Multimodal Spatiotemporal Atmospheric Data Assimilation with Latent Video Flow-matching
The study introduces a novel multimodal spatiotemporal atmospheric data assimilation framework using latent video flow-matching, departing from traditional Bayesian inference methods. The approach trains a prior on ERA5 reanalysis data (69 variables over 8 days) and employs posterior sampling to assimilate real-world observations from NOAA Integrated Global Radiosonde Archive and Integrated Surface Database. By generating continuous trajectories, the method propagates information between observed and unobserved frames, enabling filtering and smoothing tasks through frame selection. The framework achieves competitive performance in full-state ensemble forecasting directly from sparse observations, matching state-of-the-art observation-to-forecast models.
data assimilationlatent video flow-matchingera5 reanalysisposterior samplingensemble forecasting
Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes
This study systematically investigates multimodal pretraining, identifying four key mechanisms: (i) asymmetric knowledge flow between language, visual understanding, and generation; (ii) modality synergy driven by data complexity and architectural choices like shared attention; (iii) early unification's superiority over late alignment, revealing vision laziness; and (iv) efficient pretraining recipes achieving strong performance with 5% compute. Controlled experiments on synthetic and real-world datasets, validated by scaling to 13.5B MoE models trained on 2T tokens, provide empirical insights into the physics of multimodal pretraining.
multimodal pretrainingknowledge flowmodality synergyearly unificationvision laziness
Beyond Rotations: AuroOFT for Expressive Quantized Orthogonal Fine-Tuning
AuroOFT extends quantized orthogonal fine-tuning (qoft) by introducing a zero-start gated low-rank nonlinear residual to each adapted linear layer, enabling input-dependent nonlinear corrections while maintaining quantization compatibility. The method maps activations into an RMS-normalized compact latent space and employs adaptive nonlinear bases with bounded or token-dependent gating. Evaluated on 1.5B/3B Qwen2.5 settings, AuroOFT improves Macro-6 accuracy by 1.30-2.70% over qoft and by 6.52-10.62% over QLoRA, while reducing trainable parameters by 32.3-44.7% compared to QLoRA.
quantized orthogonal fine-tuningnonlinear residualrms-normalizedadaptive nonlinear basestoken-dependent gating
Beyond Full-Model Rollback: AuroSFT for Adapter-State Multi-Task Fine-Tuning
AuroSFT introduces a parameter-efficient framework for multi-task supervised fine-tuning by recasting the carried state as a compact, mergeable adapter state, addressing the limitations of full-model rollback in msft. The method freezes the pretrained backbone, trains only injected adapters, and rolls back adapter checkpoints at task-wise peaks, utilizing an AuroRA-inspired adaptive nonlinear layer applied to a low-rank weight factor. This approach ensures linear updates in the input, rank-bounded operations, and exact mergeability into the frozen projection. AuroSFT achieves 61.36% average accuracy across five backbones, outperforming msft's 59.85%.
multi-task fine-tuningadapter statelow-rank weightnonlinear layerparameter-efficient
A Unified Causal Inference Framework for the Desirability of Outcome Ranking Paradigm in Benefit-Risk Evaluation
The authors propose a unified covariate-adjusted causal inference framework for estimating desirability of outcome ranking (DOOR) probabilities in benefit-risk evaluation, applicable to both randomized trials and observational studies. The method formulates DOOR probability as a bilinear functional of marginal ordinal outcome distributions, estimates conditional distributions via sequential risk-set hazards, and derives the efficient influence function. Simulations comparing G-computation, IPW, AIPW, and TMLE with generalized linear models or Super Learner showed TMLE-SL had strongest point-estimation performance. EIF-based inference evaluations favored CVTMLE-SL for bias, distribution recovery, standard-error accuracy, and confidence-interval coverage across varied settings.
causal inferencedesirability of outcome rankingefficient influence functiontargeted maximum likelihood estimationsuper learner
Optimal Training-Time Scaling in Gradual Adaptation
The paper derives optimal per-task training-time scaling for gradual adaptation in overparameterized linear regression tasks that share a zero-loss solution. Analyzing smooth task transitions with $N$ tasks and training time $s_N$, the authors prove that final learning progress converges to a continuum curve when $Ns_N\toτ$, exhibiting $Θ(τ)$ scaling for small $τ$ and $Θ(τ^{-1})$ for large $τ$. This yields optimal scaling $s_N^\star=Θ(N^{-1})$, implying reduced per-task training as the adaptation path is divided more finely. Empirical validation on gradually rotated MNIST and Yearbook time shift datasets confirms the theoretical findings.
gradual adaptationoverparameterized linear regressiontask transitionscontinuum curvezero-loss solution
Disentangling 3D Modeling from Spatial Reasoning
The paper introduces Disentangled Spatial Reasoner (DiSR), a framework that decouples 3D perception from spatial reasoning by leveraging off-the-shelf perception models for 3D reconstruction and fine-tuning an LLM with LoRA for reasoning over explicit geometric evidence. This approach avoids large-scale 3D VQA training and complex tool-use policies while maintaining competitive performance on spatial reasoning benchmarks. DiSR offers improved interpretability, modularity, and computational efficiency compared to end-to-end methods.
spatial reasoning3d reconstructionlarge language modelslora fine-tuningmodular perception
One Qubit Can Beat One Bit: Quantum Advantage for Post-Training Quantization
The paper introduces Quantum Random Access Quantization (QRAQ), a framework that overcomes the shared-sign constraint in one-bit post-training quantization by encoding context-dependent signs in a quantum random-access code and retrieving them via Pauli measurements. Under a fresh-copy logical readout model, QRAQ produces an unbiased, context-specific binary surrogate with a tractable shot-noise penalty. Theoretical analysis shows QRAQ achieves strictly lower ideal reconstruction risk than shared-sign one-bit PTQ when optimal context-wise signs are incompatible, with finite-shot and calibrated-noise conditions preserving this advantage. Experiments validate the framework in ideal, finite-shot, noisy, and multi-context regimes.
quantum quantizationpost-training quantizationpauli measurementsshot-noise penaltyreconstruction risk
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 role-conditioned evidence views and hierarchical coordination. The method employs routing protocols, evaluator-reward policies, and state-routed memory to adapt search processes without updating LLM parameters within runs, while distilling trajectories into role-conditioned priors across runs. Evaluated on LLM-SRBench's LSR-Synth domains, A-SR improves Acc@0.01 from 25.79% to 48.30% with Llama3.1-8B and from 24.58% to 38.29% with Qwen3-4B using A-SR-LoRA. It achieves superior normalized mean squared error on 7 of 8 metrics across real-world scientific discovery tasks.
symbolic regressionrole-conditionedhierarchical coordinationevaluator-rewardstate-routed memory
Decoupling Perception from Description: Computation-Grounded Representation Alignment between Multivariate Time Series and Language
The paper introduces CGTime, a 4B-parameter computation-grounded time-series-language model that decouples perception from description to resolve the self-supervision trap in aligning multivariate time series with language. Deterministic code computes statistical features from open-source multivariate series, while an LLM verbalizes these precomputed facts, separating perceptual tasks (handled computationally) from descriptive tasks (handled by the LLM). CGTime outperforms larger general-purpose models (0.283 vs. 0.173 for GPT-4o-mini) on multivariate understanding tasks, demonstrating superior accuracy in stating verifiable numerical facts and broader statistical coverage.
multivariate time seriesrepresentation alignmentcomputation-groundedself-supervision trapin-context learning
Nonparametric Goodness-of-fit Testing under Covariate Shift
The paper introduces a nonparametric goodness-of-fit testing framework for covariate shift scenarios, where labeled data originates from a source population but evaluation targets a distinct target population. The method employs truncated importance-weighting kernel ridge regression combined with a multiplier bootstrap to construct confidence sets for the regression function, addressing instability from heavy-tailed density ratios. Theoretical analysis demonstrates nonasymptotic validity and sharpness of the confidence sets under operator compatibility conditions, with explicit error rates for coverage probability derived under specific spectral decay and density ratio conditions. Empirical results validate the theoretical claims.
covariate shiftkernel ridge regressionmultiplier bootstrapimportance-weightingnonparametric testing
PPDL: LLM-Based Flows as Probabilistic Programs
The paper introduces PPDL, a probabilistic programming language for managing uncertainty in LLM-based application flows. The method enables developers to quantify and propagate uncertainty across multiple LLM calls and tools without additional coding, while supporting experimentation with inference scaling techniques. Experimental validation includes a case study implementing a theorem-proving agent for the Rocq theorem prover, demonstrating the approach's utility in improving reliability and trust in LLM outputs.
probabilistic programmingllm-based flowsuncertainty quantificationinference scalingtheorem proving
Analysis of Numerical Localisation in LLM Translations
This study extends Tang et al. (2025) by analyzing numerical localization (times, numbers, dates) in five large language models (LLMs) on commodity hardware. A baseline accuracy was established for each model, and three improvement strategies were tested. Contrary to Tang et al., embedding localization principles into the prompt context yielded statistically significant accuracy gains over direct translation or alternative methods. The findings highlight the efficacy of context-based localization in LLMs for numerical data.
localizationllmsprompt contextcommodity hardwarenumerical translation
Quality Diversity for Reliable Data Driven Time-Use Optimization
The paper introduces a Quality Diversity (QD) framework incorporating predictive uncertainty for reliable time-use optimization in behavioral health. Using compositional data analysis on a child cohort dataset (n>1000), the method derives objective functions linking daily activity compositions to health indicators. The uncertainty-aware QD approach generates diverse, high-quality time-use recommendations by balancing expected benefits with model confidence, favoring lower-uncertainty regions while maintaining solution quality.
quality diversityuncertainty quantificationcompositional data analysisbehavioral healthtime-use optimization
📰 Industry Media (11)
Microsoft Open Sources code-testing-generator: a Polyglot Unit-Test Agent That Hits 92.1% Task Completion Versus 78.9% for Stock Copilot
Microsoft's open-source code-testing-generator introduces a polyglot unit-test agent that autonomously generates, validates, and verifies tests through a Research-Plan-Implement (RPI) pipeline. The agent resolves framework, location, and assertion ambiguities by analyzing repository context, employing three strategies (Direct, Single pass, Iterative) while avoiding production code modifications. On a 152-task benchmark, it achieved 92.1% completion (140 tasks) versus 78.9% for stock GitHub Copilot, with 63% fewer failures, particularly excelling on vague prompts (88.8% vs 66.3%) and diff-targeted tasks (100% vs 0%). It maintained comparable coverage (72.4% vs 72.2%) with 2.3% fewer tests and 5.5% faster execution.
polyglot unit-test agentresearch-plan-implement pipelinemutation testingtest framework detectionrepository-aware planning
Liquid AI Releases LFM2.5-2.6B: An On-Device Agentic Model With 128K Context, Tool Calling, And Open Weights
Liquid AI introduced LFM2.5-2.6B, a 2.69B-parameter on-device agentic model with 128K context length and tool-calling capabilities. The architecture combines 22 double-gated short convolution blocks and 8 grouped-query attention blocks, pre-trained on ~34T tokens. It outperforms larger models (Gemma-4-E4B-it, Qwen3.5-9B) on ToolSandbox (77.83 vs 65.00) and instruction-following benchmarks (IFStruct: 85.49 vs 78.50), while achieving 220 tokens/s on an M5 Max with <2.5GB memory. The model is released under the lfm1.0 license with GGUF/MLX/ONNX support.
agentic modelgrouped-query attentionon-device inferencetool-callingcontext window
Stanford Evo 2 AI model generates phages against E. coli
Stanford researchers developed Evo 2, a generative AI model that synthesizes functional bacteriophage genomes targeting E. coli. The model generates complete ΦX174 phage genomes (5,400 bp) end-to-end via left-to-right autoregressive sequencing, followed by computational screening and wet-lab validation. From 300 synthesized candidates, 16 phages demonstrated superior lytic activity versus native ΦX174, with a phage cocktail overcoming resistant E. coli strains. The open-source framework aims to expand to longer genomes (e.g., MRSA-targeting phages) while addressing genetic novelty and controllability challenges.
generative aibacteriophageautoregressive sequencinglytic activitygenome synthesis
How AI Is changing Instagram engagement without replacing the human touch
Instagram employs multiple AI subsystems (Feed, Reels, Stories, Explore) with distinct algorithms that personalize content based on user behavior, achieving a 0.7% engagement rate in 2026. The platform prioritizes human-generated content through three key signals: watch time, likes per reach, and sends per reach, with engagement velocity favoring immediate viral reactions. AI automation handles caption drafting, visual generation, and audience prediction, while human oversight manages strategy and complex DM interactions. The system penalizes fake engagement (bots, coordinated pods) and implements human handoff for sensitive conversations, maintaining a balance between AI scalability and human authenticity.
engagement velocityalgorithmic personalizationhuman handofffake engagement detectionmulti-algorithm system
Alibaba tests new business model for Qwen open-source AI
Alibaba introduces a revenue-sharing model for commercial users of its next Qwen open-weight AI model, diverging from the Apache 2.0 license used for Qwen3. The arrangement targets larger companies deploying the model as a service, requiring commercial agreements with Alibaba, though specific revenue-sharing rates remain undefined. This approach mirrors Moonshot's licensing for Kimi K3, which imposes conditions on Model-as-a-Service businesses exceeding $20 million annually. Both models employ mixture-of-experts architectures, with Qwen3.8-Max activating 95 billion parameters per request. The strategy aims to monetize large-scale deployments while maintaining open-weight accessibility, reflecting broader trends in AI licensing and infrastructure costs.
open-weightmixture-of-expertsrevenue-sharingmodel-as-a-serviceapache 2.0
Why health AI interfaces must adapt to user expertise
A Nature Medicine study by MIT researchers demonstrates that explainable AI (XAI) interfaces for dermatological diagnosis yield divergent outcomes based on user expertise. Non-experts (n=1,200) improved accuracy by 12-15% when deferring to AI predictions, particularly with LLM-generated explanations, but exhibited automation bias—wrong predictions degraded performance more than correct ones improved it. Clinicians (n=150) performed best with raw model outputs (no explanations), maintaining resilience to incorrect AI suggestions. The study tested four interfaces: confidence-only, heatmaps, similar-image retrieval, and LLM rationales, revealing that explanation utility depends on users' baseline medical knowledge and task timing (pre/post initial diagnosis).
explainable aiautomation biasdifferential diagnosisllm-generated explanationsclinical decision support
PRISM2 model uses clinical dialogue to interpret pathology slides
PRISM2 introduces a perceiver-based encoder trained jointly on 2.3M whole-slide pathology images and 685,507 GPT-4o-generated Q&A pairs from clinical reports. The two-stage architecture first learns slide-level embeddings via contrastive (BioGPT) and autoregressive (Phi-3 Mini) objectives, then freezes the encoder to fine-tune a 4B-parameter LM for diagnostic dialogue. PRISM2 achieves 0.967 AUC in pan-cancer detection (0.957 for rare cancers), outperforms clinical-grade products in breast/prostate classification (0.809 vs. 0.773 concordance in recurrence prediction), and shows 3-18% error rates in pathologist-reviewed QA. Limitations include fixed 0.5μm resolution and no spatial encoding across tiles.
perceiver encoderwhole-slide imagescontrastive objectiveautoregressive fine-tuningpan-cancer detection
Alibaba, DeepSeek push China’s AI model race towards lower costs
Alibaba released Qwen3.8-Max, a 2.4 trillion parameter mixture-of-experts model with 95B active parameters per inference, supporting multimodal inputs and 1M token context. DeepSeek's V4-Flash (284B total, 13B active) undercuts competitors with $0.14/M input tokens, while Qwen3.8-Max costs $2/M input tokens. Benchmark tests show V4-Flash averages $0.03 per task versus $10.57 for Moonshot AI's Kimi K3 (2.8T total, 104B active), highlighting cost disparities from architecture and token usage. Open-weight availability (MIT license for V4-Flash) provides deployment flexibility beyond hosted APIs.
mixture-of-expertsinference pricingcontext windowopen-weightbenchmark testing
Red Hat, NVIDIA, IBM back project turning AI policy into code
Red Hat, NVIDIA, and IBM launched asago, an open-source project automating AI governance policy translation into deployable code. The framework maps policy requirements to risk profiles using NIST AI RMF and OWASP LLM Top 10, generates tailored risk assessments, and orchestrates Kubernetes-ready controls. It aims to reduce deployment timelines from months to days while maintaining an auditable trail linking policies to runtime controls. Contributors include Microsoft, MIT Lincoln Laboratory, and The Alan Turing Institute. The project, under Apache License 2.0, is in early formation with no production benchmarks yet.
ai governancerisk mappingkubernetesnist ai rmfaudit trail
EU AI Act Article 50 transparency rules enter force
Article 50 of the EU AI Act introduces transparency obligations for AI providers and deployers, requiring disclosure when users interact with AI systems and marking AI-generated content as synthetic. The regulation mandates machine-readable marks for synthetic audio, image, video, or text outputs, with exceptions for law enforcement and minimal assistive editing. Deployers must inform individuals exposed to emotion recognition or biometric categorization systems and disclose deepfakes or AI-generated public interest text. Enforcement is divided among national market surveillance authorities, the AI Office, and the European Data Protection Supervisor, with compliance demonstrated through adherence to the Code of Practice on Transparency of AI-generated Content or alternative measures.
transparency obligationsmachine-readable marksemotion recognitionbiometric categorizationdeepfake disclosure
Why biological data matters more in AI drug discovery
GSK and Relation Therapeutics expanded their AI drug discovery collaboration with a $110M agreement emphasizing biological data generation alongside AI model development. Relation’s Lab-in-the-Loop platform integrates single-cell multi-omics, functional assays, and machine learning to identify disease targets. Recent studies highlight challenges in single-cell foundation models, noting performance plateaus beyond certain dataset sizes and issues with batch effects and data redundancy. Specialized datasets, such as Relation’s Osteomics atlas, are increasingly critical for causal and generative models in drug discovery. High-quality, disease-specific datasets are becoming a key input for AI applications in biopharma.
single-cell multi-omicsfoundation modelsbatch effectsfunctional assaysdrug discovery
Generated automatically at 2026-08-07 20:24 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.