EgoCoT-Bench - Reasoning Score: leaderboard

EgoCoT-Bench (2026) reasoning quality: on egocentric manipulation video questions, a Qwen-Max judge scores each model rationale from 0 to 5 against the annotated reference reasoning for coherence, factual consistency and grounding in the visual evidence. Mean score out of 5; higher is better.

Metric: Reasoning score (out of 5): a Qwen-Max judge rates each generated rationale against the annotated reference reasoning for logical coherence, factual consistency and alignment with the visual evidence, 0 to 5, mean over the 3,172 questions of EgoCoT-Bench; 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 Plus3.08#512
2Qwen 3.5 27B2.96#375
3Qwen 3 VL 32B2.96#622
4Qwen 3.5 122B A10B2.94#346
5Qwen 3.5 Plus2.92#242
6Qwen 3.5 397B A17B2.87#264
7GPT-5.22.85#185
8Qwen 3 VL 30B A3B2.79#784
9Qwen 3 VL 235B A22B2.78#485
10GPT-5.12.77#240
11Qwen 3 VL 8B2.73#917
12InternVL3.5-14B2.6#1504
13InternVL3.5-8B2.56#1604
14InternVL3.5-2B2.53#3284
15LLaVA-OneVision-1.5-4B2.5#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-reasoning-score · How It Works · Data refreshed daily, snapshot 2026-11-21.