Deep research agent reads whole document pages not just text
DR Pixel: Page-Native Deep Research with Vision-Language Models
Computer Vision and Pattern RecognitionComputation and LanguageInformation Retrieval
Summary
Complex papers often include important content beyond just text, like tables and images, but most computer programs only read the text part. The authors created DR Pixel, a system that looks at the full page images of papers to better find and understand information. They showed that their method works well even without training, and training it further makes it much better. This approach avoids the common step of turning images into text first, making the research process both easier and more accurate.
What this means in practice
- •For academic digital libraries: Improve search and comprehension of scientific papers by using page images to retrieve and reason over all content types, not just text.
- •For legal document review teams: Enable more accurate extraction and reasoning over visual elements of legal documents by operating directly on rendered page images.
Authors
Tiansheng Hu, Yichuan Wang, James V. Talwar, Nicholas Tomlin, Chen Zhao, Sewon Min
Abstract
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. We introduce DR Pixel, the first open-source end-to-end multimodal DR agent that uses rendered page images as its sole retrieval and reasoning unit, operating over a corpus of 1.06M pages. To evaluate this setting, we construct DRPixelBench, a benchmark on scientific papers requiring reasoning over text, figures, tables, and equations. Surprisingly, page-native DR already works well without training, outperforming OCR-based baselines across state-of-the-art models like GPT-5.4, Gemini-3.5-Flash, and Qwen3.6-27B, while requiring fewer tool calls. We next turn to DR training. A key challenge is that teacher trajectories for agentic supervised finetuning are noisy, mixing critical reasoning and action steps with redundant or failed attempts. To address this, we introduce counterfactual credit reweighting, a fine-grained credit assignment method that weights individual reasoning and action steps by their counterfactual contribution to the final answer. Our training improves retrieval recall by about 15 points and final accuracy by over 10 points on DRPixelBench, and likewise outperforms the untrained baseline on four existing benchmarks. On out-of-distribution ViDoRe-V3, DR Pixel comes within 1.8 points of an oracle that is given the gold pages directly. Our results suggest that the conventional OCR-then-retrieve pipeline is not only unnecessary but suboptimal: directly retrieving and reasoning over visual document pages makes complex DR both more effective and more efficient.