In September 2025, Microsoft Research unveiled rStar2-Agent, an open-source AI agent reasoning framework. Despite having only 14 billion parameters, this framework attained an impressive 80.6% accuracy rate in the AIME24 mathematical reasoning test. This performance eclipses that of DeepSeek-R1, which boasts a massive 671 billion parameters (48 times larger than rStar2-Agent's). Similarly, in the scientific reasoning benchmark GPQA-Diamond test, rStar2-Agent achieved a 60.9% accuracy rate, again surpassing DeepSeek-V3. Moreover, in the BFCL v3 agent tool usage task, it reached a task completion rate of 60.8%, outstripping existing benchmarks.

The core technological innovations that underpin rStar2-Agent's success are as follows:

• An isolated code execution architecture capable of supporting 45,000 concurrent tool calls with an average latency of just 0.3 seconds. This architecture ensures efficient and rapid tool execution, a crucial factor in agent performance.
• The GRPO-RoC algorithm, which significantly reduces tool error rates to below 5% and shortens reasoning length by 30% through a 'resampling when correct' strategy. This algorithm optimizes the reasoning process, enhancing accuracy and efficiency.
• An efficient training process that leverages 'non-reasoning fine-tuning + multi-stage reinforcement learning.' This approach requires only 510 training steps and 64 MI300X GPUs to complete model optimization within a week, streamlining the development cycle and reducing resource consumption.

The rStar2-Agent project has been made open source on GitHub, with the aim of expediting the industrialization of agent technology and fostering innovation in the field.