Reproduce “ReCast: Attribution-Oriented Step Representation Learning for LLM-Based Agent Systems” 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

Weilin Jin, Mingyu Wang, Taiyu Zhu, Ziqi Zhou, Wenbo Li, Haoyang Huang, Nan Duan, Yifan Wu and 2 more

ReCast adapts frozen large language model hidden states to rank candidate actions by their likelihood of being the earliest decisive error in a failed multi-agent trace. It selects layers, combines activation patterns with deviations from successful executions, and trains a contextual encoder to score steps. On four benchmark subsets, it reports the best Hit@1; benchmark rankings do not prove causal responsibility in new traces.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.