Reproduce “LLM Value Functions for Credit Assignment in 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
Kaijie Zhu, Weixiang Yan, Omar Attia, Abbas Kazerouni, Alexander Toshev, Andrew Szot
LVRL redistributes sparse task outcomes into step-level credit for long-horizon agents. It uses a frozen language model to score completed trajectories in hindsight, conditioning on same-task peer runs, and anchors estimates to group success rates and observed final rewards. On BabyAI’s OOD layouts, Qwen3-8B reaches 65.5% ± 3.0% success, while GRPO and PPO each reach 0.0%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.