CurveBench - Hard: leaderboard
CurveBench (Mohseni et al., 2026) shows a vision-language model a hand-drawn image of nested, non-intersecting closed curves and asks for the rooted tree of which regions contain which, written as parent-child edges; an answer counts only if it is isomorphic to the ground-truth tree. The Hard set is the main benchmark: 456 polygon, topographic, maze and counting images, one rollout each, with an 8,192-token generation cap. Tree-generation accuracy in percent; higher is better.
Metric: Tree-Generation Accuracy (%; exact containment tree, 456 polygon, topographic, maze and counting images, one rollout). Source: arxiv.org. Saturation forecast: Around February 2028. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 19.1 | #91 |
| 2 | GPT-5.4 | 6.6 | #121 |
| 3 | Claude Opus 4.5 | 4.2 | #152 |
| 4 | Qwen 3 VL 8B (Thinking) | 4.2 | #887 |
| 5 | GPT-5.4 Mini | 2.4 | #389 |
| 6 | Gemma 3 12B (IT) | 0.7 | #1849 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=curvebench-hard · How It Works · Data refreshed daily, snapshot 2026-11-12.