Reproduce “DeFA: Dependency-Guided Failure Attribution for LLM Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bo Deng, Xinlei Zheng, Yi Wei, Kang Zhou, Chongyang Tao, Renzhao Liang, Xuanren Chen, Lifan Guo and 1 more
DeFA traces an agent’s failure from its final answer to an earlier decisive error by linking retrieved information, its interpretation, and downstream actions. In a film-ranking example, it connects a misread table entry to a count of five shared films instead of six. On 300 selected Who&When Pro text trajectories, DeFA with GLM-5.2 achieves 80.3% exact Step accuracy, versus 72.0% for CockpitHAT.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.