LetTokenBurn
Sign in
Auto86%Harbor: pendingAgentic / Tool UseCodingMultimodal

PhysCodeGym

PhysCodeGym: Benchmarking Coding Agents for Physically Verifiable 4D Reconstruction

A benchmark of 156 tasks evaluating coding agents on physically verifiable 4D reconstruction from video using Blender scripts.

PaperRepoDataset

How to run

ltb · physcodegym
$ npm i -g lettokenburn-cli
$ ltb bench info physcodegym

External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.

Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs

Scores

9 sourced entries
HarnessSourceConfidenceLink
GPT-6 Astra0.7codexIndependentmediumOct 12, 2026arxiv.org
GPT-5.6 Sol0.6codexIndependentmediumOct 12, 2026arxiv.org
DeepSeek V4.1 Flash0.5claude codeIndependentmediumOct 12, 2026arxiv.org
GPT-5.6 Luna0.4codexIndependentmediumOct 12, 2026arxiv.org
GPT-6 Luna0.4codexIndependentmediumOct 12, 2026arxiv.org
Claude Sonnet 50.4claude codeIndependentmediumOct 12, 2026arxiv.org
Kimi K2.80.4kimi codeIndependentmediumOct 12, 2026arxiv.org
Qwen3.8 Flash0.3claude codeIndependentmediumOct 12, 2026arxiv.org
GLM-5.3 Flash0.3claude codeIndependentmediumOct 12, 2026arxiv.org

Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.

Reproduced by the community

All runs
Loading community runs…

About

PhysCodeGym comprises 156 video-to-Blender reconstruction tasks covering rigid-body, liquid, deformable, and wind-driven dynamics. It provides object checklists, event annotations, and physical requirements for systematic evaluation.

Scoring combines rule-based and LLM-based judges across six metrics, including visual fidelity and physical validity via counterfactual checks. Experiments reveal tradeoffs between artifact quality and completion, with agents often failing physical constraints despite high visual scores.