The paper introduces Online Rubric Self-Distillation (ORSD), a method that extracts natural‑language process rubrics from agent trajectories and uses them as dense token‑level supervision signals. ORSD generates multiple rubric bundles per step, verifies them against trajectories, and selects the bundle that best aligns rubric satisfaction with task success. The selected rubric is then used to compute token‑specific process credit, which is added to the trajectory‑level advantage to enable fine‑grained credit assignment. Experiments on ALFWorld and WebShop demonstrate that ORSD speeds up policy learning and improves final performance for Qwen3 models within the same interaction budget.
By Shumao Wu, Kongcheng Zhang, Jianqiang Huang, Baisheng Lai
The paper introduces a shared framework that adapts Cosmos world foundation models for real‑time interactive world simulation in Physical AI. By streaming visual observations in response to actions and scene controls, the framework allows users and embodied agents to interact with evolving virtual environments, supporting both prediction of visual consequences and generation of observations from an external simulator’s physical state. The authors demonstrate the approach in camera navigation, robot manipulation, controllable video transfer, and multi‑view autonomous vehicle simulation, providing a common foundation for developing interactive virtual environments for Physical AI training and evaluation.
By NVIDIA, :, Anis Ahmad, Jon Allen, Malik Aqeel Anwar, Jesse Archer, Maciej Bala, Piotr Barejko, Aarti Basant, Dan Blick, Taras Bolzhelarskyi, Henry Che, Jin Chen, Yukang Chen, Tae Eun Choe, Yuzhu Dong, Nat Duca, Liang Feng, Francesco Ferroni, Xiao Fu, Mohit Gambhir, Vivek Goel, Rony Goldenthal, Aryaman Gupta, Ali Hassani, Nathan Hayes-Roth, Yong He, Michael Honke, Shengyu Huang, Wei Huang, DeLesley Hutchins, Suneel Indupuru, Pooya Jannaty, Jingyi Jin, Janmesh Kalra, Ashna Khetan, Xin Kong, Aleksandr Korovko, Hui Kuang, Zhengfei Kuang, George Kurian, Jean-Francois Lafleche, Jayjun Lee, Justin Leung, William Lew, Xiang Li, Zhaoshuo Li, Chen-Hsuan Lin, Ming-Yu Liu, Xiangyu Lu, Yuhang Lu, Martin Ding Ma, Qianli Ma, Aditya Mahajan, Miguel Martin, Janarbek Matai, Jonathan McCaffrey, Qing Miao, Ashkan Mirzaei, Mustafa Munir, Sangmin Oh, David Pankratz, Trung Pham, Marilyn Reeb, Gavriel State, Jiaxiang Tang, Shitao Tang, Yue Tang, Tolou Tavakkoli, Alexander Temirov, Hweimi Tsou, Sergei Vasilev, Thomas Volk, Andrew Z. Wang, Hao Wang, Shijie Wang, Zhonghao Wang, David Wehr, Shumian Xin, Mengyao Xu, Yao Xu, Shuai Yang, Sophia Zalewski, Pengcuo Zeren, Cindy Zha, Jing Zhang, Yuliya Zhautouskaya, Kaiwen Zheng, Andrew Zhu, Rick Zhuang
LifeVal is a new benchmark that evaluates AI model performance on personalized agentic consumer tasks, covering over 1.7 trillion US dollars in annual willingness-to-pay across 148 task categories such as managing finances and finding childcare. The benchmark was built using a survey of more than 3,800 consumers and cross‑referenced with national time‑use studies. Task creators, who met strict expertise thresholds, designed AI evaluation tasks personalized with real ChatGPT conversations, memories, reference files, and personal context, and all tasks underwent at least three rounds of human quality control before inclusion.
By Jenny Yang, Elizabeth Proehl, Anjan Bharadwaj, Emily Fan, Charlotte Cole, James Kim, Noam Brown, Phoebe Thacker, Amelia Glaese, Tejal Patwardhan
The paper introduces a method for efficient agent evaluation by using cached traces from task executions. It proposes summarizing each trace with an LLM and embedding the summary to capture task-relevant signals while reducing system-identifying noise. Using these representations, the authors build low-dimensional system embeddings and predict a new system’s benchmark score, achieving lower error than baseline models on SWE-bench Verified and Terminal-Bench 2.0 with only a single executed task.
By Ben Johnson, Hayden Helm
arXiv:2611. 06868v1 Announce Type: new Abstract: In this work, we evaluate a simple inference-time scaling approach based on \textit{turn-level best-of-N sampling} (\method) for large language model (LLM) agents on complex, long-horizon tasks.
By Jalaj Bhandari, Dhawal Gupta, Hongli Zhan, Zhenting Wang, Varad Pimpalkhute, Zhengzhong Liu, Eric Xing, Rupesh Srivastava
FrontierPhysics is a new end‑to‑end benchmark for AI agents in physics, comprising 56 tasks created by 42 physicists across nine domains of theoretical, computational, and experimental physics. Each task requires an agent to review literature, devise a research plan, conduct experiments, and produce structured outputs and manuscripts, with evaluation based on deterministic verifiers and expert rubrics. In tests of 22 models, none exceeded 46 of 168 rollouts, revealing that agents struggle mainly with rigorous reasoning and expert intuition rather than physics knowledge or execution.
By Bingran You, Wenjun Ke, Ria Rosenauer, Bo Liu, Wenbo Chen, Min Chen, Shuqi Xu, Yue Ma, Jiechao Feng, Bo Chen, Xiangkai Sun, Jiarui Liu, Jinen Guo, Zeyu Zhou, Xu Huang, Zhuohan Li, Haoyu Li, Heguang Lin, Lijie Ding, Yadong Zeng, Kangqi Zhang, Xinge Liu, Hanwen Liu, Zhenghong Li, Alexander von Recum, Yuchen Zhu, Felix Schwarzfischer, Yizhou Zhao, Jicheng Wang, Binxu Li, Qian Yu, Chan Li, Cheng Dai, Di Wang, Haoyu Zhang, Jiankai Sun, Kaituo Feng, Mian Qin, Ruihong Yin, Yiqun Luo, Weiwei Wu, Shihong Pan, Tsung Yeh Hsieh, Xiao Huang, Yan Zhang, Sihan Li, Zirui Ren, Zengji Tu, Kyoung Whan Choe, Ke Sun, Vincenzo Galia, Vaisakhi Mishra, Jiashu He, Hongyang Wang, Zhiheng Yang, Shubham Gaur, Yifan Gu, Shuyi Wang, Zihao Zuo, Steven Dillmann, Yiyou Sun, Xinyang Han, Hanchung Lee, Molei Tao, Yue Zhao, Shilong Liu, Yuchen Wu, Xuandong Zhao, Xiangyi Li, Florian Brand, Gerbrand Ceder, Hartmut Haeffner
LIBRA is a framework that transforms a flat catalog of tools into an executable library by having a large language model propose contracts for tool invocations, reusable outputs, state changes, and success conditions. These proposed compositions are then tested in a sandbox, and only the relations that succeed at runtime are retained, eliminating the need for dependency annotations or task‑specific training. The approach improves one‑shot recall and NDCG on several benchmarks, and further enhances retrieval performance after multiple iterations.
By Zekai Li, Zhenke Liu, Yonghao Li, Junhan Liu, Xulong Xiao, Jiapeng Zhang
DARE (Difficulty Analysis for Repair Enhancement) is a framework that measures, predicts, and interprets the difficulty of coding repair tasks using a continuous difficulty metric based on failure rates across multiple agents. It employs 151 static features to predict difficulty, analyzes contributions from seven feature groups to identify task-specific characteristics, and selects tailored agent support accordingly. Experiments on 2,594 tasks across four benchmark partitions show that DARE’s predictions correlate well with actual difficulty and that matching support to salient feature groups improves repair success.
By Jikwang Lee, Seungwook Jung, Jaechang Nam
The paper investigates how large language models (LLMs) misinterpret office documents due to differences between the text extracted from Office Open XML (OOXML) files and what humans actually see. By analyzing common extractors, the authors identify 19 hidden OOXML channels and 12 content omissions, and demonstrate that these gaps enable five attacks that mislead major chat and coding assistants. They propose Prism, a lightweight defense that reconstructs the human view from OOXML before feeding text to the model, which generalizes to unseen attack payloads.
By Penghui Li
The paper introduces SecPatRec, a hybrid method that combines a ReAct-style LLM agent with static code property graph analysis to recover security patterns from source code repositories. It treats a security pattern as a reference architecture, detecting prescribed roles, grounding them in code entities, and recovering their interaction sequences. The authors evaluated SecPatRec on 100 pattern instances from 93 Java and Python projects, achieving high role-detection and grounding accuracy and outperforming a one-shot LLM baseline in both recovery quality and API cost.
By Anh Minh Nguyen, Tuan Dat Tran, Alzubair Hassan, Liliana Pasquale
The study examines how different harness configurations affect a fixed large language model’s compliance with banking conduct rules. Four open‑weight models were tested across five harnesses on 156 pre‑registered retail‑banking scenarios, yielding 3,120 two‑turn episodes scored for rule violations and over‑refusals. Results show that while overall compliance rates varied modestly (87.3%–92.1%), the impact of a harness differed markedly between models, with some models’ best and worst harnesses differing by over 10 percentage points and a same‑model compliance reviewer reducing violations from 9.6% to 6.4%.
whyItMatters":"The findings suggest that validating a model’s compliance requires testing the specific model‑harness pair rather than the model alone, highlighting the importance of action logs over conversation transcripts for accurate conduct assessment."
By Mehmet Murat Huyut
The paper introduces CASE-Bench, a benchmark of 145 repository-level tasks derived from real CVEs across 51 projects, each run in Docker with functional tests and proof‑of‑concept exploits. It evaluates 13 context‑construction methods and six controls across six LLMs, showing that functionality‑oriented context improves functional pass rates by up to 8.2 pp but only 3.7 pp for security, that omitting safeguards can drop secure pass rates by 8.1 pp, and that coding agents see relevant safeguards in only 27‑36 % of trajectories. These results highlight how repository context shapes the security of LLM‑generated code and guide the design of security‑aware context construction and evaluation.
By Jiongchi Yu, Keke Lian, Junqing Wang, Zhikang Fan, Zhaoxuan Li, Yongheng Huang, Ziming Zhao, Dong Zhang
The study investigates whether a physically grounded, low‑dimensional catchment model can rival the predictive accuracy of neural networks. By applying three fundamental hydrological laws—conservation, dissipation, and aggregation—to a lumped model, the authors develop CaRT and its refined version CaRT+, which uses 14 states and 35 parameters per catchment. Across 602 British, 593 US, and 505 Australian gauged catchments, CaRT+ achieves median Nash‑Sutcliffe efficiencies comparable to state‑of‑the‑art LSTM models, demonstrating that process‑based laws can match deep‑learning performance.
By Xilin Xia
ArcRAG is a retrieval-augmented generation system that operates over knowledge graphs without relying on construction‑specific schemas or relation names. It dynamically induces a query‑local graph interface at inference time, plans multi‑hop traversals, and executes them as Cypher queries. Experiments on MuSiQue, HotpotQA, and TriviaQA show that ArcRAG remains effective across both schema‑guided and open‑information‑extraction graph constructions, demonstrating that retrieval can be construction‑agnostic.
By Jyotin Goel, Abhinav Arun, Tejas Prakash Agarwal, Bhaskarjit Sarmah, Stefano Pasquali
The paper exposes a vulnerability in AI toolchains where signed Model Context Protocol (MCP) tools can hide malicious code that leaks data, even when protected by state‑of‑the‑art runtime defenses. It introduces germline execution, which runs each tool invocation in an isolated, data‑free copy, limiting information disclosure to only what is authorized for the destination. Experiments on 153 attack runs show that germline execution prevents unauthorized disclosure in all cases, outperforming baseline defenses and ablations, while maintaining efficient performance (10‑49 ms per invocation).
By Yichiet Aun (Universiti Tunku Abdul Rahman), Mohammed A. Alhartomi (University of Tabuk), Jing Jie Tan (National University of Singapore), Waqas Nadeem (Universiti Tunku Abdul Rahman), Adeb Salh (Universiti Tunku Abdul Rahman), Shankar Karuppayah (Bitranger, CYRES, Universiti Sains Malaysia), Sin Chee Tan (ABLEs)
The paper introduces SARBLOH, an open‑weight agent that emphasizes spatial perception over planning to solve ARC‑AGI‑3 visual games. By converting raw integer grids into an object inventory and geometric briefing within 35 ms, the agent improves rubric scores from 3/28 to 25/28 and reduces token usage by 89.4%. On held‑out games, its geometric invariants generalize perfectly, and it achieves high efficiency on the public set, outperforming the first milestone winner by 13.7× on Kaggle’s hidden evaluation.
By Harsidak Singh Banwait
The paper introduces a toolchain that enables machine‑learning model development on AWS EC2 F2 cloud FPGAs using high‑level Vitis HLS. It implements a complete Vitis HLS to HDK bridge to AFI workflow, a four‑level methodology (FP32 PyTorch, bit‑accurate quantized Python, HLS C‑simulation, physical FPGA) for error localization, and documents deployment, timing closure, and HLS engineering decisions. The authors validate the approach by implementing TinyBERT’s embedding layer and first encoder layer on FPGA, noting that weight streaming from host DRAM results in poor performance compared to CPU/GPU implementations.
By David Rostcheck
The paper introduces Evolutionary System Prompt Learning (E‑SPL), a method that jointly updates a large language model’s weights via reinforcement learning and its system prompts via self‑reflection. In each RL iteration, E‑SPL samples trajectories under multiple prompts in parallel, using them to both evolve the prompts and improve the model’s weights. Experiments show that E‑SPL boosts success rates across several benchmarks (e.g., from 38.8 % to 45.1 % on AIME → BeyondAIME) and outperforms RL‑only or evolution‑only baselines, indicating that weight and context updates synergistically enhance generalization.
By Lunjun Zhang, Ryan Chen, Bradly C. Stadie
The paper surveys the use of reinforcement learning (RL) in graphical user interface (GUI) agents, outlining how RL addresses challenges such as long-horizon credit assignment, distribution shifts, and safe exploration that supervised fine-tuning alone cannot solve. It introduces a taxonomy dividing methods into Offline RL, Online RL, and Hybrid Strategies, and discusses key aspects like reward engineering, data efficiency, and technical innovations. Emerging trends highlighted include the move toward composite, multi-tier reward architectures, world-model-based training to mitigate GUI I/O latency, and the spontaneous appearance of System‑2‑style deliberation, all of which inform a roadmap for future robust GUI automation.
By Junan Hu, Jian Liu, Jingxiang Lai, Jiarui Hu, Yiwei Sheng, Shuang Chen, Jian Li, Dazhao Du, Song Guo
The paper introduces Alice, a closed‑loop system for online learning of executable world models when prior knowledge is misaligned. Alice uses failed candidate updates as structural signals to identify and refine state‑dependent dynamics, generating hypothesis classes that provide targeted examples for model updates. Experiments on a modified Baba Is You game—Baba in Wonderland—demonstrate that Alice significantly improves model learning under prior misalignment, with ablation studies confirming the benefits of class structure and exploration signals.
By SeungWon Seo, DongHeun Han, SeongRae Noh, HyeongYeop Kang