FALCON is a camera‑based 3D occupancy front‑end built on the AMD Kria KR260 FPGA that provides explicit latency guarantees for autonomous driving perception. The system uses cycle‑bounded programmable‑logic compute, a measured inference envelope over 100k frames, and a watchdog to flag stale outputs, achieving all 300k proxy frames within a 100 ms deadline plus at most 10.2 ms. On the full FlashOcc pipeline the board attains 29.88 mIoU on the Occ3D‑nuScenes set, matching 92.8 % of an FP32 baseline, while maintaining tighter tail latency than competing Jetson Orin and datacenter GPUs.
By Haohua Que, Mingkai Liu, Handong Yao
The paper introduces RealtimeBench, a benchmark of 180 tasks across six game types designed to evaluate large language model agents in real-time environments where games continue running during inference. It shows that GPT-6.1-sol outperforms the faster Jev when games are paused but lags behind when they run continuously. The authors propose Two‑Timescale Offline Recursive Self‑Improvement (TORSI), a method where a slow teacher analyzes failures in paused games and writes reusable instructions for a fast student, leading to significant score improvements across all game types.
By Ningxin Pan, Hanyu Li, Yeqing Shen, Ning Ding, Yehui Tang, Wentao Zhang
The paper introduces RAC, a relational agent contract designed for engineering tasks governed by strict constraints. RAC separates semantic decisions from tool-based derivations, enabling agents to specify desired relations while specialized tools compute precise outcomes. Experiments on PCB placement, routing, repair, job-shop scheduling, and embedded-code tasks show that RAC consistently outperforms baseline methods, achieving higher clean-pass and quality pass rates across multiple models and domains.
By Kaiyuan Zheng, Fengqing Jiang, Yichen Feng, Zubin Aaron Assadian, Yuetai Li, Bhaskar Ramasubramanian, Basel Alomair, Luyao Niu, Radha Poovendran
The paper investigates the use of test‑time search, specifically Monte Carlo tree search (MCTS), to improve agent performance in terminal‑based computer‑use environments. It finds that conventional MCTS does not significantly boost task success due to limited critic information and poor intermediate state evaluation. The authors propose a new strategy called Probe‑and‑Resume, where a critic is invoked only when the agent believes it has finished, and it is given read, write, and execute access to a forked copy of the environment to probe hidden state. This approach raises success rates by 8.4 percentage points over standard MCTS and nearly matches a more expensive Codex CLI agent.
By Jiakai Xu, Tiancheng Ge, Tujie Guan, Yusheng Li, Rohan Timmaraju, Eugene Wu, Kostis Kaffes
Kolab is a new dataset comprising over 115 hours of screen recordings from 50 professionals across 40 occupations who collaborate with AI agents on realistic, long‑horizon tasks. The dataset includes participants’ judgments of collaboration quality and output, revealing varied coordination patterns and the need for agents to anticipate how experts will use or modify their contributions. Analysis shows current state‑of‑the‑art models overestimate collaboration success and struggle to differentiate high‑ from low‑quality interactions, indicating a gap in understanding long‑horizon human‑agent teamwork.
By Shirley Anugrah Hayati, Patricia Anugrah Setiani, Ruizi Wang, Zhouyi Li, Khanh Chi Le, Young-Jun Lee, Dongyeop Kang
AutoOnboarder is an autonomous system that adapts large language model agents to specific tasks by iteratively refining the agent’s foundation model, prompt, and harness code under the guidance of an intelligent coding agent called the Onboarder. The Onboarder operates with only a few examples and evaluates performance on a separate virtual machine to avoid data leakage. In experiments on genomic‑sequencing prior authorization form submission, autonomously onboarded agents matched or surpassed a Gemini 3.5 Flash baseline while reducing costs up to 29×, though performance varied and the agents were not perfect.
By Tanay Arora, Cong Liu
The paper introduces CA‑MAMBA, a constrained multi‑agent reinforcement learning method that predicts cost within a world model, enabling most policy updates to occur in imagination. Compared to a strong model‑free baseline, CA‑MAMBA requires 10–20 times fewer interactions on the median SMAC configuration and achieves 2.8–12 times higher MAMuJoCo rewards for the same budget. In controlled experiments, aligning the model’s cost estimates with a cost‑blind learner reduces that learner’s cost by more than half and satisfies a strict continuous‑control limit, indicating that accurate cost estimation is not the limiting factor.
By Raz Shmueli, Shahaf S. Shperberg
DocPilot is an agentic framework for long visual document understanding that orchestrates evidence seeking through page-level navigation and fine-grained localization, integrating visual inspection, code-based numerical reasoning, and cross-modal reasoning. It employs curriculum-based agentic training to enable small language models to learn complex document solving, progressing from single-ability learning to adaptive orchestration. Experiments on benchmarks such as MMLongBench-Doc, DocBench, FinRAGBench-V, and MADQA show that DocPilot achieves strong gains with frontier VLM backbones, and its 4B variant competes with much larger models, demonstrating effectiveness and generalization across diverse long-document reasoning tasks.
By Zhuowen Liang, Xiaotian Lin, Yuyu Luo, Haixun Wang, Nan Tang
JevAgentHarness is a new framework that benchmarks, deploys, and trains decision models to share control with large language models (LLMs) in agentic tasks. It introduces AgentDecisionBench to evaluate decision models from isolated questions to full agent control, revealing that accurate isolated decisions still miss many tasks and that naïve delegation can reduce success. By fitting confidence thresholds per choice and training a 4B decision model on states after agent mistakes, JevAgentHarness achieves high task success while cutting cost and speeding up execution, with results and resources made publicly available.
By Kunlun Zhu, Tao Feng, Xu Han, Heng Wang, Ling Yang, Xiangru Tang, Jiaxuan You, Heng Ji
The paper introduces SPRA (Structured Paper Representation for Agents), a comprehensive representation of AI research papers that organizes key research elements—Problems, Contributions, Components, ExperimentSetups, Measures, and Findings—into typed relations for easier querying, traversal, and reuse. SPRA includes paper-level citation relations that capture dependencies, component usage, and comparisons, and it has been instantiated across more than one million AI papers, yielding millions of research units and internal relations. Empirical evaluations show that SPRA improves source-conditioned scientific QA and directed cross-paper reasoning over passage retrieval while significantly reducing input and output token counts, demonstrating its practicality for scientific workflows.
By Yutong Leng, Haojiong Chen, Fukai Shang, Xiaoyang Wang, Mindong Liu, Runyuan Ma, Jiantao Qiu
ISW-Agent combines large language model dialogue with n8n workflow automation to process internal solitary wave remote sensing images. It translates natural‑language instructions into structured tasks, applies geospatial transformations, and validates outputs at each stage. In tests, GPT‑5.6 Sol and DeepSeek achieved a 94.2% task‑match rate, and the system processed 39 of 50 Sentinel‑1 scenes in a median of 48.04 seconds, while handling fault‑injection scenarios with retries and validation checks.
By Hao Zhang (First Institute of Oceanography, Ministry of Natural Resources, Qingdao, China), Hengrui Liu (First Institute of Oceanography, Ministry of Natural Resources, Qingdao, China), Lina Sun (First Institute of Oceanography, Ministry of Natural Resources, Qingdao, China), Junmin Meng (First Institute of Oceanography, Ministry of Natural Resources, Qingdao, China)
The paper derives closed‑form bounds on the performance of a drone‑mesh interception system before any learning policy is applied. It shows that, due to a separation constraint, a mesh of agents can cover at most 90.69 % of a plane (74.05 % of a volume) and that optimal hexagonal packing achieves this limit. The authors validate these theoretical ceilings with Monte‑Carlo simulations and argue that such bounds should be reported alongside any learning‑based results.
By Dmytro Babarytskyi
ClaimScope is a 72‑task benchmark that audits mathematical claims by requiring a verdict, scope, counterexample, and repair, covering areas such as optimizer‑state transport and exact $2 imes2$ linear algebra. The study evaluates Qwen3‑8B using a witness checker that only examines the model’s submitted state, showing that verifier feedback improves counterexample acceptance and safe completions compared to generic feedback. The results demonstrate that separating claim classification from executable evidence can yield gains in witness‑structure correction without necessarily improving broader reasoning.
By Mushir Akhtar
The paper introduces a framework to evaluate large language models (LLMs) on their ability to conduct full scientific inquiry cycles, combining hypothesis generation, reasoning, and exploration. By extending the Countdown task into a mathematical environment, the authors assess how LLMs probe systems, collect observations, and infer relationships, examining exploration strategies and causal discovery. Experiments across seven LLMs and various prompting methods show limited inherent inquiry skills, but structured guidance inspired by educational scaffolding markedly improves performance.
By Marta Kne\v{z}evi\'c, Seyed Parsa Neshaei, Jibril Frej, Vinitra Swamy, Tanja K\"aser
Data World proposes a multi‑agent engine that generates synthetic data environments, tasks, and evaluation rubrics by first specifying the underlying world rather than curating real files. The system produces thousands of training tasks across diverse domains and enables independent audits of target recoverability, while human curation ensures realistic question selection. Benchmark results show that models trained on these synthesized tasks still face significant challenges, yet RL‑zero training on the synthetic data yields measurable performance gains over baseline models.
By Zhanli Li, Zeyuan Xu, Chaoxu Pang, Zeyuan Liu, Anyang Ji, Te Sun, Kongcheng Zhang, Wenbo Xu, Xubin Yue, Shipeng Luo, Junjie Zhao, Yingfan Hua, Jingqing Ruan, Zhongda Su, Feng Hong, Zhizhao Zeng, Ke Zeng, Xunliang Cai, Hanjia Ye
CoWorld Reasoner is an agentic framework that leverages a Video World Model (VWM) as a unified visual reasoning engine for Vision‑Language Models (VLMs). It exposes multiple visual operations through tools and skills, allowing the VLM to decide when to imagine, select an operation, generate visual content, and interpret the resulting evidence. The framework includes a VLM‑VWM Harness Co‑Evolution mechanism that jointly refines tools, skills, and execution code based on feedback, while keeping model weights fixed. Across seven benchmarks covering spatial, temporal, causal, and perceptual reasoning, CoWorld Reasoner outperforms direct VLM reasoning and improves accuracy on all benchmarks, achieving notable gains on MMSI and ERQA.
By Dongjie Cheng, Ruifeng Yuan, Ran Wei, Ning Song, Yongqi Li, Wenjie Li
PROBE is an algorithm that trains language‑model auditors to probe deployed tool‑using agents for policy violations through realistic, multi‑turn, tool‑mediated dialogue. By framing auditing as a two‑player Markov game and using a composite reward that balances violation discovery, behavioral plausibility, conversation progress, and policy diversity, PROBE outperforms zero‑shot adversaries in uncovering broader and higher‑yield policy violations. The study compares GRPO, DPO, and GFlowNet optimizations, demonstrating that multi‑turn behavioral auditing is a distinct and necessary axis of agent evaluation, complementing existing capability and safety benchmarks.
By Kshitij Mishra, Abhijith Sharma, Nils Lukas, Salem Lahlou
X‑RIC is a new framework that embeds agentic AI into O‑RAN architectures to enhance the RAN Intelligent Controller (RIC). It combines reactive reinforcement learning with strategic reasoning via automated behavioral cloning, a telecom language model, and closed‑loop co‑adaptation, and introduces BARX, a benchmark for evaluating agentic control. In simulations and testbed deployments, X‑RIC cuts operational costs by 89 %, meets SLA compliance, and converges 75 % faster with fewer training steps and lower variance.
By Adhwaa Al-Chaab, Ayman Younis, Dario Pompili
KnowGPO introduces a reinforcement learning framework that groups search agent trajectories by the task-relevant knowledge they acquire, rather than by surface observations. By decomposing required information into atomic knowledge conditions and using the incremental knowledge gain as a process reward, KnowGPO provides finer-grained credit assignment for deep search tasks. Experiments on Xbench, BrowseComp-ZH, and GAIA demonstrate that KnowGPO outperforms outcome-only and process-supervision baselines by 3.8 to 4.2 points, with an 8B model rivaling or exceeding most 32B open-source systems.
By Siqi Chen, Zhengxi Lu, Yuchen Yan, Xingyu Wu, Qipeng Chen, Xin Zhang, Aiting Liu, Chao Deng, Jie Liu, Jin Ma, Weiming Lu, Jun Xiao, Yongliang Shen
The paper introduces DIB-Bench, a benchmark that exposes deployment‑integrity blindness in large language model (LLM) evaluation by testing 7,560 scored examples across six tamper types and a range of evaluators from regex scorers to GPT‑5‑nano. It shows that stronger evaluators are more likely to accept tampered outputs—mean acceptance rises from 88.6% for rule‑based scorers to 94.1% for GPT‑5‑nano—and that certain trigger‑conditional tampering can produce harmful outputs in 48–100% of cases, increasing with model capability. The study also demonstrates that output‑only evaluation cannot detect tampering with zero output shift, while a manifest‑hash identity channel can detect it with zero samples, and proposes a graded deployment‑identity descriptor for evaluation reports.
By Jianchang Su, Dawei Xiang, Yifan Zhang, Wei Zhang