EgoCoT-Bench - Active Object Grounding: leaderboard

EgoCoT-Bench (2026) Active Object Grounding subtask: from an egocentric manipulation video, identify the object the operator is attending to, touching or manipulating. Four-way multiple choice accuracy in percent; higher is better.

Metric: Answer accuracy (%, four-way multiple choice, strict exact match) of zero-shot MLLMs on egocentric manipulation videos of EgoCoT-Bench (351 clips, 3,172 human-verified questions built from spatio-temporal scene graphs), Active Object Grounding subtask (identify the object the operator is attending to, touching or manipulating); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 19 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 8B69.54#917
2Qwen 3 VL Plus69.21#512
3Qwen 3.5 397B A17B68.26#264
4Qwen 3.5 27B68.26#375
5Qwen 3.5 122B A10B68.26#346
6Qwen 3.5 Plus68.26#242
7Qwen 3 VL 32B67.78#622
8Qwen 3 VL 235B A22B67.54#485
9GPT-5.264.92#185
10GPT-5.164.2#240
11Qwen 3 VL 30B A3B62.44#784
12InternVL3.5-4B58.71#1930
13InternVL3.5-8B56.32#1604
14InternVL3.5-14B56.32#1504
15LLaVA-OneVision-1.5-4B55.74#1855

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

Interactive version: theaggregate.ai/benchmark?slug=egocot-bench-active-object-grounding · How It Works · Data refreshed daily, snapshot 2026-11-21.