Reproduce “When Credit Turns Against Tools: Transition-Aware Credit Assignment in Agentic Reinforcement Learning” 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
Xiangyu Lu, Zhanke Zhou, Jiazhe Ning, Chentao Cao, Jiangchao Yao, Bo Han
TACA adjusts credit on tool-call tokens, compares tool use with direct answering for the same prompt, and replays calls from correct trajectories in multi-turn reinforcement learning with verifiable rewards. It addresses how MT-GRPO assigns a trajectory’s terminal advantage across its turns. On six math benchmarks, Qwen2.5-1.5B-Instruct reaches 33.24% mean accuracy versus 28.63% with MT-GRPO, in the reported evaluation setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.