Reproduce “Rollback the World, Keep the Reflection: Rollback-Induced Reflection for Long-Horizon 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
Yi Yu, Liuyi Yao, Yaliang Li, Enshu Wang, Libing Wu
RIR helps long-horizon language-model agents recover from failed branches by reviewing execution, selecting an earlier checkpoint, restoring the environment, and retaining validated lessons while marking state-dependent claims for re-verification. With DeepSeek-V3, it reaches 69.43% average success, 6.57 points above ReAct. Across tested settings, it leads eight of nine benchmark combinations under the paper’s 24-step interaction budget, except GAIA with Qwen3.8-27B.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.