arXiv 2601.068182026-01-11cs.CL

AgentHallu: 评估基于大语言模型的代理的自动幻觉归因

AgentHallu: Benchmarking Automated Hallucination Attribution of LLM-based Agents

Beijing University of Posts and Telecommunications（北京邮电大学）；Department of Computer Science & Technology, Tsinghua University（清华大学计算机科学与技术系）；University of California, Santa Barbara（加州大学圣芭芭拉分校）；Center for Research on Intelligent Perception and Computing, NLPR, CASIA（智能感知与计算研究中心，国家工程实验室）

Xuannan Liu, Xiao Yang, Zekun Li, Peipei Li, Ran He

备注：: Project page: https://liuxuannan.github.io/AgentHallu.github.io/

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

AgentHallu提出一个评估基于大语言模型的代理自动识别幻觉来源的任务，通过高质量轨迹和多级注释评估13个模型，发现顶级模型在定位幻觉步骤上表现有限，工具使用幻觉最难识别。

摘要：

随着基于大语言模型的代理在连续的多步骤推理中运行，中间步骤中出现的幻觉可能会沿着轨迹传播，从而降低整体可靠性。与单轮响应中的幻觉检测不同，诊断多步骤工作流程中的幻觉需要确定导致初始分歧的步骤。为此，我们提出一个新的研究任务，即基于大语言模型的代理的自动幻觉归因，旨在识别导致幻觉的步骤并解释原因。为了支持该任务，我们引入了AgentHallu，一个全面的基准测试，包含：（1）693个高质量轨迹，涵盖7个代理框架和5个领域；（2）一个按5个类别（规划、检索、推理、人机交互和工具使用）和14个子类组织的幻觉分类体系；（3）由人类精心编写的多级注释，涵盖二元标签、幻觉责任步骤和因果解释。我们评估了13个领先的模型，结果表明该任务即使对顶级模型（如GPT-5、Gemini-2.5-Pro）也具有挑战性。表现最好的模型在步骤定位准确率上仅达到41.1%，其中工具使用幻觉是最具挑战性的，仅达到11.6%。我们相信AgentHallu将推动未来研究，以开发更加稳健、透明和可靠的代理系统。

英文摘要：

As LLM-based agents operate over sequential multi-step reasoning, hallucinations arising at intermediate steps risk propagating along the trajectory, thus degrading overall reliability. Unlike hallucination detection in single-turn responses, diagnosing hallucinations in multi-step workflows requires identifying which step causes the initial divergence. To fill this gap, we propose a new research task, automated hallucination attribution of LLM-based agents, aiming to identify the step responsible for the hallucination and explain why. To support this task, we introduce AgentHallu, a comprehensive benchmark with: (1) 693 high-quality trajectories spanning 7 agent frameworks and 5 domains, (2) a hallucination taxonomy organized into 5 categories (Planning, Retrieval, Reasoning, Human-Interaction, and Tool-Use) and 14 sub-categories, and (3) multi-level annotations curated by humans, covering binary labels, hallucination-responsible steps, and causal explanations. We evaluate 13 leading models, and results show the task is challenging even for top-tier models (like GPT-5, Gemini-2.5-Pro). The best-performing model achieves only 41.1\% step localization accuracy, where tool-use hallucinations are the most challenging at just 11.6\%. We believe AgentHallu will catalyze future research into developing robust, transparent, and reliable agentic systems.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。