What is rStar2-Agent
rStar2-Agent is Microsoft's open-source 14-billion-parameter mathematicalreasoning model. The model is trained using agentic reinforcement learning methods and achieves an accuracy of 80.6% on the AIME24 mathematical reasoning test, surpassing the 671-billion-parameterDeepSeek-R1. The model excels at mathematical reasoning and demonstrates strong generalization capabilities in tasks such as scientific reasoning and agentic tool use. Through three major technological breakthroughs—efficient training infrastructure, innovative algorithms, and a unique training pipeline—the model achieves high performance with low computing costs, bringing new ideas to the field of AI reasoning.
Main Features of rStar2-Agent
-
Efficient Mathematical Reasoning: On mathematical reasoning tests such as AIME24, it achieves a high accuracy of 80.6% with 14 billion parameters, surpassing models with far larger parameter counts, and can quickly and accurately solve complex mathematical problems.
-
Scientific Reasoning Ability: On the GPQA-Diamond scientific reasoning benchmark, it achieves an accuracy of 60.9%, demonstrating understanding and reasoning ability for scientific knowledge, and supporting applications in scientific question answering.
-
Intelligent Tool Invocation: Automatically invokes appropriate tools based on problem requirements, such as code execution tools, improving problem-solving efficiency.
-
Strong Generalization Ability: Performs excellently in specific domains and supports generalizing reasoning ability to a variety of other tasks and domains, with broad application potential.
Technical Principles of rStar2-Agent
-
Agentic Reinforcement Learning: The model interacts with specific tool environments, adjusts its reasoning process based on environmental feedback, and uses a reward mechanism to guide the model in learning better reasoning strategies, achieving efficient learning and improved reasoning ability.
-
Efficient Training Infrastructure: Uses an isolated high-throughput code execution service, based on a distributed architecture of 64 AMD MI300X GPUs, to achieve high-concurrency tool invocation and fast execution, ensuring an efficient and stable training process.
-
GRPO-RoC Algorithm: Integrates the Resample-on-Correct rollout strategy to optimize tool use, filters high-quality trajectories through an asymmetric sampling mechanism, reduces error rates, and makes model reasoning more accurate and efficient.
-
Multi-Stage Reinforcement Learning Training Pipeline: First conducts non-reasoning fine-tuning to cultivate basic capabilities, then uses staged reinforcement learning to gradually improve reasoning ability, completing training within one week using 64 GPUs to reach peak performance and greatly reduce costs.
Project Address of rStar2-Agent
- GitHub Repository：https://github.com/microsoft/rStar
- arXiv Technical Paper：https://www.arxiv.org/pdf/2508.20722
Application Scenarios of rStar2-Agent
- Education: Provides personalized learning tutoring for students, helps improve academic performance, and quickly grades homework and exams, improving teaching efficiency.
- Scientific Research: Assists in analyzing complex data, and builds and optimizes scientific models, providing strong support for scientific research decisions.
- Finance: Accurately predicts stock trends, provides scientific investment advice for investors, and monitors transaction data in real time, effectively preventing financial fraud risks.
- Engineering: rStar2-Agent optimizes engineering design plans, ensures high-quality project completion, and diagnoses system faults in real time, improving engineering efficiency.
- Daily Life: As an intelligent assistant, it provides personalized services and, based on user health data, formulates scientific health management plans.