AICA-Bench: leaderboard
AICA-Bench (2026) tests affective image content analysis by vision-language models: recognizing expressed and evoked emotions, explaining why an image evokes an emotion, and writing emotion-congruent descriptions. The score is the mean of the three task scores in percent, higher is better.
Metric: Overall average on a percent scale: unweighted mean of the emotion understanding average (weighted F1 of basic and CoT prompting), the emotion reasoning score and the emotion-guided content generation score; reasoning and generation are 1-5 ratings rescaled to 20-100, so the floor is 40/3, over AICA-Bench (8,086 affective images from nine public emotion datasets with GPT-4o-generated instructions; closed-source models through their APIs at standard settings, open models on A100 GPUs); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 23 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 73.49 | #276 |
| 2 | Qwen-VL-Max (AICA-Bench checkpoint unspecified) | 72.9 | |
| 3 | GPT-4o | 72.82 | #751 |
| 4 | Gemini 2.5 Flash | 71.14 | #463 |
| 5 | GPT-4o Mini | 70.81 | #1667 |
| 6 | Qwen-VL-Plus (AICA-Bench checkpoint unspecified) | 67.73 | |
| 7 | Gemini 2.0 Flash | 67.68 | #723 |
| 8 | Qwen 2.5 VL 7B Instruct | 65.78 | #1902 |
| 9 | Ovis2-16B | 64.78 | #1594 |
| 10 | Ovis2-8B | 64.29 | #2047 |
| 11 | InternVL3-14B-Instruct | 62.41 | #1801 |
| 12 | InternVL3-8B-Instruct | 62.35 | #2872 |
| 13 | InternVL2.5-8B | 62.27 | #2370 |
| 14 | MiniCPM-o-2.6 | 61.94 | #2423 |
| 15 | Qwen 2.5 VL 3B Instruct | 61.81 | #2713 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=aica-bench · How It Works · Data refreshed daily, snapshot 2026-11-21.