The study evaluates a standalone large language model (LLM) versus a four‑step agentic pipeline for generating explanations of ICU mortality predictions on the eICU Demo dataset. XGBoost achieved an AUROC of 0.855 and an AUPRC of 0.332. In a 38‑case explanation subset, the standalone LLM produced one explanation with outcome leakage, while the agentic pipeline produced none; among 14 overlapping SHAP cases, the standalone LLM had higher SHAP alignment and direction consistency, whereas the agentic pipeline showed better guideline grounding, value specificity, and plausibility.
By Di Zhu, Chen Xie, Haoyun Zhang, Zihan Wei, Ziwei Wang, Jiazhao Shi, Ziyu Wang, Qiyang Xie
The paper introduces Knowledge Cards, a new structured artefact designed to capture validated knowledge about specific concepts that AI systems use to make decisions. Unlike existing model, data, and system cards, Knowledge Cards focus on the layer between inputs and outputs, documenting entities, relationships, reasoning patterns, conditions for validity, and provenance, all grounded in a formal domain ontology and signed off by a domain expert. Prototype cards have been created in the energy and pharmaceutical domains, and the schema is released as a public draft for community engagement.
By Liliana Ferreira
The paper introduces a Latent World Model (LWM) for robot navigation that predicts action‑conditioned latent feature compatibility instead of reconstructing future observations. By exploiting the correlation between spatial proximity and latent feature similarity, the model evaluates action consequences directly in latent space and supports counterfactual training using sampled action sequences. The learned world model can supervise policy learning from unlabeled video and further improve policies via reinforcement learning entirely within the model, eliminating the need for action annotations and additional environment interaction.
By Zengmao Wang, Wei Gao, Shuhan Shen
The paper proposes a scalable, automated method for auditing candidate‑job matching systems for demographic bias. It employs large‑language‑model agents to generate neutral resumes, injects controlled demographic variations, ranks candidates with a fine‑tuned embedding model, and evaluates nine fairness metrics across counterfactual, group‑fairness, and merit‑aware families, producing a composite risk report. Experiments on a small corpus show that single‑score audits miss nuanced issues, underscoring the need for multi‑metric evaluation and LLM‑generated audits as a low‑cost complement to human reviews.
By Sai Yashwant, Shruti Bansal, Anurag Dubey, Samaroha Chatterjee, Satyam Kumar, Shreyash Gupta, Gantala Thulsiram
The study compared human and large language model (LLM) workflows for title‑and‑abstract screening in a complex scoping review. Human reviewers and two GPT‑5.4 file‑batch runs retained 42.2‑45.0% of records with 82.3‑82.9% recall, while Gemini 3.1 achieved the highest recall (83.9%) but retained 56.7% of records. Identical GPT‑5.4 runs showed 91.7% agreement yet differed on 94 records, including 29 verified eligible ones.
By Nikol Figalov\'a, Lynn Huestegge, Anne B\"ockler-Raettig
The paper argues that large language models need adaptive reasoning rather than fixed reasoning budgets. It shows that over‑reasoning leads to high computational cost without accuracy gains, while under‑reasoning results in incorrect or incomplete solutions. The authors evaluate these failure modes on MATH‑500 and the GAIA benchmark, highlighting the need for dynamic reasoning allocation in agentic AI systems.
By Md Jueal Mia, M. Hadi Amini
BekchiAI introduces a benchmark and platform for evaluating large language model agents. The benchmark comprises 13 tool‑using ReAct agents across seven task categories, totaling 2,057 deterministic test tasks with verifier‑checkable gold answers. The platform offers web‑based observability, token and latency telemetry, and remote run termination for live agents.
By Mesut Toruk
PILOT introduces a live self‑improvement framework for long‑horizon agents, enabling a supervisor to redirect or abort an active worker during execution and to distil learned procedures into reusable skills. Unlike existing single‑agent or subagent architectures, PILOT’s supervisor‑worker harness supports real‑time steering and self‑evolution. Experiments on two frozen backbones across three benchmarks show PILOT outperforming other harnesses, achieving significant gains in performance metrics and efficiency.
By Yang Xiao, Yusong Sun, Haoyi Wu, Wenyang Hui, Wen Da, Zhaokai Luo, Mu Chuan, Yao Hu, Wenjie Li, Chengyue Jiang
LiveSim is an LLM-based framework designed to simulate users in multi‑agent live‑stream ecosystems. It treats users as editable behavioral hypotheses and refines them through trajectory‑grounded interactions, using discrepancies between simulated and observed trajectories to uncover missing environmental shaping effects. The framework extracts transferable environment‑behavior patterns, stores them in a collective behavioral memory, and demonstrates improved user‑level fidelity and ecosystem‑level analysis on real‑world live‑stream risk‑control data.
By Jiaqi Xu, Yiran Qiao, Jing Chen, Qiwei Zhong, Xiang Ao, Xueqi Cheng
DuMateBench is a new benchmark for autonomous agents that uses real user sessions from a large production platform, preserving interaction history, configurations, and workspace state. It contains 200 tasks across 8 scenarios and 17 capability categories, many requiring coordination of multiple capabilities. The benchmark tests agents in Docker containers with real-world complexities—Insufficient, Unstable, and Noisy—and evaluates performance with a hybrid deterministic and LLM-as-Judge protocol, revealing significant gaps in task completion across various agent frameworks and LLMs.
By Zechun Niu, Yukun Zhao, Jiaxin Zhang, Xu Shen, Jinhua Si, Han Tian, Can Xu, Yunfan Song, Jiaxin Mao, Yansong Gao, Yuchen Li, Jianmin Wu, Lingyong Yan, Shuaiqiang Wang, Dawei Yin
The paper extends Co‑Scientist, a Gemini‑based multi‑agent system, and validates it in real‑world scientific settings. In materials science it designed a safe precursor route for MXenes and achieved single‑attempt growth of monolayer MoS₂, MoSe₂, and WS₂. In biology it predicted swarming phenotypes of engineered E. coli, and in computer science it discovered a superior inference‑time scaling architecture for HealthBench. A double‑blind study with 30 experts showed that Co‑Scientist’s reliability modules reduce hallucination and plagiarism while improving research safety.
By Samuel Schmidgall, Xiaokai Zhu, Marian Shaw, Lin Yang, Valentin Li\'{e}vin, Jingyun Yang, Yuchen Zhuang, Tim Strother, Alex Bijamov, Min Woo Sun, Anil Palepu, Justin Chen, David Steiner, Jacqueline Shreibati, Wei-Hung Weng, Yilin Zhao, Xingjian Hu, Nicholas Zahn, Sadhya Garg, Julia Kirby, Yuxiang Gan, Jiaoli Li, Divy Thakkar, Shekoofeh Azizi, David Racz, Juraj Gottweis, Vivek Natarajan, Chenglin Wu, Tal Danino, Keran Rong, Haozhe Wang, Benoit Schillings, Yong Cheng, Quoc V. Le, Tao Tu
The paper introduces a Prior‑Guided Residual Flow Matching framework for 3D multi‑person motion prediction. It uses a Deterministic Coarse Prior to anchor kinematics and a Dynamic Cross‑Interaction mechanism to synchronize inter‑agent message passing during integration, thereby improving structural consistency and social context extraction. A decoupled joint‑motion architecture with bidirectional fusion further preserves fine‑grained kinematic coherence, achieving state‑of‑the‑art accuracy on several datasets.
By Wei Wei, Yinyuan Zhao, Ruixuan Yu
The paper introduces Self-Improving Retrieval-Augmented Generation (RAG), a framework that splits document question answering into Retrieval, Reasoning, and Judge agents coordinated by an orchestrator. When the Judge scores an answer below a dynamic threshold, the system retries with broader retrieval, more careful prompting, and relaxed acceptance criteria, achieving 86% oracle-guided accuracy on FinanceBench with a 36.4% Lazarus Rate. The approach logs every decision with confidence scores, providing audit trails needed for regulated financial applications.
By Junjie Xiong, Shawheen Ghezavat, Aum Hirpara
GameWAM is the first World-Action Model designed for native closed-loop gameplay and GUI control in modern video games. It jointly generates future visual observations and executable keyboard-mouse trajectories using parallel visual and action generative processes, block-causal conditioning, and flow matching. The model predicts gameplay/GUI mode at each step, handles heterogeneous native controls, and employs block-cycle control for long-horizon interaction, achieving competitive task success with fewer native actions than prior agents.
By Yuncheng Guo, Zhanqiu Zhang, Yiwen Guo, Weijia Li
The paper proposes a contract‑centered architecture for agentic runtimes, defining four shared responsibility objects—Skill, Harness, Scaffold, and an external data substrate—to manage capabilities, runtime, control boundaries, and data governance in enterprise AI deployments. It introduces a falsifiable hypothesis (P1) about cost‑aware capability‑capacity separability and outlines six measurable design conditions, proposing a cluster‑period randomized crossover experiment to test the hypothesis. The work presents a contract‑bounded runtime architecture, a source‑preserving data substrate, and a measurement protocol, though no implementation or empirical results are reported.
By Yaxiao Liu, Pengbo Liu, Yiwen Liu, Yihua Guan, Zhenghe Hou, Jiaxing Song
The paper introduces Persona‑Execution Separation (PES), an architecture pattern that splits an LLM agent’s persona—its instructions, tone, and self‑presentation—from its execution—stateful, auditable work—by placing them in distinct trust domains linked through a governed contract bridge. PES allows the persona to evolve freely while keeping execution traceable, using mechanisms such as an approval matrix, data‑loss‑prevention exceptions, and continuous identity. A pilot implementation on a regulated digital‑employee platform demonstrated that PES successfully decouples persona drift from execution audit, preventing re‑validation or fingerprinting of hard‑asserted fields.
By Yisen Xi
BTS-AgentBench is a deterministic, replayable pipeline that transforms read‑only telemetry logs from industrial sites into executable multi‑turn agent benchmarks. The method normalizes BTS metadata and raw histories into a read‑only tool store, compiles static tasks with tool‑derived gold answers and evidence, and lifts retained tasks into typed, bounded operator‑facing episodes. The 532‑row release includes clarifications, goal revision, timestamp policy, quality‑gated reporting, and evidence attribution, and the construction process reproduces all released train/dev/test artifacts exactly.
By Jeong-Yoon Kim
SimWAM is a lightweight World-Action Model that uses future‑video prediction only during training to supervise an action expert, enabling end‑to‑end autonomous driving without costly test‑time future imagination. The architecture co‑trains a pretrained video expert and a lightweight action expert via joint flow matching, while an isolated attention mask keeps action prediction independent of future frames. This design allows the video backbone to be swapped and the action expert scaled independently, achieving 91.5 PDMS on NAVSIM, outperforming state‑of‑the‑art WAM planners with lower latency and zero‑shot transfer to nuScenes.
By Zongchuang Zhao, Xin Zhou, Tianyang Xu, Zhengyang Sun, Kaixuan Zhou, Yu Wu, Honglin Li, Dingkang Liang, Xiang Bai
FaulT-Bench is a new benchmark comprising 200 network troubleshooting scenarios across eight topologies, designed to test large‑language‑model agents on realistic, noisy user tickets that may contain false premises or incorrect fault claims. The benchmark includes 72 rewritten tickets that vary reporter confidence and detail while keeping the network state constant, allowing isolation of the impact of ticket wording on diagnosis. Evaluation of agents such as SADE, ReAct, and Claude Code shows they perform well on accurate tickets but degrade sharply on misleading or healthy‑network tickets, revealing differing failure modes and highlighting the importance of robust reasoning over unreliable input.
By Kuan-Hao Tseng, Niruth Bogahawatta, Yasod Ginige, Kunjan Patel, Kosta Dakic, Suranga Seneviratne
Behavior2Trip introduces a new task—Behavior‑Aware Travel Planning—where user preferences are inferred from past behavior trajectories rather than explicit instructions. The benchmark contains 11,400 instances from a major Chinese travel platform, each with nearly 40 recorded behaviors across 14 attributes and 5 preference dimensions. A reinforcement‑learning agent, B2T‑Agent, leverages these trajectories, external retrieval tools, and internal memory, outperforming GPT‑4.1 and other baselines on the dataset.
By Zihao Cheng, Yingyu Shan, Hongru Wang, Zeming Liu, Xinyi Wang, Xiangrong Zhu, Yuhang Guo, Wei Lin, Yunhong Wang