EgoGapBench - Three Options: leaderboard

Metric: Accuracy (%; zero-shot multiple choice on 618 three-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; predicted option string-matched to the key; chance about 33%). Source: arxiv.org. Saturation forecast: Around January 2027. 9 models tracked.

Top models

#ModelScore
1GPT-5.4 (Medium)66.3
2GPT-5.4 (Low)58.9
3gemma-4-E4B-it53.4
4Gemini 3 Flash (Preview) (Low)42.7
5Qwen 3.5 9B31.9
6GPT-4o (2024-11-20)31.1
7Qwen 2.5 VL 7B Instruct29.6
8InternVL3.5-8B25.1

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