DeFA：基于依赖关系的LLM代理故障归因
DeFA是一个新的框架，通过构建事件依赖图和故障传播图来帮助识别LLM代理执行中的根本错误原因。
01
证据
- AarXiv cs.AI一手来源2026-10-02 12:00
> Abstract: Errors in LLM agent executions and their visible consequences can be separated by many steps, making decisive-error localization a matter of understanding both step content and step dependencies.
- AarXiv cs.AI一手来源2026-10-02 12:00
Abstract: Errors in LLM agent executions and their visible consequences can be separated by many steps, making decisive-error localization a matter of understanding both step content and step dependencies. We introduce DeFA, a dependency-guided framework for agent failure attribution. DeFA first combines protocol relations and semantic dependencies into an event dependency graph spanning the traj…