PhysicsLM: Physics as Token Prediction

🤖 PhysicsLM: Physics as Token Prediction

PhysicsLM has been developed, a system that uses autoregressive language modeling to predict 2D rigid body dynamics. The LFM2-350M model was fine-tuned using the LoRA method on the PhysicsScenes dataset consisting of 900,000 scenes and supports stable simulations via WebGPU.

🌍 This demonstrates the potential of using LLM architectures to solve physical modeling tasks, turning simulation into a next-token prediction task.

👤 Physical simulations can now be run directly in the browser at high speeds, while language models are being trained to understand the laws of object motion.

Source 1: https://github.com/AlexWortega/physics-llm-paper