World Embedding Benchmark
Evaluates physical information in video embeddings using 8,000 simulation cases across fluid, solid, dynamics, and optics physics families.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info world-embedding-benchmark
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
11 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| LCO-Embedding-3B + Physics Adaptation | 4.0 | N/A | Independent | medium | Oct 2, 2026 | arxiv.org |
| LCO-Embedding-7B + Physics Adaptation | 2.9 | N/A | Independent | medium | Oct 2, 2026 | arxiv.org |
| LCO-Embedding-3B | 2.5 | N/A | Independent | medium | Oct 2, 2026 | arxiv.org |
| LCO-Embedding-7B | 2.0 | N/A | Independent | medium | Oct 2, 2026 | arxiv.org |
| LCO-Embedding-3B-2605 | 1.9 | N/A | Independent | medium | Oct 2, 2026 | arxiv.org |
| Qwen3-VL-Embedding-8B | 1.8 | N/A | Independent | medium | Oct 2, 2026 | arxiv.org |
| Qwen3-VL-Embedding-2B | 1.8 | N/A | Independent | medium | Oct 2, 2026 | arxiv.org |
| Omni-Embed-Nemotron-3B | 1.4 | N/A | Independent | medium | Oct 2, 2026 | arxiv.org |
| V-JEPA2 ViT-g | 1.0 | N/A | Independent | medium | Oct 2, 2026 | arxiv.org |
| V-JEPA2 ViT-L | 0.9 | N/A | Independent | medium | Oct 2, 2026 | arxiv.org |
| Video-CLIP-XL | 0.8 | N/A | Independent | medium | Oct 2, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
Comprises 8,000 controlled physics simulation videos with annotations, supporting text-video retrieval, property regression, and pair classification tasks.
Findings show pre-trained models have weak retrieval but recoverable physical signals via probes; contrastive training improves alignment but degrades regression.