Microsoft AI Introduces rStar2-Agent: A 14B Math Reasoning Model Trained with Agentic Reinforcement Learning to Achieve Frontier-Level Performance

Microsoft has developed rStar2-Agent, a 14-billion-parameter large language model that advances mathematical reasoning by employing agentic reinforcement learning, enabling the model to interact dynamically with a Python execution environment to verify, explore, and refine its reasoning steps. This approach overcomes the limitations of traditional Chain-of-Thought methods, which often compound subtle errors by simply "thinking longer," by teaching the model to "think smarter" through active tool use and iterative self-correction.