physics-llm-paper

PhysicsLM — ICML 2026 Paper Repository

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

Contents

Building

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:

pdflatex main.tex
bibtex main
pdflatex main.tex
pdflatex main.tex

Or with latexmk:

latexmk -pdf main.tex

Abstract

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.

Key Results

• Mean Position Error: 22.64 px (single-step, full validation set)
• Parse success rate: 100% (37/37 objects)
• Rollout stability: stable for 50+ frames (100% of scenes)
• Dataset: 900K scenes, 30 scenario types, ~582 GB uncompressed
• Browser deployment: ONNX q4f16 via WebGPU (~2x faster than q4)