The paper, “A Survey for LLM Agent Trajectory Analysis: From Failure Attribution to Enhancement,” systematically reviews 55 recent studies on the analysis of LLM agent execution trajectories. The work organizes the emerging research landscape across five dimensions: failure taxonomy, failure attribution, system enhancement and optimization, monitoring and analysis tools, and datasets and benchmarks.

The survey highlights remaining challenges including reliable step-level failure attribution, agent observability, realistic benchmarks, and the development of agents capable of learning from their own execution histories. The paper was authored by Junjie Wang, Yawen Wang, Mengzhuo Chen, Xiaofei Xie, Chunyang Chen, Fangwen Mu, Zhe Liu, and Qing Wang. More information and the publication are available here.