PhysicsLM

Anonymous submission for ICML 2026: "PhysicsLM: Autoregressive Language Modeling of 2D Rigid Body Dynamics"

PhysicsLM fine-tunes LFM2-350M (LiquidAI) via LoRA on 900K 2D rigid-body physics scenes, learning to predict next simulation states as structured decimal text.

Model details

• Base model: LiquidAI/LFM2-350M
• Fine-tuning: LoRA (r=32, alpha=64), 5-stage curriculum on PhysicsScenes
• Task: Next-frame physics prediction (autoregressive text generation)
• Format: structured decimal text encoding of 2D object states

Results (seen scenarios)

Category	PhysicsLM RMSE (px)	Copy-last RMSE	Linear extrap RMSE
Stacking	2.60	6.72	0.06
Constraint	1.35	4.99	0.06
Collision	5.37	7.69	0.09
Ramp	18.85	...	0.19
Minigame	36.14	...	0.09
Complex	109.57	...	0.04

OOD: near-distribution 0.94 px RMSE, novel OOD 24.79 px RMSE. Parse failure: 0.0%.

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tok = AutoTokenizer.from_pretrained("anonsubmiticml2026/PhysicsLM")
model = AutoModelForCausalLM.from_pretrained("anonsubmiticml2026/PhysicsLM",
                                              torch_dtype=torch.bfloat16,
                                              device_map="cuda")
# See paper for text encoding format

Dataset

Training data: anonsubmiticml2026/PhysicsScenes

Code: anonsubmiticml2026/physics-llm-paper