Chehre: leaderboard
Dominant facial-expression recognition on Chehre (Azari et al., 2026): 2,111 videos of emoji-prompted expressions re-enacted on synthetic faces, each rated by about 30 perceivers. A video language model picks the top two labels from the candidate set the raters saw, including None; top-1 accuracy in percent against the label with the highest mean human rating, higher is better.
Metric: Top-1 accuracy (%; share of the 2,111 videos where the model's first label is the label with the highest mean human rating; closed candidate label set including None; 32 uniformly sampled frames, 24 for InternVL3-38B). Source: arxiv.org. Saturation forecast: Around September 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 43.5 | #452 |
| 2 | GPT-5.6 Terra (Non-reasoning) | 42.7 | #60 (GPT-5.6 Terra) |
| 3 | Qwen2.5-Omni-7B | 33.2 | #1797 |
| 4 | InternVL3-38B | 32.3 | #905 |
| 5 | InternVL3-8B | 31.4 | #1695 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=chehre · How It Works · Data refreshed daily, snapshot 2026-11-12.