Title: PhysicsLM: Autoregressive Language Modeling of 2D Rigid Body Dynamics

You need a LaTeX distribution with the ICML 2026 style file (icml2026.sty). Download the style package from the ICML 2026 author kit and place icml2026.sty in this directory, then:

Or with latexmk:

We present PhysicsLM, a system that frames 2D rigid body physics simulation as autoregressive language modeling. Simulation frames are encoded as structured text strings, and LFM2-350M is fine-tuned via LoRA to predict the next frame token-by-token. The accompanying PhysicsScenes dataset contains 900K training scenes across 24 scenario types in six physical categories. PhysicsLM achieves 22.64 px mean position error (~3% of scene diagonal) with 100% parse rate, stable 50-frame rollouts, and first-of-kind in-browser inference via WebGPU.