EgoGapBench - Four Options: leaderboard

Metric: Accuracy (%; zero-shot multiple choice on 382 four-option 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, plus an off-vantage distractor that fits the activity but not the viewpoint; predicted option string-matched to the key; chance about 25%). Source: arxiv.org. Saturation forecast: Around March 2027. 9 models tracked.

Top models

#ModelScore
1GPT-5.4 (Medium)65.7
2GPT-5.4 (Low)57.3
3gemma-4-E4B-it49
4Gemini 3 Flash (Preview) (Low)33.5
5Qwen 2.5 VL 7B Instruct27
6Qwen 3.5 9B24.3
7InternVL3.5-8B22.5
8GPT-4o (2024-11-20)19.6

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