EgoGapBench - Direct Actions: leaderboard

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 March 2028. 9 models tracked.

Top models

#ModelScore
1GPT-5.4 (Medium)50.9
2GPT-5.4 (Low)40.9
3gemma-4-E4B-it34
4Gemini 3 Flash (Preview) (Low)27.8
5Qwen 2.5 VL 7B Instruct21.8
6InternVL3.5-8B20.5
7Qwen 3.5 9B20
8GPT-4o (2024-11-20)12.9

Interactive version: theaggregate.ai/benchmark?slug=egogapbench-direct-actions · How It Works · Data refreshed daily, snapshot 2026-09-29.