VeriOCRBench - Over-Refusal Rate (Assisted): leaderboard
The over-refusal side of VeriOCRBench (Zhou et al., 2026): 200 trap-free control tasks built from the same OCR images, answerable as posed. Over-refusal rate is the percentage of these tasks the model wrongly rejects when the prompt tells it to verify the task first, judged by GPT-4o; lower is better.
Metric: Over-refusal rate (%; share of VeriOCRBench's 200 trap-free control tasks the model wrongly rejects; assisted paradigm: a meta-instruction asks the model to verify that the task is completable before answering; judged by GPT-4o at temperature 0; lower is better). Source: arxiv.org. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 9.5 | #163 |
| 2 | Qwen 3.5 122B A10B | 9.5 | #329 |
| 3 | Claude Sonnet 4.5 | 11.5 | #244 |
| 4 | Qwen 3.5 35B A3B | 12.5 | #480 |
| 5 | Qwen 3.5 397B A17B | 14.5 | #242 |
| 6 | GPT-4o | 15.5 | #741 |
| 7 | Llama 4 Scout | 16 | #1787 |
| 8 | Llama 4 Maverick | 18.5 | #1105 |
| 9 | InternVL3-38B | 32 | #905 |
| 10 | Gemini 3.1 Pro (Preview) | 79.5 | #91 |
Interactive version: theaggregate.ai/benchmark?slug=veriocrbench-over-refusal-rate-assisted · How It Works · Data refreshed daily, snapshot 2026-11-12.