EgoIntrospect - Valuable Interaction Identification: leaderboard

Which of four proactive assistant requests the wearer accepted for a clip, asked on egocentric smart-glasses clips annotated by the wearers themselves (EgoIntrospect, 2026); native audio and video models hear the clip audio and the others read a transcript. Accuracy in percent on 1,161 four-choice questions; higher is better.

Metric: Accuracy (%) on 1,161 four-choice questions asking which of four proactive assistant requests the wearer accepted for a clip (distractors are requests the wearer did not accept), fixation-overlaid egocentric clips from 15 held-out participants with the participant profile; native audio-video models hear the clip audio, the other models read its transcript; temperature 0 (0.6 for InternVL3.5 thinking mode); higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 19 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)55.81#99
2Kimi K2.554.78#266
3Gemini 3 Flash (Preview)52.8#152
4Qwen 3.5 397B A17B51.42#264
5Qwen 3.5 Omni Plus49.96#254
6Qwen 3 VL 8B Instruct47.63#1029
7Qwen 3 Omni 30B A3B (Thinking)47.29
8Qwen3 Omni 30B A3B Instruct46.34#930
9Qwen 3 VL 8B (Thinking)46.17#910
10GPT-4o43.84#751
11InternVL2.5-8B43.58#2370
12MiniCPM-V-4.543.32#1441
13InternVL3-8B42.89#1782
14InternVL3.5-8B-Instruct41.69#1891
15Qwen 2.5 VL 7B Instruct41.09#1902

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=egointrospect-valuable-interaction-identification · How It Works · Data refreshed daily, snapshot 2026-11-21.