EmoMM: leaderboard
EmoMM (Sun et al., 2026) tests multimodal emotion recognition when video, audio and transcript disagree or one is missing: 4,000 English CMU-MOSI and Chinese CH-SIMS v2.0 clips carry separate per-modality sentiment labels, and models give a five-class sentiment label zero-shot. Accuracy in percent on the held-out test split, all subsets combined; higher is better.
Metric: Accuracy (%; all test subsets; five-class sentiment from video, audio and transcript; 20% speaker-independent test split of 4,000 CMU-MOSI and CH-SIMS v2.0 clips; zero-shot prompt). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 56.41 | #135 |
| 2 | Qwen2.5-Omni-7B | 44.53 | #1797 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=emomm · How It Works · Data refreshed daily, snapshot 2026-11-12.