pdfQA (Real): leaderboard
Question answering over whole PDF documents (pdfQA; Schimanski et al., University of Zurich and ETH Zurich, 2026): real-pdfQA re-filters human-annotated QA pairs from nine datasets that start from PDFs, mostly Tat-QA and FinQA, answered with the full PyMuPDF-parsed document in the prompt (documents under 128K tokens, 1,824 pairs). A gpt-4.1-mini G-Eval prompt scores each answer's correctness against the reference from 1 to 5; higher is better.
Metric: G-Eval correctness (1-5; gpt-4.1-mini judge; 1,824 real-pdfQA pairs with documents under 128K tokens, full parsed document as context). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 Next 80B A3B | 3.37 | #756 |
| 2 | GPT-OSS-120B | 3.3 | #699 |
| 3 | GPT-OSS-20B | 3.04 | #1150 |
| 4 | Gemma 3 27B | 2.73 | #1638 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pdfqa-real · How It Works · Data refreshed daily, snapshot 2026-11-12.