LetTokenBurnVideo2World: Benchmarking Coding Agents for Interactive World Modeling from Embodied Videos
A benchmark of 222 instances from 189 videos evaluating coding agents on autonomous video-to-simulation via geometric fidelity, dynamic fidelity, and functional correctness.
$ npm i -g lettokenburn-cli$ ltb bench info video2world
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
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Claude Fable 5.1 | 48.5 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| GPT-6 Astra | 43.5 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| Claude Opus5 | 41.6 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| Kimi K3 | 33.2 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| GPT-5.6 Sol | 31.0 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| DeepSeek V4.1 Flash | 30.5 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| Gemini 3.8 Flash | 30.1 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| Qwen3.8 Max 0902 | 26.6 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
| GLM-5.3 Flash | 25.7 | N/A | Independent | medium | Oct 3, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Video2World comprises 222 reconstruction instances derived from 189 robot and human demonstration videos. It pairs embodied videos with target simulator environments to evaluate end-to-end world modeling.
Scoring uses geometric fidelity, dynamic fidelity, and functional correctness. Evaluations show frontier agents achieve under 15% task success, revealing a gap between visual realism and factual correctness.