DR Pixel enables page-native deep research with vision-language models
Original title: DR Pixel: Page-Native Deep Research with Vision-Language Models
DR Pixel is an open-source agent that reasons directly over document page images, outperforming OCR-based methods on scientific papers.
DR Pixel is the first open-source multimodal deep research agent using rendered page images. It operates on 1.06M pages and constructs DRPixelBench benchmark. Training with counterfactual credit reweighting improves performance, coming close to oracle on out-of-distribution data.
- Outperforms GPT-5.4, Gemini-3.5-Flash on DRPixelBench without training
- Improves retrieval recall by 15 points and accuracy by over 10 points with training
- Reduces tool calls compared to OCR baselines
- Organizations
- Hugging Face
- AI models
- GPT-5.4Gemini-3.5-FlashQwen3.6-27B
From the original article
1 min read at huggingface.co
Complex documents such as scientific papers contain rich information in tables, figures, and equations; however, existing deep research (DR) agents typically reduce papers to OCR-extracted text. We instead ask whether vision-language models (VLMs) can enable page-native deep research: directly retrieving and reasoning over rendered document page images.
By Tiansheng Hu, Yichuan Wang, James V. Talwar · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.