HallusionBench
Can the model extract and reason over information in images or documents?
Vision6 ranked models6 reported values1 reportsHigher is better
Top rankings
One row per model, using its best reported score across effort settings.
| Rank | Model | Best score | Best reported setting |
|---|---|---|---|
| 1 | Qwen3.5-27BQwen | 70.0% | Reported configurationQwen report1 value · 1 reportApr 15, 2026 · source |
| 2 | Qwen3.6-35B-A3BQwen | 69.8% | Reported configurationQwen report1 value · 1 reportApr 15, 2026 · source |
| 3 | Qwen3.5-35B-A3BQwen | 67.9% | Reported configurationQwen report1 value · 1 reportApr 15, 2026 · source |
| 4 | Gemma4-31BGoogle | 67.4% | Reported configurationQwen report1 value · 1 reportApr 15, 2026 · source |
| 5 | Gemma4-26B-A4BGoogle | 66.1% | Reported configurationQwen report1 value · 1 reportApr 15, 2026 · source |
| 6 | Claude Sonnet 4.5Anthropic | 59.9% | Reported configurationQwen report1 value · 1 reportApr 15, 2026 · source |
Effort curve
Every sourced cost-linked effort value for this exact version. Lines connect complete sweeps only.
No cost-linked effort sweep for this version.
Definition and comparison boundarypublic methodology
Can the model extract and reason over information in images or documents?
Visual perception and multimodal reasoning under the reported protocol. Higher is better. The value is the percentage reported in this lab's table.
Harness, tools, prompts, attempt count, and benchmark version can materially change the result. Compare within one reporting context.
Method / source