VIA-Bench (LLM Judge) - Visual Anomalies: leaderboard
Metric: Accuracy (%) on the visual anomalies (common-sense violations such as a hand with six fingers) questions, answer letter read from the response by a GPT-4.1-mini judge; VIA-Bench's 1,004 human-verified multiple-choice questions on illusory and anomalous images, every option set including a never-correct 'Not Sure' option; zero-shot with an answer-format instruction, temperature 0.8 where supported (model defaults otherwise), mean of five runs; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 44.49 | #64 |
| 2 | Gemini 2.5 Pro | 38.74 | #145 |
| 3 | GPT-4o (2024-11-20) | 35.31 | #369 |
| 4 | GPT-5 Mini | 26.95 | #176 |
| 5 | Claude 3.5 Sonnet (20241022) | 24.44 | #286 |
| 6 | O4 Mini | 24.02 | #172 |
| 7 | Qwen 2.5 VL 7B Instruct | 22.93 | #643 |
| 8 | O3 | 22.01 | #121 |
| 9 | Claude Sonnet 4 (20250514) | 21.09 | #211 |
| 10 | Claude Opus 4.1 (20250805) | 20.67 | #126 |
| 11 | Qwen 3 VL 235B A22B (Thinking) | 20.08 | #228 (Qwen 3 VL 235B A22B) |
| 12 | Qwen 2.5 VL 72B Instruct | 19 | #364 |
| 13 | GLM-4.5V | 19 | #339 |
| 14 | Qwen 3 VL 30B A3B (Thinking) | 15.48 | #338 (Qwen 3 VL 30B A3B) |
| 15 | GPT-4o ChatGPT | 3.85 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=via-bench-llm-judge-visual-anomalies · How It Works · Data refreshed daily, snapshot 2026-10-11.