rStar2-Agent: Agentic Reasoning Technical Report

rStar2-Agent is a 14B math reasoning model trained with agentic reinforcement learning (RL) to achieve frontier-level performance. It demonstrates advanced cognitive behaviors such as thinking carefully before using Python coding tools and reflecting on code execution feedback to autonomously explore, verify, and refine intermediate steps. The model's capabilities are enabled through three key innovations: (i) an efficient RL infrastructure with a reliable Python code environment supporting high-throughput execution on limited GPU resources (64 MI300X GPUs); (ii) GRPO-RoC, an agentic RL algorithm with a Resample-on-Correct rollout strategy that mitigates environment noises from coding tools; (iii) an efficient agent training recipe starting with non-reasoning SFT and progressing through multi-RL stages. rStar2-Agent achieves average pass@1 scores of 80.6% on AIME24 and 69.8% on AIME25, surpassing DeepSeek-R1 (671B) with significantly shorter responses. The model also generalizes to alignment, scientific reasoning, and agentic tool-use tasks.