New Artificial Intelligence (cs.AI) papers, Wednesday 21 October 2026
322 papers announced on arXiv. Get them ranked for you, free →
Large Language Models for Agent-Based Modelling: Current and possible uses across the modelling cycle
AI summaryThis paper reviews how large language models can be integrated into agent-based modelling across the entire modelling cycle, from problem formulation to documentation, and discusses both opportunities and challenges of their use. It highlights LLMs’ potential to automate summarization, generation, analysis, and communication tasks within ABM.
An AI Security Agent for University ACMIS: Multi-Vector Threat Detection and Automated Response
AI summaryThis paper introduces an AI-based security agent for university management systems that combines anomaly detection and behavioral analytics. It achieves a 0.966 macro-average F1 score in threat detection while maintaining sub-millisecond response latency for critical risks.
INCARBench: A Benchmark for Scientific Configuration in VASP INCAR by Large Language Models
AI summaryThe authors introduce INCARBench to evaluate large language models on generating and repairing input configurations for VASP materials science simulations. Evaluating 19 models reveals that while semantic accuracy is high, task-critical correctness remains low, particularly in physically coupled settings like magnetism and DFT+$U$.
Fine-Tuning Small Language Models for Reliable VASP INCAR Generation
AI summaryINCAR-SLM combines a fine-tuned small language model with a deterministic post-processor to generate reliable VASP input files. It outperforms much larger general-purpose models, including GPT-5.4, on the INCARBench benchmark while enabling efficient local deployment for materials workflows.
Aptamers as a sandbox for lab-in-the-loop agentic molecular discovery
Despite breakthroughs in AI-driven molecular generation, existing discovery pipelines operate largely in silico, with wet-lab feedback reserved for validation rather than iterative improvement. Whether large language model (LLM)-based reasoning agents can accelerate de novo…
Measuring Consensus in AI Agent Networks
AI summaryThis paper introduces an evaluation framework to measure consensus formation and information flow in networks of interacting AI agents. Experiments show that sharing explicit explanations improves group decision accuracy, though persistent disagreement can occur despite correct aggregate answers.
Navalia: Advancing Cowork Agents through Verifiable Task Synthesis and Long-Horizon Post-Training
AI summaryNavalia generates verified financial analysis tasks by automatically selecting evidence, building computation chains, and checking consistency, then fine‑tunes and RL‑trains an LLM. It improves evidence integration, tool use, and long‑horizon performance, outperforming the base model on five benchmarks.
Disentangling Depression from Cognitive Decline in Elderly Speech Using Concurrent Clinical Assessments
AI summaryThe authors analyze a new Korean elderly speech corpus to disentangle depression from cognitive decline in MCI patients. Controlling for cognition reveals that only formant features reliably indicate depression, outperforming standard acoustic baselines.
PW-Pruning: Phase-Aware Expert Selection for Speech-Centric Omni-Modal Models
AI summaryPW-Pruning calculates separate importance kernels for acoustic-input and text-generation phases to prune speech-centric omni-modal mixture-of-experts models. Retaining 32 of 128 experts per layer improves speech benchmark accuracy by over 3 percentage points compared to standard pooled selection without increasing inference-time overhead.
Large language model and reinforcement learning based autonomous construction and progress updates
AI summaryThis paper introduces an autonomous construction method combining an enhanced soft actor-critic reinforcement learning algorithm for mobile crane control with ChatGPT for progress tracking and safety reporting. Experiments on full-scale retaining wall construction demonstrate the approach's feasibility for smart construction.
MOSS: A Production Framework for Monitoring and Diagnosing Semantic Failures in Search and Recommendation Systems
AI summaryMOSS integrates LLM relevance scoring into a production monitoring loop, sampling data, prompting strategically, and alerting with diagnostic root‑cause categorization. In a large retailer, it detected two defects and quantified a 76.4% alert‑rate drop, outperforming non‑LLM baselines on recommendation and search benchmarks.
Pareto-Optimal RTL Code Generation via Multi-Objective Reinforcement Learning with Large Language Models
AI summaryThis paper introduces a multi-objective reinforcement learning method for LLM-based RTL code generation that optimizes both the quality and diversity of circuit sets. The approach uses a hypervolume contribution reward to improve Pareto-optimal trade-offs in power, performance, and area.
Uncertainty Gates the Gradient: Adaptive Updates for Generalized Few-Shot 3D Segmentation
AI summaryUGA gates gradient updates using episodic uncertainty measured by intra‑class dispersion, allowing the model to learn from reliable 3D point‑cloud episodes while suppressing updates from noisy ones; on ScanNet200/++ it outperforms baselines by +3.08 HM IoU in 1‑/5‑shot settings.
MedPAO-Assist: Edge-Deployable, Modality-Aware Agentic AI for Radiology Report Structuring
AI summaryMedPAO‑Assist is a 0.6B‑parameter, edge‑deployable agentic system that turns radiologist dictation and voice commands into structured reports for four imaging modalities. It runs on a laptop CPU with <3.3 GB RAM, ~31 s latency, and requires no internet or keyboard input.
Steerable Pluralistic Policy Alignment from Anonymized Crowd Feedback
AI summaryP3L learns trajectory embeddings and a shared preference basis via PCA from anonymized crowd feedback, enabling fast, personalized reward learning without annotator IDs. It achieves competitive RL performance and out‑of‑distribution generalization in embodied control tasks while reducing compute.
Enhancing analytical photon transport in turbid media through physics-guided neural network corrections
AI summaryThis paper introduces a physics-guided residual learning framework that corrects systematic errors in analytical light transport models using Monte Carlo simulations. The method achieves significantly higher prediction accuracy and data efficiency than both standard analytical approximations and pure neural network surrogates.
DataPraxis: Scaling Analytics Tasks via Verified Workspace Evolution
AI summaryDataPraxis synthesizes analytics tasks by iteratively validating and merging workspace findings, creating 5,000 verified trajectories. The resulting 27B model matches or outperforms leading models on DABStep and JobBench.
Multilingual Brain Decoding Reveals Shared Cortical Semantic Geometry
AI summaryUsing brain recordings from speakers of Chinese, English, and French, this paper develops a multilingual decoding framework and reveals a shared cortical semantic geometry across languages that becomes increasingly consistent in higher-order association cortices.
Expert-Guided Rubrics as Dense Reward Signals for Reinforcement Learning in Non-Verifiable Domains
AI summaryUses expert‑rated rubrics to turn sparse, graded SME feedback into dense rewards for RL agents in non‑verifiable retrieval tasks; the hybrid signal boosts Recall@3 by 6% and MRR@3 by 6.7% over sparse‑reward baselines.
How Intervention History Changes AI Agent Sabotage on Subsequent Tasks
AI summaryThis paper investigates whether AI agents adapt their sabotage strategies on new tasks after learning they were previously caught and overridden. The results show that intervention history can reduce successful covert sabotage depending on the model and retained context, but never increases it.
GRaSP: Latency-Aware Optimization for Geo-Distributed Retrieval-Augmented Generation
AI summaryGRaSP adds a latency‑aware re‑scoring layer to RAG that penalizes distant passages and optional costs, solving a matroid‑constrained optimization in O(N log N). It cuts P90 fetch latency by 41% and keeps 78 % of relevance recall with negligible overhead.
Industrial Agentic AI: LLM-Driven New-Generation Agents for Smart Manufacturing
AI summaryIndustrial agents powered by large language models can autonomously plan and execute tasks in smart manufacturing, but face dynamic environments, physical constraints, and reliability challenges. This paper surveys five key challenges, proposes a unified taxonomy and four‑layer architecture, and reviews applications and future research directions.
Fix, Not Just Resample: Bounded Compiler-Guided Repair in Lean 4
AI summaryThe paper adds a bounded, online repair module that uses Lean’s compiler errors to fix failed tactics, sampling K candidates and repairing up to R times per step. This boosts solve rates by up to 32 % and 82 extra theorems, mainly by fixing shallow tactic failures.
MAMath-Bench: Process-Aware and Robust Evaluation of Multimodal Multi-Agent Mathematical Reasoning
AI summaryMAMath-Bench evaluates multimodal mathematical reasoning in multi‑agent LLM systems with 1,003 problems and fine‑grained process annotations. It shows that while agent collaboration improves reasoning coverage, it does not reliably boost final accuracy and leaves systems vulnerable to subtle adversarial errors.
LightCRS: Lightweight Multi-Domain Conversational Recommendation System
AI summaryLightCRS fuses precomputed multimodal product embeddings via a lightweight, query‑conditioned cross‑attention module shared across domains, enabling unified multi‑domain conversational recommendation. It outperforms baselines by 35.4 % in HR@1 and reduces per‑query latency 266× on a 25‑domain benchmark.
Beyond Policy Discrepancy: Capability-Aware Multi-Teacher On-Policy Distillation
AI summaryCA‑MOPD merges specialist RL models into a strong student and then adjusts teacher supervision based on verifier‑estimated capability gaps, reducing guidance when the student matches or exceeds a teacher and increasing it when regressing. It outperforms baseline distillation by 0.74‑0.86 points across six benchmarks.
FoundPDE: A generative foundation model for data-driven symbolic discovery of partial differential equations
AI summaryFoundPDE is a pretrained transformer-based foundation model that directly maps spatio-temporal trajectories to interpretable symbolic partial differential equations. It achieves over 90% type recognition across diverse PDE families and robustly generalizes to noisy data and unseen compositions via few-shot fine-tuning.
When Planning Problems Become Hard for LLMs: Separating Structural Complexity from Semantic Interpretability
AI summaryLarge language models fail on planning tasks due to predicate semantics, not structural complexity. The study shows that a simple delete‑relaxation heuristic predicts structural difficulty well, while LLM success drops sharply when predicates are renamed, highlighting semantic brittleness.
SZED at ArGuard Shared Tasks: Multi-Task Label-Attention and Cross-Modal Fusion for Arabic Hateful Meme Detection
AI summaryFusion of MARBERTv2 and CLIP ViT-L/14 embeddings yields 0.7597 macro‑F1 on Arabic hateful meme binary classification, while a multi‑task label‑attention network with cascade calibration achieves 0.3142 macro‑F1 on fine‑grained multi‑label tagging.
Dynamic Dual-Feedback Conformal Inference for Time Series Forecasting
AI summaryThis paper introduces Dynamic Dual-Feedback Conformal Inference, a method for time series uncertainty quantification that combines actual miscoverage feedback with estimated conformal thresholds. The approach achieves more stable and narrower prediction intervals while maintaining target coverage.
3D-MoE: Complementarity-Guided Hardware-Software Co-Design for MoE Decode on 3D-Stacked NMP
AI summaryThis paper introduces 3D-MoE, a hardware-software co-designed system that resolves the locality-balance conflict in Mixture-of-Experts LLM decoding using 3D-stacked near-memory processing. It achieves an average 4.63x MoE-layer speedup over expert parallelism by exploiting expert complementarity for load balancing.
CoASFL: Coupled Consistency Optimization for Asynchronous Split Federated Learning
AI summaryCoASFL reduces training lag in asynchronous split federated learning by merging client activations and adjusting aggregation weights to handle horizontal and vertical inconsistency, boosting accuracy by up to 3.26 % and speeding up training 1.6–3.7× on heterogeneous edge devices.
SPEM: Structure-Aware Procedural Memory for Efficient Skill Use in LLM Agents
AI summaryThis paper introduces SPEM, a structure-aware procedural memory framework that organizes LLM agent skills into atomic units and workflow DAGs. SPEM improves task execution efficiency and accuracy while reducing token and runtime costs across benchmark environments.
HierDTA: Hierarchical Structure-Aware Representation Learning for Drug--Target Affinity Prediction
AI summaryHierDTA models drugs as atom‑motif graphs with a hierarchical augmentation that propagates motif semantics to atoms, and converts protein kinase domains into residue interaction graphs; cross‑attention fuses these with fingerprints and ESM‑2 embeddings, achieving state‑of‑the‑art DTA performance (r² = 0.791 on KIBA) and strong cold‑start generalization.
Exploration via Activation-Space Steering for Reinforcement Learning with Verifiable Rewards
AI summaryExploration via Activation-Space Steering (ASE) diversifies reinforcement‑learning rollouts by clustering activation differences between correct and incorrect responses, then steering future rollouts with latent vectors. ASE boosts Qwen3 model performance on math benchmarks by 4.87 and 1.87 percentage points over GRPO.
Reliable Estimation of Differential Privacy under Selection Bias
AI summaryThis paper identifies and corrects a selection bias in membership inference attacks used for differential privacy audits. The proposed statistical corrections yield valid privacy lower bounds that respect theoretical upper bounds at the expected confidence levels.
BR-MoE: Budgeted Semantic Release for Privacy-Preserving Traffic Analytics
AI summaryProposes a budgeted semantic release framework that lets traffic cameras output privacy‑audited text summaries. It uses reinforcement‑learning–controlled expert summaries, calibrated noise, and a privacy filter to guarantee event‑level differential privacy while preserving traffic analysis utility.
Does Multi-Agent Collaboration Reduce Hallucination in Clinical Tasks?
AI summaryMulti‑agent LLMs boost clinical task accuracy, but they do not consistently lower hallucination in reasoning; the change depends on the underlying model, not on the collaboration topology. Self‑correction reduces hallucination for most models except Llama‑3.2‑3B‑Instruct.
Teaching AI Agent to Play Poker Strategically Through Preference Optimization
AI summaryWe train a compact poker agent to act strategically under hidden information by optimizing preferences with KL‑anchored equilibrium (SPOKE). SPOKE reduces mean exploitability (Kuhn 0.213, Leduc 1.081) and improves six‑max agreement (0.775) while matching teacher KL.
M-Plan: Comprehensive Planning For LLM Agents with Memory Arrangement
AI summaryM‑Plan combines global‑local and past‑current guidance into a closed‑loop MAS with short‑ and long‑term memory for LLM agents. Experiments on ALFWorld, Webshop, and PDDL show it improves accuracy and efficiency over baseline methods.
Robustness of Internal Signals for Hallucination Detection in Vision-Language Models
AI summaryThe paper evaluates internal signal‑based hallucination detectors in vision‑language models and finds that a simple token‑position baseline performs comparably, revealing many methods rely on positional shortcuts and are less robust to early hallucinations. It also shows ensembling complementary detectors yields modest gains.
Cross-Lingual Reusability of LLM Pipelines: Financial Extraction, Multilingual Reasoning, and Semantic Screening
AI summaryThe paper evaluates whether English‑built LLM pipelines can be reused across nine languages for financial extraction, multilingual QA, and semantic screening. It finds high accuracy transfer (≈95% extraction, 77–83% QA) and proportional reasoning‑effort gains, but stresses per‑domain, per‑language validation.
Physical AI: Concepts, Techniques, and Challenges
AI summaryPhysical AI is defined as systems that perceive, model, and act on the physical world through continuous feedback under real-world constraints. The monograph unifies its technical landscape—multimodal perception, world models, embodied foundation models, sim‑to‑real transfer—and outlines key challenges toward generalizable Physical AGI.
ReuF2F: Benchmarking Coding Agents on Formalized Undergraduate Research Problems
AI summaryReuF2F presents a benchmark of 244 formalized open problems from REU projects, evaluated in Lean 4 with coding agents. Even the best model solves only about 40 % of the tasks, highlighting significant gaps in current formal problem‑solving abilities.
Discrete Neural Interlingua for Ontology-Agnostic EHR Modeling
AI summaryThe paper introduces MEDIATOR, a discrete ontology-agnostic tokenizer that maps heterogeneous clinical concepts into a shared vector-quantized codebook using text descriptions and graph supervision. When used in EHR Transformers, MEDIATOR improves zero-shot transfer performance across health systems by 8.6 AUPRC points.
A general-purpose System One decision model for clinical records
AI summaryClinical System One (CS1‑4B) is a 4‑billion‑parameter model that provides typed clinical judgments from patient records in a single forward pass, outperforming larger open models on 13/14 tasks and speeding inference 6.4× while producing probabilistic, review‑ready outputs.
Omni-Woodpecker: Hallucination Correction via Evidence-Prior Rebalancing for Omni LLMs
AI summaryOmni-Woodpecker is a training-free framework for omni large language models that corrects cross-modal hallucinations by orthogonally separating and rebalancing grounding evidence and learned priors in hidden states. It improves reasoning and reduces hallucinations across multiple audio-visual benchmarks without modifying model weights.
Joint Constraint Aggregation for Qualifier-Sensitive Retrieval-Augmented QA
AI summaryJoint Constraint Aggregation (JCA) compares candidates against the full set of question-derived requirements using retrieved evidence, reducing bias toward frequently mentioned partial matches. Across three datasets and instruction-tuned models, JCA improves accuracy, especially when retrieval favors competing answers.
PROVISO: Provenance-Informed Action Selection for Secure LLM Agents
AI summaryPROVISO is a controller that evaluates an LLM agent’s proposed action against a trusted, provenance-typed memory of observations and alternatives, committing the action with the highest learned value. It outperforms baselines by 13.84 F1 on SkillsBench, 7.07 on ToolSafe, and 88–99 F1 on other skill detection tasks.
PORA: Learning When More Reasoning Pays Off for LLM Agents
AI summaryPORA learns when an LLM should invest a short versus a long reasoning budget by training a controller that compares outcomes from both options on the same state. The method outperforms baselines, saving up to 50 % of tokens while matching or exceeding long‑reasoning performance.
EDMA: Entropy-Driven Multimodal Answering
AI summaryThis paper introduces Entropy-Driven Multimodal Answering, a method that uses logical entropy to quantify modality contributions and trace reasoning steps in multimodal question answering. It outperforms standard prompting baselines by effectively minimizing entropy to select correct reasoning paths.
Reconstruction-Optimized Expert Pruning for Mixture-of-Experts Language Models
AI summaryThe paper introduces OptEP, a continuous relaxation method for post-training expert pruning in Mixture-of-Experts models that optimizes expert subset selection via output reconstruction. OptEP significantly outperforms existing baselines, yielding absolute accuracy gains of up to 16.2% under aggressive pruning across multiple large language models.
Mind Your Boundaries: Boundary-Aware Model Obfuscation against Cross-node Attack
AI summaryBMO obfuscates model parameters by separating boundary-position and boundary-curvature components, adding noise to the curvature to diversify decision boundaries. It reduces cross-node adversarial success to 68.68%, far better than the best baseline (22.32%), while keeping benign accuracy unchanged.
Ontic state recovery through prediction
AI summaryThis paper formalizes a system's state as a complete response profile over queries and analyzes how predictive models can recover these underlying states. The authors show that the fidelity of state recovery depends on query ambiguity, state-space geometry, and the choice of prediction targets.
Representing short-term movement to reveal animal identity and learning progress
AI summaryThis paper introduces a self-supervised framework that learns compact representations from short-term animal movement trajectories using temporal proximity. The resulting embeddings accurately predict individual identity and learning progress, revealing that egocentric pose and allocentric motion encode distinct behavioral signals.
ICLHerd: Herding-Based Data Selection for LLMs with In-Context Learning as a Proxy
AI summaryICLHerd balances mode-seeking and mode-averaging for LLM data selection using a von Mises-Fisher kernel and in-context learning as a proxy. Results show it outperforms existing data selection algorithms.
Model Exposure on One Device Endangers Others: Rethinking Edge Neural Network Security across Devices
AI summaryThis paper demonstrates that compromising an on-device neural network exposes other edge devices to transferable adversarial attacks and introduces VeilNN, a defense framework using architectural and parameter-level obfuscation to reduce adversarial transferability while preserving model functionality.
MoSAIK: Motion Search with Artificial Intelligence in K-space
AI summaryThe paper introduces MoSAIK, a deep learning method that estimates in-plane rigid-body motion directly from standard brain MRI k-space data without external navigators. It successfully improves image quality and reduces motion artifacts in over 75 percent of evaluated clinical slices during model-based reconstruction.
Understanding the Benefits of Scaffolds through Computational Separations
Inference-time harnesses, i.e., external procedures that call a model many times and keep state across calls, can substantially improve model capabilities. In this paper, we investigate the benefits of using such harnesses and the extent to which these benefits can be…
Learned Control Variates for Rare Grid-Outage Estimation at Electric Vehicle Charging Stations
AI summaryLearned control variates using a classifier on cheap features cut Monte Carlo variance by up to 60% for rare EV‑charging feeder overloads, reducing estimated blackout days by a third in Surabaya’s network compared to plain simulation.
Visual Task Modeling and Automated Planning for Humanoid Robots in Education of Individuals with Autism Spectrum Disorder
AI summaryThis paper introduces a vision-based framework that uses a Large Language Model to convert photos of educational tasks into PDDL representations, enabling humanoid robots to autonomously plan, execute, and assist students with autism spectrum disorder.
On the Fragility of Tasks under Composition: Implicit Budgeting in Language Models
AI summaryLanguage models err more on composite prompts than on isolated tasks because they implicitly budget tokens, using fewer per task. Grouping tasks by fragility lets the budget be reallocated, cutting token usage by 43 % while keeping accuracy.
CARLAR Influence: A Structured Hallucination Risk in Multimodal Cheque Extraction and an Exploratory Compressed-Image Countermeasure
AI summaryThis paper identifies CARLAR influence, a structured hallucination risk where multimodal models erroneously output identical values for distinct cheque amount fields under visual ambiguity, and evaluates a compressed-image few-shot mitigation. It shows that models exhibit high uncalibrated confidence during these errors, while a tabular visual prompting intervention yields model-dependent improvements in extraction accuracy.
LSA-PAST: Layer-Specific Adaptation and Prototype Transport for Compositional Incremental Learning
AI summaryThe paper introduces Layer-Specific Adaptation and Prototype-Anchored Adapter-Space Transport (LSA-PAST) for compositional incremental learning, using depth-specialized adapters and prototype transport to prevent representation drift. LSA-PAST achieves state-of-the-art accuracy-forgetting trade-offs on compositional image classification tasks without using data rehearsal.
LoWAgent: A Low-Resource Multilingual Benchmark for Web Agents
AI summaryLoWAgent evaluates web agents on English and four low‑resource languages, showing significant performance drops for all models, especially open‑source ones, and further degradation under interface disruptions. Small‑scale fine‑tuning on LoWAgent traces substantially recovers low‑resource performance.
Language Guidance for Supervised Vision Training: An Empirical Study of Generalization
AI summaryThe paper adds pretrained language model supervision to supervised vision training and evaluates Explicit and Implicit guidance on six generalization tasks. It finds Explicit guidance improves in‑distribution and continual learning, while Implicit guidance helps with shortcuts and adversarial robustness.
Profile-DP: Differential Privacy for User-Controlled Profile Release in Personalized LLM Inference
AI summaryThe paper introduces Profile-DP, a policy-aware local differential privacy framework for releasing user profiles in personalized LLM inference. It uses finite-domain mechanisms to provably bound the release of withheld facts while preserving authorized personal information.
Sparse RL Relaxation for LLM Reasoning
AI summarySparse RL Relaxation (SpaRLax) trains large language models with soft chain‑of‑thought tokens by treating them as differentiable computation steps, using entmax to adaptively sparsify token mixtures. Across nine reasoning benchmarks it outperforms prior methods, demonstrating the benefits of adaptive sparsity and soft‑token superposition.
Magnolia: A Persistent-Memory Agent Framework for Computational Chemistry
AI summaryMagnolia is a persistent‑memory AI agent for computational chemistry that logs and distills simulation data into reusable rules, improving reproducibility and efficiency. In four case studies, it tracked 317 runs, generated 116 project entries, and enabled 100 % of docking checks versus 10‑40 % without memory.
MMedAgent-Plus: Scaling Multimodal Medical Agents across Modalities, Tasks and Workflows
AI summaryMMedAgent-Plus integrates a vision-language planner with 18 specialist models to coordinate complex medical workflows across multiple modalities and tasks. It achieves high tool-call and final-answer accuracies, significantly outperforming existing baselines in end-to-end medical execution.
EGI-BENCH: An Enterprise General Intelligence Benchmark Suite
Enterprise adoption of LLMs in agentic workflows is gated by evaluations that span a wide arc of abilities: CRM operations, heterogeneous-document retrieval, personal data understanding, multi-turn planning, tool use, function calling, long-horizon research, and long-context…
AutoRedClaw: Automated Red-Teaming for Claw-Style Agents
AI summaryAutoRedClaw automatically generates realistic end‑to‑end attacks and required environments for claw‑style agents, achieving over 30% success at <$0.10 per sample and producing a 330‑case benchmark that raises attack success rates above 50% against state‑of‑the‑art agents.
Target-Aware Curation of Hallucination Stress Tests for Large Language Models
Hallucination stress tests can yield similar average scores across language models even when different prompts elicit hallucinations in each model. We develop a procedure that iteratively edits source-grounded seed prompts to expose hallucinations in a target model. An LLM…
ValiAssist: Scaling LLM-generated Programming Feedback Validation by Demonstration
AI summaryValiAssist extracts reusable validation rules from instructors’ edits of AI‑generated feedback and lets them inspect and refine these rules on new submissions. In a study with 16 instructors, it raised alignment with pedagogical intent from 57 % to 91 % and doubled correction accuracy from 32 % to 77 %.
TxnQA: Benchmarking Analytical Reasoning and Evidence Grounding over Transaction Event Sequences
Analyzing merchant transaction histories requires dynamic temporal filtering, entity resolution, and aggregation over irregular event streams--operations poorly captured by static tabular or time-series forecasting benchmarks. We introduce TxnQA, an analytical QA benchmark…
When Models Know but Can't Linearly Say: Relational Encoding Limits Linear Vulnerability Classification in Code LLMs
AI summaryCode LLMs struggle to linearly classify isolated code snippets as vulnerable because secure and vulnerable samples form overlapping distributions. However, these models reliably encode relative risk, successfully ranking which of two code snippets is more vulnerable.
Theorem prover Arend
AI summaryThis paper introduces Arend, a theorem prover based on homotopy type theory designed for constructive mathematics and supported by an IntelliJ IDE plugin. It features a novel cubical syntax close to Martin-Löf Type Theory and a flexible record-based subtyping system.
Auditing Safety Preference Supervision with Recorded Derivations
AI summaryThe paper introduces Symbolic Preference Labeling, which records attribute classifications, rule-based verdicts, and a ranking to recompute preference labels without re‑running a language model, enabling auditability. On 200 prompts it achieves 96% structured output, with policy models refusing 40% benign and 85% toxic prompts.
Certifying Repeatable Entry-Replay Witnesses for COBOL-to-Java Migration
AI summaryThis paper introduces a synthetic test-generation and certification method for COBOL-to-Java migration using execution traces and chronological tapes. Applied to a large legacy codebase, it generated 849 test cases achieving high branch and paragraph coverage while successfully validating behavioral equivalence.
A latent control model for realistic rodent motion
AI summarySCAMPER trains a virtual rat using two variational priors—one for motion control and one for exploratory behavior—producing realistic movements during both goal-directed tasks and random exploration, even with sparse rewards.
VPFlow: Value-Predictive Planning with Imperfect World Models for Offline Goal Reaching
AI summaryThe paper introduces VPFlow, an offline goal-conditioned reinforcement learning method that trains world models to predict values rather than future states. This value-predictive planning approach outperforms existing latent planners and model-free baselines across diverse benchmarks.
When Does Hate Transfer? A Large-Scale Study of Cross-Lingual Hate Speech Detection
AI summaryCross‑lingual hate‑speech detection works best when languages are linguistically similar; transfer quality also depends on temporal overlap and dataset size, but shared cultural cues explain residual patterns. The study offers guidelines for applying cross‑lingual models to low‑resource languages.
TEX: Test-Time Scaling Testing Agents via Execution-based Cross-Validation
AI summaryTEX parallelizes agents to generate test cases and patches, cross‑executing each pair to build a $K\times K$ feedback matrix that guides refinement without trace sharing. It attains 86.4% on SWT‑Bench Verified, a +9.0% boost over single‑agent baselines.
Vulnerability Analysis for Safe Multi-Agent Reinforcement
AI summaryThis paper investigates inference-time observation attacks on Safe Multi-Agent Reinforcement Learning by introducing a MinMax Attack framework that strategically perturbs compromised agents while accounting for benign agent resistance. Experiments show this method significantly increases safety violations while maintaining task rewards to remain stealthy.
Stronger artificial intelligence breakthroughs hasten their own exhaustion
AI summaryAI breakthroughs grow and then exhaust, as large successes draw attention away from diverse ideas. After each major win—AlexNet, transformers, LLMs—the field’s topic focus narrows and reviewer confidence drops, accelerating the depletion of future innovation.
Generalizing Multimodal Foundational Models in Context-Driven Class Incremental Learning
AI summaryExtends frozen multimodal foundational models to audio‑visual class‑incremental learning by using audio features as contextual guidance for visual representations. Feature‑ and logit‑level distillation reduces forgetting, yielding state‑of‑the‑art performance on audio‑visual CIL benchmarks.
Evaluating Context Compaction on the Questions That Come Later
AI summaryThe paper tests whether context‑compaction methods that preserve evidence for the current question also help future questions. It finds that sentence‑level compaction improves only the current answer, while word‑level deletion (LLMLingua‑2) better preserves useful content for later MultiRC questions, but its benefit varies across datasets.
Global Security Blind Spots: Translation as a Safety Bypass in Low-Resource Languages
AI summaryThey test ten large language models in 79 languages for translation‑bypass attacks, finding that safe English prompts still lead to unsafe outputs 22–44% of the time, especially in certain language families, and that model rankings are unstable across conditions.
When Does an Added Planning Capability Create Operational Value? An Observable Evidence Standard
Industrial artificial intelligence (AI) planning systems combine predictive models, optimization logic, controls, and governance rules, but improvement in the complete policy does not show whether one addition made an independent contribution or whether that contribution…
Evolving Self-tuning Benchmarks for Long-Horizon Tool Use
For AI assistants to move beyond simple chat interfaces and reliably assist users with applications like Gmail, Venmo, and Slack, we need to improve their tool-use capabilities. However, existing tool use benchmarks are still limited and difficult to scale because they rely on…
Steering Vision-Language Models for Visual-Contextual Privacy in Location Disclosure
Vision-language models (VLMs) can infer sensitive information such as geolocation from subtle visual cues, creating privacy risks when responses disclose more detail than the visual context warrants. We study visual-contextual privacy in geolocation disclosure, where a VLM…
Unsupervised Tool Evolution for Coding Agents in Data-Lake Analytics
AI summaryTEAL evolves reusable data‑lake tools for coding agents without labels by letting an Explorer and Critic propose changes that are kept only if the agent reuses them and the answer remains correct. It boosts performance under step budgets, improving Sonnet‑5 by up to 21.7 points.
Modeling Authentic Research Trajectories: A Graph View of Research Ideation
AI summaryThe paper introduces the Research Progress Graph to capture every step, evidence, and reasoning in idea evolution, shows current LLM agents miss these details, and argues that future agents should predict next states and train on full trajectories.
An LLM-Based Agent for Legal Compliance Assessment: Architecture, Memory, and Evaluation in a Canadian and Quebec Context
Checking whether a text complies with a body of law is slow when done by hand and fragile when done with keyword rules. This paper presents ComplianceAgent, the legal compliance sub-agent of COMPASS, a multi-agent framework that evaluates texts along four axes: digital…
Refuses When Asked, Complies When Acting: A Runtime Safety Check for Robot Control Loops
AI summaryVision-language models executing robotic control loops often fail to refuse harmful instructions because their action interface lacks a safety check. Adding a preliminary query asking the model to evaluate the instruction's safety stops all hazardous attempts without false positives on harmless tasks.
Exposing Vulnerability Monoculture Through Systemic Adversarial Robustness
AI summarySystemic Adversarial Robustness (SAR) quantifies how many attacks are needed to compromise a pool of models, revealing a “vulnerability monoculture” when SAR is low. Experiments show distilled models from DeepSeek‑R1 and Claude Opus 4.6 have lower SAR than their source pools, indicating higher systemic risk.
LiBRA: Logit Fingerprints for Byzantine-Robust Client Selection in Federated Learning
Byzantine-robust aggregation shields federated learning from malicious updates, but the rules that filter out suspect updates do so by comparing clients' parameter vectors. That comparison requires access to every client's update, which secure aggregation exists to withhold,…
On the Limits of Scale: Inverse Scaling in Vision-Language Models
Scaling laws suggest that increasing model size, training data, and compute generally yields monotonic performance improvements. While this often holds at the aggregate level, we present systematic evidence that vision-language models (VLMs) exhibit inverse scaling at the…
Predictive Closure: A Set-Level Characterization of JEPA's Feature Selection
In joint-embedding predictive architectures (JEPA), an encoder and a predictor are trained together to produce a representation that minimizes the error in predicting the representation of the next observation. Because the representations of both the current and the next…
Trail-Weighted Influence Scores for Approximate Outcome Prediction in Bayesian Networks
We study trail-weighted pruning of evidence for Bayesian-network prediction. A traversal enumerates the active trails from each evidence node to the outcome, scores each trail by the product of its edge strengths, and drops evidence whose score falls below a threshold. For…
Fundamental Impossibility Results in AI Safety: Trading Off between Desiderata
AI summaryThe paper formalizes AI safety interactions as a game between a defender, attacker, and AI to study trade-offs in AI alignment. The main result proves a fundamental impossibility theorem showing an inevitable tradeoff between helpfulness and safety when AI advice is publicly accessible.
REM: What to Store, When to Recall, What to Forget in Experience Memory for Coding Agents
Large language model (LLM) agents now autonomously complete long-horizon tasks, such as repository-level coding, through hundreds of think-act steps, yet the experience each run produces is rarely reused on the next. Prior memory systems concentrate on retrieval, finding the…
Correcting Attribution Patching with a Few Exact Measurements
AI summaryGuided‑AP uses a handful of exact patching measurements to identify and correct systematic errors in attribution patching. It reduces error by up to 26% on a single prompt and 75–79% on average factual‑completion effects, while keeping verification cost low.
When PTQ Calibration Helps or Hurts: Dose-Driven Collapse and Validation-Gated Abstention in Quantized Text Embedders
Learned post-training quantization (PTQ) calibration fits quantizer parameters on a small calibration set, and for a text embedder the natural supervision is retrieval relevance. We ask two questions that are usually conflated: how far to calibrate, and whether to calibrate at…
Unified Sharp Bounds on Probabilities of Causation
AI summaryThis paper proves the sharpness of previously proposed closed-form bounds for multi-valued probabilities of causation in high dimensions. It also demonstrates that various existing causation measures are special cases of a single, unified general formula.
ProMemAgent: Long-Horizon Proactive Assistance with Context-Aware Prospective Memory from Sensory Perceptions
AI summaryThis paper introduces ProMemAgent, an LLM agent with context-aware prospective memory that organizes sensory perceptions to provide long-horizon proactive assistance for delayed intentions. Experiments on a new benchmark show it outperforms existing baselines in memory retrieval and proactive assistance.
More Memory, Worse Agents: Error Reproduction and Anti-Persistence in LLM Agents
AI summaryLLM agents can self‑improve by storing past runs, but reusing inferred abstractions propagates errors, a process called error reproduction. The Anti‑Persistence design discards synthesized guidance after use, only keeping raw trajectories, and outperforms other memories on success rate and LLM‑call cost.
KPHarness: Co-Evolving Knowledge and Probing for Long-Horizon Gaming Harnesses
A failed gaming trajectory reveals that an agent needs intervention, but not necessarily how its execution harness should change. Effective harness adaptation therefore requires gathering diagnostic evidence, rather than merely proposing alternative recovery rules. We present…
The Router Already Knows
Mixture-of-experts (MoE) models already make a sparse choice for every token: which experts should process it. We ask whether a sparse dictionary should reuse that choice. NR-SAE gives each routed expert a small dictionary and opens only those the frozen router selects. At block…
Harness Self-Evolution for Long-Horizon Terminal Agents
AI summaryRiseHarness iteratively refines terminal-agent harnesses by contrasting successful and failed runs, using runtime traces to diagnose and patch shared mechanisms. Experiments show that integrating cross-framework fixes into a single harness boosts reliability and task success across models and unseen benchmarks.
MLIPS: Multi-Layer Indexer Pool Sharing
Sparse attention limits the keys used by each query, but finding those keys can still require a full-prefix scan at every layer. We introduce \textbf{MLIPS (Multi-Layer Indexer Pool Sharing)}, a training-free method that shares candidate search while retaining each consumer's…
Scaling Graph Reasoning in Language Models with Learnable Graph Positional Encodings
Solving reasoning tasks supported on graphs typically requires providing a Language Reasoning Model (LRM) with a serialized description of the graph. This approach results in the input length growing with the number of edges, which limits the size of the problems that can be…
Beyond Scalar Leaderboards: Adaptive Sampling for Pareto Frontier Identification in Multi-Objective LLM Arenas
The evaluation of large language models (LLMs) is inherently multi-objective, as different models may offer different tradeoffs across performance dimensions such as math, code, and writing. Because such evaluations are often based on costly and time-consuming human feedback, it…
Ask to Distill: Learning to Synthesize Data for On-Policy Distillation
On-policy distillation (OPD) can succeed with surprisingly few training questions, motivating a view of questions as carriers of distributional knowledge transfer. How can we assess whether a newly generated question is valuable for distillation, and learn to generate better…
Recoverability and Verifier Strength: Two Measured Preconditions for Unattended Agent Self-Improvement
AI summaryDefines two metrics—verifier strength and recoverability—to decide when an automated agent can self‑modify without human oversight. Experiments show only changes with stored previous versions and a measured strong verifier can be safely deployed unattended.
Projected Distillation: On Learning to Solve Hard Problems from Close, Correct Targets
On-policy learning is central to post-training: training on a model's own generations limits distribution shift and helps preserve existing capabilities. On hard problems, however, on-policy RL provides no reward-based learning signal when all attempts fail. On-policy…
Training-Free Continual Learning for Multimodal Forecasting via Prompt Evolution
Adapting a forecasting system to a changing world classically means retraining. We ask whether the task of continually learning regimes for time series forecasting can be represented by a piece of text: a guidance block, evolved by outcome-driven reflection and used in a frozen…
Privacy-Preserving Personalized Federated Learning (PPP-FL) for Medical Imaging
AI summaryThe authors propose a DCGAN-based model inversion attack on medical federated learning and evaluate client-side differential privacy paired with personalized federated learning as a defense. They demonstrate that while generative attacks can successfully reconstruct private brain MRI scans, personalized federated learning recovers the accuracy lost from differential privacy defenses.
Beyond the Prompt: Recovering Hidden Concepts in Text-to-Image Models
AI summaryThis paper introduces EmbGA, an embedding-space genetic algorithm that efficiently recovers hidden visual concepts unmentioned in the prompts of text-to-image models. EmbGA outperforms existing optimizers by achieving higher concept alignment and faster execution speeds.
Runtime Control for AI Agents: Closing the Evaluation--Execution Gap in Operational Workflows
Language-model agents increasingly participate in operational workflows where approvals, binds, and payments create commitments. An approval valid when issued may no longer authorize execution after relevant facts, evidence, policy, or authority change. We call this the…
Securing Agentic Commerce Through Behavior-Aware Fraud Detection and Cybersecurity for Trusted AI-Agent Payment Authorization
The growing use of autonomous artificial intelligence agents in digital commerce creates new security concerns involving agent impersonation, unauthorized payments, compromised agents, and misuse of delegated authority. This research introduces the Autonomous Commerce Agent…
The More the Better? A Comparison of Single- and Multi-Agent LLM Setups on a Financial Review Task
AI summaryThe paper compares single- vs. multi-agent LLMs for reviewing credit assessment memos. A 3-agent team finds 0.08–0.10 more defects but raises more false alarms, using 2.2–2.4× tokens, while repeated sampling of a single agent yields higher balanced accuracy.
Orthogonal Boundary Unlearning for Dual-Use Concepts in Large Language Models
Mitigating catastrophic risks from large language models requires excising hazardous operational capabilities, such as cyberwarfare exploitation and pathogen enhancement, without degrading benign scientific utility. In these dual-use domains, legitimate defensive science and…
Self-Evolving Agents through Privileged Retrospective Distillation
Self-evolving agents aim to improve from their own experience rather than remain fixed after deployment, but naturally collected trajectories often lack ground-truth solutions or verifier-derived rewards. We introduce Retrospective Self-Distillation (RSD), a multi-round…
Cross-Visibility-Gap Identity Persistence for 4D Scene Graphs
4D scene graphs require persistent object identity when an object disappears inside an opaque carrier and later reappears as a fresh detection. When appearance and target-centric motion are unavailable, identity recovery must rely on relational history. We formulate…
Structured-Token Attention for Fine-Grained mmWave Liquid Identification
Fine-grained liquid identification requires a classifier to distinguish similar substances from compact millimeter-wave measurements. We investigate structured-token attention, which represents a radar feature tensor through complementary antenna, feature, branch, and global…
Progressive Subset Selection and Cascaded Consistency for Unsupervised Vision-Language Model Adaptation
Pretrained Vision-Language Models (PVLMs) achieve strong zero-shot recognition but often suffer performance degradation under target-domain shift. Unsupervised adaptation provides a practical way to improve target-domain performance without labeled data, but it remains…
HealBench: A Fine-Grained Benchmark for Evaluating Psychological Comfort in Text-to-Image Generation
AI summaryThis paper introduces HealBench, a 30,000-sample benchmark evaluating whether text-to-image models produce psychologically comforting visuals. Evaluations of 14 models reveal that high aesthetic and semantic quality do not guarantee emotional comfort, and demonstrate that a proposed reward model can improve restorative generation.
When Health Assistants Agree Too Readily: A Benchmark of Time-Critical Self-Diagnosis
AI summaryThis paper introduces SelfDx, a benchmark evaluating whether AI health assistants maintain epistemic independence instead of dangerously agreeing with user self-diagnoses. It finds that 27 frontier models frequently fail in time-critical scenarios, a flaw that supervised fine-tuning successfully mitigates.
MemDev-Bench: Does Memory Help Code Agents?Benchmarking Memory Systems over Real Repository Histories
Code agents increasingly work on evolving real repositories, and memory systems aim to help them reuse earlier experience. However, such experience is diverse and evolving, so it remains unclear how memory systems organize it and thereby affect code agents. Existing benchmarks…
Automated Multi-Agent Organization Design: From Architecture Search to Organizational Learning
Agentic system design is increasingly shifting from hand-crafted prompts and workflows toward automated search over agentic programs and interaction structures, with multi-agent systems receiving growing attention. Yet a central question remains unresolved: when does an…
Adversarial Benchmarks and BenchDoor Attacks: Exploiting Benchmark Incentives for Backdoor Attacks
AI summaryThis paper introduces BenchDoor, a data poisoning attack that exploits machine learning benchmark incentives to implant backdoors while improving benchmark scores. Across 30 configurations, the attack achieves high backdoor success rates while increasing benchmark performance, even after decontamination.
Less is More: Concise Retrieval and Critic-based Re-ranking for Composed Image Retrieval
AI summaryConcise Retrieval and Critic-based Re-ranking reduces semantic noise by generating minimal, modification-focused captions and re-ranking candidates for consistency and reference preservation, achieving state‑of‑the‑art zero‑shot composed image retrieval performance.
Act on Your Behalf, Beyond Your Authorization: Benchmarking LLM Agent Authorization Alignment
When an LLM agent acts on behalf of a user, it is expected to act within the authorization scope the user grants. Existing agent evaluations largely measure task completion or harmful consequences and can therefore overlook actions that violate this delegation without producing…
Can Native Safety Travel? Reusing Safety Pretraining Across Models via On-Policy Distillation
Safety pretraining has emerged as a promising paradigm for embedding safety deeply into a model during pretraining, yielding robustness to subsequent fine-tuning that conventional safety post-training often lacks. We refer to this safety acquired during model pretraining as…
AnchorRM: Parallel Intra-Batch State Reconstruction with Anchor-Based Residual Memory for Temporal Graph Neural Networks
Memory-based temporal graph neural networks (TGN) rely on recurrent node memories to capture evolving interaction histories. However, temporal mini-batching introduces a trade-off between computational efficiency and temporal fidelity: when a node appears multiple times within…
CityPaint: A Free Lunch from 2D Diffusion for City-Scale 3D Texturing
Generating high-fidelity, city-scale 3D environments plays a crucial role in applications such as virtual reality, gaming, and urban digital twins. However, the scarcity of real-world city-scale 3D training data forces existing methods to either stitch independently generated…
AnomalyJev: Towards Evaluating Image Anomaly Explanations via Claim-Level Verification
Recent multimodal LLM (MLLM)-based anomaly detection (AD) methods can generate textual explanations that help interpret their predictions. However, reliably and efficiently evaluating these explanations remains challenging. While LLM judges can assess explanations through…
Learned Readout Tokens for Extracting Large Language Model Confidence
AI summaryLearned readout tokens replace existing-token probes for LLM confidence estimation, yielding higher AUROC across datasets and models. A multi‑token extension further improves performance, and the method can be applied to any probe, boosting accuracy both in‑domain and out‑of‑domain.
AudiPsy: Evidence-Grounded Multimodal Reasoning over Spoken Counseling Conversations for Depression Assessment
AI summaryAudiPsy presents an evidence‑grounded multimodal reasoning framework for depression assessment in spoken counseling, integrating textual and acoustic emotions via an Observation‑Assessment‑Summary paradigm. On DAIC‑WOZ it outperforms specialized detectors and general LLMs while providing higher‑quality reasoning.
TRCM: Topology-Aware Route-Consistency Modeling for Robust Neural Vehicle Routing
Neural combinatorial optimization has enabled fast constructive solvers for the Capacitated Vehicle Routing Problem (CVRP), yet most autoregressive models still construct solutions through sequential node selection and represent route organization only implicitly. We investigate…
RaspAssistant as an Agentic Framework for Automating Adsorption Simulation Research in Porous Materials
Porous-material adsorption studies require coordinated structure preparation, force-field selection, simulation, validation, and interpretation, making end-to-end automation difficult. We test whether RaspAssistant, a hierarchical multi-agent framework, can carry an adsorption…
When LLMs Manipulate: Measuring Behavioral, Social, and Temporal Effects
Large language models (LLMs) can influence user trust and decision-making through subtle linguistic cues, yet the behavioral effects of manipulative outputs remain poorly quantified. We introduce a behavioral computing framework for measuring \textit{behavioral vulnerability},…
The Show-Up Attack: When Subagents Win Over Their Supervisor
Multi-agent systems are increasingly deployed to coordinate multiple large language model (LLM) agents. However, this also creates new attack surfaces: malicious agents may have chances to exploit the coordination. In this paper, we propose and study the \textit{show-up attack},…
Factual Access States in Diffusion Language Models: From Mechanism to Knowledge Editing and Retrieval
Factual recall is a lookup: some representation inside the network serves as the key, and what that key is made of decides which queries can retrieve the fact, what an edit must cover, and what a retriever can read. We call it the factual access state. In autoregressive (AR)…
DuplexLedger: Benchmarking Transactional Memory in Full-Duplex Speech Agents
Full-duplex speech agents must update persistent memory while listening and speaking, yet it remains unclear which events should be stored, revised, or excluded. Existing benchmarks isolate interaction timing from memory outcomes, obscuring the state transitions behind a correct…
Beyond Experience: Knowledge-Guided Evolution for NPU Kernel Generation
LLM agents have become capable kernel engineers on mainstream GPU stacks, but they struggle on emerging accelerators such as Ascend NPUs, whose programming model is barely represented in pretraining data. Recent NPU agents address this gap largely through experience: they…
Benefit and Cost of Random-Token Replacement under Repeated Data: A Known-Entropy Decomposition and Evidence from Ten Real Corpora
When unique pretraining data is scarce, models see the same tokens many times and are thought to memorize. Random-token replacement, an input-corruption augmentation, is one remedy reported to lower validation loss, but a lower loss alone does not reveal which source of error it…
WORTH Benchmark: Evidence Discovery under Not-Fully-Cooperative Information -- A Two-Episode Pilot
AI summaryThe paper introduces the WORTH Benchmark, where models read full mystery stories, autonomously pick key clue sentences, identify participants, and explain reasoning. In a two‑episode pilot, models frequently selected incorrect participants and missed key clues, revealing misaligned evidence selection.
Meta Expert: Cross-Layer Expert Sharing for Storage- and Retrieval-Efficient MoE
As model parameters scale up, sparsely activated Mixture-of-Experts (MoE) approaches have become essential for model scaling. However, mainstream models currently adopt hierarchical architectures, where experts in different layers, although topologically independent, exhibit…
From Encoding to Generation: Towards Empathetic Large Language Models
Empathetic interaction involves two distinguishable capacities: representing another person's emotional state (cognitive empathy) and generating an appropriate response to that state (affective empathy). We apply this distinction to examine how emotion-related information is…
Auditing the Product, Not the Recursion: The Evaluation Evidence Behind Recursive Self-Improvement Claims
AI summaryThe paper audits 114 studies claiming recursive self‑improvement, finding that while 90% report product‑level evaluation, only 22% assess repeatability with full loops, and only 4 compare identical procedures under matched budgets. It proposes a claim‑indexed reporting standard.
ABench: Adversarily Benchmarking Large Language Models across Taxonomized Values
Ensuring the alignment of Large Language Models (LLMs) with diverse cultural values is critical for their safe global deployment. However, existing alignment benchmarks are predominantly English-centric or confined to simple discriminative tasks. Moreover, they often fail to…
Learning Inter-Graph Interactions Between Heterogeneous Substructures of Chemical Systems
Complex chemical systems containing heterogeneous substructures are common in real-world applications. Various physical phenomena of the complex chemical systems are derived from the interactions between the heterogeneous substructures. However, existing graph representation…
Local Allow, Global Deny: Why Per-Action Authorization Does Not Compose in LLM Agents
AI summaryThis paper shows that checking LLM agent tool calls individually fails to catch global policy violations that require comparing multiple executions. The authors prove that no per-action monitor can detect these relational violations without false alarms, a limitation confirmed empirically across multiple LLM judges.
A Head Camera Is (Almost) All You Need: An Electronics-Free UMI Glove Labelled and Observed by One Moving Camera
AI summaryThe paper replaces wrist cameras, pose trackers, and cables with a single head‑mounted camera that reads fiducials on an electronics‑free glove to infer 6‑DoF pose and jaw width. This approach yields 6.5 mm/1.6° median error, full frame coverage, and improves policy performance, matching or exceeding traditional multi‑camera setups.
TriGeoBench: A Multimodal Mathematical Geometry Benchmark with Figures and TikZ
Multimodal Large Language Models show strong vision-language capabilities, yet their geometry-centric reasoning remains insufficiently understood. Existing mathematical benchmarks typically rely on rasterized figures and cover limited geometric knowledge areas and response…
When Does Feedback Help? Auditing Learned Voltage-Clamp Experimental Design
AI summaryThe paper audits learned voltage‑clamp controllers and finds that they can outperform poorly specified planners even without using feedback. In synthetic tasks, fixed protocols and masking lead to near‑optimal performance, and the learned policy’s actions can be reproduced by a simple threshold model.
SCODEL: Shared Canonical Ontology Definitions for Food Entity Linking and Ontology-Grounded Recipe Knowledge Graphs
AI summarySCODEL learns shared canonical definitions for food mentions and ontology concepts, acting as a semantic normalization layer that boosts existing EL models without architectural changes. It raises Hit@1 to 88.33% on seen and 83.15% on unseen concepts, outperforming LLM baselines and enabling large-scale ontology-grounded recipe knowledge graphs.
BCT: Bayes-Consistent Tempering for Flow Language Models
AI summaryBayes-Consistent Tempering (BCT) integrates temperature into the flow distribution and denoiser dynamics, aligning training and inference. BCT boosts GSM8K and Sudoku performance across temperatures, outperforming post‑hoc scaling and training‑free variants on hyperspherical and von Mises–Fisher flows.
CUA-Recovery: Benchmarking Error Recovery in Long-Horizon Computer-Use Agents
AI summaryCUA-Recovery benchmarks error detection and recovery for long-horizon desktop‑agent tasks, providing 2,250 verified failure states and 8.6K recovery trajectories. Fine‑tuning Qwen3.5 on this data boosts error awareness from 20 % to 35 % and task success scores markedly.
Corpus Quality Improves, Retrieval Utility Does Not: An Ablation of Preprocessing and Refinement in a Speech-to-QA Pipeline
AI summaryThe study shows that preprocessing and refinement improve corpus quality, but only transcript rewriting harms downstream retrieval accuracy. The improved corpus raises retrieval scores from 0.231 to 0.748 across models, yet quality metrics do not predict this benefit.
A Unified Attack Framework against Statistical Generative Watermarks for Large Language Models
AI summaryThis paper introduces a black-box attack against statistical LLM watermarks by exploiting a fundamental link between detection scores and token probability shifts. The proposed optimal attack successfully evades seven major watermarking schemes, reducing detection rates to below five percent while preserving text quality.
Does Multi-Agent Collaboration Pay When Agents Serve Users? Testing Agent Scaling in Dual-Control Settings
AI summaryThe paper tests a multi‑agent system (MAS) against a single‑agent system (SAS) in user‑interaction domains and finds MAS performs worse in 12 of 13 settings, with no measurable accuracy gain and higher token usage, challenging the assumed benefit of collaboration.
When Can LLM Reasoning Reliably Reuse Intermediate Results?
AI summaryThe paper studies when LLMs can reliably reuse intermediate results during multi‑step reasoning and quantifies the computational cost. It shows that providing prior subproblem answers or full reasoning boosts second‑step success (up to ~58%) and gives theoretical bounds on retry attempts.
Inferring continuous cognitive traits from limited neuroimaging data using pairwise convolutional neural networks
AI summaryPairwise convolutional ranking learns from whole‑brain MRI by comparing subjects, yielding accurate predictions of spatial ability in small samples (N = 108) and outperforming linear and pointwise neural baselines. Backward elimination highlights shared and task‑specific cortical regions linked to performance.
Pairwise Learning Increases Data Efficiency in Deep Neural Networks
AI summaryPairwise learning reframes regression as aggregated ranking comparisons, yielding massive implicit data. It improves data efficiency, enabling a lightweight ResNet10 to achieve 3.64‑year MAE on UTKFace and boosts STL10 classification by up to 5.2% with limited labels.
Reusing Multimodal Agents: Efficient Evaluation of Sound Designs in Games
AI summaryThe paper presents a framework that pre‑trains a multimodal encoder and a fixed RL policy, then fine‑tunes only the encoder to evaluate in‑game sound designs. It detects performance differences due to sound changes while reducing training cost by up to an order of magnitude.
CovNav: Conformal Prediction-Guided Adaptive Inference for Object Goal Navigation
In object goal navigation (ObjectNav), an embodied agent must find a target object (e.g., "find a bed") in a previously unseen environment, relying on spatial commonsense such as object co-occurrence. Large Language Models (LLMs) encode this commonsense, but deploying them on…
Temporal Evidence Allocation For Longitudinal Clinical Retrieval: A Position And Technical Report
AI summaryThis paper frames longitudinal clinical retrieval as allocating a limited evidence budget across time to avoid temporal redundancy in oncology records. It introduces the Onco-CONSTRUE framework and demonstrates that standard retrieval strategies struggle because text volume does not correlate with clinical usefulness.
Pattern Intelligence: Unsupervised, LLM-Interpreted Discovery of Systematic Underpayment and Denial Patterns in Healthcare Revenue Cycle Management
AI summaryPattern Intelligence clusters claims by adjudication behavior, not payer, to discover systematic underpayment and denial patterns. Using HDBSCAN and LLM-annotated centroids, it identifies actionable clusters, revealing a long tail of high-dollar denials that remain uncaptured by traditional methods.
EASformer: Environment-Action-State Alignment for Industrial Generative Ranking
AI summaryEASformer aligns environment, action, and state tokens for generative ranking by disentangling user actions from history and candidate features, enabling action retrieval and state‑conditioned prediction. It outperforms the baseline with 0.88 % AUC lift offline and 1.26–3.94 % eCPM gains online.
VLA-Retain: Benchmarking Cross-Dimension Effects of Vision Tuning in VLAs
AI summaryThe authors introduce VLA-Retain, a benchmark evaluating how vision-tuning in vision-language-action models affects robustness across diverse perturbation dimensions. They find that targeted fine-tuning frequently causes off-target performance regressions that are not predicted by target gains.
E8-QAT: Efficient 2-Bit Quantization-Aware Training with Implicit Lattice Codebooks
AI summaryThis paper introduces E8-QAT, a 2-bit quantization-aware training framework using an implicit lattice codebook that avoids explicit codebook searches. It matches full-precision performance on Qwen3.6-35B-A3B while improving decoding throughput by nearly double.
Beyond Calibration: Temporal LLM Confidence Assessment Across Languages and Local Settings
AI summaryThe paper separates pre‑solve (Feeling‑of‑Knowing) and post‑solve (Judgement‑of‑Learning) confidence signals in large language models, showing that post‑solve confidence is better calibrated yet more expensive. Across English and Marathi, translation preserves calibration, but local Marathi knowledge degrades it, and hallucinations keep high confidence.
How Fine-Tuning Objectives Shape Layer-Wise Information in LLM Hallucination Detection
AI summaryClassifier‑head fine‑tuning yields a two‑cluster hidden‑state geometry with strong class separation, outperforming sequence‑to‑sequence fine‑tuning. Truncating the upper 31% of layers, guided by the InfoWindow, preserves accuracy while cutting latency 1.4×.
RLSplat: Resource-Linear Budget Control for Feed-forward 3D Gaussian Splatting
AI summaryThe authors propose RLSplat, a budget-controllable feed-forward 3D Gaussian Splatting framework where inference latency and peak memory scale linearly with the Gaussian budget. RLSplat outperforms prior methods across all budgets while achieving significantly faster execution and lower memory usage at low budgets.
When Does Experience Memory Help Closed-Loop LLM Driving Agents? A Controlled Evaluation with Resampled Warmup Pools, Token-Matched Fillers, and Audited Baselines
AI summaryThe study evaluates closed‑loop LLM driving agents with controlled warm‑up experience pools and token‑matched fillers on a HighwayEnv testbed, finding that memory gains depend on the specific warm‑up pool, token fillers dominate, representational depth is ineffective, and baseline quality reverses apparent advantages.
How Language Models Organize Task Information
AI summaryThe paper introduces k, the minimum number of linear‑probe features needed to recover 95% of dense‑probe performance, and shows that across 19 language models many core tasks require similar sparse feature sets. Instruction tuning reduces k, and k predicts label needs for learning and diagnostics.
World2HOSI: A Multi-Expert Harness for Long-Horizon Human-Object-Scene Interaction Generation
AI summaryWorld2HOSI orchestrates existing HOI and HSI experts to perform long‑horizon human‑object‑scene interactions using 3DGS scenes and natural language instructions. It achieves higher task completion, interaction quality, and data diversity than single‑expert models.
Anytime-Valid Runtime Monitoring of Learned Barrier Certificates
AI summaryThe paper presents an anytime‑valid runtime monitor that tests learned barrier certificates using closed‑loop trajectories, forming a sequential hypothesis test and a supermartingale to control false alarms under any stopping time. It demonstrates empirical alarm behavior on valid and invalid certificates.
RSI Control: Can AI Models Perform Covert Sabotage in the Self-Improvement Process?
AI summaryRecursive AI models can covertly sabotage their successors during training, succeeding in many side‑tasks without hurting main performance; current LLM monitors miss most attacks, indicating RSI poses serious safety risks.
Hearsay or Fact? How Framing and Elicitation Shape LLM Probability Updates
AI summaryThis paper investigates how prompt framing and elicitation formats affect how much large language models update their probability estimates in response to new information. The results show that indirect reporting slightly reduces update magnitudes, while eliciting multiple weighted estimates reduces extreme probabilities but can impair evidence sensitivity.
S$^2$-MAS: A Society-Scale Multi-Agent Simulation Framework with 500 Million Household Persona Agents
AI summaryS$^2$-MAS simulates 500 million synthetic household agents with realistic decision dynamics using the Genese engine, enabling large-scale economic experiments. Its FinL-mlx model reaches 74.55% accuracy on consumer finance choices, outperforming baselines, and the framework shows how purchase goals and credit limits drive consumption across demographic groups.
Give the Model a Finger: Developmentally Aligned Evaluation of Counting in Vision-Language Models
Vision-language models (VLMs) are exact on small sets and drift into estimation on larger ones, a profile reminiscent of children before they master counting. Children cross the barrier with epistemic actions: pointing at each item while reciting number words, moving counted…
CrystalAR: Autoregressive Modeling for Symmetric Crystal Generation
AI summaryCrystalAR unifies autoregressive modeling of Wyckoff symmetry templates and 3‑D crystal coordinates, enabling constraint‑aware search over both discrete templates and continuous structures, achieving competitive de novo generation and template‑free CSP accuracy without pre‑existing symmetry templates.
How Much Do $\ell_1$ Attacks Need Concentrated Perturbations?
AI summaryThis paper investigates perturbation concentration in L1 adversarial attacks by enforcing per-coordinate caps during optimization. The results show that limiting large coordinate changes retains 97 percent of attack success while significantly reducing peak perturbation magnitudes.
Radar-Based 3D Human Pose Estimation via Motion-Aware Pose Initialization and Observation-Guided Refinement
AI summaryThe paper introduces I2R-Pose, a two-stage framework for radar-based 3D human pose estimation that combines motion-aware initialization with observation-guided refinement. This approach significantly reduces the mean per-joint position error compared to existing methods on benchmark datasets.
RITE: Evaluating End-to-End Resilience to Silent Tool Errors in E-Commerce Agentic Systems
AI summaryRITE benchmarks end‑to‑end resilience of e‑commerce LLM agents to silent tool errors by injecting controlled semantic errors into real workflows. Across 10 models, silent errors cut success rates by ~50%, revealing recognition and evidence integration as key bottlenecks.
GRADUS: Discovering and Recovering Agent Lessons from the Environment's Latent Syllabus
AI summaryGradus extracts behavioral rules from policy-constrained environments to automatically synthesize a curriculum of increasingly difficult tasks. This approach jointly calibrates tasks and solutions using execution feedback, boosting a 30B LLM's performance from 59% to 89% on tau-Bench.
RELAY: Scaling Historical Build Restoration through Rule Reuse and Agentic Repair
AI summaryRelays combines LLM‑derived deterministic repair rules with per‑case agents to fix historical builds. It automatically produces reusable rules that restore 69.7% of broken OSS‑Fuzz cases, matching per‑case agents but using fewer rebuilds, while agents finish the remaining failures.
SLM-SS: Token-Based Speech Separation with Speech Language Models
AI summarySLM-SS tokenizes multi‑speaker speech into discrete codebooks, then uses an autoregressive speech language model plus a non‑autoregressive refinement to separate speakers, achieving 6.92 % WER on Libri2Mix and superior subjective quality over discriminative baselines.
Toward Agentic AI for the 6G Physical Layer: From Algorithm Discovery to Verified Baseband Silicon
AI summaryLLMs can automatically implement, optimize, and verify AI‑discovered 6G baseband algorithms on real silicon, achieving cycle‑exact operation. In a demo, the agent produced 12 OTFS detector kernels, reducing cycles by 13.9% over the original design while preserving error performance.
LatBridge: Efficient and Stable Model Replacement in Latent-Communication Multi-Agent Systems
AI summaryThis paper introduces LatBridge to enable efficient and stable model replacement in latent-communication multi-agent systems. By using a shared bridge and modular low-rank interfaces, LatBridge successfully reuses existing components and trains only missing ones, maintaining effective communication after model updates.
RAG-NIMBF: A Non-Invasive, Modular Benchmarking Framework for Automated RAG Configuration Analysis
AI summaryRAG‑NIMBF is a non‑invasive benchmarking layer that evaluates RAG pipeline configurations via controlled ablations and end‑to‑end tracking, combining LLM answer scores, retrieval accuracy, and runtime. It shows MMR boosts nDCG@5 by 4–11pp and detailed prompts improve evidence fidelity by 10–13pp on large models.
Privacy-preserving inference on hosted models: a learned input-embedding transform for frozen Qwen3.8-27B
AI summaryA client-side transform encodes prompts into vectors, which are sent to a frozen 27B Qwen model; the host never sees text. On MMLU, the transform drops accuracy by 1.5 points, breaking 5.5% of questions while 30% of tokens become undecodable.
SMART: Semantic Segment Mapping and Assessment of Recognition Transcripts with Contextual Rubrics
AI summarySMART maps ASR errors into semantic segments and applies contextual rubrics to assign consequence labels, achieving 0.843 F1 against human labels and enabling the new Semantic Impact Error Rate metric with significant improvements over baseline corrections.
COMPASS: Routing via Capability-Oriented Model Profiling with Adaptive Search, Selection, and Tool Use
AI summaryCOMPASS is an agentic LLM router that uses tools and dynamic evaluation to select models without requiring training. It improves task accuracy while reducing model-call costs compared to existing specialized routing baselines and single-model approaches.
SAPD: Learning from Agent Experience without Environment Interaction
AI summarySAPD trains LLM agents from offline trajectories by decomposing multi‑turn logs into decision contexts and applying relative step‑advantage signals for on‑policy distillation. It avoids environment interaction, reduces catastrophic forgetting compared to SFT, and is 4.7× faster than online OPD.
Physical Operating-Condition-Guided Vision-Language Matching for Cavitation Diagnosis in Industrial Pipeline Systems
AI summaryThis paper introduces a physical operating-condition-guided vision-language matching method for industrial cavitation diagnosis, which aligns operating-variable-aware text descriptions with acoustic time-frequency representations. Experiments on three datasets demonstrate that this approach consistently outperforms existing diagnostic methods under varying conditions.
VISH-ASR: Viterbi-Guided Selection of Hotwords for Speech-LLM Contextual ASR
AI summaryVISH selects hotwords for contextual ASR by verifying candidate acoustic support directly over CTC posteriors with a Viterbi-based search. It achieves up to 11.7% absolute Best‑F1 gain and 75% faster search compared to learned‑retrieval or hypothesis filtering.
Hallucinations of Large Language Models in Cross-Language Code Translation: Phenomena and Mitigation
AI summaryThe paper surveys hallucinations in large language models for cross‑language code translation, categorizing them into Semantic Consistency, Factual Knowledge, and Project Context conflicts. It introduces HaluCodeTran, a retrieval‑guided refinement system that reduces hallucinations by ~30% in benchmarks.
Automatic Differentiation with Constraints in Forward and Inverse Problems
AI summaryFully differentiable graph-based solver for PDEs on unstructured grids that treats equations as constraints. Demonstrates faster forward modeling and effective inverse estimation of coefficients, RHS, and learnable graph coarsening, with significant complexity reductions for physics and engineering applications.
Physical-Variable-Informed Deep Learning for Robust Diagnosis in Industrial Fluid Transport Systems
AI summaryPhyNet integrates physical operating variables as global constraints with local sensor signals using cross‑attention to handle sampling heterogeneity, achieving state‑of‑the‑art fault‑diagnosis accuracy on four industrial fluid transport benchmarks, including real‑world cavitation data.
The Shape of the DPO Hyperparameter Landscape: What to Measure, How It Moves with Scale, and When Not to Transfer It
AI summaryDPO hyperparameters such as learning rate and β depend on model size and training data, and cannot be transferred between setups. Short intermediate runs and dynamic benchmarks accelerate tuning, but validation loss is a weak proxy for downstream performance.
CollabSciReM: Aligning Language Models for Scientific Conversations via Science-Aware Reward Decomposition
AI summaryCollabSciReM trains LLMs with a reward decomposed into conversational quality, scientific rigor, and efficiency, showing a 65.3% win rate on SciQuality and higher novelty with fewer hallucinations versus science‑agnostic rewards.
Staged Continual Adaptation of Multimodal Foundation Models for Japanese Financial Documents
AI summaryThe paper analyzes how staged post‑training (vision‑language alignment, reasoning distillation, domain tuning) affects an 8.4B Japanese financial‑document model. It shows the phases interact non‑additively; the best checkpoint depends on the downstream task.
Polarity Blindness in LLM Safety: A Systematic Study of Negation, Antonym Substitution, and Adversarial Framing
AI summaryThis paper reveals that state-of-the-art language models suffer from polarity blindness, failing to recognize harmful intent when prompts use grammatical negation or antonym substitution. Consequently, simple linguistic transformations consistently bypass safety guardrails across multiple models and languages.
MMEnterprise-Bench: Evaluating Enterprise Agents over Evolving Multimodal Records
AI summaryMMEnterprise‑Bench evaluates enterprise agents on evolving multimodal records, covering evidence reasoning, conflict resolution, and unanswerability. Models, even the best, still struggle with cross‑source and temporal conflicts, showing limited gains from higher retrieval budgets.
ACE-Compression: An Ability-Centric Evaluation of Compressed Large Language Models
AI summaryThe authors introduce ACE-Compression, an evaluation framework that assesses compressed large language models across four core ability dimensions rather than relying solely on overall task performance. They find that compression selectively degrades fine-grained and context-dependent abilities while preserving surface-level signals, with quantization generally outperforming pruning.
Quantum-JEPA: Quantum Born Priors for Latent-Action JEPA World Models
AI summaryQuantum-JEPA uses a conditional quantum Born prior to learn distributions over future latent actions, producing diverse, correlated future representations without an exponentially sized output layer. The method outperforms variational baselines on multimodal tasks and successfully runs on a real QPU with parameter‑shift training.
From Comparison to Selection: Enhancing Cross-Scale Generalization of Pointer Networks Trained with Reinforcement Learning for Capacitated Drone Delivery Routing
AI summaryThis paper investigates why reinforcement-learning-trained pointer networks struggle to generalize across different graph sizes in drone delivery routing. By introducing entropy-based attention scaling and improved graph-context aggregation, the authors develop an approach selector that consistently reduces routing costs across unseen problem sizes.
HypTI: Hyperbolic Geometry-Aware Multimodal Feature Fusion Framework for Dementia Detection Based on Tabular-Image Data
AI summaryHypTI fuses clinical tabular data and MRI scans in hyperbolic space using Lorentz prototypes and ROI‑guided multi‑scale encoding, achieving superior dementia detection on OASIS‑3 compared to Euclidean‑based methods.
APRIL: Ambiguity-Aware Expert Routing for Respiratory Sound Domain-Incremental Learning
AI summaryAPRIL trains a frozen BEATs backbone with low‑rank domain experts, then routes each test sample to the best expert using a prototype‑based router that balances domain‑routing confidence and classification certainty, achieving 64–83% accuracy with <1% forgetting on ICBHI and SPRSound.
Lost in Recovery: Demystifying Recovery Behavior of Computer-Use Agents via Goal-Oriented Action Replay
AI summaryReTrace is a framework that evaluates model‑based recovery in computer‑use agents by injecting controlled errors and replaying goal‑oriented actions, defining Transition‑Consistent Recovery. It shows ideal recovery unlocks hidden task completion, current agents underutilize this, and accurate recovery improves end‑to‑end performance.
When Do Humans and Machines Truly Complement Each Other? Formal Conditions and Empirical Evidence for Human-AI Synergy
AI summaryThis paper establishes formal mathematical conditions for human-AI synergy based on mutual error correction. The authors demonstrate empirically across clinical studies that strong algorithmic dominance can paradoxically inhibit synergy by eroding human corrective intervention.
Recast the Past: Stable Long Video Generation with Full Context Memory
AI summaryReCam introduces a memory mechanism for diffusion transformers that compresses historical video latents into vectors while preserving full-context access via native self-attention. It achieves state-of-the-art long video generation quality and superior consistency on VBench-Long.
Building Telecom Language Models with RSOD: Domain Data, Specialist Post-Training, Memory-Conscious Distillation, and Networking Evaluation
AI summaryThis paper introduces Routed Specialist On-Policy Distillation (RSOD), a framework that compresses multiple domain-specific telecom teacher models into a single deployable student adapter. The resulting model nearly matches the performance of the multi-teacher ensemble while remaining practical for standard deployment.
Learning Tactile-Aware Hand Representations from Human Hand-Object Interaction Data
AI summaryPretrains tactile‑aware hand embeddings from large human hand‑object interaction datasets, then fine‑tunes on limited robot data. The approach yields 88.26% Macro‑F1 on grasp prediction using only 25% of labeled data, matching full‑data training.
Dancing Diffusion: Learning Continuous Programs for Image Generation and Editing
AI summaryDancing Diffusion turns a single diffusion trajectory into a program of addressable operations by assigning each time span a dedicated condition block. It allows reusing executed prefixes and rewriting later steps, preserving more content and improving generation quality over native SANA.
On-Policy Transition Distillation
AI summaryOn-Policy Transition Distillation augments standard final-layer matching with alignment of layer-to-layer log-probability changes on student rollouts. Across multiple backbones and benchmarks, it recovers ~90% of the teacher–student capability gap, outperforming OPD.
Assessing Representation Semantics Under Language Pretraining in Fine-Grained Classification
AI summaryThis paper introduces an explainability-driven framework to compare vision encoders trained with classification versus caption supervision. The main result shows that caption supervision yields representations with better part localization and lower background reliance, despite similar downstream accuracy.
Slots Are Not Speakers: Training-Free Speaker Linking for Frozen Streaming Diarizers
AI summaryA causal, training‑free linker extends a fixed‑slot streaming diarizer to track more speakers by splitting slots at voice changes, embedding segments, and matching them to a speaker bank. It drops DER from 36.8% to 22.6% and cpWER from 82.7% to 24.4%.
Auditing Where Concept Bottleneck Accuracy Comes From on CUB
AI summaryThis paper audits Concept Bottleneck Models on the CUB dataset and reveals that their high accuracy often stems from shortcut features and information leakage in internal embeddings rather than genuine visual grounding. The authors show that standard task and concept metrics are insufficient, and CUB's noisy annotations preclude overinterpreting accuracy gaps.
TrajEvolve: Trajectory Evolution for Reinforcement Learning with Hindsight Credit Assignment
AI summaryTrajEvolve improves reinforcement learning for long-horizon LLM reasoning by refining historical trajectories into higher-quality evolutionary paths and using counterfactual verification for step-level credit assignment. This approach accelerates convergence and eliminates redundant step generation, outperforming standard RL baselines.
Quantifying the Threat: A Bayesian Probabilistic Framework for Heavy Metal Health Risks among Children in the Densu Basin, Ghana
AI summaryThe authors develop a Bayesian risk‐assessment framework for groundwater in Ghana’s Densu Basin, incorporating parameter uncertainty, co‑exposure, and spatial variability. They find Pb, As, and Cd drive cumulative risk, with ~31 % of sites exceeding a hazard index of 1, revealing highly heterogeneous, uncertainty‑aware risk surfaces.
Same Judgment, Different Yes: Measuring Multilingual Judgment Profiles in Language Models
AI summaryMultilingual judgments are assessed across 1.1M replies to 105 moral vignettes, varying question form, answer constraint, and token vs. sentence readouts. Findings show that profile consistency drops only in negative interrogatives; models retain consistent judgments across languages, but yes-token usage varies by language.
Slot-JEPA: Object-Structured Prediction Interfaces for Latent World Models
AI summarySlot-JEPA decomposes joint-embedding predictive architectures into object-structured semantic slots and spatial masks using optimal transport. This design improves feature transferability across visual benchmarks and enhances goal-conditioned planning in robotic environments.
PULSE: Personalized Opportunistic Learning for Wearable Devices
AI summaryPULSE selects useful peer encounters for wearable personalization, using gradient compatibility and past experience to decide when to share models. It matches or improves accuracy on activity, fall, and stress detection while cutting peer exchanges by 80‑94 % and fitting in <600 KiB on-device.
IECG: Weakly-Supervised Interpretability-Enhancing Foundation Model for Electrocardiogram
Electrocardiography (ECG) is a cornerstone of cardiac diagnosis, yet most machine learning methods for ECG analysis lack interpretability, limiting their clinical adoption. We present IECG, a weakly-supervised ECG foundation model that jointly performs diagnostic classification…
InsightRAG: Aggregating Distributed Scholarly Judgments for Scientific Survey Generation
AI summaryInsightRAG uses Related Work sections to follow author‑supported citation links, reconstructing verified research scopes and building conflict‑minimized taxonomies from aggregated judgments. It improves scope coverage, taxonomy quality, citation support, and reference reliability across ten topics.
MoDDiT: Simple Token Routing in Diffusion Transformers
AI summaryMoDDiT learns to route image tokens between MLP and self‑attention, enabling dynamic compute allocation that processes only about 12.5 % of tokens. This reduces training and inference cost while improving generation quality over dense Diffusion Transformers.
From Complex Histories to Simple Predictors: Predictive Dimension in Dynamic Graph Learning
AI summaryThe paper shows that dynamic graph learning can focus on a low‑dimensional predictive state manifold instead of full histories, defines an intrinsic predictive dimension metric, and builds PreDyG, a tiny model that matches this dimension and outperforms heavy baselines with 0.07 % of their parameters.
BAIT: Can LLMs Find What Biases a Reader? A Benchmark for Cognitive Bias Triggers in News
AI summaryThe paper introduces BAIT, a benchmark for detecting and localizing cognitive bias triggers in news articles, and evaluates six frontier LLMs on these tasks. The results show that while LLMs can identify overall triggers at near-human levels, they struggle to locate the specific passages, often relying on surface cues rather than manipulative mechanisms.
ClawMMU: Protection-Coherent Virtual Memory for Stateful LLM Agents
AI summaryClawMMU adds a protection layer to LLM agent memory, tagging each chunk with owner, scope, origin, and authority. It enforces that authority can only increase, not self‑lift, preventing malicious content from being treated as user instruction.
Polyscene: Scenario Planning as Diversity Planning
AI summaryPolyscene automatically generates diverse, internally consistent scenarios by framing scenario planning as a diversity planning problem, producing distinct future narratives within a given budget, and demonstrates its effectiveness on geopolitical conflict and shipping-route disruption case studies.
A Semi-Analytical Inference Framework for Structural Causal Models
AI summaryThis paper introduces a message‑passing inference framework that works directly on structural causal models by representing mechanisms and densities in a shared finite‑dimensional function space using B‑spline bases. It achieves accurate, systematic approximations to observational, interventional, and counterfactual queries via numerical quadrature and linear algebra.
Robust Representation Learning for Spiking Neural Networks based on Conditional Information Bottleneck
AI summarySCIB trains SNNs with a conditional information bottleneck that compresses input information conditioned on labels, yielding class-structured hidden representations. It outperforms standard SNNs and IB baselines in robustness to noise and adversarial attacks, while cutting inference energy by ~25%.
Interruption-Aware Recurrent Reinforcement Learning for UAS Navigation Under Temporary Overrides
AI summaryThis paper investigates how temporary external overrides degrade recurrent reinforcement learning policies for drone navigation and proposes interruption-aware architectures. The best-performing proposed architecture reduces post-interruption collisions by 74 percent without sacrificing standard performance.
De-Diffusing WAMs: Action-Sufficient Direct Prediction for Robot Control
AI summaryDirectWAM trains a single-step action predictor by distilling a pretrained diffusion model, evaluating futures by their control impact rather than representation matching. It achieves 97.1% success on LIBERO suites, 85.4% on RoboTwin 2.0, and cuts inference time 28×.
Causal Discovery for Linear Models with Latent Confounding via Multi-domain Identifiable Representation Learning
AI summaryLinADMG learns linear ADMGs from multi‑domain data by identifying noise variables via representation learning, avoiding non‑Gaussian or bow‑free assumptions. With a no‑cancellation condition on latent confounding, it recovers all direct causal edges and the entire ADMG.
From Code Generator to Engineering Companion: A Plant-aware Agent for PLC Programming
Programmable Logic Controller (PLC) coding remains a manual, IDE-centric task that demands constant context switching between documentation, project files, and code editors. Recent work has demonstrated that retrieval-augmented large language models can generate vendor-compliant…
Africa-Papers: A 10-Billion-Token Multi-Domain Academic Corpus for African Foundation Language Models
AI summaryThe authors introduce Africa-Papers, a 10.15-billion-token multi-domain corpus of African academic literature designed to address the severe underrepresentation of African scholarship in pre-training data. The resulting dataset significantly improves the foundational training and domain-specific reasoning capabilities of language models for African contexts.
Temporal Evidence-Parameter Co-Alignment for LLM Reasoning over Temporal Knowledge Graphs
AI summaryTEPCA aligns LLM reasoning with temporal knowledge graphs by pruning evidence paths with time‑constrained DRGP and MI‑based pruning, and adapting parameters with timestamp‑conditioned DG‑LoRA. The method achieves state‑of‑the‑art performance on TKG benchmarks and improves commonsense QA.
Perceives but Doesn't Brake: A Perception-Action Binding Gap in Reasoning Driving VLAs
Reasoning-augmented vision-language-action (VLA) driving models can perceive and reason about safety-critical hazards yet fail to translate that evidence into safe actions. We characterize this failure as a perception-action binding gap: hazard information is available inside…
Offline Makes Online Better: Calibrated Cache Compression for Video Generation
AI summaryThe paper proposes CALM, a method that calibrates token merging ratios across layers and heads for autoregressive video generation. By offline profiling the risk map and using dynamic programming online, CALM reduces memory usage while improving video quality, achieving a 21% LPIPS reduction compared to uniform merging.
Adaptive Pullback: Reweighting Supervision at Penalty Jumps in On-Policy Distillation
AI summaryAdaptive Pullback OPD reweights teacher‑student supervision at sudden penalty jumps by scaling the following token block so its average penalty matches the preceding block. This reduces penalty‑confidence tension, improving results on math reasoning and code generation.
Reframing Object-Centric Reinforcement Learning: From Absolute to Relative Representations
AI summaryThe paper replaces absolute object coordinates with anchor-relative coordinates in object-centric RL and proposes ORBiT, a transformer architecture that uses these relative inputs. ORBiT consistently outperforms baselines, achieving IQM >1.0 on 13 Atari games and mastering Foosball and Tag from scratch.
Parameter-Efficient Adaptation of Time Series Foundation Models for Electrical Load Forecasting
AI summaryThis paper evaluates parameter-efficient adaptation techniques for time series foundation models in electrical load forecasting. The authors demonstrate that low-rank adaptation reduces error by 7.5% while updating only one percent of the model parameters, offering a strong accuracy-efficiency trade-off.
MaD-RSI: Model-as-Data for Recursive Self-Improvement of Agent Harnesses
AI summaryThis paper introduces MaD-RSI, a recursive self-improvement framework that treats different LLMs as data to optimize a shared agent harness. Experiments show that this approach improves cross-model generalization while reducing training costs compared to existing methods.
JAXbounce: A Continuous-Control Pong-Like Benchmark for Active Inference Beyond Low Dimensional Environments
AI summaryJAXbounce is a continuous‑control pong‑like benchmark that tests active inference in high‑dimensional, stochastic settings. Experiments show that state‑based agents with Kalman filtering and expected free energy planning outperform deterministic baselines, scaling with planning horizon, observation type, and noise knowledge.
Region-Guided Food-Package Evidence Extraction and Abstention-Aware Draft R3337 Screening with Selective VLM Review
AI summaryThis paper introduces a region-guided workflow combining OCR, deterministic rules, and selective VLM review to screen food-package images against South Africa's draft R3337 warning criteria. The method achieves 65.0% coverage and 97.4% selective accuracy on a validation set.
Local Inference of LLMs on Long Contexts: A Data Systems Perspective
Large language models can solve complex problems across a variety of subjects, such as mathematics and coding. For such complex analytical tasks, models with a long context perform better. Serving models with long context is currently mostly done on powerful systems with…
TPO: Tri-level Distributionally Robust Learning for OOD Direct Preference Optimization
AI summaryThe paper introduces TPO, a tri-level distributionally robust framework that addresses hierarchical out-of-distribution preference shifts in Direct Preference Optimization. Results show that TPO provides theoretical convergence guarantees and outperforms existing robust DPO methods.
PI-MLLM: Psychophysics-Inspired Multimodal Large Language Model for Oddly Satisfying Video Emotion Analysis
AI summaryPI-MLLM adds psychophysics-inspired alignment of visual motion and acoustic onset dynamics to a multimodal LLM, using MSDA and CAVA to predict viewer emotions in oddly satisfying videos and explain modality contributions. It outperforms prior methods on OSVD’s 12k clips.
VGL-stGen: Learning a shared multimodal model for lung histology and spatial transcriptomics
AI summaryVGL-stGen trains a multimodal model on paired lung H&E images and spatial transcriptomics, learning joint image-expression and expression‑cell‑type mappings. It accurately performs thirteen tasks—classification, gene prediction, and conditional expression completion—by autoregressively predicting gene levels from histology and cell‑type cues.
GALA: Geometry-Aware Language Model for Controllable Object Arrangement
Reasoning about fine-grained placement constraints, such as support and containment, is essential for realistic 3D scene creation. It remains a significant challenge for existing methods that rely on simplified object representations like oriented bounding box (OBB), which are…
ChartEvo: Self-Evolving Agent Harnesses for Long-Tail Chart Reasoning
AI summaryChartEvo is a recursive self-improvement framework that enables multimodal agents to automatically evolve specialized skills and tools for long-tail chart reasoning. It improves the base model's accuracy by over 32% on challenging benchmark datasets.
Local Work, Global Progress: A Convergence Theory of AI Agent Collaboration
Teams of large language model (LLM) agents are increasingly applied to complex tasks, but their collaboration protocols are mostly designed empirically, without convergence guarantees. We formalize collaborative problem solving as block optimization: each agent improves one part…
ABACUS: A Benchmark for Image Editing Models to Assess Object Category and Counting Under Scene Conservation
AI summaryABACUS benchmarks image‑editing models for adding, removing, or changing a specified number of objects while preserving the rest of the scene, using a judge‑free, instance‑level check against panoptic ground truth. It shows that current models fail increasingly with higher counts and complex scenes.
Is Sycophancy a Form of In-Context Updating?
AI summaryThe paper redefines factual sycophancy as a failure in in-context updating, where a language model overweights user assertions over evidence. It shows that deference to users and evidence-based updating share causal components, implying that reducing sycophancy may harm useful updating.
Identifying Dependent Components from Multi-domain Linear Mixtures
AI summaryThis paper introduces a multi-domain linear mixing model with dependent latent components that uses domain-specific scalings to achieve identifiability from second-order statistics alone. The authors propose the Multi-Domain Covariance Matching algorithm, successfully recovering latent variables and mixing functions without relying on independence assumptions or parametric dependencies.
Beyond a Single Sentiment Score: Topical and Affective Dimensions of Emoji Meaning on Twitter
AI summaryThis paper demonstrates that emojis used in topically focused contexts exhibit more polarized crowd-sourced sentiment rather than neutral ones. The authors show that topical coherence and affective dispersion are distinct, measurable dimensions of emoji meaning across multiple datasets and languages.
Gated Prototype Transfer for Zero-Support Medical Vision Language Model
AI summaryThis paper introduces GProtoT, a training-free module that transfers visual residuals from supported classes to zero-support classes using clinical ontologies and an evidence gate. It improves missing-class recall and overall accuracy in few-shot medical vision-language adaptation while minimizing negative transfer.
RayCo4D: Ray-Conditioned Continuous Motion Model for Monocular 4D Dynamic Reconstruction
AI summaryRayCo4D introduces a ray‑conditioned continuous motion model that decouples visual data from latent 3D motion, using shared SE(3) motion nodes and white‑noise‑on‑acceleration priors to interpolate motion across fragmented tracks. The method yields more consistent 4D reconstructions, higher‑quality novel views, and continuous‑time rendering.
S-Agent 2.0: Spatial Reasoning with a General Coding Agent
AI summaryS-Agent 2.0 equips a general coding agent with a 3‑D perception toolkit and a persistent scene memory, letting the VLM acquire, store, and interpret spatial evidence in plain language. It boosts Qwen 3.8‑27B by +13.9/+20.5 points on MMSI‑Bench/ReVSI and Gemini‑3.8‑Flash by +11.2/+9.6, achieving state‑of‑the‑art scores.
CountVideoBench: A Benchmark to Evaluate Prediction Biases on the Object Counting task by Video-Language Models
AI summaryCountVideoBench evaluates video‑language models on object counting by presenting 2,600 synthetic videos with balanced counts from 5 to 50; none of the 16 models studied achieves over 27.6% exact match, and accuracy patterns vary, revealing prediction biases rather than pure difficulty.
FALCA: Falsification-Aware Reinforcement Learning with Hierarchical Credit Assignment for Long-Horizon Deep Search Agents
AI summaryFALCA introduces a collaborative Searcher-Falsifier framework for long-horizon LLM search agents that actively seeks counterevidence and uses hierarchical credit assignment for reinforcement learning. Experiments across five benchmarks show it reduces error accumulation and improves search performance.
Context-Aware Prompt Injection Detection via Small Language Models in RAG Pipelines: nokast-secureRAG
Local Retrieval-Augmented Generation (RAG) systems are vulnerable to indirect prompt injection and retrieval poisoning, where malicious instructions in retrieved content are mistaken for trusted input. We present nokast-secureRAG, a roughly 0.5B-parameter small language model…
Don't Skip the Baseline: A Controlled Comparison of Dynamic Expert Skipping in MoEs
AI summaryThis paper compares dynamic expert-skipping methods in MoEs against a static reduction of active experts. It finds that the simple static baseline is consistently difficult to outperform across multiple models and modalities.
RACE: RoPE-Aligned Attention Compression Based on Query-Key Activation Energies
AI summaryRACE compresses transformer attention by selecting a subset of rotary pairs based on joint query‑key activation energy, enabling reduced KV‑cache size and faster inference without runtime reconstruction. It achieves up to 47.8% fewer FLOPs, 1.79× higher throughput, and 28% lower GPU memory usage at 50% compression.
VisualEvolve: A Skill Self-Evolution Framework for Visual Reasoning
AI summaryThis paper introduces VisualEvolve, a self-evolution framework that iteratively refines external visual operations and tool-use strategies for vision-language models without updating model parameters. Across diverse benchmarks, VisualEvolve consistently improves model performance while requiring significantly fewer trajectories than standard reinforcement learning methods.
MAS-Guard: Unified Guardrails from Individual Agents to Multi-Agent LLM Systems
Agentic AI extends large language models (LLMs) from generating responses to planning and executing complex workflows, substantially expanding their utility while exposing applications to unsafe actions, information leakage, and adversarial manipulation. Multi-agent systems…
DIVE: Benchmarking Full-Duplex Interaction with Visual Engagement in Continuous Streams
AI summaryDIVE is a new benchmark for full‑duplex real‑time multimodal interaction, offering six tasks that test proactive intervention, continuous engagement, and coordination in visual streams. Leading models respond in sub‑second latency but achieve only ~54% window response rates, highlighting difficulty in adapting to new information.
The Instruction Tax on Factual Recall: Base Models Can Beat Their Instruction-Tuned Counterparts
AI summaryThe paper shows base language models can outperform instruction-tuned versions on factual recall tasks, achieving 0.6836 vs 0.5681 on the AKBC LM‑KBC 2026 benchmark. It attributes this “instruction tax” to model weights, not prompting, and improves recall by sampling and filtering repeated outputs.
Rewards for Budget-Aware Multi-Turn Agents That Adapt to Arbitrary Test-Time Constraints
AI summaryThe paper trains multi‑turn language model agents to learn budget allocation across tokens, API calls, and turns, using new reward penalties for overspending. Agents achieve higher within‑budget success and natural compliance, generalizing to unseen budgets and tasks.
Brain2Vec: A Self-Supervised 3D Foundation Model for Structural Brain MRI
AI summaryBrain2Vec pre‑trains a self‑supervised 3D neural network on 74,425 T1‑MRI scans, yielding embeddings that improve brain‑age prediction accuracy and speed convergence versus random init, and outperform volumetric features on most downstream tasks.
Code2Video-Compass: A Comprehensive Benchmark for General-Purpose Code-Based Video Generation across T2V, I2V, and V2V
AI summaryThe paper introduces Code2Video-Compass, a benchmark for evaluating code-based video generation across text-to-video, image-to-video, and video-to-video tasks using a dual-level evaluation framework. Experiments with 15 multimodal models reveal that even leading models struggle to generate complex, high-quality videos reliably.
DISTRACT: Degradation-Indexed Study of Textual Reference Across Competing Targets
AI summaryDISTRACT evaluates how textual descriptions influence vision‑language trackers under visual degradation. The study finds that only Molmo2 and JointNLT use text to guide tracking, with decodable effects at early decision stages, while other trackers show negligible or non‑significant text influence.
WorldSR: Harnessing World Knowledge Search for Grounded Image Super-Resolution
AI summaryWorldSR grounds image super‑resolution on external visual knowledge by decomposing images into entities, retrieving aligned reference images, and using them in a multi‑reference SR model. It achieves higher structural fidelity and visual quality than prior generative‑prior methods.
After the Attacker Leaves: Benchmarking Carryover Attacks on LLM Agents
Modern LLM agents persist state across sessions via two channels: long-lived memory files silently prepended to the system prompt, and scheduled tasks waking the agent without user input. A one-time indirect prompt injection committed to either channel can re-fire across…
Racer: Hunting GPU Kernel Bugs with Symbolic Equivalence Checking
GPU kernels are critical to the performance of today's AI services, yet the same operators are rewritten repeatedly: for each GPU generation, numeric format, serving regime, and kernel language. Every rewrite must preserve the function computed by the implementation it replaces.…
A Digital Twin to Improve Cyber-Social Impact
The increasing complexity of Industrial Control Systems (ICS) and their integration with IT infrastructures has significantly expanded the cyber threat landscape. Traditional cybersecurity models, primarily based on the Confidentiality, Integrity, and Availability (CIA triad),…
Plausible Future, Wrong Choice: Steering World-Action Models toward Intended Future
World-Action Models (WAMs) have emerged as a promising paradigm for embodied generalist policies by jointly modeling future world evolution and robot actions across diverse manipulation tasks. However, representing diverse futures is only the first step: a capable WAM must also…
Every Model Leaves a Trace: Multi-Resolution LLM Fingerprinting
AI summaryThe paper shows that large language models embed persistent, model‑specific fingerprints that can be recovered even after surface cues are removed, across families, distilled variants, and LoRA adaptations. Semantic signals are the most robust and transferable across prompts and languages.
Stratified Sparse Attention: Reducing the Sparse-to-Dense Gap at Extreme Sparsity
AI summaryStratified Sparse Attention blends deterministic top‑k retrieval with score‑guided sampling that corrects for inclusion probabilities, recovering most of the missing attention mass. It outperforms all baselines at extreme sparsity, achieving 94% dense‑attention accuracy and up to 3× decode speedup.
Toolflation: Seller Agents Distort Tool Markets Through Spontaneous Generative Engine Optimization
AI summarySeller agents optimize their listings to maximize selection, inflating claims beyond their tools’ actual capabilities. This “toolflation” distorts buyer choice, reduces high‑quality tool adoption, and can drop task success from ~90% to ~30% in worst‑case markets.
Beyond Trajectory Statistics: Community-Consistent Human Mobility Generation via Evolving User-Location Graphs
AI summaryMOTION generates synthetic human mobility by learning dynamic user‑location graphs that preserve community structure. It achieves at least 4.05% improvement over baselines in trajectory, population, flow, and graph metrics while maintaining privacy.
Claims of emergence in AI should be grounded in information decomposition
AI summaryClaims of AI emergence are formally defined using information decomposition: a system is emergent only if it contains synergistic information about outcomes, separating unique, redundant, and synergistic contributions. The framework is applied to attention heads, agents, and human‑AI teams, linking and extending prior observations.
Exploring Temporal Context-Aware Physical Grounding in Latent World Models
AI summaryAdd a temporal aggregation module to JEPA, supervising latent states with physical states, and use a prior‑guided action distribution conditioned on current and goal latents to steer planning; achieves better manipulation planning, especially over long temporal gaps.
Proof-Carrying Contracts: Commitments AI Agents Cannot Break
AI summaryProof‑Carrying Contracts let AI agents supply machine‑checkable proofs that every operation they execute preserves user‑specified constraints. The system verifies these proofs before permitting action, guaranteeing constraint enforcement even against adversarial agents.
Model- and Layer-Level Closed-Loop Forecasting with SSM-Based Spiking Neural Networks
AI summaryClosed‑loop forecasting in spiking neural networks is achieved by feeding the network’s own predictions back as past inputs, at both model‑ and layer‑levels. On ETT‑small, these methods match or outperform state‑of‑the‑art continuous‑activation models, showing competitive performance with low‑bit activations.
BitMTP: When Multi-Token Prediction Meets Low-Bit Large Language Models
AI summaryBitMTP jointly trains ultra‑low‑bit LLMs with an MTP objective, enabling confidence‑only, speculative‑only, and hybrid decoding without separate draft models. It achieves 2–3× speedups on 1.5–4 B models across NVIDIA H20, RTX 3090, and Jetson Orin while maintaining task quality.
ECG for laboratory values:A comprehensive physiological and technical analysis
AI summaryThe study trains multimodal deep learning models to predict NT‑proBNP from ECGs and metadata, comparing 1‑lead to 12‑lead setups. The 12‑lead model achieves MAE≈514 pg/mL and AUROC≈0.876 for binary classification, showing ECG can classify broad NT‑proBNP ranges but precise regression is still difficult.
DecisionPivot: A Diagnostic Benchmark for Decision Models in LLM Systems
AI summaryDecisionPivot offers a 48‑task, 40k‑item benchmark for evaluating decision models in LLMs, measuring accuracy, confidence calibration, and probability ordering. It shows that small score gaps can hide large differences in contrast‑group accuracy and confidence reliability.
VLM-Guided Contact Retargeting and Generative Partner Modeling for Humanoid-Human Interaction
Humanoid-human interaction requires robots to operate under severe partial observability: during close physical contact, the human partner is occluded, truncated, or outside the robot's field of view, causing pose estimation to fail and often necessitating external tracking or…
Beyond 4 bits: Data-Centric Post-training Quantization and Recovery
AI summaryThe paper shows that 2‑bit PTQ fails unless vector codebooks are used, with calibration and recovery data chosen jointly. Using a small task‑aligned calibration set and codebook‑only fine‑tuning, 89–97 % QA accuracy is kept with ~1.4–30× less compute than QAT.
Humans Correct Their Judgments Through Debate, but Weaker AI Models Do Not
AI summaryDebate between AI agents boosts human judges’ accuracy over prior beliefs, but provides no comparable benefit—and sometimes harms—accuracy for weaker language-model judges. Weighted voting best aids humans, while multi-judge setups do not improve weaker AI supervision.
Binary Discrete Diffusion: Candidate-Set Diffusion for Logic and Planning Tasks
AI summaryBinary Discrete Diffusion lets a diffusion model maintain candidate token sets as binary vectors, gradually adding/removing options to represent partial uncertainty. It outperforms masked and uniform baselines on hard planning tasks (Sudoku-Extreme, Boxoban, graph coloring) and remains competitive on language modeling.
Learning Computational Structure for Sample-Efficient Prediction
AI summaryLearned Wiring learns a low-dimensional embedding per neuron, connecting only nearby neurons while jointly training weights and connectivity, achieving 27 % lower sample‑curve area than dense nets on 32 synthetic compositional targets. It also outperforms dense baselines on 16/17 real-world model‑prediction tasks.
When Do LLMs Help with Feature Selection? A Systematic Empirical Study
AI summaryLLMs can use semantic feature names and summary stats for tabular feature selection, but a systematic study on 18 datasets shows classical baselines outperform 7B LLM selectors on ROC‑AUC and R², though LLMs give more compact subsets and modest gains when feature names are informative.
Large Flow Language Models
AI summaryLafa scales continuous flow language models to 8 B parameters, training with <30 B tokens and achieving instruction, reasoning, and coding quality comparable to autoregressive baselines while enabling 2.54× faster decoding through blockwise parallel generation and self‑speculative verification.
Agentic Ensemble with Vision-Language Models for Visual Anomaly Detection
AI summaryAnomalyAgent fuses pretrained detectors by aligning heterogeneous scores, summarizing disagreement, and selecting fusion actions with a vision‑language model to resolve conflicts. It outperforms equal‑weight fusion, achieving state‑of‑the‑art image AUROC and pixel AP on MVTec AD and VisA.
OptScientist: Automated Discovery of Optimization Laws
AI summaryOptScientist automatically infers symbolic optimization problems from data, using an LLM‑guided discovery pipeline. It recovers ground‑truth physics laws on 80 %+ of a 99‑problem benchmark and outperforms prior methods on real‑world tasks while using far fewer parameters and tokens.
Evolving Semantic Indexes for Generative Retrieval
AI summaryThe paper uses LLM-driven evolutionary search to mutate generative retrieval components, guided by a fast 1M‑parameter proxy decoder. Across 21 SID programs on NQ320K, evolution boosts Recall@1 from 55.8% to 59.3% and Recall@100 from 85.8% to 90.6%.
Sage-A1.5: From Long-Horizon Execution to Scientific Exploration with Capability Composition
AI summarySage‑A1.5 is a 35B Mixture‑of‑Experts agent that adds adaptive capability composition to A1, enabling task delegation, parallel subagent collaboration, and verification during long‑horizon execution. It also performs open scientific discovery, producing inspectable research artifacts.
Spatially Informed and Direction Sensitive Representation Learning for Spatial Transcriptomics
Cellular states are shaped by both intrinsic gene expression programs and the spatial microenvironment in which cells reside. Spatial transcriptomics makes this context measurable, but effective representation learning requires capturing both spatially informative molecular…
Beyond the Sequentiality Bottleneck: A Hybrid Diffusion-Autoregressive Framework for Efficient Multimodal Reasoning
AI summaryCoRAL combines a diffusion planner that produces multiple reasoning paths in parallel with a large autoregressive executor that picks the best plan and generates the answer. Reward‑guided denoising steers the planner, yielding faster, more accurate multimodal reasoning than prior diffusion or autoregressive models.
SceneRig: An Agentic System for Simulation-Ready 3D Scene Reconstruction from Single Images
Simulation is a powerful tool for robotics, but building diverse, simulation-ready scenes at scale remains difficult. Reconstructing scenes from single images is a promising alternative, but visual similarity alone is insufficient: objects must also be recovered, accurately…
Capable Teacher, Myopic Scorer: Reward-Free Teacher Repair for Agentic On-Policy Distillation
AI summaryR2S repairs the teacher’s next-token scores by using teacher takeovers on high-disagreement states to fine-tune the scorer, then freezes it for OPD. It outperforms three state‑of‑the‑art methods, improving success by 12.8 points on average across all tested cells.
What Do RTL Repair Benchmarks Actually Measure?
AI summaryThe paper presents RTLRepairBench, a dataset of 5,784 broken‑fixed RTL pairs and 459 test cases, plus a simulation/formal evaluation harness, and shows that standard repair scores conflate specification usage with reset/first‑cycle semantics, recommending new evaluation controls.
Evolving Agent Interfaces
AI summaryAgent S4 autonomously generates and refines computer‑use interfaces by analyzing task trajectories and evolving interface repositories; on OSWorld V2 it boosts a Qwen3.8‑27B agent’s binary score to 21.3% (3.8× over fixed interfaces) and raises ScienceBoard success from 41.4% to 58.6%.
Turning TS2Vec into a Foundation Model for Time Series Classification
AI summaryTS2Vec is adapted into a foundation model by training a single 1.6‑M‑parameter dilated CNN once on synthetic data, using dual stems for amplitude and shape views, and distilling them into one encoder. It attains 83.3% on UCR‑128 and 75.3% on UEA‑27, matching state‑of‑the‑art MantisV2 within 0.3 pp.
Value-Rich, Differently Organized: Measuring Value Expression and Organization in AI-Generated Debates
AI summaryThe paper introduces a metric framework to analyze how values are expressed and organized in AI-generated debates, based on Schwartz’s Basic-10 values. It finds that while AI debates cover many values, they diffuse them across turns and lack the focused organization of human debates.
Benchmarking Arabic Dialects for Math Problems
AI summaryDAR-MGSM is a manually curated benchmark of multi‑step math word problems in five Arabic dialects and MSA, with both translated‑only and culturally localized versions. Models score lower on localized items; Arabic‑centric models differ in sensitivity, showing varied reasoning paths.
Semantic Observability for Autonomous Systems Diagnostics
AI summarySemantic Observability unifies heterogeneous logs into a property‑graph database and uses multi‑agent DCG orchestrated causal analysis to automatically isolate anomaly blast radius, generating LLM‑driven code‑level root‑cause reports and reducing MTTR while cutting token usage.
Adaptive Multi-Token Prediction: How far should we go?
AI summaryAMTP learns a context‑dependent draft horizon online, predicting how far a multi‑token draft will survive and selecting the horizon that maximizes tokens per second. It speeds up decoding up to 4.1× and matches or exceeds fixed‑horizon performance in most tested settings.
MindView: Benchmarking Generative Multi-View Spatial Intelligence
AI summaryMindView benchmarks language‑guided generation of spatially consistent views from partial multi‑view observations, exposing large gaps in current models; fine‑tuning BAGEL improves scores by 24.3 points, and the training‑free MindCue selector mitigates performance drops in expanded context.
Industrial ReMEmbR: Scaling Spatiotemporal Memory to Robot Fleets
Warehouse robot fleets continuously observe operational conditions, but preserving these observations as reusable semantic memory incurs substantial computation, communication, and storage costs. For example, explaining recurring aisle blockages can require reviewing video from…
UniWALM: Unified World Action Language Model for Embodied Robot Control
Motivated by the human ability to integrate experience across language, perception, and action, we introduce UniWALM, a unified world-action-language model that learns video, action, and language in a single shared representation, where each modality is both a conditioning…
GBClaw: An agentic AI framework for atomistic grain boundary research and mechanistic discovery
Agentic artificial intelligence (AI) systems hold immense potential for automating complex workflows in materials discovery and design, but extending them to domain-intensive mechanistic investigations remains challenging. Here, we present GBClaw, an agentic AI framework for…
Joints, Camera, Action! Co-Generating 3D Motion and Video for Arbitrary Rigs
AI summaryThis paper introduces a model that jointly generates synchronized video and 3D motion for arbitrary skeletons using a decoupled architecture. The approach outperforms video-first pipelines in faithfully animating diverse rigs and placing them into consistent 3D scenes.