DR Pixel Lets Research Agents Read Papers as Pages, Not OCR Text

Introduction

Scientific papers are not simply collections of sentences. A result may depend on the relationship between a paragraph and a figure, a value in a table, an equation, or the layout that connects several pieces of evidence. Yet many deep research systems still convert documents into OCR text before retrieval. That conversion makes documents searchable, but it can also discard visual structure and information that is difficult to express as plain text.

The paper “DR Pixel: Page-Native Deep Research with Vision-Language Models” explores a different design. Instead of using extracted text as the main research unit, it lets a vision-language model retrieve and reason over rendered page images directly.

Key points

• Pages are the retrieval unit: DR Pixel uses page images for search, evidence inspection, and answer generation, rather than treating OCR output as the sole representation of a document.
• A benchmark for mixed document evidence: The authors introduce DRPixelBench, with questions requiring reasoning over prose, figures, tables, and equations. The system operates over a corpus of about 1.06 million pages.
• Strong results without additional training: The page-native approach outperforms OCR-based baselines across the advanced models listed in the paper, while making fewer tool calls. This suggests that visual document input can be a primary research interface, not merely an auxiliary feature.
• A finer method for agent training: Demonstration trajectories often mix essential reasoning with repeated searches, failed attempts, and unnecessary actions. Counterfactual credit reweighting assigns different weights according to how much each reasoning or action step contributes to the final answer under a counterfactual test.
• Training improves both search and answers: The reported training procedure raises retrieval recall by roughly 15 points and final accuracy by more than 10 points on DRPixelBench. It also beats the untrained version on four other benchmarks. On the out-of-distribution ViDoRe-V3 evaluation, DR Pixel comes within 1.8 points of an oracle given the correct pages directly.

Why it matters

The broader contribution is a challenge to the standard document pipeline. OCR is useful for indexing, but it does not fully preserve two-dimensional layout, chart semantics, mathematical notation, or the visual relationships that connect evidence. A page image keeps text and its spatial context together, allowing the model to inspect the document closer to the way a human researcher does.

This also changes what retrieval must optimize. The system is not only looking for matching terms; it must identify pages that contain visual evidence capable of supporting a particular claim. That distinction is important for scientific literature review, patent analysis, and technical-report investigation.

The approach is not without costs. Image encoding can be expensive, long documents still require careful management, and users need reliable evidence tracing. The results therefore do not prove that OCR pipelines are obsolete. They do show, however, that page-native retrieval is a credible and potentially more efficient foundation for research agents as vision-language models become cheaper and more capable.

Source: Hugging Face Daily Papers