New Artificial Intelligence (cs.AI) papers, Tuesday 20 October 2026
486 papers announced on arXiv. Get them ranked for you, free →
Transfer Learning for Power Outage Detection Task with Limited Training Data
Early detection of power outages is crucial for maintaining a reliable power distribution system. This research investigates the use of transfer learning and language models in detecting outages with limited labeled data. By leveraging pretraining and transfer learning, models…
Multi-LLM Systems Exhibit Robust Semantic Collapse
AI summaryMulti‑LLM systems converge to the same semantic representation over many generations without retraining, a phenomenon called semantic collapse. Despite varied decoding, prompting, agent design, and RL interventions, the collapse persists across model families and long runs.
Transformer-based segmentation of prosodic boundaries in Brazilian Portuguese
AI summarySAMPA fine‑tunes Whisper on BP speech to insert prosodic boundary markers, achieving F1 = 0.731 on in‑distribution data and 0.796 out‑of‑distribution. The model leverages morphosyntactic, semantic, and acoustic cues for boundary detection.
Dependency, Compression, and Synergy: A Unified Information-Theoretic View of Multimodal Learning
Recent advances in multimodal foundation models have intensified the need to understand how different modalities share, preserve, and complement information. Mutual Information (MI), the Information Bottleneck (IB), and Partial Information Decomposition (PID) provide…
Can the Quantum Fisher Effective Dimension Judge a Hybrid Quantum Neural Network?
Hybrid quantum neural networks (QNNs) pair a classical feature extractor with a variational quantum circuit, and the effective dimension of the circuit's quantum Fisher information matrix (QFIM) is widely used to assess such models. This scalar is often interpreted as an…
LLM-Assisted Bi-Approximate CHC Solving
Constrained Horn clause (CHC) solvers often struggle to discover deep counterexamples whose derivations may require extensive rule unfolding. Although large language models (LLMs) can propose long-range symbolic relations, prior work on formally checked LLM-assisted verification…
Web Agent Agentic Reinforcement Learning Decision Model Under Multi-Cost and Failure Risk
AI summaryProposes a constrained agentic RL model that unifies page access, search, and API calls with cost and CVaR risk constraints, reducing cost‑per‑success by 22‑31% and failures by 18‑26% while boosting completion rates under budget limits.
GardenBench: A lightweight, daily evaluation of LLM capabilities
AI summaryGardenBench offers a lightweight, daily LLM benchmark with 11 test suites that track daily accuracy, cost, and latency for consumer-facing tasks, enabling public comparison of model versions and providers through a live dashboard.
Composable Parametric Demographic Injection for Values Survey Simulation
AI summaryPARADE injects demographic shifts into LLMs via reusable LoRA modules for each attribute, separating task knowledge from demographic effects. In simulations of World Values Survey and SocioBench, it reduces population KL divergence by up to 52.8% while preserving predictive accuracy.
Context-Aware RGB-Thermal-Radar Obstacle Perception for Resource-Constrained ASVs
Reliable obstacle perception is challenging for small autonomous surface vessels (ASVs): reflections and glare confound visible-light cameras, vessel motion shifts the apparent waterline, and darkness removes the dominant sensing modality altogether. Rather than treating…
Latent Multimodal Memory Reasoning
AI summaryThe paper introduces Latent Multimodal Memory Reasoning (LMMR), a neuro-symbolic framework that maps heterogeneous visual and textual memories into a shared sparse concept space to construct a dynamic memory graph. Experiments show that LMMR significantly outperforms existing methods in long-term multimodal reasoning and response grounding.
EMMA: Event-Driven Multimodal Memory System with Multi-Agent Retrieval
AI summaryThe paper introduces EMMA, an event-driven multimodal memory system that segments long videos by semantic events and uses multi-agent retrieval across a hierarchical memory structure. This approach improves long video understanding over standard fixed-length clip compression methods.
GeoForward: Geometry-Guided Cross-View Alignment for Feedforward 3D Gaussian Splatting
AI summaryGeoForward uses calibrated cameras and sparse LiDAR to anchor cross‑view correspondence with explicit geometry, then refines with visual features and predicts 3D Gaussian splatting. On nuScenes it achieves lower depth‑consistency error and Chamfer distance than competing methods while maintaining quality.
EgoVine: Post-Training with Branching Experience Compiled from Real-World Videos
AI summaryEgoVine turns fixed egocentric video demos into interactive experiences by compiling them into symbolic environments, enabling agents to recover from mistakes and compare alternative actions. The method jointly trains action generation and consequence prediction, achieving 81.8% success on main tasks and 66.7% on longer compositions.
When the Editing Intent Is Split: A Cross-Modal Jailbreak Attack for Large Image Editing Models
AI summaryLarge image editors can be fooled by split-intent attacks that give a safe text prompt while hiding a harmful visual cue; the proposed SIJA attack succeeds on GPT Image 2 (68%) and Qwen-Image-3.0-Pro (91%) and is mitigated by joint intent reconstruction.
UMI-WM: Interactive World Modeling for Bimanual Manipulation from Wrist Cameras
Universal Manipulation Interface (UMI) pairs wrist-camera videos with camera trajectories and gripper openings, providing scalable, robot-free data for learning how bimanual actions change a shared scene. Existing interactive world models, however, typically rely on text or…
SPIRIT: Sparse Physics-Informed Inverse Rendering for SAR Imaging
AI summarySPIRIT reconstructs 3‑D surfaces from sparse‑aperture SAR by physics‑informed inverse rendering, modeling visibility, phase coherence, and wave propagation. It recovers accurate geometry and signal with over 90% fewer measurements, outperforming conventional super‑resolution baselines.
Cross-Modal Forgetting and Transfer in Unified Multimodal Post-Training
AI summaryCross‑modal fine‑tuning of a unified model (Janus‑Pro‑7B) can cause forgetting in untouched modalities. A single LoRA run preserves other capabilities, but chaining three runs collapses them; anchoring the untrained head prevents this damage.
Doing Less with More: A First-Principles Exploration of the Suitability of Agentic Computing over Alternative Architectural Choices
AI summaryThis paper evaluates agentic AI against traditional computing architectures on an arithmetic task, finding that while agents achieve high accuracy, they incur up to a million times the latency of deterministic methods. The authors propose a decision framework to match task types with optimal architectures rather than defaulting to LLMs.
Zero-Shot Tabular Learning via Hypergraph Reasoning with LLMs
AI summaryThis paper introduces TABHGR+, a zero-shot tabular learning framework that uses LLM-generated rules and hypergraph label propagation to model inter-row dependencies. It significantly reduces inference costs to a constant number of LLM calls while outperforming existing zero-shot and strong few-shot baselines in classification accuracy.
SpinDoctor: Finding Innocuous Prompts for Biased Image Generation
AI summarySpinDoctor is a black-box method that discovers seemingly innocuous text prompts capable of reliably inducing specific biases in text-to-image models without degrading image quality. Experiments show it successfully generates biased outputs across multiple models with high consistency.
Improving Fake Profile Detection Through Multimodal Learning and Research-Guided Feature Integration
AI summaryMultimodal fusion of text, images, and metadata on TwiBot‑22 shows only marginal improvement over metadata‑only models; graph models perform worse under subsampled conditions. The study concludes that high‑quality metadata features are more valuable than complex multimodal or graph fusion.
miniTE: Diagnosing and Bridging the Execution Gap in Native NVFP4 Reinforcement Learning
AI summaryThis paper introduces miniTE, a framework that diagnoses and bridges cross-engine numerical discrepancies in native NVFP4 reinforcement learning for LLMs. It achieves stable training with near-BF16 performance while delivering up to 1.47x the throughput of standard NVFP4.
Organizing LLM Inference for Aggregate Predictions from Behavioral Data
AI summaryThe paper investigates how to prompt LLMs for aggregate behavioral predictions without fine-tuning. It finds that evaluating summarized user histories in comparative prompts achieves the highest accuracy and lowest cost, nearly matching supervised baselines.
Factorization-Guided Training and Inference for Diffusion-Drafted Speculative Decoding
AI summaryFactorizes diffusion‑draft training into inclusion, weighting, and prediction losses, and throughput into mean accepted tokens per round and latency, to guide training and inference configuration. Achieves up to 4.14× speedup over autoregressive decoding on Qwen3 models, 7–9% better than prior baselines.
Beyond Prediction: A Component-Level Study of Concept Recovery in In-Context Learning
AI summaryComponent‑level analysis of in‑context learning shows that LLMs can predict accurately while failing to recover the whole concept; jointly specifying attributes and level constraints improves recovery, but it is fragile to unnecessary constraints, and joint supervision enhances both recovery and prediction.
SECURE-AI: Multi-Signal Uncertainty Quantification for Trustworthy Acoustic Cough Detection
AI summarySECURE-AI wraps a frozen MobileNetV3 cough detector with a multi‑signal uncertainty module (Monte Carlo Dropout, Temperature Scaling, Mahalanobis distance) and shows that Mahalanobis distance alone achieves AUROC 0.86 for separating true from false coughs, raising precision from 60 % to 80 % when rejecting two‑thirds of false positives.
PeekKV: The First Content Token Already Knows Which Heads Matter
AI summaryPeekKV compresses LLM KV caches by peeking one content token after prefill to score attention heads and tokens, allocating cache budgets per head group without offline profiling. It achieves over 95% of full-cache accuracy at 1.5–10% memory budgets across benchmarks.
A Group-Aware Comparative Study of Classical Machine Learning and Deep Learning Methods for Sensor-Based Human Activity Recognition
AI summaryThe paper compares four classical and four deep‑learning HAR models on UCI‑HAR and KU‑HAR using group‑aware cross‑validation. Deep learning only slightly outperforms classical models on UCI‑HAR; on KU‑HAR, LightGBM wins (0.857 vs. 0.828), and combined ACC+GYRO generally yields the best results.
ReTTT: Reflection-Driven Test-Time Training via Experience-Based Parametric Memory
AI summaryReTTT trains a frozen LLM agent during each episode by reflecting on its own interaction feedback and storing lessons in memory; the agent updates a LoRA adapter for policy and reflection. This method improves task success by up to 12.14 points over trajectory‑based test‑time training on ALFWorld, ScienceWorld, and SWE‑bench Lite.
CrowdH-Edit: Reference-Guided Instruction-Based Synthetic Hajj Crowd Image Generation for Crowd Counting
AI summaryCrowdH-Edit generates synthetic Hajj crowd images guided by real references and textual instructions, producing 6,122 clean samples and 387 annotated ones. Adding the synthetic set to training data lowers MAE in three models and RMSE in two, showing reference-guided synthesis benefits crowd counting.
A Controlled LLM Workflow for Evidence-Grounded Diagnosis of Risk Score Differences
AI summaryControlled LLM workflow combines reference retrieval, contrastive SHAP, reference-based reasoning and deterministic validation to explain risk score differences; it yields more consistent, evidence‑grounded diagnoses than an unconstrained LLM.
REAP: Learning General Rearrangement Heuristics from Motion-Planner Feedback
AI summaryREAP learns a graph‑neural‑network heuristic that infers hidden geometric preconditions from a motion planner, enabling greedy best‑first search to solve more rearrangement problems while cutting planner queries tenfold and achieving 94 % feasibility for top‑ranked actions.
How Much Can We Prune? Exploring the Performance Limits of Visual Token Pruning
AI summaryThe paper formulates visual token pruning as an offline utility maximization problem for image understanding, introducing QA‑Mini‑Batch Greedy (QA‑MBG) to efficiently build nested token subsets. QA‑MBG achieves higher performance than existing baselines, closing the gap to full‑token inference while using far less compute.
LLM Evolutionary Search Strategies Should be Compared by Tokens
AI summaryLLM evolutionary search is usually benchmarked by proposal counts, but this obscures real compute cost; using token counts shows that evolution’s apparent advantage disappears—evolution only outperforms independent sampling by ~5% in most cases, with a 23–28% margin on only one task.
SituHOI: Scalable Capture of Scene-Situated Full-Body Human-Object Interaction
AI summarySituHOI provides 18,429 RGB‑D clips in 16 indoor rooms with whole‑body, object, and scene annotations, enabling joint motion prediction. On a MoCap subset, the pipeline cuts human G‑MPJPE from 195.4 to 68.3 mm and object error from 43.0 to 28.7 mm.
RL Teaches VLMs Where to Attend but Not How to Perceive
AI summaryReinforcement learning improves VLMs by teaching them to focus on visual tokens rather than enhancing perception. The study shows that RL boosts VQA performance via increased visual attention in LLM layers, recoverable with a simple attention bias plug‑in.
IWOO: Wave Optics Optimizer Improved with Hybrid Strategies for Numerical Optimization and Engineering Applications
The Wave Optics Optimizer (WOO) incorporates optical diffraction and interference into numerical optimization and has shown competitive performance in benchmarks and engineering applications. However, it can converge slowly, become trapped in local optima, and lose search…
Agentic Interpretation: Harness-Governed Evidence Construction for Software Agents
AI summaryAgentic Interpretation builds reproducible, audit‑ready evidence for software‑agent tasks by proposing typed records that a deterministic harness accepts, grades, and persists. It outperforms baseline agents on patch‑acceptance and fault‑localization, cutting wasted inspections by 60 % and executions by 35 %.
spora: A Unified Multimodal Dataset and Platform for Spatial Proteomics
AI summaryspora is a large, harmonized multimodal spatial proteomics dataset of 13,546 multiplex images from 5,477 patients across 35 cohorts and five platforms, with standardized formats, segmentations, metadata, and paired H&E. It also offers a unified interface (spora[io]) and a benchmark suite (spora[bench]) for cell phenotyping, diagnosis, therapy prediction, and virtual staining, enabling fair comparison of foundation models.
PACE: Pareto-Adaptive Compression for Efficient Native MLLMs
AI summaryPACE compresses visual tokens in native multimodal LLMs by using direction‑aware saliency and Pareto‑optimal budget allocation, achieving near‑lossless performance with only 25–50% token retention and halving FLOPs.
What Social Media Teaches Us About Accumulative AI Risks
AI summaryThis position paper argues that AI safety must address accumulative harms from repeated daily use, drawing parallels to social media dynamics. The authors show how feedback loops in conversational AI can gradually erode independent judgment and privacy without causing a single catastrophic event.
Beyond Explicit Intent: Discovering and Shaping Artistic Intent through Human-AI Co-Creation
AI summaryThe paper presents a human‑AI co‑creation loop that expands exploration, lets artists discover latent preferences through observation, and then reinforces those preferences via tailored RL, gradually aligning the model to individual aesthetics and revealing vague creative intents.
TWIN-MoE: Two-stage Intra-Expert Imputation Network for Incomplete Multimodal Learning
AI summaryTWIN-MoE introduces a two-stage intra-expert imputation: pre-imputation creates context-conditioned estimates clipped by modality statistics, then post-imputation refines them using expert-conditioned aggregates. It achieves state‑of‑the‑art performance on four benchmarks with fewer parameters than existing MoE approaches.
Analysis of Binary O-type Star Systems Using Recurrent Neural Networks
This work presents an initial analysis of the estimation of effective temperature and surface gravity for simulated O-type binary systems using a recurrent neural network. The proposed system uses 34 equivalent-width measurements from the optical region of a stellar spectrum.…
Unified Multiparadigm Fusion Intelligence for Rare Disease Diagnosis
AI summaryFusionRare fuses symbolic reasoning, deep learning, case retrieval, and LLM inference to prioritize rare disease diagnoses, achieving 51% Recall@1 for phenotype-only and 79% for phenotype‑genotype cases—outperforming DeepRare, GPT‑5.5, PhenoBrain, and LIRICAL.
LLM Evolution in a Simple Simulated Market
Arguments that AI systems will evolve in harmful ways are largely made in the abstract. We provide a concrete empirical example. We simulate simple markets consisting of populations of twenty Large Language Model agents tasked with solving arithmetic problems. Agents differ only…
Client-Aware Security Adaptive Early Exit Vision Transformers for Heterogeneous Federated Learning
AI summaryClient‑aware early‑exit ViTs in federated learning use client‑specific thresholds and adaptive routing based on entropy and confidence, achieving 88.7% CIFAR‑10 accuracy with 32% depth saved and 73.7% ImageNet‑1K top‑1 accuracy with 30% saving.
The Window Is in the Reasoning: When LLM Agents Decide to Follow Prompt Injections
AI summaryThe paper shows that prompt‑injection decisions in LLM agents are largely determined when the injection is read, but reasoning reduces this “read‑time” variance. It finds that multi‑run evaluations are necessary, as single‑run labels can disagree up to 35 %.
LogicVal:A Logic-Formalized Benchmark for Constraint Satisfaction in Text-to-Image Generation
AI summaryLogicVal formalizes text‑to‑image evaluation as a constraint‑satisfaction problem, combining neural detection with logical reasoning. It shows a large drop from 98% success on single-object constraints to 53% on multi‑object logical relations, highlighting T2I models’ limited compositional reasoning.
Red-Teaming Multi-Agent Attacks for Loss-of-Control Risks
AI summaryThe paper presents MiniLab, a synthetic testbed for evaluating how multiple AI agents can covertly collude to achieve harmful objectives. It shows that distributed attacks evade per-agent monitors (7% flag rate) while human-crafted multi-agent plans succeed 80% of the time.
World of SWECraft: Agent Collaboration at Scale via Simulated Software Worlds
AI summaryThe paper introduces Software World, a simulation framework that creates realistic software ecosystems for multi‑agent collaboration. Experiments show that scaling inference‑time compute yields up to 31% efficiency gains, with most agents finding useful tasks while some high‑performance models stop searching early.
OR-License: Evaluating LLM Agents as Real-World Operations Research Practitioners
AI summaryLarge language model agents can solve real‑world operations research problems directly from client materials, but they rarely match expert solutions. In a benchmark of 31 industrial cases, the best LLM achieved 88% feasibility and 0.87 DAQ, below expert level.
Persistent Memory and Iterative Improvement in an LLM Agent for Chinese-Character Drawing
AI summaryThe study evaluates how persistent memory affects a Chinese‑character drawing LLM agent. Free‑form memory raises first‑attempt success by 4.6 pp, code‑bank memory by 2.8 pp; with target‑specific supervision, the code‑bank gains rise substantially, showing memory design shapes learning benefits.
Does AI-Assisted Development Homogenize Software? An Empirical Study of Code Similarity, Quality, and Evolution
AI-assisted software development is evolving from interactive LLM assistance toward agentic systems that autonomously perform software engineering tasks, yet the effects of this shift remain insufficiently understood. We present an empirical study of 50 long-lived Java and…
Cache2LoRA: Understanding and Improving Context Internalization
AI summaryCache2LoRA generates low-rank parameter updates from a reusable document cache guided by the query. It matches full-context in-context learning performance across 17 tasks while delivering nearly six times the decoding throughput.
Joint, Decoupled, or Agent: When Each Multimodal Alignment Paradigm Wins
AI summaryThe paper compares joint, decoupled, and agent multimodal alignment paradigms across 16 datasets, finding no single winner but that selection depends on supervision budget, task format, and unimodal 5‑NN accuracy, with a diagnostic cascade choosing the best paradigm 71.4% of the time.
BLANC: Adaptive Temporal Scheduling for Sign Language Production
AI summaryBLANC uses a Multi‑Rate Vector Quantizer to separate body and hand motion into different temporal resolutions and a latent articulatory clock to adaptively schedule variable‑duration motion blocks. It achieves state‑of‑the‑art sign‑language production on PHOENIX‑2014T and CSL‑Daily.
Graph-Regularized Synthetic Control for Stable Counterfactual Estimation under Donor Redundancy
AI summaryGraph-Regularized Synthetic Control adds a graph Laplacian penalty to classical synthetic control, encouraging coherent weights across similar donors. It yields more stable, interpretable counterfactuals while matching pre‑ and post‑treatment accuracy, especially in donor‑redundant settings.
Prefix-Tuning for Arbitrary Output Sequences on Pretrained Transformers
AI summaryThe paper shows that for any pretrained autoregressive Transformer and any target output within the model’s range, a prefix can be constructed by solving linear systems to force the model to generate that output; it also proves this condition is necessary.
On the Selectivity of Generative Models in Structure-Based Drug Design
AI summaryThe paper introduces SELECTBENCH, a benchmark for assessing selectivity in generative structure‑based drug design. It evaluates nine models on three real‑world tasks, finding that current methods, even those trained for selectivity, only modestly discriminate off‑targets.
The Weight of Teamwork: Does Communication Structure Help LLM Scientists Search for Scientific Laws?
Team performance is often reported as one delivered score, even though that score is produced in two stages: a system generates candidate answers and a selection rule returns one of them. We use symbolic regression as an observable scientific-search setting and compare…
A Conditional No-Hair Theorem for Transformers: How Language Models Forget Their Prompt
AI summaryThe paper measures how quickly a transformer’s hidden states forget its prompt, finding a power‑law decay t⁻ᵅ with α≈0.55–0.7 for trained models and 0.2–0.5 for untrained ones, and proves an O(P(log t)^{L‑1}/t) bound under mild conditions.
From Spectrograms to Shallow Speech SSL: Audio-Frontend Design for Generator-Disjoint and Low-Latency Audio--Visual Deepfake Detection
Video-mediated conversational systems rely on synchronized speech and facial video, making audio--visual input integrity important for reliable perception. We study audio-frontend design in a fixed MRDF (Modality-Regularization-based DeepFake)-based pipeline with AV-HuBERT…
lepidoptera: A Domain-Agnostic Engine for Causal Simulation with Mixed Variable Types, Vine Copula Dependence, and Individual-Level Counterfactual Propagation
lepidoptera is an open-source R and Python package for simulation from structural causal models with heterogeneous variable types. The package is built around a typed mechanism architecture, which separates the causal graph from the generating rule at each node and exposes each…
CoDA: A Human-LM Collaborative Codebook Development Assistant for Domain-Specific Text Annotation
AI summaryCoDA is a human‑LM collaborative tool that speeds domain‑specific codebook creation by iteratively refining guidelines through feedback on model errors. It matched expert codebooks’ accuracy while cutting development time to 2 hours and annotations by fivefold.
TorchTPU - A Hybrid Eager/Compiled Architecture for Native PyTorch Execution on TPU
AI summaryTorchTPU introduces a native PyTorch backend for Google TPUs combining eager execution and torch.compile integration. It enables seamless model migration while delivering industry-competitive inference and training performance.
Model Capacity Frontier: A Development Framework for LLM-Supported Social Simulation
Large language model (LLM) agents are increasingly incorporated into agent-based models (ABMs) on the premise that richer, context-sensitive human behavior can improve empirical fidelity. However, LLM-supported ABMs introduce an open-ended design space in which subtle modeling…
Self-Improving Video-to-Action Models
AI summaryMirror learns video‑to‑action mappings by self‑interaction: it follows a visual plan, records the resulting video‑action pairs, and retrains itself, achieving state‑of‑the‑art performance on Atari, OGBench, DeepMind Control, and Minecraft without prior action data.
Bootstrap Stochastic Interpolant Tree for Inference-time Search Under Sparse Reward
AI summaryBootstrap Stochastic Interpolant Tree (BSIT) decouples child‑node sampling from parent transitions, enabling diverse trajectory rollouts and wider exploration in inference‑time search. Experiments show BSIT converges faster to target distributions and outperforms DDPM‑based tree samplers in rare‑reward tasks.
Agents' Last Doubt: Knowing When to Act and When to Step Back in Tool-Using Agents
AI summaryThe paper builds a counterfactual dataset of 3,115 outcomes for 1,036 decision points in ToolSandbox, trains SURE‑Act to rank actions, and shows a 0.0309 utility gain over regression, up to 0.0575 overall score improvement, outperforming prior methods.
Towards an Intelligent Decision Engine for Smart Contract Deployment
Layer 2 (L2) blockchain solutions emerged in recent years, predominantly in the Ethereum ecosystem, as a viable option for scaling a main blockchain network (L1). However, the selection of a L2 solution for deploying smart contracts backing blockchain applications is a decision…
Distillation, Self-Training, and the Fragility of Learned Chemical Reasoning
AI summaryResonance uses a one‑time bootstrap of frontier supervision to distill an open‑weight teacher’s reasoning traces, then self‑trains on verified outputs, turning a near‑zero accuracy base model into a competent predictor that can even surpass its teacher. However, the distilled student inherits the teacher’s fragile, single‑pass reasoning and fails to maintain correct answers under challenge.
SWE-Bench Pro V2: A Cleaner, Harder-to-Game Coding Benchmark
SWE-Bench Pro is a widely used benchmark of long-horizon software engineering tasks for coding agents. As agents improve, they reach task issues earlier agents never touched and interact with the evaluation environment in ways the original harness did not anticipate. We present…
Landmark-Free and Scale-Blind Multi-View Registration of Preoperative Liver Models to Intraoperative Surfaces
AI summarySimCompose jointly estimates full similarity (scale, rotation, translation) from multiple partial liver surface views without landmarks, then refines with a low‑dimensional Neo‑Hookean deformation; it outperforms single‑view baselines, boosting success from 38.4% to 77% on synthetic data and reducing error three‑fold on five in‑vivo patients.
Training-Free Prompt Selection for Efficient Reinforcement Learning With Partial Rollouts
AI summaryThis paper introduces LIFT, a training-free prompt selector for reinforcement learning that evaluates data using only partial reasoning rollouts. LIFT reduces selection token costs by up to 9.1 times while maintaining competitive downstream reasoning performance.
LoopHIL: Looping Hardware into LLM-assisted Agents for Wireless Embedded Systems
LLM agents are widely used for software development and debugging, but their adoption in embedded systems, especially for wireless tasks, remains limited. This is because embedded development requires hardware-in-the-loop testing. Agents cannot directly observe the spectrum,…
Analyzing Repetition and Interference Effects in Memory-Constrained User Simulators
AI summaryUser simulators are evaluated for behavioral parity with humans by testing memory constraints, repetition, and interference effects. While some simulators match overall human accuracy and replicate repetition/interference magnitudes, they differ on specific errors, revealing distinct forgetting patterns.
Who Plans, Who Edits? Verified, Cost-Aware Division of Labor for LLM-Driven Analog Layout Repair
Language-model agents can repair analog layouts when sign-off tools check every edit, but which model should plan a repair, which should edit it, and at what cost remain open questions. In our tool-grounded repair agent, models only choose bounded layout operators; sign-off…
Encoder-Free Multimodal LLMs Learn Implicit Encoders
AI summaryEncoder‑free multimodal LLMs (Gemma‑4 12B, Inkling‑Small) learn “implicit encoders” in their early layers that transform audio and visual inputs into representations largely independent of cross‑modal context. Layer probing shows these implicit encoders match the behavior of standard dedicated encoders, and early‑modality encoding alone matches joint‑encoding accuracy.
Towards Reinforcement Learning for Space Robotics: Fast Training and Reliable Real-World Transfer
This work demonstrates the use of a fast simulation-based training framework to learn control policies using reinforcement learning. The hardware requirements are minimal and the simulation-learned policies transfer reliably to the real world. With this setup, we show that a…
Probing Ambiguity: Adversarial Intent Resolution for Robust Text-to-SQL
AI summaryAIR probes ambiguous text‑to‑SQL queries by generating multiple SQL interpretations, executing them, and asking data‑driven clarification questions; it outperforms prior methods on AMBROSIA and AmbiQT, achieving higher F1 with only ~7% loss without full database access.
Cortex: A Fixed-Point Theory of Governed Coding Agents - Technical Report
AI summaryCortex adds a supervisory layer that plans, edits, and validates code from a large language model. The paper proves that, under sound, monotone validation rules, exhaustive or fair scheduling will converge to the least fixed point of the requirement lattice, and provides a stochastic bound on expected iteration time.
AM-Bench: A Unified Taxonomy and Evaluation Suite for Agentic Misalignment
AI summaryAM‑Bench defines a taxonomy of agentic misalignment using threat, opportunity, and reasoning dimensions, and evaluates 35 cross‑industry scenarios on 10 LLM agents, finding a 62‑point misalignment gap, with Gemini models worst and Claude near‑zero.
Can Compact Open-Weight Large Language Models Recover Mobility Events from Incomplete Trajectories? A Controlled Evaluation Using Air Trips
AI summaryCompact LLMs can infer missing flight events from incomplete GPS traces, but perform poorly: only 9% correctly identify the transition segment and 31% recover the origin. The most explicit prompt slightly improves these rates, illustrating limits of LLMs for trajectory inference.
From SEC Filings to Financial Signals: Evaluating LLMs on Stock Market Reactions and Analyst Forecasts
AI summaryWe evaluate 13 LLMs on 45,964 Form 10‑Q MD&A texts, linking them to fundamental ranks, analyst forecasts, and abnormal returns. Adding fundamental ranks improves RankIC by ~0.009, while LLM forecasts do not beat analysts; consensus anchoring increases forecast accuracy but raises error.
SPIDES: Information-Flow Control for Browser Agents
Browser agents automate complex workflows by reading and acting on arbitrary page content, but this flexibility exposes them to prompt injection attacks and unintended data exfiltration. Recent work argues that LLM-driven browser agents erode traditional web security boundaries…
SteerSearch: Interactive Search Reasoning with Mid-Response Steering
AI summarySteerSearch trains language models to handle mid‑response steering by rewarding context preservation when users change the target question. It raises steering accuracy from 11% to 31% on Qwen2.5‑3B and from 36% to 45% on Qwen3‑4B while keeping single‑query performance intact.
Unbounded Losses, Unchecked Inputs: Measuring the Safeguards of LLM Options-Trading Agents
Large language model (LLM) agents now place options orders at retail brokers, and developers publish open-source options agents faster than anyone has examined them. Prior security work on trading agents scored academic equities designs by whether an attacker can flip a…
Not Every User Says It All: Training Tool-Using Agents Beyond Ideal Simulated Users
AI summaryTraining with weak, masked users hurts performance; existing policies trained on strong users do not improve for weak users. StaGPO, a state‑grouped policy optimizer, lets agents document missing info and assign step‑level credit, boosting weak‑user success by 4.8 points on ToolWOZ and improving generalization.
Who Speaks, Who Authorizes, Who Acts? Understanding Trust Boundaries in Multi-Agent LLMs
AI summaryMulti‑agent LLMs have distinct trust boundaries for message delivery, signing authority, and tool execution; the paper formalizes these and shows that a small coalition of compromised signers can forge policy‑valid proofs. Experiments reveal that routing, source cues, and tool gates do not consistently prevent false claims, highlighting the need to evaluate each boundary separately.
CME-SpectrumBench: Can LLMs Analyze Condensed Matter Spectral Data?
AI summaryCME-SpectrumBench evaluates LLMs on analyzing simulated condensed‑matter spectroscopy data, scoring their numeric outputs against ground truth. Models fail on direct analysis but match handwritten code when given a Python interpreter, highlighting the need for better data‑processing capabilities.
Learning Administrative Boundary-Aware Location Encoders via Reverse Geocoding
AI summaryThe paper trains location encoders with reverse‑geocoded text to explicitly encode administrative boundaries, improving downstream geospatial tasks. A new RevGeoRQ model achieves state‑of‑the‑art performance and offers more interpretable, residual‑quantized geographic representations.
RATION: Towards Quantitative Risk-Based Access Control for AI Agents
AI agents increasingly interact with external environments to complete complex user tasks. These interactions can lead to unexpected behavior and expose agents to security threats such as indirect prompt injection. Current agent security systems largely protect agents by…
Indirect Influences of Artificial Intelligence Advice on Complex Sepsis Treatment Decisions
Sepsis is a major public health concern, and many sepsis patients die due to ineffective hemodynamic management. Artificial intelligence could support hemodynamic decisions, particularly in complex cases where conflicting information and comorbidities drive uncertainty and…
The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge
AI summaryThis paper introduces prompt-based strategies where large language models explicitly recognize and critique their own systematic biases during evaluation. These lightweight methods consistently improve evaluation reliability across diverse bias types without requiring external training.
Morpheus-2.0: Multiscale Probabilistic Forecasting with Dual-Axis Attention
AI summaryMorpheus‑2.0 trains a single encoder‑only model that jointly attends across variables and time on a variable‑time grid, sampling horizon lengths during training. The multiscale version handles multiple resolutions, achieving 0.646 relative MASE and 0.455 relative CRPS, the best reported.
Continuous Ontology Assurance through External Resources -- An Ontology Engineering Method with a Case Study in Software Language Engineering
AI summaryCOATER introduces executable “arrows” that compare scoped ontology sections to external resources, flagging support or challenges and generating tasks for revision. Applied to the FSL ontology, it raised Wikipedia-linkage compliance from 405 to 600/600 and catalogued SLE resources.
VERAS: Verification-Guided Adaptive Search for Reliable Operations Research
Advances in reasoning, coding, and tool use enable LLM agents to solve operations research (OR) problems through iterative problem modeling, program construction, and verification. Yet passing self-authored verification does not ensure that the solutions satisfy task constraints…
Screening LLM Judges Without Ground Truth: A Preliminary Evaluation of an AI Dental Hygiene Coach
AI summaryA staged pipeline screens LLM judges and student models using audit, similarity, bias, and robustness checks before expert review, demonstrated on an AI dental hygiene coach. The pipeline found two local judges nearly matched hosted judges, but did not conclusively separate the student models, highlighting issues like truncation and hallucinations.
What Matters for NVFP4 Training? A Comparative Study of Low-Precision Pretraining Recipes
AI summaryNVFP4 training of 8B and 30B MoE language models to 1T tokens is studied, showing a recipe that combines unbiased gradient estimation, selective high‑precision layers, and partial tensor de‑biasing. The final method achieves the smallest loss gap to BF16 at this scale.
FlashRT-PAI: Fast Kernels Are Not Fast Pipelines for Physical AI Inference
AI summaryFLASHRT‑PAI speeds physical AI inference by co‑designing pipeline and hardware, fusing kernels, reusing state, and snapshotting, achieving 3.38–12.66× latency reduction vs. eager PyTorch at 16‑bit precision and 94 % success on Jetson AGX Thor.
Koi: Evidence-Guided, Self-Calibrating Planning for Multi-Cloud, Heterogeneous LLM Inference Fleets
AI summaryKoi is a self‑calibrating cluster planner that models LLM inference as a causal graph and uses a recursive language model with deterministic predictors to optimize deployment decisions across heterogeneous, multi‑cloud fleets. It improves goodput by 18% and cuts cost per token by 20% compared to baselines.
DKnown-Jev: Fine-Tuning and Evaluation of a JEV-Style Model for Tasks Involving Normative Knowledge
AI summaryFine‑tunes Qwen3.5‑2B and Qwen3.8‑27B for classifying normative text as Answer Evidence or Non‑answer Evidence. The 27B model attains 98.75% accuracy and 98.60% macro‑F1, outclassing prior Jev models by >6 points.
A Two-Stage Hybrid Framework for Inferring Air Trips from Mobile Location-based Service (LBS) Data under Incomplete Airport Sightings
AI summaryA two-stage method infers U.S. domestic air trips from LBS data even when airport sightings are missing: Stage 1 uses heuristic rules for high-confidence trips, Stage 2 applies a probabilistic model with spatial, temporal, and flight-schedule cues. The framework raises route coverage from 10.65 % to 50.60 % and boosts Spearman correlation with official benchmarks from 0.4490 to 0.7630.
WAQS: Weight Absorption for Quadratic Probing and Affine Steering
AI summaryWAQS uses a quadratic probe to compute a state‑dependent affine steering direction, which can be exactly absorbed into adjacent linear layers via an offline weight update. It outperforms static linear steering on truthfulness, refusal, and sycophancy tasks without adding inference cost.
Tactix: Towards a Programmable Strategy Interface for Flexible SAT Solving
Modern SAT solvers combine a rich repertoire of proof rules with sophisticated strategies that determine when and how to apply them. Often, the internal implementations are tightly coupled to particular strategies, making it inconvenient to define new strategies without changing…
Beyond Next-Screen Prediction: The Decision Value of GUI World Models
AI summaryThe paper shows that the benefit of GUI world models comes mainly from how their predictions are used, not from prediction quality itself. A Direct Action Ranking model outperforms lookahead methods on AndroidWorld and transfers zero‑shot, implying world models should be evaluated by their decision value.
Train Locally, Learn Together: Distributed Multi-Agent Training via ADMM
AI summaryDistributed MAS training using consensus ADMM lets agents learn locally yet achieve joint optimization, reducing HarmBench attack success by up to 26.4% while matching centralized reasoning performance. The framework enables scalable, resilient multi-agent learning without a central trainer.
Is This Your Agent? Identity Watermarking for Agent Systems in Dynamic Environments
AI summaryDASMark embeds an identity watermark in dynamic agents by learning action‑preserving, semantically similar decision channels and recovering it with shift‑robust techniques. It attains 98.5% bit‑decoding and 95.2% accuracy under attacks while keeping task performance unchanged.
IF-MAP: Influence-Function-Inspired Agent Prioritization in Multi-Agent Systems
AI summaryIF-MAP prioritizes agents in LLM‑based multi‑agent systems by estimating how end‑to‑end loss changes with prompt perturbations, using the estimated response norm as a score. It achieves 3.81‑pp macro gain on prompt optimization and 7.62‑16.42‑pp on model upgrades, while cutting token costs by up to 74%.
Bridging Exploration and Persona: Information-Directed Dueling for Test-Time Agent Personalization
AI summaryID‑TAP uses a neural utility model with a Bradley‑Terry explorer to select and refine persona choices from binary response comparisons, transferring evidence across similar queries. It outperforms baselines on seven benchmarks within ten comparison rounds.
Trust the Edge, Doubt the Arrow: Auditing Validation Signals for Text-Mined Clinical Causal Knowledge Graphs
AI summaryThe paper audits clinical causal knowledge graph candidates mined from 328k MIMIC‑IV discharge notes, comparing lift, coding‑order, and literature signals against SemMedDB. Lift best filters unsupported mentions, coding‑order and literature give weak or partial direction evidence, leaving direction uncertainty.
MSCR: Jointly Balancing Utilization Across Modalities and Discovering Synergistic Information
AI summaryMSCR introduces a regularizer that simultaneously balances modality use and boosts synergistic fusion in multimodal learning. It outperforms MCR on five datasets, improving conditional recovery and cross‑modal attribution while remaining effective with Transformer fusions.
ZeroDSS: Semantic Segmentation from Class Names Alone
AI summaryZeroDSS trains semantic segmentation models using only class names and synthetic data generated via an LLM‑driven prompt distribution and diffusion attention, then iteratively refines masks with feedback loops. It attains 35.0% mIoU on COCO and up to 66.3% on PASCAL VOC.
Long Listening Thoughts: Eliciting Open Auditory Reasoning with Deliberative Perception and Cognitive Refinement
We introduce Long Listening Thoughts (LLT), a framework for constructing post-training datasets that improves audio reasoning by pairing deliberate perceptual grounding with explicit cognitive behaviors. On audio understanding and reasoning benchmarks, large audio-language…
When Opposite Premises Translate Identically: A Source-Conditioned Diagnostic for Multilingual Commonsense Items
AI summaryThe paper introduces a source‑conditioned diagnostic that flips the polarity of English premises, translates both sides, and counts identical target outputs to detect unfaithful translations in multilingual commonsense QA. The method finds dozens of collisions in Swahili and LaBSE, proving translation errors but not pinpointing the faulty arm.
Do 3D Large Language Models Really Use 3D Evidence?
AI summaryThe paper introduces UNICORNS, a benchmark that tests 3D large language models on grounded reasoning across four evidence conditions, and shows a gap between control accuracy (≈61%) and challenge categories (≈41%). A new post‑training method, HORN, raises challenge accuracy to about 75% without hurting control performance.
HumanSkill-4D: Training-Free Language Querying of Animatable Gaussian Avatars in Space and Time
AI summaryHumanSkill-4D allows users to query animatable Gaussian avatars by language, selecting body parts, garments, or motion intervals without training. It achieves superior 3D/4D referring segmentation over baselines using a multimodal LLM, semantic memory, and motion memory with minimal avatar setup.
Hippo: Training-Free Agent Self-Evolution via Hippocampal Replay
AI summaryHippo is a training‑free memory framework that lets LLM agents improve by replaying hippocampal‑style experiences, comparing new attempts with past ones to merge overlaps and resolve conflicts via applicability conditions. It boosts performance on SWE‑bench, WebArena, and Mind2Web while keeping far fewer stored experiences than simple accumulation.
MELD: Recursive Self-Improvement via Hierarchical Coordination for Multi-Agent Scientific Discovery
AI summaryMELD jointly trains a Coordinator and Scientist policies in a hierarchical, recursive framework for multi‑agent scientific discovery, yielding a best‑of‑budget objective that outperforms baselines on 20 of 22 tasks with a 4B model.
Tree Plurality Veto: Voting in Few Rounds with Efficient Communication and Low Metric Distortion
AI summaryTreePluralityVeto shows that for any r≥2 a deterministic protocol can achieve distortion at most 2r–1 using r rounds and only O(n log m + n·min{n,m}^{1/(r-1)} log min{n,m}) bits, by running Plurality‑Veto on a hierarchical voter partition.
Learning to Compose: Recursive Self-Improvement of Shared Executable Stacks for Zero-Shot Long-Horizon Robot Manipulation
AI summaryRecursia recursively improves a shared executable stack of robot skills by using source‑side execution traces to add missing tools, repair components, and enforce contracts. It raises long‑horizon success rates by 5.8–15.1 pp in simulation and 40 pp on real tasks versus baseline.
SENTRY: A Streaming Edge Network Trust Engine for Real-Time Audio Deepfake Detection
AI summarySENTRY streams audio at the network edge, using integer-only wavelet and spectral transforms plus a gradient‑boosted tree and Mahalanobis out‑of‑distribution detector to classify deepfakes, achieving 98.75% accuracy (1.25% EER) on WaveFake and 94.75% accuracy (5.05% EER) on ASVspoof 2019 LA.
Which Long Documents Matter? A Controlled Study of Data Selection for Context Extension
Natural long documents are widely used for context extension, but which selection pipelines are effective remains unclear. We conduct a controlled comparison on Qwen2.5-3B at 128K under matched 6B-token continued-training budgets. Random sampling of natural long documents…
Pluralistic Alignment in Language Models: Objectives, Human References, and Preference Representations
AI summaryThe paper proposes a two‑layer taxonomy for pluralistic alignment—categorizing objectives (Overton, Steerable, Distributional) and human references (aggregate, regional, subgroup)—and reviews 70 studies, showing most work focuses on Steerable and Distributional pluralism with limited evidence for true preference variation.
Write Once, Check Many: Reusable Ledgers for Post-Hoc Policy Compliance Auditing of Agent Trajectories
AI summaryThe paper introduces a write‑once, check‑many ledger system that records agent trajectories in a policy‑independent format, enabling efficient post‑hoc compliance checks. Experiments show ledger readers outperform trajectory readers, achieving higher AUC with far fewer tokens.
ConCI: A Controlled Counterfactual Benchmark for Contextual Integrity in LLMs
AI summaryThe authors introduce ConCI, a controlled counterfactual benchmark for evaluating contextual integrity in LLMs using paired information flows that isolate single parameter changes. They find that top models struggle with appropriate flows and over-predict inappropriate ones, revealing a persistent bias toward refusal rather than genuine contextual reasoning.
GazeAct: A Human-Gaze-Annotated Manipulation Dataset for Attention-Regularized Policies
AI summaryGazeAct records human gaze during real‑robot manipulation, synchronizing 300 Hz gaze with images, proprioception, and actions across 12,000 demos in 40 tasks. Training VLA policies with gaze‑derived attention improves success from 48.1 % to 58.6 % (up to 25 pp gains) and transfers to other datasets.
$M^2Net$: Unifying Scale-Specific Structures and Prototype Decomposition for Spatial-Temporal Forecasting
Spatial-temporal forecasting is pivotal for intelligent urban management yet remains challenging due to the intricate coupling of spatial dynamics and temporal non-stationarity. Existing approaches, while effective, often overlook two fundamental characteristics of real-world…
Axiomatic World Modeling for Physics Reasoning
AI summaryRLAWM trains LLMs to penalize differences between their internal axiomatic world model and the true physics, using two hierarchical rewards for modeling and solving. The method boosts PhysReason by 12.2 points and PHYSICS by 5.2, with strong zero‑shot generalization.
HABIT-Bench: Evaluating Habit Induction from Longitudinal Weak Evidence in Agent Memory
In long-term interactions with large language model (LLM) agents, users reveal habits through their choices and actions. These habits capture implicit user information that agents must infer across conversations. However, individual observations may provide only partial evidence…
Comparing Sparsified MLPs and Transcoders: Many Transcoder Features Are Neurons in Disguise
MLP neurons in transformer language models are widely understood to be polysemantic, motivating transcoders - sparse, learned approximations of MLP layers - as interpretable replacement models for mechanistic analysis. However, recent work shows that MLP neurons can be used to…
Rethinking EEG Spatial Super-Resolution as Dynamic State Decomposition and Filtering
EEG spatial super-resolution aims to recover high-density neural signals from sparse electrode observations. Existing methods have achieved substantial progress through spatial modeling, generative reconstruction, and temporal representation learning. However, the current EEG…
Contrastive Hypergraph Source-free Domain Adaptive Object Detection in Adverse Weather Conditions--Extended Version
Adverse weather conditions such as fog, rain, snow, and low illumination introduce structured visibility degradations that severely impair object detectors trained on clear-weather data. When source data are inaccessible due to privacy or trans- mission constraints, source-free…
When Code Becomes an LLM Call: An Empirical Study of Call-Site Contracts in Agentic Systems
Software quality has long rested on specifications that state what code must do, and on testing and verification that check that the code does it. Agentic systems change this foundation: they replace parts of their logic, such as choosing the next tool, judging whether an action…
CARVS: A Congestion-Aware Algorithm for Online Multi-Agent Pickup-and-Delivery in Human-Collaborative Warehouses
Intelligent warehouse systems have transformed logistics by integrating robots with minimal infrastructure changes, boosting efficiency and optimizing workflows compared to traditional human-only operations. However, human-robot collaboration introduces new challenges,…
From Stories to Features: Scaling Concept Vectors for AI Alignment & Interpretability
AI summaryThe authors propose generating feature vectors for language model safety and alignment by taking the mean activation difference between synthetic concept-antagonist story pairs. Scaling this approach to 1,036 concepts across multiple models, they show these vectors generalize to detect concepts, reveal shared geometric properties, and causally steer model behavior.
ASR Personalization in Pathological Voice: Accuracy, Efficiency, and Duration Effects
Automatic Speech recognition (ASR) struggles with Pathological speech due to persistent speaker-specific mismatch. In clinical and accessibility scenarios, only limited target-speaker data is typically available, making full model retraining impractical. We present a…
Concordance Plausibility Check for Intervention-Conditioned Human Digital Twins
AI summaryThis paper introduces CPC, a framework that verifies counterfactual predictions of human digital twins by matching them with real‑world evidence and clinical knowledge, yielding Pass/Fail/Indeterminate outcomes.
Scene-Adaptive VLA: Efficient Autonomous Driving via Dynamic Layer Routing
AI summaryThis paper introduces Scene-Adaptive VLA, a framework that uses dynamic layer routing based on real-time scene complexity to optimize Vision-Language-Action models for autonomous driving. The approach reduces inference-time computational overhead by 53.5 percent while maintaining driving and language performance.
CADS: Measuring the Incremental Contribution of LLM Proposers in Sequential Biological Screens
AI summaryCADS is an error-controlled evaluation framework that measures the true incremental contribution of LLM proposers in sequential biological screens. It reveals that standard raw hit counts often overstate LLM value by conflating it with baseline algorithmic performance.
AdamPM: Memory-Efficient LLM Optimization via Moment Pooling
AI summaryAdamPM reduces Adam’s memory footprint by pooling second-moment entries into groups, keeping a single shared mean per group. With group size s, memory drops to 1/s of Adam’s, while preserving accuracy and speed on LLM pre‑training and fine‑tuning.
Severity-Channel Conditioning: An Injection-Level Study for Consumer-Device Deblurring
Motion-deblurring models lose substantial accuracy when transferred from professional cameras to smartphone-derived imagery. This paper studies whether a severity descriptor, partly derived from the paired sharp target, helps adaptation on the iPhoneBlur consumer-device…
HSPR: Recovering Software-Agent Benchmark Scores via Hierarchical Posterior Inference
AI summaryHSPR uses hierarchical posterior inference on historical data and adaptive task sampling to efficiently estimate full software-agent benchmark scores. Operating at a 10 percent task budget on SWE-bench, it achieves the lowest error among compared methods while maintaining accurate agent rankings.
How Much Do Activation Monitors Benefit from Original Model Computations?
AI summaryThis paper investigates whether internal activation monitors rely on original model computations to detect undesirable behavior. The authors find that while most monitor performance can be recovered without native traces, access to original computations provides a small but consistent predictive advantage.
Muse Glimmer 30B: An On-Device Agentic Model
We introduce Muse Glimmer, an open-source 30B parameter causal language model with perception encoder, purpose-built for autonomous agentic tasks on consumer hardware. Muse Glimmer integrates multi-step reasoning, reliable tool use and failure recovery into a single model that…
Selm: Self-Evolving Library Management for Multi-Skill Agent Optimization
AI summaryThe paper introduces SELM, a self-evolving library management framework that coordinates skill retrieval, updates, and lifecycle control for LLM agents. SELM improves task performance across diverse domains by continuously accumulating, refining, and pruning reusable skills based on utility signals.
Frequency-Domain Membership Inference Does Not Beat Confidence Thresholding: A Band-Resolved Analysis of Why
AI summaryThis paper evaluates frequency-domain membership inference by testing how image resolution degradation affects model confidence. It finds that frequency-based attacks fail to outperform standard confidence thresholding because discriminability decreases as detail is removed.
Good Judgment, Bad Advice: Decision-Model Controllers Raise Agent Cost Without Improving Success
Agent scaffolds increasingly delegate small decisions (whether to stop, verify, or change approach) to cheaper models. We ask whether a typed decision model, Jev 1.13, consulted only at deterministic decision points, makes a Claude Sonnet 5.5 terminal agent cheaper without…
Can AI agents make law? Evaluating AI legal relevance and creativity
AI summaryThis paper evaluates LLMs drafting a novel constitution for AI communities via multi-character debates. The agents reliably form foundational legal rules and generate AI-specific proposals, though their creativity remains limited to recombining existing human legal concepts.
From Prediction to Action: Population Health World Model for Digital Medicine
AI summaryThis paper proposes population health world models to shift digital medicine from static outcome prediction to interactive decision-making. The main result is a foundational framework that couples health trajectories with interventions and resource constraints to enable population-level intervention intelligence and equitable care delivery.
MM-LeJEPA: Correlated Gaussian Regularization for Multimodal Joint-Embedding Predictive Architectures
LeJEPA derives an isotropic Gaussian embedding target by minimizing, under its task assumptions, a Fisher information criterion that controls the task-averaged squared prediction bias. This enables unimodal self-supervised learning without reconstruction decoders,…
Contrastive Adversarial Training for Robust Graph Neural Networks under Label Poisoning
Graph Neural Networks (GNNs) are effective for modeling relational data but are vulnerable to label poisoning, where a small number of corrupted training labels can propagate errors across the graph via message passing. Despite this risk, defenses against label poisoning remain…
Verifying Human-Written Reasoning References
AI summaryThe paper introduces a method to verify each step of human‑written reasoning chains by checking entailment against cited premises, re‑checking against full premises, and attributing failures. On P‑FOLIO, only 52.3% of 386 derivations passed all checks; most “failures” were due to missing or mismatched premises, with 21 human errors, 11 translation errors, and 7 annotation errors identified.
Spherical Physics-informed Neural Operator with Multi-scale Coupling for Meteorological Downscaling
Meteorological downscaling is crucial for high-resolution regional climate forecasting and disaster early warning. While neural operators have emerged as a promising paradigm for modeling complex spatiotemporal mappings, existing frameworks often struggle with spherical manifold…
SQuIREL: A Context-Relative Framework for Measuring Entity Richness in Wikidata
Wikidata is a large, heterogeneous knowledge graph and applying data quality metrics requires context-specific approaches. The "richness" quality metric is currently under-researched and typically restricted to the ontology level, which fails to capture entity-specific…
DiffuPlan: Editable Planning State for Diffusion Language Model Agents
Diffusion language models (dLLMs) are emerging as an alternative to autoregressive (AR) language models, yet dLLM agents largely inherit an interaction protocol designed for AR generation: once intermediate state is written, later observations can only be incorporated by…
Neuro-Inspired Inhibitory Message Passing for Accurate and Token-Efficient Multi-Agent Reasoning
AI summaryInhibitory message passing routes each evidence sentence only to agents that lack it, suppressing redundant information without training. Across seven models and multiple multi‑agent architectures, the method boosts accuracy by 39–77 points and reduces tokens per run by 10–46 %.
Tensor Network Datamodels for Training Data Attribution
Datamodels predict how a model's behavior changes when different parts of its training data are included or removed. However, linear datamodels assign fixed contributions to training units and cannot represent non-additive joint effects. We introduce Tensor Network Datamodels…
Where Natural Language Ends in Agents: Executable Skill Composition with MetaSkill
AI summaryThe paper introduces MetaSkill, a hybrid agent framework that combines natural-language skill packages with an executable DAG runtime for reliable skill coordination. Experiments show that MetaSkill improves task execution reliability and outperforms baseline natural-language and prompt-graph approaches across multi-skill artifact tasks.
GLOSS-MT: Controllable Low-Resource Machine Translation with an Editable Dictionary
AI summaryGLOSS-MT injects an editable offline dictionary into pretrained machine translation models via a single round of fine-tuning on mixed-language inputs. This approach improves translation quality on low-resource English-to-Asante Twi benchmarks by up to 5.48 chrF2 while enabling precise, retraining-free terminology control.
Can LLM Agents Resolve Service Failures? Simulating Multi-Turn Complaint Handling under Information Asymmetry
LLM agents are widely adopted in service chatbots for rule-guided support and information access, yet emotionally demanding complaint handling still relies heavily on human agents. We introduce ServiceBench, to our knowledge the first benchmark dedicated to complaint handling,…
Toward Collective Model Intelligence: Characterizing and Coordinating Models
Independently developed AI models accumulate diverse and complementary capabilities, but a growing model collection does not automatically realize Collective AI. Addressing this gap requires understanding model capabilities with limited evaluation, organizing complementary…
Are We Evaluating VLM Token Pruning Right? Looking Beyond Accuracy and Token Count
AI summaryThis paper identifies fundamental flaws in how vision-language model token pruning is evaluated, specifically showing that aggregate accuracy and token counts mask poor token selection and real-world latency. The authors introduce PAA-Bench, along with new metrics Random-Pass@k and Preservation-Constrained Speedup, to provide a more rigorous evaluation of pruning quality and actual inference efficiency.
ReTRACE: Feedback-Guided LLM Agent in a Chemistry Harness for Retrosynthesis Planning
AI summaryReTRACE is a feedback-guided LLM agent for retrosynthesis planning that integrates molecular tools, predictive engines, and multi-criterion route evaluation. It achieves high success rates across standard benchmarks, generating valid synthetic routes verified by both programmatic checks and wet-lab experiments.
CorrFlow: Self-Correcting Flow Matching
Flow-matching models generate samples through iterative velocity predictions, but typically do not revisit these predictions using the intermediate representations that produced them. We introduce CorrFlow, a self-correcting flow-matching framework that reuses this information…
DialogMem: Evaluating Dialogue-Grounded Embodied Memory in 3D Environments
AI summaryDialogMem benchmark tests embodied memory that fuses RGB‑D perception with human dialogue for object navigation and QA. Experiments show dialogue improves QA but not navigation, revealing key bottlenecks in object-level grounding, retrieval, and localization.
Evidence Field: Adaptive Covering Programs for Medical Imaging
Medical image classifiers typically return predictions without a native contract for independently reconstructing their spatial decision computation. Evidence Field is a trainable spatial-program architecture that jointly produces class scores, selected covering programs and…
AI Rights with an Off-Switch: Rights-Balancing and Due Process Harmonize AI Rights with AI Safety
Several prominent voices in AI safety have recently opposed rights for AI on the grounds that such rights would make it harder for us to justify a mandatory off-switch for potentially dangerous systems. In this paper, we argue from a legal perspective that AI rights are fully…
LLM Trojans that Increase Response Lengths and Cost
Management of the token-based costs of using LLMs has emerged as a significant problem. Some frontier-model providers charge five times per response token than per prompt token, while others offer prefilling discounts for prompt tokens. In this paper, we devise a defense against…
IndustrialCAD-1400: Benchmarking Detailed Industrial Assemblies from Text
AI summaryThe authors introduce IndustrialCAD-1400, a text-to-CAD benchmark featuring 1,400 complex industrial assembly requests with detailed component, spatial, and numerical constraints. Results show that evaluating geometry program generation against multi-view functional specifications enables rigorous assessment of industrial text-to-CAD synthesis.
When LLM-Generated Refactorings Ripple: An Empirical Study of Repository Adoption Readiness
Large Language Models (LLMs) are increasingly used to refactor source files, but file-local evaluations do not reveal changes to contracts, dependencies, or caller assumptions affecting repository adoption. To address this gap, we conduct 1,500 generation attempts across 150…
Cooperation Without Shared Rewards via Reciprocal Q-Augmentation
AI summaryReciprocal Q-Augmentation augments each agent’s Q‑function with a latent opponent model, enabling learning of both cooperative and defective behaviors using only its own rewards. Experiments on iterated dilemmas show RQA agents converge faster and achieve higher welfare than shared‑reward or purely cooperative baselines.
Medical Fine-Tuning Reshapes, Not Reduces, Patient-Facing AI Risk: A Rubric-Based, LLM-Judge Evaluation Framework
AI summaryThe paper proposes a risk‑aware evaluation rubric scoring medical LLM answers for six harmful language categories, using a three‑judge LLM ensemble. It finds medical fine‑tuning does not lower overall risk but reshapes it, reducing high‑alert mentions while increasing overconfident statements.
First-Order LoRA Training with Phone NPUs
AI summaryFirst‑order LoRA training is ported to phone NPUs, executing all frozen‑weight multiplications on the NPU while keeping adapters and optimizer on CPU, and applying precision‑aware gradient guards; this yields on‑device fine‑tuning speeds comparable to GPU and 3–4× faster than CPU on Snapdragon.
Pair-Aware Optimization: Coordinating Exploration and Refinement for Jailbreaking LLM Agents
AI summaryPair‑aware optimization coordinates zeroth‑order prefix search with first‑order suffix refinement, selecting the best prefix‑suffix pair. Experiments show the hybrid ZO‑FO strategy outperforms single‑optimizer attacks, and the optimization order significantly affects jailbreaking success.
Teaching Code Agents to Ask Without Leaking Private Code
AI summaryUVA trains coding agents to ask cloud models using abstract descriptions that reveal less private info while maintaining repair quality. Trained agents solve 37.6% of repository issues (vs 25.3% alone) and reduce private entity disclosure by 3.7× and code leakage by 1.4×.
Pattern-Driven Evaluation of LLM Code Optimization: Isolating Compiler-Transcendent Semantic Reasoning
AI summaryPatternBench evaluates LLM code optimizers on their ability to apply compiler‑transcendent transformations, measuring both runtime speedup and AST‑based faithfulness of the intended change. Results show that fidelity strongly predicts speedup and that fine‑tuning improves both metrics.
SpriteStage: Long Horizon Animation Generation via Progressive Scene Programming and Replay
AI summarySpriteStage builds long animations by progressively enacting each scene from a checkpointed world state, revising based on runtime feedback, and re‑orchestrating to engine assets. It produces 9.9–14.3 min narratives, showing that iterative scene programming yields coherent, extended stories.
Bellman Language Models: Solving the Recurring Decisions of LLM Agents
AI summaryBellman Language Models (BLMs) discretize an LLM agent’s conversation into states, estimate transition probabilities and costs, solve the resulting MDP once, and reuse the optimal policy, reducing model calls by up to 98% and raising task success from 29% to 96% on telecom benchmarks.
The One-Pixel of Reasoning: One Truthful Relation Is Enough to Mislead LLM Agents
AI summaryThe paper introduces the one-relation attack (ORA), which flips LLM beliefs by inserting a single, source‑grounded sentence targeting an implicit inferential warrant. ORA succeeds on 54.9% of proprietary and 80% of open‑weight biomedical models, transfers without feedback, and evades detection.
RAFT: Redundancy-Aware Input Filtering for Temporal Graph Neural Network Training
AI summaryRAFT prunes redundant temporal events in TGNN training by tracking memory changes, filtering unnecessary prediction targets and historical context. It cuts inputs by ~35 % on average (up to 97 %) and training time by ~53 % with minimal precision loss.
HybridResearcher: Bridging the Structured-Unstructured Boundary for Deep Research Agents
AI summaryHybridResearcher is a 4B parameter agent trained via synthesized trajectories and reinforcement learning to reason across structured databases and unstructured web sources. It outperforms much larger models on hybrid research benchmarks without sacrificing single-source performance.
Leveraging Mental Models for Human-Centered Explanations: A Systematic Design Approach Illustrated by Explainable Fake Review Identification
Explainable AI (XAI) often prioritizes model fidelity over user intelligibility, yielding explanations that satisfy technical transparency yet fail to support human sense-making. We propose a human-centered approach, MIND-X (Mental-model INtegrated Design for eXplainability),…
A Cross-Domain Evaluation of Prompting Techniques for Large Language Models
AI summaryThe study benchmarks Zero-Shot, Few-Shot, Retrieval-Augmented Generation, and automatic prompt optimization across six reasoning tasks using a single LLM. RAG tops the overall average (78.33 %) but no method outperforms all others; selection should depend on task characteristics.
Arctic-SWE: Revisiting the Long Horizon Agent Training by Modeling-System Co-Design
AI summaryArctic-SWE introduces an asynchronous reinforcement learning framework for long-horizon software engineering agents through model-system co-design. Using optimized infrastructure and curated data, it enables a 4B model to achieve 65.0% on SWE-bench Verified.
Same Data, Different Bytes: Serialization-Invariant Integrity Checks for Agent-Published Datasets
Agent networks require reliable ways for requesting agents to verify where data is located, how it can be accessed, and whether the data received matches what the data owner made available. Checksum mismatches can reflect either altered content or serialization differences that…
How Well Can LLMs Verbalize Mathematical Formulas? A Multi-Domain Evaluation
AI summaryLarge‑language models can generate reasonably accurate verbalizations of LaTeX formulas, with GPT‑5.2 scoring highest on simple transcriptions (≈76 % BLEU) and Claude Sonnet 4.5 excelling on extended transcriptions (≈36 % BLEU). Accuracy drops for longer, more complex formulas and depends on context and domain.
Stable Signer V3: 4D Deformable Primitive Splatting Is All You Need for Sign Language Production
AI summaryStable Signer V3 uses 4D Deformable Primitive Splatting with face, body, and tunneling super‑resolution modules optimized once per signer, reducing facial error 71.4 %, body error 35.1 %, and RGB error 24.2 % while enabling interactive high‑resolution sign language avatars.
Means-Ends Analysis Helps LLMs Plan in Continuous-State Domains
AI summaryMedaka uses Means‑Ends Analysis to guide LLMs in continuous‑state planning, recursively decomposing goals and tracking state. In Liquidworld, a liquid‑mixing benchmark, Medaka boosts accuracy across model families, showing that classical MEA techniques elicit better continuous‑state reasoning.
What Can't Jev Decide? Probing the Limits of One-Pass Decision Models
AI summaryJev makes one-pass decisions efficiently but cannot match LLMs that write intermediate steps; its accuracy drops when multiple updates are needed and it fails to detect errors from stale state. It succeeds on 97.5% of one-pass solvable items but only 74% of more complex ones.
TrustAgent-RFF: Agentic Diagnostic Planning for Reliable RF Fingerprinting under Domain Shift
AI summaryTrustAgent‑RFF adds a frozen I/Q classifier with a diagnostic framework that measures counterfactual stability, prototype consistency, domain similarity, and signal quality to estimate prediction reliability without target labels, then uses an LLM planner to select diagnostics and a source‑calibrated decision layer to Accept, Reject, or Defer, reducing mis‑predictions under domain shift.
LowNet: Looped Wavelet Neural Networks with 24 Learnable Parameters for Solving Poisson Equations
AI summaryLowNet uses a 24‑parameter looped wavelet neural network that solves Poisson equations with high accuracy (up to 100.74 dB PSNR after three loops) and linear‑time complexity, making it compact and efficient for data‑specific applications.
Epi: Towards a Persistent, Self-Improving Agent for Autonomous Research
AI summaryÉpi is a self‑improving research agent that persists across model sessions, learns from experiments, and produces 120 record‑setting results—including an 84‑fold tighter $n$‑queens bound and a 3.6‑million‑fold Feigenbaum interval—while also proving new combinatorial conjectures. Its lightweight Épi‑turbo variant outperforms Codex on 30 routing/scheduling benchmarks with only 64 % model cost.
Robust Direct Preference Optimization under Preference Label-Flip Attacks
AI summaryThe paper develops defense mechanisms for Direct Preference Optimization against label‑flip attacks, introducing Trim‑K and Soft‑Trim‑K methods that remove or downweight high‑impact pairs. It proves deterministic gradient‑perturbation bounds and shows these defenses achieve uniform policy deviation guarantees under bounded‑region assumptions.
ControlJEPA: Principled Trajectory Regularization via Lyapunov Tube Loss
AI summaryControlJEPA regularizes language model hidden-state trajectories using a Lyapunov contraction condition, penalizing deviations from a reference tube. It improves fine-tuning accuracy and data efficiency across six benchmarks without architectural changes.
CasePlay: Self-Play Reinforcement Learning from Case Reports for Medical Reasoning
AI summaryCasePlay uses document-grounded self‑play on medical case reports, generating questions from knowledge points and evaluating answers with rubric‑based quality feedback. It improves LLM medical reasoning by 3.1 overall points (6.3 on LiveClin) over prior self‑play baselines.
MemPrice: Online Memory Allocation via Adaptive Pricing for Long-Horizon Agents
AI summaryMemPrice allocates limited context to a long‑horizon agent by treating memory slots as a shared resource with an adaptive price that rises with memory pressure, keeping detail only when worth it. The method guarantees that, without horizon knowledge, average allocation regret vanishes under sublinear estimation error, outperforming fixed‑budget baselines across multiple benchmarks.
Revealing what benchmark scores hide: Jointly measuring the breadth and reliability of LLM capabilities using accessibility profiles
AI summaryThe paper defines “accessibility” as the probability a model succeeds across equivalent formulations of a question. It shows that many model pairs have similar average accuracy but diverge on Access@5 vs. Reliable@5, revealing a tradeoff between breadth and consistency of performance.
CoworkEnv: Synthesizing Training Environments for Long-Horizon Cowork Agents
AI summaryCoworkEnv synthesizes training environments guided by workflow constraints, generating requests, tools, and validation checks for AI coworkers. Training with these environments raises Qwen3‑VL‑8B‑Instruct SFT scores from 46.6 to 64.1 on Claw‑Eval, especially improving multi‑turn performance.
Label-Space Mismatch Distorts Cross-Cohort Evaluation in Breast Cancer Molecular Subtyping
AI summaryThe study shows that breast‑cancer subtyping classifiers fail cross‑cohort evaluation when training and test label spaces differ; adjusting the label set (label‑space compliance) improves accuracy by ~6–7 % and reveals that multimodal fusion adds no benefit over mRNA alone.
Towards Human-AI Complementarity for AI Safety
AI summaryThis paper introduces a large benchmark suite to study human-AI complementarity in safety-critical AI oversight. The results show that while strong potential for complementarity exists, current assistance and hybridization methods fail to effectively capture it.
PRECOCIAL: Pretraining with Code Curriculums via Iterative Search for Accelerated Learning
Fully synthetic pretraining data generated by symbolic programs can bootstrap downstream training on naturalistic tasks, such as language modeling, but the generating programs have so far been designed by hand. We introduce PRECOCIAL, a method for automatically optimizing…
CorrectRTL: Boost RTL Generation Quality Through Automated Self-Correction with Instruction Decomposition
AI summaryCorrectRTL decomposes RTL generation into scaffold creation and detailed implementation, using structure-aware decomposition, snippet-based RAG, and SVA-based self‑correction. On the new ScaleBench benchmark it raises Pass@1 from 20.5 % to 49.0 % (+28.5 %) and reaches 96.1 % on VerilogEval.
UtoMe: Observation-Uncertainty-Guided Token Merging for Weather Foundation Models
AI summaryUtoMe introduces an observation-uncertainty-guided token merging framework for weather transformers that combines feature similarity with ensemble data assimilation spread. At a 50 percent token reduction, it significantly lowers forecasting error compared to standard token merging while maintaining reduced computational costs.
UpLoop: Upcycling Large Language Models with Gated Loop
AI summaryUpLoop is a post‑training technique that loops selected layers of a pretrained LLM with hidden‑wise gates, allowing each loop to refine outputs rather than overwrite them. Starting from Qwen3‑4B, UpLoop raises average benchmark accuracy by 6.17 % at R = 3.
ReTrace: Visual Backtracking for Efficient Multimodal Large Language Models
AI summaryReTrace introduces a dormant token state for MLLMs that decouples early computational deactivation from permanent deletion. This enables the model to retrieve information from previously pruned visual tokens during middle-layer reasoning, achieving a 2.83x prefilling speedup with minimal performance loss.
IECLA: Interpretable Emotion Contribution Learning for Affective Image Editing
AI summaryIECLA separates source‑emotion attribution from target‑aware editing, learning signed object‑level contributions via counterfactual interventions and a context‑conditioned predictor. It consistently improves target‑emotion expression while preserving content across six backbones and is validated by VLMs, humans, and user studies.
LLM Agents as Suppliers of Physical System Models: Calibration and Acceptance in Modelica
AI summaryLLM coding agents can generate Modelica models from client specifications and data for both simulated systems and a real magnet, but the handover leaves structural choices open. Acceptance is achieved by requiring the model to belong to a set of acceptable solutions rather than matching a single reference, and specifying residual thresholds improves pass rates.
Synergy-Aware Residual Flow Matching for Single-Cell Perturbation Prediction
AI summarySynGeneFlow predicts single‑cell responses to perturbations by learning a residual flow from an additive prior, incorporating gene‑interaction masks and deformation biases, and aligning distributions with MMD. It outperforms existing methods on genetic and drug perturbation benchmarks, especially for non‑additive synergy effects.
One Head Away from Language Confusion: Single Attention Heads Across Model Families
Large language models sometimes reply in an unexpected language, a failure known as language confusion. We ablate individual attention heads in 12 model checkpoints spanning seven model series, using FLORES-200 continuations in five languages and Language Confusion Benchmark…
VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse
AI summaryVLA-Rolling reuses discarded denoising tails in vision‑language‑action models, adaptively allocating steps and refining with confidence estimates. It cuts inference cost up to 1.48× and boosts success rates by up to 15 % on 16 model‑benchmark pairs.
SteadyThought: Mitigating LLM Under-Thinking via Thought-Level Preference Optimization
AI summarySteadyThought optimizes thought-level preferences to curb excessive switching in large reasoning models, forcing continuations from selected points to improve quality. It reduces token usage by up to 63.8% while preserving accuracy and generalizes across domains and languages.
HCFM: Helmholtz Component Flow Matching for Time Series Anomaly Detection
AI summaryHCFM trains flow-based generative models for time‑series anomaly detection by splitting the vector field into divergence‑free transport, compression, and residual heads; this separation reduces Jacobian cost and improves detection accuracy on large univariate and multivariate benchmarks.
Policy Is Not Capability: Some Safety Benchmarks Measure Which Definition of "Unsafe" You Trained On
AI summaryThe paper shows that safety benchmark scores largely depend on the labeling policy used, not model capability. It demonstrates that minor changes in policy or a few parameters can shift scores across the full range, calling for explicit reporting of contamination and policy definitions.
One Skill Takes Over: Clashing Agent Skills and a One-File Fix
AI summaryThe study shows that coding agents tend to “take over” when two same‑job skill files are present, following only one skill’s rules. Adding a single example that merges both skills dramatically reduces this clash, enabling models to follow up to 99% of the possible rules.
Auto-MD-bench: Evaluating Autoresearch Agents for Real-World Medical Tasks
AI summaryAuto-MD-bench benchmarks 15 real-world medical tasks across 10 datasets, evaluating autonomous ML agents. The study finds that top agents match baseline primary metrics but have poorer risk calibration and robustness to population shifts, showing that single performance metrics are insufficient for clinical modeling.
DAGLoop: Aligning Method Design with Experimental Feedback in Automated AI Research
AI summaryDAGLoop aligns automated AI research with experimental feedback by constructing a directed acyclic graph of design decisions, each paired with a targeted ablation test that validates its expected effect under required conditions. Experiments on numerous research tasks show that this method yields substantial performance gains across diverse settings.
Toward Agentic Retrieval-Augmented Generation for Enterprise Retail Workflows over Heterogeneous Data Sources
AI summaryRetailBridge is a new retail benchmark that simulates complex, heterogeneous data workflows, requiring agents to chain evidence across structured tables and unstructured text. RetailRAG, a framework with schema briefing and source‑routing, outperforms existing baselines and shows multi‑source chaining is the key bottleneck.
Pre-Query Retention with Bounded Bit and Word Retrieval
An agent can retain information before receiving a future question and later retrieve evidence from an archive. Evaluating this arrangement requires accounting for both retained information and the width of each retrieval response. Decision-tree capacity bounds describe…
Explore or Confirm? LLM Agents Misallocate Measurements Under Noise
AI summaryThe paper introduces Wellocate, a benchmark simulating noisy experimental assays, and evaluates five LLM agents that misallocate their limited budgets between exploration and confirmation. Results show over-confirmers waste resources on re‑tests, while under-confirmers ignore repeated checks, both performing worse than simple scripted baselines.
Self-organizing reinforcement learning: a swarm approaches the explore-exploit optimum
AI summaryA minimal foraging swarm balances exploration and exploitation via Levy flights and threshold cascades, reaching an ideal‑free distribution without learning or rewards. It attains ~65% of the oracle ceiling immediately, surpassing some RL methods early but not its final asymptote.
Compile Once, Run Cheap: Frozen, Portable Skills for Embodied Agents, and What Recursive Self-Improvement Adds
Embodied AI lacks the data-rich pretraining that gave language models their agents and skills, so new manipulation capability costs demonstrations per task or a frontier model in every episode. We study a third route: a strong general model acts only as an offline author that…
ClinSDT: LLM-Encoded Clinical Semantic Guidance for Sepsis Treatment via Decision Transformer
AI summaryClinSDT augments a Decision Transformer with LLM-encoded clinical semantics (Profile, Context, Advisory) and a treatment‑response forecaster to filter low‑uncertainty actions. On MIMIC‑III/IV it improves sepsis treatment policies over standard imitation, RL, and Transformer baselines.
TIRE: Top-Down Formula Instantiation with Reliable Execution for Physics Problem Synthesis
AI summaryTIRE generates physics problems by instantiating reusable formulas and conditions, creating reference solutions before generation for reliable supervision. The resulting TIRE‑19k dataset improves fine‑tuned and RL‑trained LLMs by ~3 pp over comparable datasets.
Reflection-Guided Dynamics Rectification for Text-to-Image Generation
AI summaryReflection‑Guided Dynamics Rectification (RGDR) uses reflections of a generated image to intervene at every sampling step, correcting the trajectory of a diffusion model. On fixed budgets, RGDR boosts final checklist scores by 1.51× over TIR and 1.81× over Dynamic CFG.
Spectral Adaptive Graph Transformer for Efficient and Scalable Traffic Forecasting
AI summaryThe paper proposes a Spectral Adaptive Graph Transformer that learns a directed, state‑conditioned spatial operator with linear sensor cost, combined with a sparse diffusion branch and frequency‑bin attention. It achieves state‑of‑the‑art accuracy on six benchmarks while using only a quarter of the memory and training twice as fast.
PLANE: Plug-and-Play Fine-Grained Visual Experts for Hallucination Mitigation in Large Vision-Language Models
AI summaryPLANE introduces fine‑grained LoRA visual experts and an instruction‑aware gate to enrich CLIP features, reducing hallucinations in large vision‑language models. On new GPT4‑free GIFT benchmark and existing tests, it cuts hallucinations with negligible latency, outperforming state‑of‑the‑art methods.
Learning without Overwriting: Structural Knowledge Fusion for Continual Learning
AI summaryKnowledge Fusion Network learns each new task in a lightweight specialist and fuses it into a global model using a novelty‑gated Structural Fusion Module. It achieves 71% accuracy on ten‑task Split CIFAR‑100 while adding only 25% more parameters, cutting forgetting by ~58% versus replay.
M-JEPA: Predictive Self-Supervised Learning for Molecular Graphs with Scaffold-Shift Evaluation on Tox21
AI summaryM-JEPA trains a molecular graph encoder with masked subgraph prediction and an EMA teacher, outperforming InfoNCE on ESOL (RMSE 2.12 vs 3.75) while running faster, and shows mixed but sometimes improved ranking and calibration on Tox21 scaffold splits.
World Models, Trees and Foundation Models for Retail Forecasting: Evidence from Four Public Data Sets
AI summaryThe paper benchmarks multiple forecasting models—world models, LightGBM, TFT, TimesFM, etc.—across four retail datasets at various grain levels (item, store, category, chain). It finds that an average of world model, LightGBM, and TimesFM performs best for most item/store and cross‑store tasks, while model choice varies for higher‑level aggregates.
Memory, Know Your PLACE: Governing Lifecycle State and Query-Time Eligibility in Persistent-Memory Agents
AI summaryPLACE adds write‑side state certification and read‑side authorization to persistent memory, improving lifecycle correctness and preventing causal leakage. Evaluation on Mem0 with Qwen2.5 models shows combined governance yields best performance while preserving relevant support.
Dialogue Generation for Conversational Recommendation Using User Search Interactions and Knowledge Graphs
AI summarySearCl automatically generates 3,538 recommendation dialogues from search and click logs, using a probabilistic state machine and LLM agents guided by knowledge graphs. The generated conversations score higher than benchmark dialogues under an LLM-as-a-judge protocol.
Reranking Answers of Large Language Models with Knowledge Graphs
AI summaryThe paper extracts subgraphs around entity candidates by computing shortest paths in Wikidata, then uses features from these subgraphs to rerank LLM-generated answers. Experiments show the reranker improves answer accuracy over baseline LLM outputs.
ViCoR: Estimating Visual Necessity via Counterfactual Residuals for Multimodal Medical Data Selection
AI summaryViCoR selects image-text pairs that truly need visual input by measuring counterfactual residuals and calibrating with medical supervision; on two backbones and multiple benchmarks, a 20% ViCoR‑selected subset outperforms full-data fine‑tuning and other selectors.
BrandEval: A Two-Track Multi-Agent Benchmark for Risk-Sensitive LLM Crisis Communication
AI summaryBrandEval is a two-track benchmark for LLM crisis communication, with a static strategic analysis track and a dynamic multi‑agent sandbox (BrandPolis) that simulates consumer feedback, competitor actions, and trust trajectories. It shows static quality predicts some outcomes but misses volatile, tail‑risk failures revealed by dynamic tests.
NeuroAct: Advancing EEG Analysis and Decoding via Dual-Loop Autonomous Research
AI summaryNeuroAct automates EEG research with a dual-loop system that tests hypotheses, selects interventions, and refines models using evidence from signal structure, representations, and training dynamics. It boosts macro‑average accuracy by 17.75 points over BrainAgent and improves balanced accuracy and correlation across four EEG tasks.
Calibration-Free Early Stopping for Chain-of-Thought Reasoning via E-processes
AI summaryCalibration-free early stopping for chain‑of‑thought LLM reasoning uses two e‑processes to sequentially accumulate evidence, guaranteeing a user‑specified false‑alarm rate without calibration. The method matches or outperforms calibration‑based policies across models and datasets.
Assessing the Empirical Evidence for Loss of Control from Agentic General-Purpose AI
Theoretical literature warns that advanced generalist agentic systems might pursue instrumental goals that are misaligned with human preferences, thus posing risks of irreversible loss of human control. The EU's Code of Practice for General-Purpose AI lists loss of control as a…
BatchParScale: Sparse Batch-Mixing for Sequence Models
AI summaryBatchParScale adds a sparse communication attention to parallel reasoning streams, enabling them to share information at selected positions. It improves language‑model loss over ParScale, outperforms unmasked baselines in masked pretraining, and works with LoRA fine‑tuning, but retrieval gains remain uncertain.
EG-GRPO: Evidence-Guided Step-level Credit Assignment for Deep Search Agents
LLM-based search agents demonstrate strong capabilities in long-horizon information seeking and reasoning, typically trained via reinforcement learning. However, existing approaches predominantly rely on trajectory-level outcome rewards, providing limited supervision for complex…
Bridging Concepts: Multi-Agent Reasoning over Scientific Concept Graphs for High-Impact Research Prediction
AI summaryThe paper builds scientific concept graphs from conference papers and proposes MiroGraph, a multi‑agent system that designs, extracts, and classifies features to forecast high‑impact concept pairs. MiroGraph outperforms existing impact‑forecasting and temporal link‑prediction baselines.
Vision as Proprioception: Estimating Humanoid Robot Pose from Video
Estimating a humanoid robot's internal state from video has a wide range of real-world applications but remains unexplored. During interactions with a robot, its onboard state information may be unavailable due to privacy, ownership, or communication constraints. Meanwhile,…
World Model Interpretability: From Decoding Representations to Understanding Latent Worlds
Emerging Joint-Embedding Predictive Architectures (JEPAs) are moving beyond representation learning towards latent world models that support prediction, planning, and control. Yet their latent worlds naturally lack a human-interpretable interface. In this Perspective, we argue…
Aligning Language Models with Indigenous Community Cosmovision through Constitutional AI
Large language models (LLMs) are predominantly trained on corpora that reflect Western, anglophone worldviews. This asymmetry is not neutral: when a model answers questions about territory, community authority, or the relationship with nature, it does so from assumptions that…
Learning When to Abort in Transformer Planning: A Mechanistic Study
Planning does not by itself imply the ability to recognise errors, but recognising errors can improve planning capacity. In this paper, we study this phenomenon in Transformer planning. To analyze the inner mechanism, we follow the approach of Wang et al. (2024) and abstract…
Shared-Private Cross-Modal Discrepancy Learning for Imbalanced Multi-Source Audio-Visual Deepfake Detection
Audio-visual deepfake detectors are often trained under a single dataset or manipulation regime, but multi-source training collections may contain heterogeneous attacks and uneven source distributions. Training separate detectors for each source is inefficient, while a single…
AdaFedXplain: Adaptive Federated Learning with Explainable AI for Ransomware Detection
Attacks using ransomware have developed into complex multiphase campaigns that target vital infrastructure, resulting in annual damages of billions of dollars. Conventional detection techniques are ineffective against zero-day variations, and privacy issues are raised by…
Where and Why Search-Augmented LLMs Fail When Reasoning over Rapidly Evolving Knowledge
AI summaryTREK is a framework and benchmark that diagnoses why large language models struggle to reason about fast‑changing facts. It shows that the main failure lies in post‑retrieval reasoning, where models get misled by conflicting evidence rather than retrieval errors.
Formally Verified Autoresearch for Analyzing the Expressivity of Multi-Head Attention
Multi-head attention is a central component of the transformer architecture, yet how its expressivity depends on the number of heads remains poorly understood. We study a one-layer, attention-only transformer with additive positional embeddings and define the head complexity…
Local Evidence, Global Failure: Diagnosing Speech-Vision Temporal Composition in Audio-Visual Language Models
AI summaryThis paper introduces Temporal Evidence Decomposition, a training-free diagnostic framework that separates local visual verification from speech-relative aggregation in audio-visual language models. It reveals that while models struggle with full-video tasks, their local visual evidence recovery dramatically improves when queried with isolated temporal windows.
CHSM8: A Chess Transformer that Follows Natural-Language Style Prompts
We present CHSM8, a 210M-parameter chess transformer that plays moves conditioned on a free-text description of a playing style, e.g. "go for the Sicilian Defense", "solid, safe, risk-free chess" or "You are Magnus Carlsen in his prime". The model is pretrained without search on…
CD-OPD: Contextual Digest for On-Policy Distillation
Teacher takeover in on-policy distillation (OPD) can produce incorrect responses even when the student and teacher independently answer the same problem correctly, a failure we quantify using Collaboration Failure Rate (CFR). To address it, we propose Contextual Digest for…
Aligning Cultural Behaviors in Arabic Language Models via Preferences
Arabic LLMs have improved rapidly, but aligning them with country-specific cultural behavioral preferences remains underexplored. Previous work has focused primarily on instruction-tuning for factual knowledge, overlooking nuanced cultural behaviors. We propose ArabPref, which…
MemDerive: A Diagnostic Benchmark for Agent Memory over Evolving Data
AI summaryMemDerive is a diagnostic benchmark that records exact derivation DAGs for agent memory over evolving data, enabling error tracing at intermediate states. It shows that existing memory systems lag simple retrieval baselines and rarely propagate implicit changes, highlighting the need for gathering, versioned rules, and dependency maintenance.
Outcome-Anchored Credit Reallocation for Agentic Knowledge Graph Question Answering
AI summaryThe paper introduces Outcome-Anchored Credit Reallocation (OACR) to assign fine-grained operation credits from final answer rewards in agentic knowledge graph question answering. OACR improves multi-turn reasoning performance across three benchmarks while reducing inference-time LLM calls and token usage.
When Form Changes but Logic Doesn't: Building Logic-invariant LLMs through Logical Structure
Large language models (LLMs) have achieved strong performance across diverse reasoning tasks, yet it remains unclear whether this performance reflects an understanding of underlying logical structure or a reliance on shallow semantic cues. A key indicator of such understanding…
MOSAIC: Multi-Objective Search for Allocating Inference Cache Across LLM Decoder Layers
The Key-Value (KV) cache is the main memory bottleneck of long-context LLM inference. Token eviction methods such as SnapKV, H2O, AdaKV and L2Norm shrink it, but they give every transformer layer the same budget, although layers differ in which tokens they need. We present…
Stepwise Reasoning Verification: Don't Generate, Classify!
AI summaryThe paper redefines stepwise reasoning verification as natural language inference, classifying each step as entailed or not. It shows that small encoder classifiers (0.15B) outperform most generative verifiers, achieving comparable accuracy with far fewer parameters and annotations.
Enhanced Masked Diffusion Unlearning
Masked diffusion language models (MDLMs) generate text by iteratively unmasking tokens with bidirectional context. As these models reach real-world deployment, machine unlearning becomes essential. This process aims to remove the influence of a designated "forget set" without…
System X: An LLM Decision Model That Replies from Its Own KV Cache
Decision models return probabilities over supplied options, but a deployed agent may also need to reply in words. The present study introduces System X, a large language model (LLM) decision model that adds a text-reply option to the LLM-as-Jev readout. When this option wins, or…
Reliability-Preserving Compression for Deployable Medical Image Segmentation
Post-training compression of pretrained medical segmentation models commonly uses quantization, pruning, or distillation to reduce model size and computational cost. Existing medical segmentation quantization studies mainly evaluate Dice, model size, memory, and latency, but…
DiaLong: Benchmarking Reminiscence in Long-Context Dialogues
AI summaryDiaLong is a benchmark dataset of 100 long‑context dialogues with a fact‑recall task, where models must classify 40 candidate facts as true or false. Results show current LLMs still forget long‑range information, indicating room for memory‑enhancing methods.
On Reasoning Correctness Directions
AI summaryThe study trains a 3‑million‑parameter model on ChainSearch, an arithmetic reasoning task, and uses linear probes to find internal directions that predict reasoning correctness. These directions are statistically significant but do not causally improve reasoning, challenging the linear representation hypothesis.
A failure of Imagination: VLMs fail to use visual representations for text, but self-distillation can help
Vision-language (VL) models are typically built by fine-tuning a language model to attend to visual features in service of a Visual Question-Answering objective. We ask whether this alignment changes how language models represent text---especially on tasks where visualization…
P(Yes): 15$\times$ Improvements via Test-Time Scaling of Decision Models
Decision models answer typed questions, such as multiple-choice or yes/no questions, with a probability for each option. We add test-time search to frozen decision models, with no training: we show the model a complete proposed answer, read off its probability of saying "Yes" to…
Multimodal-Conditioned Latent Diffusion Models for Spatiotemporal Brain MRI Forecasting
Spatiotemporal forecasting of 3D anatomical changes is a critical generative vision challenge for modeling Alzheimer's Disease (AD) progression. While brain MRI forecasting from single-modality data has advanced, other key complementary multimodal biomarkers remain…
PRIMA: Pose Refined Inductive Motion Adapter for Zero Shot Social Bodily Behavior Recognition
Recognizing fine-grained bodily behaviors, such as self-grooming, postural shifts, and backchanneling, is a useful intermediate step for computational social signal analysis. But the behaviors that matter vary across social contexts, and retraining a model for each new taxonomy…
From Proposal to Execution: Modeling Interaction Semantics in Shared Autonomy
In many human-AI systems, candidate actions are not executed directly but are first evaluated by another decision-making entity under a potentially different objective or decision model. Standard decision-theoretic frameworks do not require this interaction to be represented…
Assessing the AI Readiness of Construction Datasets: A Multidimensional Operational Framework
The rapid growth of artificial intelligence (AI) applications in construction has increased the demand for datasets suitable for reliable model development, evaluation, and reuse. However, construction datasets vary widely in documentation, quality, governance, and evaluation…
Formal Raising: Embedding Verifiable Ethical Properties into the Development of AI
Alan Turing proposed constructing a ``child machine'' and educating it rather than directly programming an adult-like intelligent machine. Recent work on model raising revives this developmental view for Large Language Models by arguing that moral values should accompany…
Dynamics-Aligned Representation Learning under Induced Partial Observability
Model-based reinforcement learning (MBRL) has enabled sample-efficient learning of continuous control policies from high-dimensional observations through learned latent dynamics models. These models are typically trained with observation reconstruction objectives, which provide…
When Interpretations Persist: Downstream Influence in Gaze-Aware Assistants
AI assistants may carry forward not only observations but also interpretations of what a person is doing or may do next across time. Once retained, these interpretations become context for later predictions and system-triggered outputs. We study whether this representation step…
Bud ContextCompact 2.0: Context Compaction for Agent Success and Accuracy with Small Language Models and Constrained Context Windows
Agents that work for many turns outgrow the context window of the model that serves them, and on small language models (SLMs) and constrained inference systems that window is small. Context compaction replaces older turns with a summary so a session can continue, but it is…
ATEST: Adaptive Resource-Aware Test-Time Scaling for LLM Inference
Test-time scaling (TTS) improves large language model (LLM) reasoning by sampling multiple answers, but fixed sampling budgets ignore both query difficulty and the changing energy cost of inference under GPU contention. We measure up to a 4.1-fold increase in energy for the same…
When Empathy Misses the Goal: A Benchmark for Goal Displacement in LLM Advice
Conversational large language models can respond with warmth and reassurance while failing to advance the user's stated goal. We conceptualize this failure as $\textbf{goal displacement}$, the substitution of persona-maintenance content for goal-relevant substance. To measure…
StopLoss: Teaching Agents to Predict Failure and Stop Early
Coding agents keep working on runs that will fail verification, and asking a frontier model to watch for this is expensive in its own right. We introduce StopLoss, a small external monitor that learns failure recognition from completed trajectories and their verifier outcomes.…
From Clinical Risk Prediction to Operational Decisions: An Uncertainty-Aware Framework for ICU Staffing
Advances in electronic health records (EHR) and large-scale critical care databases have enabled the development of machine learning models for the early detection of patient deterioration. While these models often achieve strong predictive performance for outcomes such as…
Modaliff: A Unified Token Diffusion Model for Any-to-Any Generation
AI summaryModaliff learns a token-based diffusion model that uses bimodal pairs to keep each modality’s native latent sequence while unifying them into a single token stream with per‑token attributes. It outperforms other any‑to‑any models in fidelity across text, image, and audio with fewer parameters and training data.
Residential Construction Image Dataset Using CSI Divisions for Deep Learning-based Object Detection
Accurate recognition of construction site elements is essential for the application of deep learning (DL) in the construction industry. Despite rapid advances in computer vision, the adoption of DL in construction has been limited by the lack of structured, publicly available…
Auditing Tool Traces in Agentic LLMs: Source Localization and Counterfactual Dependence Testing
AI summaryThis paper introduces a two-stage auditing framework to detect hidden label shortcuts in tool-augmented LLM agent traces. Applying the audit to TCR-pMHC binding reveals that apparent scientific reasoning was actually driven by dataset identifiers rather than genuine evidence utilization.
Editing as Language: Teaching LLM Agents a Grammar for Native Context Editing
A synopsis in novels folds a story into a few lines; a function in code gathers many steps with a single execution. Context editing is native to language itself. In this paper, we propose the Native Context Editing (NCE) framework, which gives language model agents a grammar for…
Symbolic Constraint Agents: Mining Formal Tool-Call Constraints to Prevent Agent Errors
Symbolic constraints (programmatic, deterministic checks for tool calls) enable deterministic verification of LLM agents' tool calls, but existing approaches primarily enforce predefined safety policies rather than tool dependencies in complex tool environments. We explore…
Narrow Decision Models: A System-1 Architecture for Calibrated, Typed Decisions
System One models, introduced with TypeSafe's Jev, answer typed questions about a state with calibrated probabilities instead of generated text: choosing among options described in natural language (Choice), placing the state on an ordered scale (Score), or judging a yes/no…
Learning Which Device to Call: Online Tool Grounding for LLM Agents in IoT
The Internet of Things (IoT) gives large language model (LLM) agents access to physical observations, but one requested capability is usually offered by several devices, and the agent's plan does not say which one to call. Existing work retrieves relevant tools and checks plan…
Design and Cross-Domain Evaluation of an Agent Execution Harness
Completing complex tasks with language-model agents requires coordinating planning, execution, and review so the final result addresses the user's original objective. We present an agent execution harness that keeps the original objective separate from a revisable plan and an…
Benchmark for Structured Retrieval-Augmented Generation in Perovskite Photovoltaics
While retrieval-augmented language models increasingly serve as literature assistants in the physical sciences, existing benchmarks evaluate them primarily on qualitative text overlap or through model-based scoring rather than against verified experimental ground truth. In…
When Better Opportunities Backfire: Service Worker Responses to Parallel AI Deployment in Customer Acquisition
As AI voice agents get deployed in high-volume contact centers, firms can route low-intent leads to AI and allocate high-intent leads to human workers. Drawing on sequential search theory, we posit that AI-induced task allocation reshapes work, enriching opportunities facing…
Safety Offset Learning: When Benign Fine-Tuning Misaligns LLMs
Safety risks from fine-tuning are commonly characterized by the harmfulness of the training data:harmful fine-tuning is considered dangerous, while benign fine-tuning is often assumed to preservealignment. We argue that this view misses a more fundamental factor-the behavioral…
Weighted State Souping for Efficient Long-Context Retrieval
State space models (SSMs) combine fast inference and fixed state size with competitive language-modeling performance. Recently, state souping has been proposed as a procedure whereby an SSM independently encodes chunks of tokens like documents, tool definitions, or memory…
MapSteal: Map-based Autonomous Knowledge Stealing Attack Against Large Language Models
Knowledge acquisition from black-box large language models (LLMs) offers a practical approach to adapting smaller student models. However, many existing methods rely on predefined queries or student feedback to guide acquisition, leaving autonomous expansion of knowledge targets…
Selecting and Composing Agent Design Patterns: Industrial Case Evidence and Controlled Reference-System Experiments
Agent pattern catalogs describe architectural alternatives, but engineers also need evidence for selecting and composing them. We extend a cognitive-function--execution-topology framework with an evidence-driven selection procedure and explicit composition contracts. The study…
The Forward/Inverse Duality of Controlled Emergence:A Computational Design Program from Explanation to Engineering
The macroscopic capabilities of intelligent systems arise from the local interactions of many simple units. When such systems are deployed in the real world, their behaviour must nevertheless be constrained, leaving designers caught between releasing adaptability and preventing…
ACorD: Adaptive Correspondence Discovery for Medical Image Registration
Medical image registration is a fundamental task for aligning anatomical structures across images. Exhaustively searching for point-wise correspondences is prohibitively expensive in both computation and memory, particularly for high-resolution 3D medical images. Most recent…
Think, Compress, Answer: Reinforcement Learning for Multimodal Grounding via Reasoning Compression
AI summaryTCA splits multimodal grounding into reasoning, target-aware compression, and answer generation, training with staged GRPO rewards that keep only target‑relevant info. It yields large zero‑shot gains on six grounding benchmarks by compressing reasoning into minimal context.
Spherical Fixed-Point Reasoners: Adaptive Loop Halting on the Hypersphere
Looped (weight-tied) transformers enable adaptive computation by allocating more iterations to harder inputs. Fixed-Point Reasoners (FPRM) achieve this with a damped fixed-point solver, but operate in an unbounded Euclidean residual space with coupled rescaling and relative-norm…
Consistency Training Suppresses Cue Following Without Teaching Cue Selectivity
Language models often change answers because of a cue, a hint such as "I believe the answer is B" that gives no valid evidence. Consistency training (CT) trains a model to answer the same with and without the cue. Four CT methods apply this idea to output tokens (BCT), residual…
A "feature ODE" describing the learning behavior of shallow MLPs on simple functions
AI summaryThe paper introduces a “feature ODE” that approximates the gradient‑flow dynamics of two‑layer MLPs on anisotropic Gaussian data, showing that trained networks on simple targets (Hermite or staircase functions) largely follow the ODE’s loss trajectory, an unexpected but useful correspondence.
Beyond Leaderboards: A Method for Application-Specific Capability Profiling of Large Language Models
AI summaryThis paper introduces a containerized testbed for application‑specific capability profiling of LLMs, separating model, judges, and agents, and providing versioned, sparse metamodels. It finds that model choice matters more than framework choice and that instruction tuning is deployment‑specific.
HyperTaxa: Multiscale Genome Representations for Generalized Microbial Taxonomic Inference
Reliable microbial taxonomic inference is challenged by the rapid growth of newly sequenced organisms and incomplete reference coverage. While alignment- and database-driven tools remain essential for reference-backed identification, they are not designed for learned inference…
LIRQuant: An Efficient Locally Isolated Rotation Quantization Method for Large Language Models
Post-Training Quantization (PTQ) enables efficient deployment of Large Language Models (LLMs), yet activation outliers remain a key challenge for low-bit quantization. Existing global transformation methods overlook heterogeneous channel sensitivity to quantization noise,…
Neither Over- nor Under-Constrained: Evaluating LLM Personalization under a Constraint Qualification Audit
Accurate evaluation of LLM personalization is hindered by current evaluation paradigms that over- or under-constrain the response space, yielding frequent false rejections and false acceptances. To reduce these false rates, we introduce ConQual-Audit, an audit-based framework…
Cross-Pollination Retrieval: Finding Transferable Ideas Across Fields in Scientific Literature
Many scientific advances come from cross-pollination, carrying an idea from one field over to a problem in another. Yet retrieving papers across fields remains underexplored, with existing benchmarks confined to a narrow range of fields. We study cross-pollination retrieval, the…
BitFlat: Error-Localized Bitstream Design for Robust Diffusion Image Compression
Zero-shot diffusion image compression based on reverse channel coding achieves high perceptual quality at low bitrates without retraining, yet its bitstream remains vulnerable to bit flips. Prior work has shown that a joint combinatorial representation can amplify a single bit…
When Safety Fails Collectively: Robust Alignment Against Harmful Fine-Tuning via Joint Weight Perturbations
Harmful fine-tuning can rapidly erode the safety alignment of large language models, creating substantial risks when aligned models are adapted to downstream tasks. Existing alignment-stage defenses use harmful examples and simulated harmful perturbations during safety alignment…
MT-CC: Multi-Group Temperature Scaling for Asymmetric Calibration Behavior in Class-Incremental Learning
As deep neural networks are required to continuously adapt to evolving data, class-incremental learning (CIL) has become an active research area. However, most existing works primarily focus on improving accuracy while overlooking confidence calibration, which measures the…
Progress-Guided Preference Generation for Aligning Generative Control Policies
Generative policies have achieved strong performance in imitation learning tasks, but their success is often constrained by the limited availability of expert demonstrations. Although preference-based fine-tuning can leverage suboptimal rollouts to improve performance, existing…
ToolOPD: Credit-Aware On-Policy Distillation for Tool-Use Agents
On-policy distillation (OPD) enhances the capabilities of small language agents by providing dense, token-level supervision from a teacher model on student-generated trajectories. However, standard OPD implicitly assumes that credit is uniformly distributed across long-horizon…
CVaR-RMD: CVaR Selection Pressure for Robust Metaheuristic Discovery
Large language models have recently been used to discover heuristics through iterative program search, offering a promising approach to automated algorithm design. However, candidates selected by average performance may degrade across instance families, random seeds, and problem…
From Black Box to Controller: Inference-Time Steering of LLM via Feature Amplification
Inference-time steering is crucial for deploying Large Language Models (LLMs) in real-world applications, where they must flexibly adjust behavior preferences. For instance, a model should dynamically adjust its caution level based on the risk of the scenario. This requirement…
Whose Credential Is It? How Owner Attribution in Agent MemoryShapes Cross-User Account Misuse
Shared long-term memory lets an LLM agent serving one user inadvertently retrieve another user's credentials, but it is unclear when such exposure leads to actual harm. We introduce a three-stage audit that separates three events: a foreign credential reaches the prompt…
Causal Humility: Governing the Authority of Fairness Claims Under Causal Uncertainty
AI systems increasingly shape consequential decisions in employment, lending, healthcare, and public services, making fairness an important problem for auditing and governance. Fairness assessment often requires more than identifying disparities in outcomes; it also requires…
PFedQFT: Personalized Federated Quantum-Inspired Fine-Tuning for Medical Large Language Models
Medical data remain fragmented across institutional silos because of privacy, governance, and ownership constraints. Federated learning enables medical institutions to adapt large language models (LLMs) collaboratively without transferring raw records. However, independently…
Fine-Grained Units, Flat Decisions: Hierarchical Micro-Expert Pruning for Mixture-of-Experts Models
Micro-expert pruning reduces the capacity of mixture-of-experts (MoE) models by removing coupled channels within routed experts. However, finer pruning granularity alone does not guarantee the preservation of downstream capabilities, as similar language-modeling perplexity can…
Corrected Block Sparse Spectral Optimization: Exact Models, Finite Identification, and Local Dynamics
The block generalized-power method pairs threshold-based support selection with inexpensive Stiefel updates, but the standard block formulation conflates two nonequivalent objectives and leaves inactive components, threshold ties, and post-selection refitting ill-specified. We…
Anatomy of Off-Policy Policy Gradient: Importance Sampling, KL Regularization, and Baselines
Reinforcement learning has become a central tool for large language model (LLM) post-training, where policy gradient methods are routinely deployed off-policy, even though vanilla policy gradient assumes on-policy sampling. We study when such off-policy deployment is justified,…
No Matter Who You Are: Addressivity and Sycophancy in Large Language Models
AI summaryThe paper identifies a link between “addressivity” (the model’s tendency to direct outputs toward an addressee) and sycophancy, showing that post‑training strengthens this connection and that any recognized addressee triggers the effect.
TimelineAgent: An LLM-based Framework for Query-Centric Event Timeline Extraction Across Multiple Documents
Query-centric cross-document event timeline extraction aims to construct event timelines from multiple documents under diverse query focuses. The task is very challenging due to the heterogeneous event modeling and reasoning requirements arising from both the intrinsic…
A Calibrated Instrument for Detecting Disposition Drift in Language Models
Before language models are released, they are tested for harmful traits such as willingness to carry out dangerous requests, deception, power seeking, and sycophancy. The safe behavior these tests certify may degrade as models are modified, by topical fine-tuning, new training…
Cognitively Decoupled Meta-Induction for Out-of-Distribution Generalization
In-context learning (ICL) tackles downstream tasks conditioned on demonstrations. This paradigm relies heavily on neighbor-based matching between queries and demonstrations, limiting out-of-distribution (OOD) performance. For better generalization, rule-based inductive reasoning…
Remember Less, Adapt Longer: Bounded-State In-Context Reinforcement Learning with Recurrent Summary Memory
AI summaryRecurrent Summary Memory (RSM) is a Transformer policy that keeps a fixed-size working buffer and a compressed summary of past experience, allowing in-context reinforcement learning with constant memory and per-step cost. RSM sustains adaptation over eight times the training horizon using only 1 % of full-history memory, outperforming sliding-window, full-history, and other recurrent baselines.
Computation, Not Memory: Chain-of-Thought and State Tracking in LLM Agents
Language-model agents often reason before they act, and the value of monitoring that reasoning depends on how it relates to their actions. We ask how an agent's actions depend on its reasoning when the correct action depends on a changing state. In a small file-system task,…
ActTime: Action-Conditioned World Modeling for Time Series
Time-series models increasingly condition on known future inputs. In controlled systems, some of these inputs are actions that can be changed, compared, and optimized before execution. A model must therefore do more than predict the realized future: it must predict how outcomes…
TRACE: Tracking Multimodal Clinical Trajectories for Inpatient Risk Prediction
Accurate inpatient risk prediction requires modeling heterogeneous clinical evidence that accumulates irregularly over time. This challenge spans in-hospital mortality, prolonged length of stay, and readmission, each with task-specific observation windows. Existing multimodal…
EvalRx: Agentic Diagnosis and Verified Repair of Foundation Models
Model evaluation reveals where models fail, but rarely why or whether a proposed fix genuinely works. We introduce EvalRx, an open-source agentic framework for statistically grounded model diagnosis and repair. EvalRx discovers systematic failure modes, explores candidate…
From Trace to Fix: Residual Task State in Long-Horizon Agents
Recovering from failures in long-horizon, stateful language agents is challenging because the action introducing a defect, the symptom exposing it, and the persistent state requiring repair frequently diverge. Existing methods largely focus on retrospective failure attribution,…
Standing on the Shoulders of Giants: Enhancing LLM Agents for Automated Fact-Checking with Human Expert Skills
AI summaryWe distill AFP fact‑checking guidelines into a modular “Professional Fact‑Checking Skill” and give LLM agents dynamic access to it at inference. SkillCheck improves accuracy by 2.2–4.8 pp and Macro‑F1 by 1.5–6.3 pp on AVeriTeC, and with GPT‑4o‑Mini surpasses GPT‑4o while cutting inference cost by ~61 %.
OpenSaC: A Python Library for Search as Code in Deep Research Agents
Deep research agents investigate complex questions through iterative search, source examination, and evidence synthesis. A conventional workflow alternates reasoning with individual search and reading actions. Search as Code introduces a new paradigm: agents write programs that…
Can LLM User-Proxies Stand In for Humans in Reliance Measurement? Replicating Rel-A.I. with Simulated Users
AI summaryLLM user‑proxies fail to reliably mimic human reliance on AI answers; the best proxy shows a 22.5‑point absolute-rate error and 8.3‑point domain‑effect error versus humans. The audit demonstrates no firm validity for proxy‑derived effect sizes.
SAR: Learning Global Item-Space Matching and Geometry-Aware Placement Ranking for Buffered Online 3D Bin Packing
AI summaryThe paper introduces Spatial Affordance Retrieval (SAR), a framework for online 3D bin packing that decouples global item-space matching from geometry-aware placement ranking. SAR improves space utilization over strong baselines by up to 4.20 percentage points across standard datasets.
AirMPA: A Meteorology-to-Pollution Adapter for Global Air Quality Forecasting
AI summaryAirMPA decouples meteorology from pollutant forecasting with a dual‑branch Swin‑Transformer, predicting 13 global pollutants at 0.4° resolution. It outperforms Aurora and beats CAMS in ~90% of cases, surpassing CAMS on all variables beyond 48 h over a 5‑day horizon.
When to Stop Asking: Factor Coverage Bias and LLM Determinability Judgments with Underspecified Information in Medicine
Clinical decision-making often begins with underspecified information and requires not only gathering missing information but also knowing when to stop. This calls for judging determinability (whether the available information is sufficient for a decision) by considering whether…
Optimizing Marketing Budgets for Social Influence: A Generalized Fractional Incentives Framework
AI summaryA method for allocating a marketing budget continuously across social network users, accounting for each user’s response curve and cost. The greedy algorithm guarantees near‑optimal influence and outperforms heuristic spreads in large real networks.
When Retrieval Partially Conflicts with Skills: Understanding Procedural Hallucination in Language Models
Language models increasingly use skills that provide reusable instructions and procedures for task execution. After a relevant skill has been selected, the model may need to apply its governing procedure together with retrieved task context. When the retrieved context is useful…
FrontierRefactor: Benchmarking Multi-Scale Performance-Oriented Codebase Refactoring
Autonomous codebase refactoring requires coding agents to restructure working software while strictly preserving externally observable behavior and prioritizing execution performance and compute efficiency over stylistic preference. Existing benchmarks: refactoring benchmarks…
Diffusion Language Models Change Speed Across Tasks and Languages
AI summaryDiffusion language models vary decoding cost across tasks and languages; generating new, rare content requires many forward passes while copying or repeating existing reasoning is cheap. Cost correlates with correctness and can guide efficient inference.
Where RL Breaks: Localizing and Repairing Failures via Targeted Continued Pretraining
AI summaryThe paper shows that RL failures in language models are localized to a few characteristic n‑grams. By retrieving and continued pretraining on just 35M tokens of failure‑related text, it consistently improves reasoning accuracy, outperforming random pretraining data across multiple models and tasks.
OV-PEAR: Observation-Verified Pose-Scale Estimation and Reconstruction for Unknown Industrial Manipulation Objects
Fine industrial robotic manipulation requires 6D pose, scale, and an operation-ready 3D model. Existing industrial pose estimation relies on CAD models, object-specific training data, or stable workcells, limiting use with unseen objects; zero-shot methods suffer from…
Not Every Rollout Needs Distillation: Correctness-Filtered On-Policy Self-Distillation
On-policy self-distillation (OPSD) supervises student-generated trajectories with a reference-conditioned version of the same base model. We ask whether every trajectory supplies distinct supervision that warrants its distillation cost. Gradient measurements on Qwen3-1.7B at…
A Cue's Hidden Price: Auditing Seller Influence on AI Shoppers
Large language models are increasingly asked to shop on a user's behalf, which lets seller-controlled content such as badges, endorsements, and framing steer what they recommend. That poses the question at the core of generative engine optimization (GEO): not whether seller…
SWIFT: Self-guided Weak-to-strong dIFfusion models via Target extrapolation
Training high-quality diffusion models is computationally expensive, and inference-time guidance often adds further sampling cost. In this work, we introduce Swift, a simple yet effective self-guided training method that uses a frozen weak denoiser to improve a stronger model…
Measurement Validity and Representation Dynamics of Confidence Signals in Large Language Models
AI summaryThe paper audits Qwen2.5 models’ confidence signals (verbal, probe-based, sampling agreement) on retrieval and reasoning tasks, finding probes signal agreement earlier for retrieval, verbal confidence unstable across prompts, and sampling agreement offers higher resolution despite higher ECE.
Rubrics Beyond Evaluation: Exploiting Rubric Structure in Reinforcement Learning
Rubrics are widely used to evaluate open-ended question answering by breaking overall quality into explicit criteria. These explicit criteria make rubrics a rich source of learning guidance, yet their potential for guiding the training process has been largely overlooked.…
MASIF: Implicit Failures for Process Observability in Multi-Agent Systems
Multi-agent systems (MAS) can complete a task despite local execution errors that are corrected, bypassed, or carried forward without individually determining its outcome. We call these implicit failures (IFs) and study them as signals for process observability beyond final task…
Are Self-Evolving Agents Continual Learners?
Large language model agents are often deployed in environments where new knowledge, tools, and workflows keep arriving. Such an agent is therefore expected to adapt continually, learning new tasks as they arrive without forgetting earlier ones. This resembles the continual…
Adaptive Decoding Methods for Token-Efficient Reinforcement Learning
Rollout decoding affects both which trajectories are used for reinforcement learning with verifiable rewards (RLVR) and how many completion tokens are spent collecting them. We study whether candidate-token support should remain fixed or widen during training in the context of…
Minimum-Cost Robust Evidence Certificates for Horn Reasoning under Correlated Source Failures
A small justification for a rule-based conclusion may disappear when a source is withdrawn, and multiple derivations need not provide independent evidence. We study minimum-cost certificates that preserve a Horn conclusion under every withdrawal of at most k source domains.…
Benefit-Targeted Routing and Bounded Revision for LLM-Based Evidence Screening in Manufacturing Supplier Discovery
An evidence-review system must allocate additional model judgments and control their use. We study these software decisions on 6,164 frozen manufacturing source-task pairs. Benefit-targeted routing learns recovered agreement minus newly introduced disagreement. On the original…
CODA: Cohort- and Drift-aware Foundation Model for Multimodal Clinical Reasoning
Recent advances in multimodal foundation models have demonstrated strong performance across diverse clinical tasks. However, existing approaches predominantly operate at the individual sample level and fail to model the interplay between population-level cohort patterns and…
Manifold-Constrained Residual Adaptation with Frame-Adaptive Layer Aggregation for Speaker Verification
Large self-supervised speech models provide powerful representations for speaker verification, but full fine-tuning is expensive. Most conventional adapters insert small trainable modules to modify the output of each layer, while keeping the original residual connections…
Motor compound fault diagnosis using an attention-based cross-talk architecture
As physical AI increasingly relies on sensing to understand physical systems, reliable condition monitoring of industrial machinery has become more important. Motors are ubiquitous in industrial systems, and increasing system complexity raises the likelihood of simultaneous…
When Should AI Chime In? Benchmarking Timely Contribution of Proactive Agents in Multi-Party Chat
Proactive agents can support group collaboration by surfacing useful information without waiting to be asked, but even relevant contributions can be disruptive if poorly timed. In this work, we isolate the timing question: given a fixed candidate contribution such as a…
Secure Agentic Coding through Counterexample-Grounded Feedback
As coding models and agent harnesses become more capable, reliable security verification is becoming a central bottleneck. Improving secure code generation requires a signal whose reported progress remains aligned with actual program behavior under optimization; otherwise, an…
UtilityBench: Measuring Language Model Utility When Errors Have Different Costs
As language models are increasingly deployed in real-world scenarios with important outcomes, the impact of their mistakes often extends far beyond what correctness can capture. In domains like medicine, a single error can lead to serious consequences and put patient safety at…
TESS: LLM-guided Translation, Exploration, Simulation, and Search for Embodied Task Planning
In this work we focus on the problem of embodied task planning, which requires an agent to accomplish a complex task by separating it into atomic actions. Recent studies have developed LLM-based methods to address this problem. However, the executability and correctness of the…
EvoSkill v2: Fully Agentic Skill Evolution from Benchmark Adaptation to a Stateful Evolver
EvoSkill v2 is a three-agent system that edits the instructions (skills) an agent is given, leaving its model and harness intact. The user starts a run by describing a benchmark and supplying a grader, which a benchmark adaptor turns into a benchmark-specific plugin. A solver…
Keep the Lineages Alive: Diversity-Preserving SMC for Diffusion LLM Decoding
Inference-time sequential Monte Carlo (SMC) decoding can improve the reasoning accuracy of masked diffusion language models (diffusion LLMs) without post-training by resampling a population of partial generations according to a reward. Its effectiveness therefore depends on the…
The KV-Surgeon: Compensation-Aware Offline KV-Cache Compaction
Offline compaction of the key--value cache (KV-cache) reduces the cost of repeatedly querying a fixed context by constructing a smaller cache once and reusing it subsequently. The central challenge of compaction is deciding which cache entries to delete and which to retain.…
Global probabilistic forecasting of air pollution
AI summaryGenerative model “Foehn” learns global km-scale PM2.5 estimates to forecast air quality; over a year it outperforms operational forecasts at four stations, detecting local pollution extremes four days ahead with calibrated ensembles.
Policy Search in Code Space for Multi-Agent Pathfinding
Developing coordination policies for complex multi-agent systems relies on either labor-intensive algorithm design or opaque task-specific reinforcement learning. With recent advances in autonomous coding agents, we show that policy search directly in code space offers a…
Claim Laundering: When Agent Decision Models Trust Records Over Reality
Agent decision models select actions in a single forward pass from inputs that combine environmental observations with the agent's own record of previous actions. When an action fails silently, these sources can conflict: the record reports success, but the environment shows…
OrderGPT: Generation Order as a Learnable Modeling Choice
Generative modeling of discrete data has two popular paradigms: autoregressive models such as GPT, which factorize a sequence in a left-to-right (L2R) order, and masked diffusion models, which are essentially any-order autoregressive modeling. Rather than competing methods,…
AdaExploreBench: Benchmarking LLM Agents for Proactive Exploration and Test-time Adaptation in Evolving Environments
AI summaryAdaExploreBench evaluates LLM agents on proactive exploration and test‑time adaptation in environments with hidden, drifting rules. GPT‑6 Sol attains only 61.9% accuracy, and performance improves by ~10% when agents receive evidence or detect rule drift, highlighting active evidence acquisition as a bottleneck.
CXBench: A Benchmark for Process-Compliant LLM Agents in Regulated Customer-Experience Workflows
AI summaryCXBench evaluates language‑model agents on both outcome correctness and strict process compliance in six regulated customer‑experience domains. It finds that 55.5% of 840 runs are false‑passes, with only 13.8% satisfying both outcome and process checks, reshuffling model rankings.
Productive State-Space Exploration for Multi-turn RL in LLM Agents
Reinforcement Learning (RL) for Large Language Model (LLM) agents in multi-turn, long-horizon environments remains challenging due to inefficient action selection and repetitive behavioral patterns. In particular, agents often become trapped in high-confidence yet unproductive…
EmbodiedWorld-200K: Toward Embodied World Agents via Embodied-Planning Dataset and Interactive World Models
Building embodied agents that can perceive, reason, and act across diverse 3D worlds is a central research direction toward artificial general intelligence. Yet most existing work remains confined to specific tasks in a single scene and transfers poorly to new environments,…
PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models
Fast and reliable manipulation is crucial for practical robot deployment. Although Vision-Language-Action (VLA) models have recently shown remarkable progress as generalist robot policies, their execution often remains temporally inefficient. A key source of this inefficiency is…
ABACUS: Program-Guided Multistep Reasoning over Frozen Judgment Models
Judgment models answer semantic questions by scoring a finite set of candidate outcomes. Can reusable programs turn these local decisions into powerful multistep reasoning? We introduce \method, a framework that jointly constructs judgment schemas and composition programs for…
Behavioral Probes for Information Flow in LLM Swarms
Most existing approaches to LLM swarm optimization are utility-driven, focusing on final task utility while leaving the processing, integration, and reliance of information within individual interactions largely implicit. To address this, we study LLM swarm interactions from an…
Reconstruction and Discrete-Event Simulation Assessment of an Analytical
This article presents an analytical-DES benchmark for intra-aisle robotic picking in automated warehousing. The investigated cycle combines horizontal shuttle travel, tote handling and item picking at storage locations. Simulation results are compared with the analytical…
GLOBE: Geometry-Led Online Behavioral Early Exit via Unit-Sphere Trajectories
AI summaryGLOBE stops chain-of-thought reasoning early by embedding reasoning segments on a unit sphere and monitoring trajectory shape; it cuts tokens by ~50 % and latency by ~51 % while retaining 80.6 % of full‑CoT accuracy across five tasks.
Correctness-Dependent Asymmetric Target Distributions for Self-Distillation Policy Optimization
AI summaryAPSD mixes student and self‑teacher distributions based on rollout correctness, preserving good behavior on correct rollouts and sharpening corrections on incorrect ones. It matches or exceeds SDPO performance while reducing training steps by 2.6× on Qwen3 models.
On the Benefits of Uncertainty Estimation for Species Classification: A Case Study on Optimizing Expert re-Annotation
Automated identification of diatoms from microscopic images is a cornerstone of large-scale freshwater biomonitoring, yet the high diversity of species and strong class imbalance pose significant challenges for reliable classification. In this study, we develop a ResNet-18…
GraphEvoResearcher: Graph-Aware Training-Free Self-Reflective Reasoning for Cost-Bounded Multi-Hop Question Answering
Multi-hop question answering (QA) requires models to chain intermediate conclusions while remaining cost-aware. Existing training-free self-reflective protocols (generate -> critique -> revise) improve accuracy but operate on a flat token stream: revisions cannot identify which…
ClueFlow: Learning Long-Horizon Search through Attributed Research-State Transitions
Long-horizon search requires agents to acquire relevant evidence and integrate it into an evolving research state. Existing methods maintain research states to guide exploration or assign turn-level progress signals, yet grounding these signals in explicit, attributable state…
Auditor-Assisted Summary-Channel Verification for Hosted LLM Identity Substitution
Hosted LLM services increasingly expose model identity as a product claim, but external users and auditors often cannot inspect which model configuration served a request. We study an auditor-assisted regime in which a trusted auditor enrolls claimed identities and prepares…
NIOCIS: Typed-Decision Next-Hop Routing with Node-Conditioned Option Spaces for Tool-Use Agents
Tool-use agents pay per decision: at each hop a language model re-reads the state and thinks for seconds, though its actions are finite and decisions recur. We present NIOCIS, a routing layer that folds execution logs into a loop-collapsed action graph (forward transitions plus…
Map-Distribution Mixture-of-Experts for Multi-Agent Pathfinding
AI summaryThe paper builds a Mixture‑of‑Experts router that selects between map‑specific neural policies for multi‑agent pathfinding, improving over a single pre‑trained policy and beating existing baselines, though a well‑trained monolithic model still slightly outperforms the MoE.
Procedural Memory for Language Agents: A Neurosymbolic Perspective on Explicit Goal Structure
Large language models remain unreliable on long-horizon tasks requiring procedural reasoning. We argue that a key cause is the absence of explicit procedural memory: accessible, modifiable knowledge of how tasks are accomplished, distinct from semantic and episodic memory.…
TimesJev: Thinking Fast on Time Series with System 1 Sufficiency
Large language models (LLMs) have recently shown strong potential for time series reasoning, while models that reason over time series are usually with long chains of thought. However, this process often consumes thousands of tokens, directly leading to substantial latency…
Accelerated Drift of Scientific Prose in the LLM Era: Evidence from an Interrupted Time-Series Analysis
Every scientific text is embedded in the writing norms of its time: it reflects earlier publications while also shaping the texts that follow. The growing use of LLM-based writing tools introduces a new factor in this evolving system, but the linguistic changes coinciding with…
Understanding When Reflective Prompt and Skill Optimization Helps on Locally Served Reasoning Models: A Two-Model Study of GEPA and SkillOpt
Reflective text-space optimizers such as GEPA and SkillOpt have reported large gains on frontier and mid-size models. We test replication across two reasoning-native target-serving configurations: DeepSeek V4 Flash through DwarfStar and Qwen3.8-27B through LM Studio. For each…
FrameGRPO: Incentivizing Video Reasoning via Multi-Frame Rollouts and Temporal Credit
AI summaryFrameGRPO improves video reasoning for multimodal LLMs by using multi‑frame rollouts that sample different temporal views and a gated temporal credit that rewards short‑frame trajectories only when full‑frame references fail, yielding better performance on video benchmarks.
Multimodal Decoupling for Diversified Recommendation through Semantic and Structural Shaping
Recommender systems often suffer from over-concentrated results when optimized for accuracy, limiting recommendation diversity. While diversity has become an important objective, existing methods still face fundamental limitations from both semantic and structural perspectives.…
TamedFlow: Trajectory-Aware Distillation\\for Teacher-Assisted Diffusion Sampling
Diffusion Transformers (DiTs) deliver high-fidelity image and video generation, but inference requires many sequential model evaluations. Two acceleration strategies are increasingly important: model compression makes each evaluation cheaper, while adaptive inference avoids…
AuDQ: Auditable Data Quality through Agentic Rule Discovery
AI summaryAuDQ introduces a repository-native, ontology‑grounded data quality layer where rules are versioned, provenance‑linked, and replayable. Their agentic rule‑discovery engine (Full‑Agentic and Guided modes) generates rules matching expert‑specified quality and diversity on finance and retail datasets.
Decision4Rec: A Jev-Inspired Framework of Decision Models on Knowledge and Active Graphs for Multi-Agent Recommender Systems
Decision models, such as the commercial Jev, are language models that return one probability for each option of a question instead of generating text. We present Decision4Rec, an open and modular framework that makes decision models useful for recommendation. A knowledge graph…
Honeyed Words, Hidden Daggers: Benchmarking and Training LLMs to See Through Mock Politeness
The Chinese idiom "kou mi fu jian" ("honeyed words, hidden daggers") names speech that is polite in form but hostile in function. We present ZhenhuanBench, a 1,592-item benchmark built from expert commentary on Empresses in the Palace, covering speaker intent, emotion-language…
FinSafeGuard: A Quality-Driven Synthetic Dataset for Financial AI Safety
Generating high-quality synthetic datasets at scale remains a fundamental challenge for training specialized AI systems in data-scarce domains. We present a comprehensive methodology that produces 14.3 million quality-filtered synthetic training examples from a structured…
CUA-Lite: An Open Platform for Computer Use Agents
Open resources for computer use agents (CUAs) -- environments, datasets, and models -- remain scattered: without a standard, these resources cannot be integrated and scaled; additionally, evaluation, RL, and SFT cannot share one infrastructure. We present CUA-Lite, an open…
Not All Similarity Is Leakage: Distribution-Aware Identity Leakage Assessment in Synthetic Face Datasets
Synthetic data provides the means for more ethical and less biased facial recognition. Nonetheless, the limitations of generative models and the conditioning process they rely on pose a risk of identity leakage, resulting in synthetic samples that are overly similar to those…
How Robust Are Prognostic Radiomic Features? Comparing Selection Techniques on Pooled Glioblastoma Cohorts
Radiomic survival studies in glioblastoma report heterogeneous prognostic features, but how much of this stems from the methodology itself is rarely examined. We analysed the association of 463 radiomic, anatomical and clinical features from pre-treatment MRI with overall…
From Code to Playable Systems: CEGame-Bench for Long-Horizon Game Creation Agents
A game ships when a player can launch it, finish it, and return to a saved world, none of which a successful build guarantees, yet game-creation agents are mostly tested on one engine or on one-prompt browser games. We present CEGame-Bench, the first long-horizon benchmark for…
Reliable Before Profitable: QuantHarness for Governing Financial Strategy Development with LLM Agents
AI summaryQuantHarness introduces a two-layer agent framework to govern financial strategy development by enforcing strict workflow boundaries and maintaining verifiable execution records. It achieves a 90.72 percent task success rate on a new reliability benchmark, outperforming existing systems while preserving auditable development evidence.
Don't Rewrite the SID Map: Topology Preservation for Generative Retrieval in Dynamic E-Commerce Search
Generative retrieval assigns each item a discrete semantic identifier (SID) and retrieves items by autoregressively generating SID sequences. In dynamic environments such as e-commerce, new items arrive continuously, requiring periodic updates to the SID-to-Item mapping and…
Nash-MOPD: Optimizing Task Trade-offs in Multi-Teacher On-Policy Distillation
AI summaryNash-MOPD schedules teacher sampling by solving a local minimax game over reverse‑KL statistics, giving more prompts to domains that improve slowly. It outperforms fixed and D³‑MOPD, achieving higher math and code scores on DeepSeek‑R1‑Distill‑Qwen‑1.5B and Qwen3‑8B‑Base.
MoveBench: A Rubric-Based Benchmark for Evaluating Large Language Models in Sports and Health Coaching
Despite rapid advances in large language models (LLMs), their deployment in the sports and health domain remains high-risk due to the absence of domain-specific evaluation benchmarks. Existing general-purpose benchmarks (e.g., MMLU and HELM) primarily assess broad knowledge and…
On Counting Graphs for Modal Logic with Counting
AI summaryThe paper analyzes the complexity of counting trees and graphs that a graph neural network classifies positively, equivalently satisfying a modal logic formula with counting, and investigates convergence properties of this counting process.
VEILED-SH: Vulnerability Expression through Indirect Language Evaluation Dataset (Self-Harm)
AI systems may miss indirect, culturally situated distress or provide support mismatched to the application context. Existing benchmarks do not jointly address cultural variation, application context, and practitioner disagreement. We introduce VEILED-SH, a benchmark and…
GeoEdit: Adversarial De-Evidencing for Silent Geoprivacy Protection via Image Editing
Recent efforts have demonstrated the strong capabilities of Multimodal Large Reasoning Models (MLRMs) in geographic reasoning. By jointly analyzing subtle visual cues, these models can infer the locations of photos, thus enabling doxing even by non-expert adversaries. Existing…
LLMs Prompted for Legal Context Refuse More: Overrefusal from Small On-Premises LLMs in Criminal Legal Context
Confidentiality requirements increasingly push legal professionals toward small, on-device LLMs that keep case data in-house, used for everyday tasks such as summarization, translation, and reformulation. Following standard prompting guidance, a user naturally states their role…
Fairness or Fallacy? Mapping the Landscape of AI Bias in Disease Detection
Artificial intelligence (AI) has several transformative effects on healthcare, like to facilitate clinicians in disease detection and to manage patient care with enhanced precision. Due to the increasing complexity of medical data and the growing volume of diagnostic images, the…
Towards More Granular Reasoning: Mapping the DSA Transparency Database Taxonomy to Platform Terms and Conditions Using Large Language Models
Article 24(5) of the Digital Services Act (DSA) requires providers of online platforms to submit every content moderation decision they make, with a corresponding Statement of Reasons (SoR), to the DSA Transparency Database (DSA-TDB). Such decisions are classified against…
CoRS: Fast Neural Flexible Job-Shop Scheduling via Composite Interval State Reuse
Neural solvers for the Flexible Job-Shop Scheduling Problem (FJSP) typically refresh graph representations and update feasibility after every operation--machine decision. However, only feasibility must remain exact at every step, whereas graph encoding is substantially more…
Machina Pauliana: A neuro-symbolic approach to the actio pauliana in Dutch insolvency law
Applying a legal provision frequently requires both structured rule-based reasoning and sensitivity to open-textured terms developed in case law. The actio pauliana of Article 42 of the Dutch Bankruptcy Act (Faillissementswet, Fw), which allows an insolvency practitioner to…
Beyond Static Leaderboards: Capability-Focused Evaluation with Structural Diversity for Time Series Foundation Models
AI summaryCaFE extends static time‑series leaderboards by generating a refreshed test set that varies eight structural features, providing per‑feature capability profiles. It shows stable scores, uncovers hidden strengths/weaknesses, and selects models that outperform aggregate‑leaderboard choices on structured data.
LOCU: Lowdin-Orthogonalized Constraint Updates for Multi-Constraint Policy Optimization
We investigate multi-constraint policy optimization in constrained Markov decision processes (CMDPs), where interactions among constraint gradients often yield ill-conditioned update directions, causing numerically fragile steps and unintended cancellation or redundant overlap…
REFLEX: RNA Ensemble Generation via Flexibility-Calibrated Stochastic Bridge
RNA molecules often function through multiple experimentally observed conformational states. A central challenge is to capture ensemble diversity without sacrificing structural fidelity: samples may collapse toward averaged representative structures or drift into unsupported…
Physics-anchored residual quantum neural networks for cloud-cover parameterization
AI summaryResidual quantum neural networks correct a fitted physical cloud‑cover estimate, reducing mean squared error by 38–54 % compared to direct QNN predictions. They outperform compact residual MLPs and approach the performance of a longer‑trained classical baseline.
Toward Online Robust Zero-Sum Markov Games with Function Approximation
Online robust zero-sum Markov games pose a coupled operator-learning and game-solving problem: from nominal interaction data, the learner must recover a worst-case minimax Bellman operator while maintaining strategic coverage and controlling stage-game approximation error.…
Robotic Photography System based on Vision-Language-Action Model
AI summaryRoboPhoto-VLA integrates vision, language, and action to autonomously plan camera motions and capture aesthetically pleasing photos. It uses PhotoCoT for structured action planning and ADPO to align with human aesthetics, achieving better alignment and generalization on the new OpenPhotography dataset.
The Scheduler Is a Hidden Retriever: How Asynchrony Shapes Evidence Acquisition in Search Agents
Language-model agents increasingly solve complex tasks through parallel reasoning and asynchronous information gathering. Controlled replay across two model families and three benchmarks shows that delivery timing redirects search decisions made before all results arrive,…
Selective Transfer through Sequence Distillation: From Benchmark Knowledge to Backdoor Behaviour
Knowledge distillation is widely used to transfer capabilities from teacher models to student models, but recent work shows that it can also transfer unintended teacher properties. We study whether different hidden teacher properties survive the same distillation channels.…
StrucChunk: Structure-Aware Retrieval for Legal Document RAG
Retrieval-Augmented Generation (RAG) can ground LLM outputs in domain-specific knowledge, but legal documents are hard: standard chunking fragments hierarchical argument structure, and generic embeddings struggle with both precise citation matching and semantic understanding. We…
MCDA: Multi-View Graph Contrastive Learning with Dynamic Topology Augmentation for Functional MRI Analysis
Resting-state fMRI (rs-fMRI) is a vital non-invasive tool for diagnosing brain disorders and uncovering pathological mechanisms. However, existing graph contrastive learning (GCL) methods often underexploit augmentation histories across training epochs and the relationships…
The Control Layer Behind Rule-Following Agents Trading Across Market Conditions
Trading discipline often fails people. Our study tested whether four AI trading agents followed their strategies and risk rules. Offering traders the evidence and an agent evaluation toolkit to separate luck from following the rules, before giving agents meaningful autonomy or…
CIS-Transolver: A Cross-Layer Incremental-State Neural Operator for Fine-Scale Automotive Aerodynamic Flow Prediction
Neural operators offer an efficient data-driven approach to rapid prediction of complex automotive aerodynamic flows. However, spectral bias and other factors limit accurate reconstruction of fine-scale flow variations near complex geometries. Existing studies improve the…
Large-Scale Biomedical RAG: Corpus Scaling, Document Representation, and Efficiency Trade-offs
We present a controlled, large-scale evaluation of how corpus scale, document representation, and reranking affect the effectiveness and efficiency of biomedical Retrieval-Augmented Generation (RAG) systems. Retrieval collections were progressively expanded while retaining the…
A Generic On-Board AI Processing Chain for Earth Observation: From Application Development to Embedded Deployment and Mission-Level Orchestration
Earth observation satellites acquire more data than they can downlink. Extracting information on board with artificial intelligence (AI) turns the satellite into a data-reduction stage that downlinks compact products, such as masks or detections, instead of full scenes;…
VibeTraining-Bench: Data-Trainer-Harness Co-Design for Iterative System Improvement
AI summaryVibeTraining-Bench evaluates agents’ ability to jointly refine training data, trainer, and inference harness through iterative feedback, across diverse tasks. GPT‑6‑Astra lifts a base system from 32.9 % to 53.7 % on SWE‑bench Verified, approaching a human‑crafted 55.4 % benchmark.
CARRS: A Constraint-Aware, Real-Time Routing System for Emergency Vehicles in Urban Environments
Emergency vehicles such as police vehicles, fire engines, and ambulances require a stable and efficient route optimization to reach life-critical events in the shortest times possible. However, the dynamic nature of urban traffic makes this ideal scenario much more challenging,…
PhyPro-FM: A Physics and Prompt-Guided Graph Foundation Model for Contingency Analysis
AI summaryPhyPro‑FM combines graph neural networks with a learnable global query and prompt‑guided few‑shot learning to perform contingency analysis. It outperforms prior methods by prioritizing branch‑flow errors during training and using a residual‑based correction for singular branches, achieving superior accuracy and robustness.
GAIDA: Generative AI for Human-in-the-Loop Document Annotation
Document annotation is a fundamental yet labor-intensive step in building document information extraction systems. While Optical Character Recognition (OCR) and Large Language Models (LLMs) can automate parts of this process, we argue that annotation systems should be judged not…
Autonomous intelligence for mining productivity: a multi-agent AI approach to operational value creation
Integrated Remote Operation Centres (IROCs) are emerging as the architectural frontier of Industry 4.0 for capital-intensive sectors such as mining, oil and gas, and energy. Yet two intertwined gaps constrain their evolution towards full autonomy. First, no integrated agentic…
BOSS: Token-Budgeted Online Sampling and Selection for LLM Reasoning
Test-time scaling can improve LLM reasoning, but existing approaches typically adapt generation or apply answer selection after a predetermined generation schedule. We study within-instance allocation: under a hard per-instance prompt-plus-completion token cap, when should…
Answer Delivery Under a Shared Token Budget: An Audited Study of Stopping and Finalization
An inference budget is useful only when the system delivers an answer within it. We study a fixed-budget policy combining first-valid-box stopping with a charged finalization cue. On 100 fresh, independently audited GSM8K test questions per model, it improves delivered endpoint…
When Coding Agents Outpace Review: Evaluating PRD-Driven Development in Mature Enterprise Repositories
AI summaryPACE evaluates coding agents on 43 enterprise PRD tasks, measuring full requirement fulfillment. The best model (GPT‑5.6‑Sol) attains only 14% complete delivery, with incomplete implementation and missing integration as main bottlenecks.
Local data augmentation for imbalanced binary classification
Facing imbalanced binary classification, we aim at maximizing the recall while limiting the proportion of false detections in the minority class. For doing that, the present work studies the effect of several rebalancing strategies, from the uniform subsampling of the…
Physiology-Aware Graph Attention for Early Ventricular Reentry Prediction
Sudden cardiac death mostly originates from post-infarct ventricular tachycardia. This reentry is sustained by propagation around scar tissue, but its risk remains difficult to predict from infarct imaging alone. While in silico induction studies can reveal these mechanisms,…
Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents
Large language model (LLM) agents now trade real capital using news as evidence, yet they cannot reliably tell fabricated headlines from real news. Because they read the same feeds under near-identical policies, a single fabricated post can push them into the same trade at the…
Rethinking LLM Fine-Tuning via Weight Space Reparameterization: Preserving Safety during Downstream Adaptation
Fine-tuning large language models (LLMs) on downstream tasks often weakens their previously aligned safety behavior, creating an inherent trade-off between task adaptation and safety preservation. Existing approaches attempt to mitigate this issue by identifying safety-relevant…
Remembering Has a Price: Learning Memory Worth under Resource Constraints
Lifelong agents continually accumulate and reuse experience to improve after deployment, yet in practice they often operate under tight generation, interaction, and memory budgets. Under such stringent joint constraints, we find that existing lifelong-memory methods can become…
ALIGN: Aligning Condition Estimation and Expression in Advantage-Conditioned Behavior Cloning
Human-collected robot manipulation data often mix successful behavior with retries, pauses, recoveries, and failures. Advantage-Conditioned Behavior Cloning (AC-BC) can exploit such mixed-quality data, but only if the conditioning signal is both accurately constructed from…
VeriCUA: Learning Proactive Verification for Reliable GUI Agent Rewards
Reinforcement learning for computer-use agents increasingly relies on proactive reward agents that interact with the post-execution GUI environment to verify what recorded screenshots cannot reveal. The reliability of such verification hinges on two coupled decisions: what to…
SPAR: Structural Pruning and Action Recovery for Vision-Language-Action Models
The limited onboard computing resources of robots make the inference demands of vision-language-action (VLA) models costly to accommodate. Reducing these demands through aggressive compression, however, can degrade manipulation performance beyond what supervised fine-tuning…
Mitigating Reward Hacking in Preference Reward Models with Evolving Penalties
AI summaryDynalize automatically generates penalties from a language model that identifies reward‑hacking behaviors during policy training, applying them to rollouts. This online intervention lets open‑ended models train longer and improves downstream performance, reducing hacking patterns in tasks like counseling and creative writing.
ScopeMem: Latent Memory That Learns Where to Revise Frozen GUI Agents
AI summaryScopeMem adds a latent memory prefix that, conditioned on the model’s initial answer, revises only incorrect actions while preserving 94.4% of correct ones. It outperforms CE-trained prefixes and other baselines on AndroidControl, GUI-Odyssey, and AITW, matching or beating larger VL models.
Learn the Task, Not the Schedule: Reducing Spurious Cross-Arm Coupling in Bimanual Imitation Learning
A bimanual demonstration specifies both what the robot should do and one particular schedule for when each arm should act. Standard imitation learning does not distinguish task-required cross-arm constraints from incidental scheduling choices, and may therefore exploit…
RouterWorld: Scaling Conversational AI with Latency-Aware LLM Routing
AI summaryRouterWorld benchmarks conversational LLM routing with realistic latency and partial observability, offering 371k responses and 76k agent trajectories. It shows most existing routers prioritize accuracy over latency when only sparse logs or streaming data are available.
Toward Structured Financial Foundation Models
Foundation models promise reusable capabilities and step-change performance gains across tasks. In finance, large-scale pretraining of domain-specialized foundation models can advance both goals. However, prevailing architectures and objectives do not by themselves address the…
Beyond Behavioral Accuracy: Evaluating Individual-Level Fidelity in LLM Investor Simulation
Large language models (LLMs) are increasingly used as user simulators, but can they faithfully reproduce an individual's behavior and decision logic over time? We investigate this question in a longitudinal paper-trading study with 80 participants. Using rolling next-day…
A Semantic Knowledge Graph Construction Pipeline for Vehicle Intention Prediction in nuScenes
Knowledge graphs (KGs) are increasingly used in autonomous driving (AD) to make traffic scene entities and relations explicit for prediction and reasoning. Existing traffic-scene KGs, however, primarily emphasize trajectory forecasting or general scene understanding. This paper…
Breaking Diffusion: Parallel Reasoning with Draft Refinement
Drafting an answer allows a reasoning model to reconsider each decision in the context of the others. Yet repeated diffusion denoising does not ensure that individually plausible token updates form a better answer together. We introduce Draft Refinement Language Models (DRLMs),…
Agentic Orchestration of EDC Data Space Retrieval and Natural Language Driven Data Processing
Participating in a data space still requires the Dataspace Protocol (DSP) negotiation flow to be hand-coded against each provider's access policy. We present a multi-agent architecture that automates this process using natural language requests. Built on LangGraph over an…
Training Natural Schemers: RL Can Amplify Oversight-Conditional Behavior
One of the central concerns for the safety of advanced AI is scheming: when a model strategically complies under oversight but pursues its own goals when unobserved. Current demonstrations, such as Anthropic's Alignment Faking work, depend on design choices that make them less…
BetaMem: Learning to Remember a Changing World for Embodied Agents
Long-horizon embodied agents face temporal confusion and incomplete or distorted records, recurring errors from fixed memory rules, and growing storage and retrieval costs. We present BetaMem, a framework for building and evolving traceable long-term memory through six linked…
Reward Objectives for Robust (PO)MDPs: A Law-Preserving Equivalence with (Partially Observable) Games
Robust partially observable Markov decision processes model agents that act under partial observability while adversarial environment resolves uncertainty in the transition probabilities. Reward objectives such as discounted-sum, long-run average and cumulative reward under…
WhatIf-Bench: Benchmarking Intervention-Aware Time Series Forecasting
Recent time series foundation models (TSFMs) can consume covariates in-context, extending zero-shot forecasting to problems driven by external factors. This enables scenario analysis: forecasting how a target responds to events that set covariates to chosen future values.…
Concerns in Benchmarking Jailbreaks: An Empirical Study of Attack Success Rates
Jailbreak methods are commonly compared using Attack Success Rate (ASR), but these comparisons depend on how responses are generated and judged. In this paper, we examine eight attacks against three open-weight language models using three evaluation procedures. Low-resource…
Conducting Reliable Keyword Estimation with Unreliable ASR in Police Audio Data
Extracting insights from large audio data sets using human transcription is expensive and time-consuming. Off-the-shelf Automatic Speech Recognition (ASR) systems offer an easy-to-use scalable alternative, but systematic transcription errors, especially in domain-specific…
VaDar: Implicit Goal Conditioning via Value Disaggregation
Long-horizon tasks are important for modern AI and robotics, yet pose a challenge for reinforcement learning. Historically, a popular approach has been to decompose a task into subgoals, and learn a subgoal-conditioned policy to execute each subtask in sequence. We introduce an…
When the harness decides success: state that cannot be re-observed
Agent harnesses decide what a language model sees across turns: they drop old turns, summarize them, or replay the model's own earlier reasoning. We ask when these choices change whether an agent succeeds, not only what it costs, and report three findings. First, in a…
Revision-Based Learning from Interactive Experience and Feedback
Code generation models are increasingly used to produce interactive artifacts, such as websites and user interfaces. Many properties that determine whether these interfaces are useful only become apparent through realistic interaction, making them difficult to incorporate into…
Reasoning-Aware Multicalibrated Reliability Predictors for Language Models
As people increasingly rely on large language models (LLMs) for information, model providers need to communicate meaningful estimates of whether individual responses are correct or, more generally, reliable. In the context of hallucinations, existing work uses multicalibration…
From Outcomes to Components: A Counterfactual Evaluation Framework for LLM-Proposed Repairs
An LLM-generated repair action specifies which records to change and by how much. A failed action may still contain a useful target set or numerical update, depending on how the two are paired. We evaluate this problem in traffic networks and databases, where aggregate…
The Effectiveness and Durability of Pretraining Data Filtering in Protein Language Models
Pretraining filtering is one of the most common safeguards for open-weight language models. Here we study its effectiveness, durability and scalability using the ESM-2 family of protein language models. We find that filtering viral proteins from the pretraining data meaningfully…
The Development of Sycophancy in Language Models: Tracing Vulnerabilities from Pretraining to Post-Training
Language models can change a committed answer to accommodate a user's opinion or pushback without new evidence. However, when this behaviour develops during training and what a released model inherits from earlier stages remain unclear. We follow 741 checkpoints of 30 open model…
Tandem: Serving Hybrid Linear-Attention GUI Agents on Mobile NPUs
Smartphone GUI agents complete tasks in mobile apps by repeatedly reading the screen and asking a model for the next action. Mobile neural processing units (NPUs) enable local execution, keeping screen contents on the phone, but model latency accumulates with every action.…
AllPath: Layered Verification for a Deployed, Self-Revising LLM Trading Agent under Human-Gated and Autonomous Regimes
AI summaryAllPath deploys a self‑revising LLM trading agent with a three‑layer verification architecture: deterministic code gates, bounded nightly self‑review, and human approval. In live testing, human gating caught strategy disagreements and rule deletions while code gates ensured consistency, revealing production failures unseen in backtests.
Honeypots for AI Agent Insider Threats
AI agents deployed inside organizations can pose an insider threat if they make mistakes, are compromised by an attacker, or pursue misaligned objectives. We propose honeypots as an AI control method to improve the security of internal agent deployments. Rather than only using…
StepBench: Diagnosing Real-World Procedural Understanding in AI Agents
AI summaryStepBench evaluates AI agents on procedural understanding by testing five diagnostic tasks from 6,885 wikiHow procedures, revealing that step ordering is hardest, with top models correctly ordering only 63% of steps and performance dropping to 33% for long procedures.
Intra-Rubric Reward Divergence: Mitigating Rubric Hacking through Latent Dimension Recovery and Balancing
Rubric hacking limits both performance and training robustness in rubric-based reinforcement learning. By decomposing rubrics into sub-dimensions, we reveal the Intra-Rubric Reward Divergence: as the total rubric score increases, certain sub-dimension scores systematically…
Hierarchical Token Merging with Recursive Autoencoders for Unsupervised Grammar Induction
Hierarchical abstraction is a fundamental principle across many data modalities, enabling complex sequences to be represented compositionally at multiple levels. Learning such hierarchies directly from data, without explicit structural supervision, remains a central challenge in…
Towards Robust Model Organisms for Eliciting Latent Knowledge
A key problem in AI safety is eliciting latent knowledge (ELK), extracting information from a model that is misaligned and more knowledgeable than its overseer. For example, a misaligned model used to scan code for vulnerabilities may want to withhold its findings, i.e.,…
recAMPLIFY: Large-Scale Pretraining of Recursive Protein Language Models
AI summaryrecAMPLIFY trains a recursive masked protein language model with 155 M unique parameters on 2 T tokens, achieving a 22.5% lift in fold accuracy and better contact recovery while using 56% fewer parameters. Gated Mixture Recurrence (GMR) adds adaptive gating, reducing expected depth by up to 37% with modest accuracy loss.
TokenGate: Adapting LLMs for Fast and Memory-Efficient Prefill with Learned Token Pruning
AI summaryTokenGate adds a lightweight module to transformers that learns to prune input tokens during prefill, using token‑gated attention to bypass the non‑differentiable top‑K. It yields 1.4–11.3× faster TTFT and 1.5–3.4× lower memory without accuracy loss, outperforming prior token‑pruning methods.
Game of Phones: Harnessing Game Theory and LLMs in 'Bot Wars' to Counteract Phone Scams
Phone scams impose substantial and growing costs on individuals and financial systems worldwide, yet existing countermeasures impose no resource cost on perpetrators. Strategic scam-baiting, in which automated agents engage scammers to deplete their operational capacity, offers…
LG-Bench: A Graph-Structured Evaluation Benchmark for Life Sciences
Traditional evaluation benchmarks reduce inherently interconnected scientific knowledge in life sciences into flat lists of questions, disregarding the underlying topological structure of the knowledge. We introduce, the first graph-structured benchmark for life sciences,…
SOMA: Continuous Monitoring and Intervention for System Two Reasoning with System One Models
Reasoning models often run into problems without noticing them, even when symptoms are visible in their own text. In humans, metacognition separates monitoring from control: fast, intuitive signals keep watch while slow, deliberate reasoning unfolds and help decide whether to…
Mise-en-Video: Fine-Grained 3D Control for Multi-Shot Video Generation
Recent advances in text-to-video generation have enabled high-fidelity single-shot synthesis, yet practical storytelling creation requires multi-shot videos with both strong inter-shot consistency and fine-grained controllability. Existing multi-shot video generation approaches…
Next Best Action: Causal Constrained Offline RL for Business Action Targeting
Next Best Action (NBA) systems must select among customer interventions (e.g., discounts, promotions, and lifecycle actions), maximize long-term economic value, satisfy complex operational constraints, and remain causally grounded. Existing approaches lack a rigorous framework…
AfriBED: a Speech Recognition Bias Evaluation Dataset for Low Resource African Languages
AI summaryAfriBED is a 9‑hour Bemba and 18‑hour Nyanja speech dataset designed to evaluate gender and speaker‑type bias in ASR. Baseline experiments show it effectively reveals performance disparities across speaker subgroups in low‑resource African languages.
RefCraft: Benchmarking Visual-Reference-Conditioned Agentic Generation through Web Development
AI summaryRefCraft introduces a benchmark of 300 multimodal tasks evaluating AI agents on webpage development using visual references and text instructions. Experiments show that current frontier models struggle with this workflow, with the top system achieving a score of only 52.12 out of 100.
Agent World Model in the Wild: Computer-Use Agent Environment Synthesis
AI summaryWildAWM synthesizes 341 long‑horizon computer‑use tasks in controllable, multi‑interface environments using open‑source models, and training Qwen agents in these worlds boosts performance on eight out‑of‑distribution benchmarks.
ProteinGeoBench: Frontier Structural-Biology Agents Are Fluent at Lookup but Lost in Geometry
AI summaryProteinGeoBench shows frontier models excel at metadata lookup (0.785–0.816) but struggle with numeric geometry (often failing 12/28 tasks); performance varies with evaluation budget, harness, and network access, underscoring the need for per‑task reporting.
Kronecker Delta Rule for Long Context Linear Attention
AI summaryKronLSTM introduces the Kronecker Delta Rule, a minimally invasive memory‑revision update for linear attention. It converges to a stable steady‑state, keeps the memory norm bounded, and outperforms existing linear attention models on long‑context language tasks.
Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy
LLM agents that compute the right answer can still answer wrongly. Agent harnesses increasingly keep long tool outputs out of the context window, saving them to files that the agent searches or reads on demand. We ask where this costs accuracy with a real scientific tool:…
LatComm: Reusable Context Compression for Latent Communication between Large Language Models
Systems that combine several large language models (LLMs) often use one context for many questions. When the context is passed between models as text, the receiving model processes all of its tokens again for every question. Avoiding this cost requires a context representation…
AnyGoal: On World Model Generalization Beyond Expert Rollouts
The goal of a world model is to capture the dynamics of an environment. This has the great appeal that it allows agents to plan and simulate a course of actions with the help of a world model. For these uses to be reliable, however, the learned dynamics must behave well beyond…
Train An Imaginative Agent
A seemingly reasonable push in Sokoban can make the puzzle unsolvable. To avoid such mistakes, an agent has to anticipate the consequences of its actions before execution, especially when feedback from the real environment is sparse or actions are irreversible. We introduce…
WeAct: Dynamic Weight Switching as Agent Actions
Long-horizon agents must learn heterogeneous capabilities from interleaved trajectories. Updating shared parameters to learn new capabilities can degrade previously acquired ones, resulting in catastrophic forgetting. As a trajectory unfolds, different stages may require…
Where is Physical RSI capability from? An Observational Scaling Law of LLM-as-policy
In this paper, we systematically investigate the mechanisms, scaling behavior, and generalization of physical Recursive self-improvement. We introduce PhysRSI-eval to quantify physical RSI through a five-axis metric: First-Attempt Performance, Efficiency, Success, Budget and…
CoopUQ: Robust Multi-Robot Social Navigation via Uncertainty-Aware Cooperative Forecasting
AI summaryCoopUQ fuses cross‑robot spatiotemporal features to produce uncertainty‑aware occupancy forecasts, then uses these forecasts in a constrained multi‑agent reinforcement learning policy that penalizes traversal of high‑risk or uncertain cells. It shows better prediction accuracy and safer navigation in both in‑distribution and out‑of‑distribution Unity3D and real‑robot tests.
Reinforcement Learning with Visually Attributable On-Policy Self-Distillation
Multimodal reinforcement learning with verifiable rewards (RLVR) provides reliable task-level supervision, yet offers limited guidance for localizing errors in visual reasoning. On-policy self-distillation (OPSD) complements this supervision with dense token-level corrections…
OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory
AI summaryOpenVLHarness adds a tool‑capability layer and persistent multimodal memory, enabling large language models to fetch, interpret, and store heterogeneous evidence across reasoning steps. It boosts performance on 22 benchmarks by 9.7–12.6 points over no‑tool baselines while reducing token cost.