EgoOmniWorld
EgoOmniWorld: Unifying Egocentric Understanding via Ego-Action-World Interactions
A unified benchmark of 24,166 QA groups for egocentric video understanding, organizing questions into Node, Edge, Chain, and Loop interaction structures with visual and auditory evidence.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info egoomniworld
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
10 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Gemini 3.8 Flash | 61.5 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen 3.8 Omni Flash | 54.6 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen3.5-Omni-Plus | 49.1 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen3-Omni | 46.4 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| SALMONN2+ 72B | 44.9 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen2.5-Omni-7B | 36.4 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen2.5-Omni-3B | 36.2 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| SALMONN2+ 7B | 35.3 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| MiniCPM-o 2.6 8B | 34.4 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| VideoLLaMA2 7B AV | 22.8 | N/A | Independent | medium | Oct 13, 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
EgoOmniWorld contains 24,166 evidence-grounded QA groups derived from Ego4D and HD-EPIC datasets. It uses an interaction-first pipeline to structure questions around Ego-Action-World loops, spanning diverse modalities and single/multi-turn formats.
Scoring relies on accuracy across interaction structures. Experiments show current MLLMs have uneven limitations, but training on EgoOmniWorld improves in-domain understanding and enables zero-shot transfer to other egocentric benchmarks.