CulMind-R - ReaScore: leaderboard
CulMind (2026) tests multimodal models on 50 Chinese cultural heritage tasks built from more than 100 museum collections. CulMind-R is its 24-task reasoning subset; this board is ReaScore, a 0 to 100 task-adaptive rating of the reasoning process against expert-defined dimensions; higher is better.
Metric: ReaScore (0-100): task-adaptive score of the reasoning process, each task's relevant dimensions weighted and scored against expert reasoning, on the 24-task CulMind-R reasoning subset, the model writing a structured reasoning process before its answer; higher is better. Source: arxiv.org. Saturation forecast: Around April 2028. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 64.7 | #144 |
| 2 | GPT-5.5 | 59.4 | #49 |
| 3 | Qwen 3.5 27B | 58.5 | #359 |
| 4 | Qwen 3 VL 8B Instruct | 58.4 | #981 |
| 5 | Qwen 3.5 9B | 57.7 | #803 |
| 6 | Qwen 3 VL 32B Instruct | 57.2 | #585 |
| 7 | GLM-4.5V | 55.2 | #744 |
| 8 | Gemini 2.5 Flash | 55.1 | #453 |
| 9 | GPT-5.4 Mini | 53.3 | #389 |
| 10 | InternVL3-8B | 48.2 | #1688 |
| 11 | InternVL3-14B | 48 | #1182 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=culmind-r-reascore · How It Works · Data refreshed daily, snapshot 2026-11-13.