Reproduce “AttnSR: Attention Step Reward for Long-Horizon Agentic RL” 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
Mengsong Wu, Tianchen Zhou, Zhaochen Su, Zhenhua Liu, Zhe Cui, Aimin Zhou
AttnSR distributes additional rewards across an agent’s earlier steps by reusing the policy’s attention during replay, treating attention as a relevance heuristic rather than evidence of causation. On held-out ALFWorld tasks, a matched comparison reports success rising from 91.24% to 97.08%, with 20 tasks won and 4 lost; evidence comes from one training run per method.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.