EgoGapBench: leaderboard

Metric: Accuracy (%; zero-shot multiple choice on 1,000 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 (618 three-option and 382 four-option items, the latter adding an off-vantage distractor); predicted option string-matched to the key; chance about 30%). Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.

Top models

#ModelScore
1GPT-5.4 (Medium)66.1
2GPT-5.4 (Low)58.3
3gemma-4-E4B-it51.7
4Gemini 3 Flash (Preview) (Low)39.2
5Qwen 3.5 9B29
6Qwen 2.5 VL 7B Instruct28.6
7GPT-4o (2024-11-20)26.7
8InternVL3.5-8B24.1

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