EgoCoT-Bench - State Retrospection: leaderboard

EgoCoT-Bench (2026) State Retrospection subtask: from an egocentric manipulation video, recall an earlier state of an object. 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), State Retrospection subtask (recall an earlier state of an object); higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 19 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3.5 27B72.83#375
2Qwen 3 VL 235B A22B71.32#485
3Qwen 3.5 122B A10B70.72#346
4Qwen 3.5 397B A17B69.81#264
5Qwen 3 VL Plus69.7#512
6GPT-5.267.55#185
7Qwen 3.5 Plus67.55#242
8GPT-5.166.04#240
9Qwen 3 VL 30B A3B65.15#784
10Qwen 3 VL 32B64.53#622
11Qwen 3 VL 8B64.02#917
12LLaVA-OneVision-1.5-4B61.51#1855
13InternVL3.5-2B58.11#3284
14LLaVA-OneVision-1.5-8B57.36#1864
15InternVL3.5-14B56.98#1504

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-state-retrospection · How It Works · Data refreshed daily, snapshot 2026-11-21.