P1: Mastering Physics Olympiads with Reinforcement Learning

Overview

Physics reasoning is central to understanding and shaping the real world. Top contests like the International Physics Olympiad (IPhO) set a high bar for complex reasoning and deep physical understanding — a benchmark for evaluating AI's grasp of reality.

P1 is the first open-source model series designed to tackle Olympiad-level physics reasoning through multi-stage reinforcement learning (RL) and a co-evolutionary multi-agent system (PhysicsMinions). It achieved gold medal-level performance on IPhO 2025. We release two model versions:

• P1-30B-A3B: A 30B parameter model that surpasses larger closed-source models, demonstrating exceptional efficiency
• P1-235B-A22B: A 235B parameter model achieving gold medal performance on IPhO 2025, rivaling top closed-source models

Results

P1 models demonstrate top-tier physics reasoning across all HiPhO contests.

P1’s physics reasoning transfers effectively across other STEM domains.

STEM Benchmarks

🧮 HiPhO Benchmark

HiPhO (High School Physics Olympiad) is the first benchmark focused on recent Olympiad-level physics contests with human-aligned evaluation.

📚 It compiles 13 competitions (IPhO, APhO, EuPhO, etc.) from 2024–2025, using official rubrics and fine-grained scoring aligned with medal cutoffs.

Co-Evolution Multi-Agent System: PhysicsMinions

To go beyond single-model limits, P1 introduces PhysicsMinions — a co-evolution multi-agent system that iteratively refines solutions through self-verification and reflection.

Failures trigger a feedback loop to improve the reasoning process — resulting in stronger robustness and reliability.

Acknowledgements

We are grateful to the open-source community for their invaluable contributions. Special thanks to:

• Qwen3 - for providing the foundational base models that powered our research
• slime - for their innovative work on efficient reinforcement learning framework that powered our training pipeline
• verl - for the versatile reinforcement learning framework that enabled our training pipeline
• sglang - for the efficient LLM serving and inference infrastructure
• Megatron-LM - for the large-scale model training framework

We also thank colleagues and collaborators who supported the development of P1 models, the accompanying datasets and visual assets.

🧾 Citation

If you find this work useful, please cite:

@misc{p12025,
  title={P1: Mastering Physics Olympiads with Reinforcement Learning},
  author={P1 Team},
  year={2025},
  url={https://prime-rl.github.io/P1/}
}