DR Pixel is an open-source research agent that skips OCR and works directly on rendered images of scientific paper pages.
Setup: It retrieves and reasons over page images from a corpus of 1.06M pages. The paper also introduces DRPixelBench, a benchmark of questions over text, figures, tables and equations.
Without training: The authors say it beats OCR-based baselines on GPT-5.4, Gemini-3.5-Flash and Qwen3.6-27B while using fewer tool calls.
Training: A method called counterfactual credit reweighting lifts retrieval recall by about 15 points and accuracy by over 10 points on DRPixelBench.
Out of distribution: On ViDoRe-V3, the authors report it comes within 1.8 points of an oracle that gets the gold pages directly.
Sign in to comment.
TL;DR of the paper, generated automatically.
DR Pixel is an open-source research agent that skips OCR and works directly on rendered images of scientific paper pages.