DR Pixel Ditches OCR to Read Documents as Images, Boosting Accuracy by 10%
A new arXiv paper argues the OCR step in document research agents is both unnecessary and worse than feeding page images straight to a vision-language model.
- New arXiv paper introduces DR Pixel, a deep research agent that retrieves and reasons over rendered page images instead of OCR text.
- Operates over a 1.06M-page corpus and beats OCR baselines on GPT-5.4, Gemini-3.5-Flash, and Qwen3.6-27B with fewer tool calls.
- Introduces counterfactual credit reweighting, which masks unhelpful teacher steps during agentic supervised fine-tuning.
- Training lifts retrieval recall by ~15 points and final accuracy by >10 points on the new DRPixelBench.
- On out-of-distribution ViDoRe-V3, comes within 1.8 points of an oracle given the gold pages directly.
- Paper: arxiv.org/abs/2611.03097. No public code or weights linked yet.
A new arXiv preprint introduces DR Pixel, a document-research agent that searches and reads rendered page images instead of OCR-extracted text. Across a corpus of 1.06 million scientific-paper pages, the page-native approach outperformed OCR-based baselines on the authors’ benchmark while using fewer tool calls.
The results suggest that document representation can limit retrieval before an agent begins reasoning. Preserving figures, equations, tables, and spatial relationships gives vision-language models access to evidence that text extraction often flattens or discards.
Pixels replace the OCR pipeline
Document-research agents typically search a corpus, inspect retrieved material, and repeat the process until they can answer a question. Most systems prepare PDFs by extracting embedded text or applying optical character recognition, then dividing the result into chunks for indexing.
Scientific PDFs expose the weaknesses of that pipeline. OCR can recover words while losing the structure of a table, the position of a plotted value, or the relationship between an equation and its surrounding labels. Those details may contain the evidence required to answer a question.
DR Pixel renders each page as an image and uses the page as the retrieval and reasoning unit. A vision-language model, which accepts both text and images, searches the page collection, opens candidate pages, and reasons over their complete visual contents.
| Component | Conventional pipeline | DR Pixel |
|---|---|---|
| Document representation | Extracted or OCR-generated text | Rendered page images |
| Retrieval unit | Text chunk | Complete page |
| Visual evidence |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
