CDH-Bench (Multiple-Choice 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), choosing among four shuffled options: the answer the image shows, the commonsense alternative as an explicit distractor and two other plausible distractors; 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) | 81.67 | #54 |
| 2 | Qwen 3.5 Plus | 74.37 | #123 |
| 3 | Gemini 3.1 Flash Lite (Preview) | 71.67 | #186 |
| 4 | Kimi K2.5 | 64.67 | #139 |
| 5 | Doubao-Seed-1-8-251228 | 64.67 | #146 |
| 6 | Qwen 3 VL 32B (Thinking) | 60.46 | #287 (Qwen 3 VL 32B) |
| 7 | GPT-5.4 | 58.86 | #76 |
| 8 | GPT-5.4 Mini | 53.33 | #202 |
| 9 | Qwen 3 VL 8B (Thinking) | 50 | |
| 10 | Qwen 3 VL 4B (Thinking) | 47.33 | #471 (Qwen 3 VL 4B) |
| 11 | Qwen 3 VL 32B Instruct | 46 | #276 |
| 12 | Qwen 3 VL 8B Instruct | 42 | #401 |
| 13 | Qwen 3 VL 4B Instruct | 34.33 | #506 |
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-multiple-choice-qa · How It Works · Data refreshed daily, snapshot 2026-10-11.