Chiaroscuro: leaderboard
Contrastive emotion recognition grounded in appraisal theory: minimally different situations that flip the appraised emotion. Macro-F1, higher is better. Source: arxiv.org/abs/2609.03394.
Metric: Macro-F1 (%). Source: arxiv.org. Saturation forecast: Around 2030. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 67.3 | #33 |
| 2 | Qwen 3.6 Plus | 66.9 | #124 |
| 3 | DeepSeek V4 Pro | 66.5 | #112 |
| 4 | Qwen 3.5 27B | 66.3 | #192 |
| 5 | Llama 3.3 70B | 66.3 | #578 |
| 6 | Gemini 3.5 Flash | 64.1 | #66 |
| 7 | Qwen 3.5 9B | 59.9 | #391 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=chiaroscuro · How It Works · Data refreshed daily, snapshot 2026-11-05.