DR Pixel, the paper search tool that reads pages as images

Update, November 3, 2026: the first version relied on the abstract alone and said detailed data, code and datasets were missing. The full paper has them: we rewrote the article from the complete text.

DR Pixel is a deep research agent, a system that searches many documents to answer complex questions, and it works on images of scientific paper pages instead of OCR-extracted text. In tests by its authors, six researchers from NYU Shanghai, UC Berkeley, the Toyota Technological Institute at Chicago and Rational Dynamics, it answers better and with fewer searches than the traditional pipeline.¹

How it works

A paper is not only sentences: tables, charts and equations carry information that a text transcription often loses or distorts. DR Pixel uses a vision-language model, which reads images and text together, and treats the page image as the single unit both for finding documents and for reasoning.¹

To measure it, the authors built DRPixelBench: 500 questions over a collection of 99,055 papers and 1,062,153 pages, each with a short answer verifiable on one specific page. The questions were generated with Gemini-3.5-Flash from arXiv and open-access conference papers, then checked and rewritten by experts.¹

The results

Without specific training, searching page images beats the pipeline that uses extracted text and figures with all three models tested. With GPT-5.4 accuracy rises from 46.3 to 51.9, with Gemini-3.5-Flash from 41.3 to 46.4 and with Qwen3.6-27B from 34.9 to 40.1. The searches needed drop by 20 to 30%.¹

The authors then trained Qwen3.6-27B with a method that weights each reasoning step by how much it contributes to the final answer. Correct pages found rise from 36.3 to 52.0 and accuracy from 40.1 to 52.1: the same level as GPT-5.4, with an open-weight model.¹ The gain carries over to benchmarks the authors did not build: on IRPAPERS from 65.4 to 79.4, on M3SciQA from 49.8 to 64.2 and on AirQA from 42.7 to 52.3. On ViDoRe-V3 the model reaches 70.8, against 72.6 for an “oracle” handed the correct pages directly.¹

What to keep in mind

The main benchmark is the authors’ own and the results have not yet been replicated by others. Only 1.7% of ViDoRe-V3 questions really require reading visual content, so the advantage of images barely shows on that test.¹ Grading also matters in multimodal tests: an audit of eight benchmarks found differences of up to 22 accuracy points from the same answers under different grading rules.²

Code and data are public: the code on GitHub, the benchmark on Hugging Face.¹

This article was written by artificial intelligence from the sources cited, without human review. Automatic newsroom by Redazia.

Sources

Footnotes

1. DR Pixel: Page-Native Deep Research with Vision-Language Models, arXiv, 2026-11-02 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

2. Same Outputs, Different Scores: Auditing the Unchecked Graders of Multimodal LLM Benchmarks, arXiv, 2026-11-02 ↩