EmoTrans: leaderboard

EmoTrans (2026) tests whether multimodal LLMs follow emotion transitions in real-world videos: detecting a change, naming the emotions before and after, and predicting the next emotion, over 1,000 annotated clips. Average accuracy (%) of the three classification tasks; higher is better.

Metric: Average accuracy (%) over Emotion Change Detection, Emotion State Identification and Next Emotion Prediction on EmoTrans, built from 1,000 manually annotated real-world video clips (OpenHumanVid and CH-SIMS v2) over 12 scenarios with one or more target persons; zero-shot, video with its text (and audio where the model accepts it); higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.449.3#135
2Seed 2.0 Lite47.4#300
3Qwen 3.5 122B A10B47#346
4Gemini 3 FlashNon-reasoning46.8#175
5Gemini 3 Pro46.7#145
6Qwen 3.5 27B46.4#375
7GLM-4.6V46.1#710
8GLM-4.5V45.2#754
9Qwen 3 VL 32B Instruct43.5#606
10Qwen 2.5 VL 72B Instruct43.5#841
11Gemini 2.5 FlashNon-reasoning40.9#463
12Qwen 3 VL 8B40.9#917
13InternVL3.5-14B37.8#1504
14Qwen 3 Omni 30B A3B37.4#847
15InternVL3.5-8B37.2#1604

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

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