AIBullisharXiv – CS AI · 14h ago7/10
🧠
Solving Physics Olympiad via Reinforcement Learning on Physics Simulators
Researchers demonstrate that physics simulators can generate synthetic training data for large language models, enabling them to learn physical reasoning without relying on scarce internet QA pairs. Models trained on simulated data show 5-10 percentage point improvements on International Physics Olympiad problems, suggesting simulators offer a scalable alternative for domain-specific AI training.