Announcement
DR Pixel is an open-source research agent that reads page images without OCR
DR Pixel's researchers report that page-native research outperforms OCR-based baselines without training and requires fewer tool calls.
TLDR
Built for scientific papers, the agent retrieves and reasons over page images from a 1.06-million-page corpus instead of converting them to OCR text. The researchers report that training improved retrieval recall by about 15 points and final accuracy by over 10 points on DRPixelBench, which tests reasoning across text, figures, tables and equations.
Combined views
3.8K
9 Sources, first seen ago
