Reproduce “LongRCA Bench: Root-Cause Localization in Long-Horizon Agent Trajectories” 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
Yunfei Zhang, Boyu Feng, Changhua Pei, Zexin Wang, Zhihuang Peng, Xinlong Liu, Hengyue Jiang, Difeng Ma and 9 more
RCTA retrieves original records from long agent histories and compares candidate errors with upstream handoff instructions to localize root causes. LongRCA Bench evaluates it on 1,140 human-annotated failed trajectories from five sources. RCTA achieves 24.1% exact root-step accuracy, versus 13.2% for ECHO. Fewer than one quarter of roots are recovered exactly, and accuracy does not establish causal correctness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.