EgoGapBench - Indirect Actions: leaderboard
Metric: Accuracy (%; zero-shot multiple choice on 350 indirect-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 a passive response such as watching or waiting); predicted option string-matched to the key; chance about 30%). Source: arxiv.org. Saturation forecast: Around December 2026. 9 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5.4 (Medium) | 94.3 |
| 2 | GPT-5.4 (Low) | 90.6 |
| 3 | gemma-4-E4B-it | 84.6 |
| 4 | Gemini 3 Flash (Preview) (Low) | 60.3 |
| 5 | GPT-4o (2024-11-20) | 52.3 |
| 6 | Qwen 3.5 9B | 45.7 |
| 7 | Qwen 2.5 VL 7B Instruct | 41.1 |
| 8 | InternVL3.5-8B | 30.9 |
Interactive version: theaggregate.ai/benchmark?slug=egogapbench-indirect-actions · How It Works · Data refreshed daily, snapshot 2026-09-29.