VisuLogic: leaderboard
Visual logic reasoning benchmark requiring models to perform logical deduction from visual inputs without relying on language shortcuts.
Metric: Overall Accuracy (%). Source: visulogic-benchmark.github.io. Saturation forecast: Around 2031. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Median Human | 51.4 | |
| 2 | Qwen 3 VL 235B A22B (Thinking) | 34.4 | #238 (Qwen 3 VL 235B A22B) |
| 3 | O3 | 29.5 | #124 |
| 4 | Gemini 2.0 Pro (Preview 02-05) | 28 | #307 |
| 5 | InternVL3-78B | 27.7 | #352 |
| 6 | InternVL2.5-78B | 27.3 | #431 |
| 7 | ERNIE-4.5-VL-424B-A47B | 27.3 | #391 |
| 8 | InternVL3-38B | 27.1 | #405 |
| 9 | GPT-4o | 26.3 | #343 |
| 10 | Qwen 2.5 VL 72B Instruct | 26.2 | #377 |
| 11 | Qwen 2.5 VL 7B Instruct | 26 | #654 |
| 12 | ERNIE 4.5 VL 28B A3B | 26 | #695 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=visulogic · How It Works · Data refreshed daily, snapshot 2026-10-20.