Datasets:

SourajaKundu123
/

PhySimCode

license: mit
pretty_name: PhySimCode
language:
  - en
size_categories:
  - 1K<n<10K
task_categories:
  - video-text-to-text
tags:
  - physics
  - simulation
  - video
  - code-generation
  - multimodal
  - benchmark

PhySimCode: A Benchmark and Evaluation Method for Physics Video to Code Generation

Website · Paper · Code

Can a multimodal LLM watch a 2D or 3D physics simulation, work out the involved physics laws and parameter values, and write code from scratch that regenerates it?

This dataset is the PhySimCode evaluation set: 2,430 physics simulation videos, each paired with the Python code that rendered it, its true parameter values, and a step-by-step physics explanation (chain-of-thought). It is the set on which all models in the paper are evaluated.

Overview

162 physics phenomena across 9 physical domains (rigid body, spring / wave, articulated, constrained, pendulum, real-life mixed, rolling, rope / granular, gravitation / fluid)
15 samples per phenomenon, each with randomly sampled physical parameters
Two engines: analytic SciPy simulations (2D) and PyBullet rigid-body scenes (3D)
Drawn from the full PhySimCode corpus of 160,614 procedurally generated samples

Engine	Folder	Phenomena	Samples	Video
SciPy (2D)	`scipy/`	87	1,305	360×180, 10 FPS
PyBullet (3D)	`kubric/`	75	1,125	240×240, 10 FPS
Total		162	2,430

Structure

scipy/<experiment>/<sample_id>/
    video.mp4            # input to the model
    simulation_code.py   # ground-truth code that renders video.mp4
    cot.json             # ground-truth chain-of-thought
    params.json          # sampled parameter values
kubric/<experiment>/<sample_id>/
    ...                  # same four files

Every cot.json contains simulation, physical_observation, physical_law (name, statement, LaTeX formula), ode_system, derivation_steps, numerical_method, code_validation, expected_behavior and parameters (value, unit and description for every parameter).

SciPy samples also include physical_constants_derived (quantities computed from the parameters).
PyBullet samples also include category, seed, modeling_assumptions, impulse_model, pybullet_numerical_model and visual_scene.

Task

Given only video.mp4 and the engine name (scipy or kubric), a model must output:

a chain-of-thought JSON with simulation, physical_observation, physical_law and parameters, and
a self-contained Python script that regenerates the video.

Models are scored on physical-law correctness and equivalence (three LLM judges), parameter recovery within ±20%, code compilation / execution / video output, CodeBLEU, and DINOv2 / VideoCLIP similarity between the regenerated and input videos. The prompt is given in Appendix A.4 of the paper, and the evaluation code is on GitHub.

Download

pip install -U huggingface_hub
hf download SourajaKundu123/PhySimCode --repo-type dataset --local-dir benchmarking_data

from huggingface_hub import snapshot_download
snapshot_download("SourajaKundu123/PhySimCode", repo_type="dataset", local_dir="benchmarking_data")

Citation

@misc{kundu2026physimcode,
  title         = {PhySimCode: A Benchmark and Evaluation Method for
                   Physics Video to Code Generation},
  author        = {Kundu, Souraja and Gupta, Aditya and Julin, Joel and
                   Zhao, Yizhou and Xie, Liuyue and Jeni, Laszlo A.},
  year          = {2026},
  eprint        = {2610.15722},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.15722}
}

Contact

Souraja Kundu, sourajak@cs.cmu.edu