Reproduce “Trajectory Unlearning on LLM-based 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
Yingdan Shi, Ren Wang
GiRPO reduces repetition of designated action trajectories while preserving an agent’s ability to complete tasks. It adds the unwanted sequence as a complete pseudo-rollout during policy optimization and estimates advantages using real rollouts alone. On ALFWorld Clean with post-trained Qwen2.5-3B-Instruct, exact match fell to 1.4%, while success remained 88.4%; results cover sampled behavior, not robust removal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.