Solving Physics Olympiad via Reinforcement Learning on Physics Simulators
2026-04-13 • Machine Learning
Machine LearningArtificial IntelligenceComputer Vision and Pattern RecognitionRobotics
AI summaryⓘ
The authors show that physics simulators can help train large language models (LLMs) to reason about physical problems, especially when real-world question-answer data is limited. By creating fake physics scenarios and generating questions and answers from them, they teach LLMs using reinforcement learning. Their models trained only on these simulated data perform better on real physics tests like the International Physics Olympiad. This approach offers a new way to improve physical reasoning in AI without needing vast internet QA datasets.
Large Language ModelsPhysics SimulatorsReinforcement LearningSynthetic DataPhysical ReasoningZero-shot TransferInternational Physics OlympiadQuestion-Answer PairsSim-to-Real Transfer
Authors
Mihir Prabhudesai, Aryan Satpathy, Yangmin Li, Zheyang Qin, Nikash Bhardwaj, Amir Zadeh, Chuan Li, Katerina Fragkiadaki, Deepak Pathak
Abstract
We have witnessed remarkable advances in LLM reasoning capabilities with the advent of DeepSeek-R1. However, much of this progress has been fueled by the abundance of internet question-answer (QA) pairs, a major bottleneck going forward, since such data is limited in scale and concentrated mainly in domains like mathematics. In contrast, other sciences such as physics lack large-scale QA datasets to effectively train reasoning-capable models. In this work, we show that physics simulators can serve as a powerful alternative source of supervision for training LLMs for physical reasoning. We generate random scenes in physics engines, create synthetic question-answer pairs from simulated interactions, and train LLMs using reinforcement learning on this synthetic data. Our models exhibit zero-shot sim-to-real transfer to real-world physics benchmarks: for example, training solely on synthetic simulated data improves performance on IPhO (International Physics Olympiad) problems by 5-10 percentage points across model sizes. These results demonstrate that physics simulators can act as scalable data generators, enabling LLMs to acquire deep physical reasoning skills beyond the limitations of internet-scale QA data. Code available at: https://sim2reason.github.io/.