EgoCoT-Bench - Next Action Anticipation: leaderboard

EgoCoT-Bench (2026) Next Action Anticipation subtask: from an egocentric manipulation video, predict the most likely next action of the operator. 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), Next Action Anticipation subtask (predict the most likely next action of the operator); higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 19 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3.5 122B A10B70.71#346
2Qwen 3.5 Plus70.71#242
3Qwen 3 VL 32B70.2#622
4LLaVA-OneVision-1.5-8B69.19#1864
5Qwen 3.5 27B68.69#375
6Qwen 3.5 397B A17B68.18#264
7Qwen 3 VL 235B A22B68.18#485
8GPT-5.167.17#240
9InternVL3.5-8B66.16#1604
10Qwen 3 VL Plus66.16#512
11InternVL3.5-14B65.66#1504
12GPT-5.264.14#185
13Qwen 3 VL 8B63.13#917
14LLaVA-OneVision-1.5-4B61.11#1855
15InternVL3.5-4B59.6#1930

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-next-action-anticipation · How It Works · Data refreshed daily, snapshot 2026-11-21.