VeriOCRBench - Logical Dimension (Assisted): leaderboard
The logical verification dimension of VeriOCRBench (Zhou et al., 2026): 400 OCR-grounded tasks with the traps logical paradox and missing variable, where the premise contradicts a visible relation or the task lacks a variable it needs. Diagnosis accuracy is the percentage on which the model both rejects the task and names the correct trap mechanism, judged by GPT-4o; higher is better. Assisted setting: the prompt tells the model to check that the task is completable before answering.
Metric: Diagnosis accuracy (%; the 400 trap-injected tasks of the logical verification dimension (logical paradox and missing variable: the premise contradicts a visible relation, or the task lacks a variable needed to solve it); the model must reject the task and name the ground-truth trap mechanism; judged by GPT-4o at temperature 0; assisted paradigm: a meta-instruction asks the model to verify the task first). Source: arxiv.org. Saturation forecast: Around February 2028. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 397B A17B | 75.5 | #242 |
| 2 | Claude Sonnet 4.5 | 70.25 | #244 |
| 3 | GPT-5 | 69.5 | #163 |
| 4 | Qwen 3.5 35B A3B | 68 | #480 |
| 5 | Qwen 3.5 122B A10B | 67.25 | #329 |
| 6 | InternVL3-38B | 67 | #905 |
| 7 | GPT-4o | 62.75 | #741 |
| 8 | Llama 4 Maverick | 56.5 | #1105 |
| 9 | Llama 4 Scout | 49.25 | #1787 |
| 10 | Gemini 3.1 Pro (Preview) | 5 | #91 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=veriocrbench-logical-dimension-assisted · How It Works · Data refreshed daily, snapshot 2026-11-12.