EgoGapBench - Direct Actions: leaderboard
EgoGapBench (2026) asks multimodal LLMs to pick the action available from an assigned viewpoint in multi-agent COCO scenes without first-person body cues, where copying another visible agent is a structural error. Accuracy in percent on the 650 items whose correct answer is an active intervention, higher is better.
Metric: Accuracy (%; zero-shot multiple choice on 650 direct-action single-image items built from COCO multi-agent scenes: choose the action available from the given viewpoint, with distractors copying the singleton action of another visible agent (the correct answer is an active intervention); predicted option string-matched to the key; chance about 30%). Source: arxiv.org. Saturation forecast: Around June 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (2026-03-05) (Medium) | 50.9 | #87 (GPT-5.4 (2026-03-05)) |
| 2 | GPT-5.4 (2026-03-05) (Low) | 40.9 | #87 (GPT-5.4 (2026-03-05)) |
| 3 | gemma-4-E4B-it | 34 | #1469 |
| 4 | Gemini 3 Flash (Preview) (Low) | 27.8 | #144 (Gemini 3 Flash (Preview)) |
| 5 | Qwen 2.5 VL 7B Instruct | 21.8 | #1833 |
| 6 | Qwen 3.5 9B | 20 | #803 |
| 7 | GPT-4o (2024-11-20) | 12.9 | #866 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=egogapbench-direct-actions · How It Works · Data refreshed daily, snapshot 2026-11-13.