DR Pixel: Page-Native Deep Research with Vision-Language Models

DR Pixel. Page-Native Deep Research with Vision-Language Models. By Tiansheng Hu and five others, from NYU Shanghai, the University of California, Berkeley, and three other institutions. Published on the second of November.

Scientific papers carry much of their information in tables, figures and equations. Yet existing deep research agents typically reduce a paper to text extracted by OCR, and whatever lives in the layout and the images is at risk of being lost along the way. A chart becomes a scatter of numbers, an equation a string of symbols, a table a jumble of cells. The authors ask a direct question: can vision-language models make deep research page-native, retrieving and reasoning over rendered images of the document pages themselves?

Their answer is DR Pixel, described as the first open-source, end-to-end multimodal deep research agent that uses rendered page images as its only unit of retrieval and reasoning. There is no text extraction step at all. What the agent searches is pages, what it reads is pages, and what it reasons over is pages, exactly as a person would see them. The agent operates over a corpus of one point zero six million pages. To evaluate it, the authors build DRPixelBench, a benchmark on scientific papers whose questions require reasoning over text, figures, tables and equations.

The first result comes before any training. Page-native deep research already works well out of the box, outperforming OCR-based baselines across state-of-the-art models such as GPT five point four, Gemini three point five Flash, and Qwen three point six, and it needs fewer tool calls to do so. The authors call this surprising, and the reason is plain: these models were never trained for this setting, yet looking at the page beats reading extracted text, and does it with less work.

Training is the next step, and it has a specific obstacle. Teacher trajectories for agentic supervised finetuning are noisy. Critical reasoning and action steps are mixed in with redundant or failed attempts. Imitate a trajectory like that wholesale and the student learns the dead ends along with the steps that actually led to the answer. To handle this, the authors introduce counterfactual credit reweighting, a fine-grained credit assignment method. Each individual reasoning or action step is weighted by its counterfactual contribution to the final answer, so the steps that mattered count for more in training than the ones that did not.

On DRPixelBench, training raises retrieval recall by about fifteen points and final accuracy by more than ten points. The trained agent also outperforms the untrained baseline on four existing benchmarks. On ViDoRe-V3, which lies outside the training distribution, DR Pixel comes within one point eight points of an oracle that is handed the gold pages directly. In other words, on that benchmark, finding the right pages on its own costs the agent very little compared with being told which pages they are.

The authors draw a pointed conclusion. In their reading, the familiar pipeline of OCR first and retrieval second is not merely unnecessary but suboptimal. Retrieving and reasoning directly over the visual page, they argue, makes complex deep research both more effective and more efficient, because the page was the information all along.