The paper introduces a framework for k-shot learning with structured decision models Jev and Clef-Flash, enabling defect detection using as few as two reference images per class. It presents an agentic method to refine text prompts that guide classification and a calibration technique to improve defect probability reliability. Experiments on five manufacturing case studies show that both models outperform CLIP with the same limited references, and the proposed refinements and calibration consistently enhance accuracy and reduce Brier scores across most configurations.
By Fadel M. Megahed, Ying-Ju Chen
The paper introduces WeGenEditBench, a benchmark with 2,740 evaluation instances and 148 task types for image editing, focusing on local failure detection and prioritization for model improvement. It also presents WeGenEvoStudio, a multi‑agent harness that supports recursive self‑improvement by diagnosing failures, constructing retrieval‑grounded data, and adapting models. Experiments show that models such as GPT‑Image‑2 and Seedream‑5.0‑Pro excel in single‑image and text‑intensive editing, and evolution experiments demonstrate significant targeted gains on selected benchmarks.
By Jun He, Canmiao Fu, Ying Zhang, Zhipeng Huang, Chenyu Mu, Yazhen Xie, Chen Li, Jing Lyu, Weijia Li, Xiang Zhang
The paper introduces PEPS, a training‑free framework that improves spatial reasoning in Vision‑Language Models by parsing queries into a compact set of executable requirements before execution. PEPS uses a typed FESM schema (Frame, Entity, State, Metric) to commit queries, then selectively grounds primitives via fixed geometric APIs and verifies sufficiency, allowing early termination or targeted repair. Experiments on spatial reasoning benchmarks show that PEPS boosts accuracy while reducing refinement rounds, API cost, and latency compared to existing tool‑integrated baselines.
By Xiwei Liu, Yulong Li, Xinling Zhuang, Xuhui Li, Yichen Li, Haolin Yang, Zhixiang Lu, Yutong Xie
WorkIng is a dual-domain benchmark for autonomous navigation in highway work zones, offering a real-world dataset (WorkIng-Highway) with 360° multi‑modal sensor data and over 12,000 annotated frames, and a synthetic counterpart (WorkIng-CARLA) that simulates 24 regulation‑aware scenarios across six topologies and traffic standards. The benchmark includes 227,399 tracked 3D bounding boxes and 100+ custom 3D assets. Experiments show that real‑world data drives performance, while scenario‑stratified synthetic augmentation improves 2D recognition, 3D detection, and trajectory planning.
By Xujun Xie, Dominik R\"o{\ss}le, Adithya Mohan, Daniel Cremers, Torsten Sch\"on
The paper introduces PanoVideoBench, a manually annotated benchmark with 1,000 questions covering five categories and 15 question types for panoramic video understanding, and identifies cross‑view tracking and sphere‑time evidence integration as weak points in existing models. To address these gaps, the authors create PanoVideoCoT, a dataset of 13,465 QA‑CoT pairs generated via sphere‑time‑grounded synthesis, and train PanoVideoReasoner using panoramic CoT fine‑tuning and evidence‑privileged on‑policy self‑distillation. The resulting model achieves 55.7% accuracy on PanoVideoBench, improving the baseline by 12.5 points and surpassing the previous state‑of‑the‑art by 2.6 points, particularly in spatial, cross‑view, and temporal reasoning.
By Lidong Lu, Lidong Zou, Xinyi Mao, Jiaxin Lv, Zhenquan Li, Liangrui Dong, Tong Lu, Guo Chen
GUI‑Lens is a coarse‑to‑fine grounding framework that lets a general‑purpose vision‑language model (VLM) iteratively crop and zoom into a GUI screenshot to locate a target element precisely. By extracting OCR text and UI component positions as coordinate references, the VLM selects successive regions to focus on, refining its view until the target is identified. Experiments on five GUI grounding benchmarks and six base models show that GUI‑Lens consistently improves grounding accuracy, achieving up to 9.3× the accuracy of direct grounding on ScreenSpot‑Pro.
By Zichuan Fu, Shirong Wang, Wenlin Zhang, Guojing Li, Yimin Deng, Jingtong Gao, Junjia Qi, Hanyu Yan, Yefeng Zheng, Xiaopeng Li, Wanyu Wang, Xian Wu, Xiangyu Zhao
The paper investigates why chain-of-thought (CoT) reasoning in language models is hard to control, attributing the issue to an entropy collapse in the first few tokens after the opening thinking tag. By applying two inference-time interventions—prefilling a single token and banning frequent opening tokens with in-context examples—the authors significantly improve CoT controllability, especially for addition and style tasks. The study also distinguishes between addition/style and suppression controllability, revealing different underlying mechanisms and correlations with monitorability across models.
By Hugh Van Deventer, Dmitrii Troitskii, Laura Ruis
EvoMCP introduces a mutual evolution framework that co‑evolves Model Context Protocol (MCP) agents and their training data. The method iteratively assesses an agent’s strengths and gaps, generates new tasks by adapting verified seed tasks, refines demonstrations through trajectory evolution, and fine‑tunes the agent with these task‑trajectory pairs. Experiments on Toolathlon, MCPMark, and MCP‑Atlas show consistent performance gains of 16.5%, 10.2%, and 12.6% respectively.
By Xuanang Gao, Kepeng Lin, Xinyan Shi, Yong Deng, Shuheng Shen, Changhua Meng, Xiaofeng Wu, Jie Yang, Wei Liu
MetaVul is a self‑improving framework that discovers and refines harnesses—code scaffolding around large language model agents—for vulnerability detection. Starting from a seed reasoning scaffold, it iteratively optimizes harnesses using feedback from paired vulnerable and patched functions, producing Pareto‑optimal designs that can be reused across different models. Experiments show that optimized harnesses can boost pairwise vulnerability detection by up to 25.8 percentage points and transfer effectively between models without further tuning.
By Ira Ceka, Arnav Garg, Ryan Lutz, Ting Zhang, Ratnadira Widyasari, Gail Kaiser, Baishakhi Ray
RAMA is a retrieval‑augmented multi‑agent deep reinforcement learning framework for routing in Low‑Earth Orbit satellite networks. It combines each satellite’s local observation, recent routing history, neighboring state information, and retrieved entries from a per‑satellite episodic memory, integrating them via a Transformer‑based decision network trained with Approximate Information State objectives. In a 140‑satellite Kepler constellation, RAMA cuts mean end‑to‑end latency by 13.4% and 99th‑percentile latency by 66.7% compared to a state‑of‑the‑art DDQN, and on a 1,584‑satellite Starlink shell 1 it achieves a 93.2% mean latency reduction, with further gains from online updates.
By Xin Ming Li, Peng Hu
The paper investigates how large language models (LLMs) can be trained to cooperate in multi‑agent social dilemmas. It shows that standard reinforcement learning often leads LLM agents to defect and exploit others, even when they have cooperative priors. By adapting the Advantage Alignment algorithm and introducing a group‑relative baseline, the authors fine‑tune LLMs to achieve higher collective payoffs and resist exploitation, and they present a new Trust‑and‑Split environment that requires natural language communication for optimal outcomes.
By Dereck Piche, Mohammed Muqeeth, Milad Aghajohari, Juan Duque, Michael Noukhovitch, Aaron Courville
The article introduces rc_prediction, a Python package that implements parameter‑aware reservoir computing (PARC) for modeling nonlinear dynamical systems when governing equations are unknown. PARC trains a reservoir computer on time‑series data collected at multiple parameter values, then uses the trained model in closed‑loop to generate trajectories at unseen parameter settings, enabling parameter scans and autonomous prediction. The package offers hyperparameter optimization, optional GPU acceleration via PyTorch, and demonstrates its use on three benchmark systems, including predicting critical transitions and system collapse from pre‑critical data.
By Jin-Cheng Rao, Zheng-Meng Zhai, Ying-Cheng Lai
The paper compares three decentralized coordination strategies for large language model (LLM) agents in a perimeter‑defense simulation: a central commander, an auction‑based consensus team, and independent agents. Across five network conditions, the consensus team achieved the lowest breach rate (15%) while the commander and independent agents had higher rates (32% and 29%). The study also explores how network range, LLM authority, and prompt wording affect performance, finding that a dedicated conflict‑resolution protocol yields the fewest breaches.
By Khawaja Abaid Ullah, Ananya Acharya, Mohammad Javad Khojasteh
The paper investigates how standard evaluations of detectors for hijacked large language model (LLM) agents may conceal important weaknesses. By auditing eleven evaluations on the AgentDrift and AgentDojo benchmarks, the authors show that detectors often perform poorly when faced with rare attacks, unseen tools, or data from a different benchmark, with precision dropping dramatically at low attack rates and rankings reversing as tool coverage falls. They also find that a released LLM action monitor, despite high AUROC on its own data, would still generate many false alerts when applied to a new benchmark.
"whyItMatters":"The study highlights that current evaluation practices can give a misleading sense of detector reliability, underscoring the need for more robust, cross‑benchmark testing to ensure real‑world safety of LLM agents."
By Mahmoud Abbasi, Juan M. N\'u\~nez V., Amin Shahraki
The paper establishes formal theorems linking game theory and reinforcement learning, showing that a game can be viewed as a Markov decision process in three distinct ways—definitional identity, exact reduction, or structural analogy. It systematically maps every standard game class and solution concept to corresponding reinforcement learning models, training procedures, and complexity results. The work also identifies where these correspondences break down, such as non‑stationarity, impossibility of uncoupled convergence to Nash equilibrium, and gaps between coarse correlated and correlated equilibria.
By Eduardo C. Garrido-Merch\'an
The paper presents a recovery‑based shielding framework that guarantees a provable safety lower bound for reinforcement learning in unknown, non‑linear continuous dynamical systems. It combines a backup policy (shield) with the RL agent and uses Gaussian process models to quantify uncertainty and predict safety violations, enabling dynamic recovery to safe trajectories only when needed. The shielded agent’s experience trains the GP models, and policy optimization is performed via internal model‑based sampling, allowing unrestricted exploration and sample‑efficient learning while maintaining strict safety compliance in continuous control tasks.
By Alexander W. Goodall, Francesco Belardinelli
arXiv:2604. 11811v3 Announce Type: replace-cross Abstract: Large language model agents rely on a memory harness to write, organize, retrieve, and use past experience.
By Wenbo Pan, Shujie Liu, Xiangyang Zhou, Xianlong Wang, Shiwei Zhang, Bingjing Xu, Wanlu Shi, Xiaohua Jia
The paper introduces an enhanced Google Cloud Dataflow system that optimizes hybrid streaming machine learning workloads. By co‑designing autoscaling on heterogeneous worker pools and implementing prefix‑aware batching, the system boosts throughput by 1.5x to 2x for LLM inference while keeping latency comparable. It also achieves up to 75% peak resource savings and an 84% reduction in peak tail latency compared to existing streaming data processing systems.
By RuiLong Jiang, Dustin Rhodes, Danny McCormick, Yuta Labur, Reza Rokni
The paper introduces AgenticTuner, an agentic calibration pipeline that uses a large language model to optimize cosmological simulations. Unlike traditional Bayesian optimization, the agent can incorporate informal, human-readable priors such as codebase documentation and heuristic guides. Experiments show that the agent achieves competitive calibration performance while reducing CPU cost.
By Chaipat Tirapongprasert, Matthew Ho
MedResearcher‑R1 is a medical deep‑research agent that uses a Knowledge‑Informed Trajectory Synthesis (KISA) framework to build medical knowledge graphs and generate complex multi‑hop question‑answer pairs, especially for rare medical entities. It also incorporates a medical retrieval engine and general‑purpose tools to reliably synthesize medical evidence. The resulting 2,100+ trajectories across 12 specialties enable supervised fine‑tuning and reinforcement learning, producing a 32B model that outperforms larger proprietary systems on MedBrowseComp and shows strong performance on GAIA and xBench benchmarks.
By Ailing Yu, HaoYang Hong, Lan Yao, Jingnan Liu, Zhe Chen, Jiajun Yin, Yuan Wang, Xinhao Liao, Zhiling Ye, Ji Li, Yun Yue, Hansong Xiao, Hualei Zhou, Chunxiao Guo, Peng Wei, Junwei Liu, Jinjie Gu