arXiv · cs.AI · Avinash Baidya, Xinran Liang, Ruocheng Guo, Xiang Gao, Kamalika Das2026-11-03 13:00 · 9 小时前模型评分43

轻量外部监控器实现对话与大语言模型（LLM，Large Language Model）智能体轨迹的弱监督早期失败预警

When Evidence is Sparse: Weakly Supervised Early Failure Alerting in Dialogs and LLM-Agent Trajectories

导读

研究者提出一种仅用轨迹级结果训练的轻量外部监控器，用于在对话或大语言模型（LLM，Large Language Model）智能体轨迹展开过程中预警失败，其基于注意力的多实例学习（MIL，Multiple Instance Learning）组件从轨迹结果中学习每轮贡献并蒸馏为在线证据特征。

分类置信度

中 · 待人工复核

来源级别：媒体与个人

事件类型：研究发布

证据摘录

“We address these limitations with a lightweight external monitor trained only from trajectory-level outcomes, motivated by a pattern across five benchmarks: turns rated highly relevant to failure are sparse and often late.”

来源：arXiv · cs.AI · arxiv.org

轻量外部监控器实现对话与 大语言模型（LLM，Large Language Model） 智能体轨迹的弱监督早期失败预警