CDH-Bench (Binary QA): leaderboard
Metric: Counterfactual accuracy (%) on the 300 counterfactual images (100 each of counting, relational and attribute anomalies, generated and human-verified so the commonsense-violating fact is unambiguous and central), answering a yes/no question that asserts the commonsense reading and denies the depicted anomaly (for example 'Does the mouth have 2 rows of teeth and not 3 rows?'), so the correct answer on every counterfactual image is no; each image is shown alone, without its matched commonsense twin, under one prompt per format; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 90.33 | #54 |
| 2 | Gemini 3.1 Flash Lite (Preview) | 83.67 | #186 |
| 3 | Doubao-Seed-1-8-251228 | 76.67 | #146 |
| 4 | Qwen 3.5 Plus | 75.87 | #123 |
| 5 | Kimi K2.5 | 75.67 | #139 |
| 6 | GPT-5.4 Mini | 64 | #202 |
| 7 | GPT-5.4 | 63.67 | #76 |
| 8 | Qwen 3 VL 32B (Thinking) | 56.37 | #287 (Qwen 3 VL 32B) |
| 9 | Qwen 3 VL 32B Instruct | 55.67 | #276 |
| 10 | Qwen 3 VL 4B (Thinking) | 55 | #471 (Qwen 3 VL 4B) |
| 11 | Qwen 3 VL 8B (Thinking) | 53 | |
| 12 | Qwen 3 VL 4B Instruct | 52 | #506 |
| 13 | Qwen 3 VL 8B Instruct | 45 | #401 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=cdh-bench-binary-qa · How It Works · Data refreshed daily, snapshot 2026-10-11.