Deepak Pathak

What if AI learned physics the way Newton did – by experiencing it? We built Sim2Reason: train LLMs inside virtual worlds governed by real physics laws, zero human annotation. Result: +5–10% improvement on International Physics Olympiad, zero-shot. 🧵

Excited to share Sim2Reason -- training LLMs in simulation to learn Olympiad-level physics (mechanics)! Today, LLMs learn science by reading what humans have already written, absorbing distilled knowledge from textbooks and the internet. But human-annotated physics data is fundamentally scarce, and that bottleneck isn't going away. Analogy to robotics: Sim2Real transformed robotics, where we train in simulation and deploy zero-shot in the real world. We do not try to teach robots by describing physics to them, but they have to experience it. Approach: Our Sim2Reason makes the same bet we made in robotics -- skip the descriptions, go straight to the source. Let models learn directly from simulated worlds, observing how objects move, collide, and interact, much like scientists build intuition through experiment. Result: Models trained purely on simulated experience develop transferable physical reasoning skills, improving even on problems that were never simulated. Zero-shot gains on IPhO, IIT JEE Advanced, OlympiadBench — problems the model never saw during training.

Excited to share Sim2Reason -- training LLMs in simulation to learn Olympiad-level physics (mechanics)! Today, LLMs learn science by reading what humans have already written, absorbing distilled knowledge from textbooks and the internet. But human-annotated physics data is fundamentally scarce, and that bottleneck isn't going away. Analogy to robotics: Sim2Real transformed robotics, where we train in simulation and deploy zero-shot in the real world. We do not try to teach robots by describing physics to them, but they have to experience it. Approach: Our Sim2Reason makes the same bet we made in robotics -- skip the descriptions, go straight to the source. Let models learn directly from simulated worlds, observing how objects move, collide, and interact, much like scientists build intuition through experiment. Result: Models trained purely on simulated experience develop transferable physical reasoning skills, improving even on problems that were never simulated. Zero-shot gains on IPhO, IIT JEE Advanced, OlympiadBench — problems the model never saw during training.