MedRepBench - Item Name Recall (OCR Text Only): leaderboard
Structured extraction from photographed, screenshotted and electronic Chinese laboratory reports, by text-only language models given the recognized OCR text; MedRepBench (Shang et al., 2025). Recall in percent of the reference items' item names, with items matched by normalized name within a budget of as many items as the reference lists; higher is better.
Metric: Field-level recall (%; item name of each ground-truth lab item, matched by exact normalized item name within the model's top-K items, K = number of reference items; text LLM on the OCR text). Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3 | 96.48 | #705 |
| 2 | Qwen 3 235B A22B 2507 | 96.42 | #613 |
| 3 | Qwen 3 30B A3B | 96.08 | #1124 |
| 4 | Qwen 3 14B | 95.75 | #1235 |
| 5 | Qwen 3 32B | 95.66 | #960 |
| 6 | DeepSeek R1 Distill Qwen 14B | 95.5 | #2594 |
| 7 | Qwen 3 235B A22B | 94.96 | #622 |
| 8 | Qwen 3 8B | 94.93 | #1838 |
| 9 | Qwen 3 4B | 94.64 | #2297 |
| 10 | DeepSeek R1 Distill Qwen 32B | 92.93 | #1780 |
| 11 | Qwen 3 1.7B | 92.88 | #5551 |
| 12 | DeepSeek-R1-Distill-Qwen-7B | 91.68 | #8045 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=medrepbench-item-name-recall-ocr-text-only · How It Works · Data refreshed daily, snapshot 2026-11-12.