DR Pixel: Page-Native Deep Research with Vision-Language Models
DR Pixel beat OCR-based research agents using page images alone, with fewer tool calls, the authors report.
The open-source agent retrieves and reasons over rendered document pages from a corpus of 1.06 million pages. On the authors’ DRPixelBench, which tests reasoning across text, figures, tables and equations, it outperformed OCR-based baselines even without training. A training method that weights reasoning and action steps by their contribution to the answer improved retrieval recall by about 15 points and final accuracy by more than 10 points. On the separate ViDoRe-V3 benchmark, it finished within 1.8 points of an oracle given the correct pages directly.
HF Daily Papers' note
The open-source agent retrieves and reasons over rendered document pages from a corpus of 1.06 million pages. On the authors’ DRPixelBench, which tests reasoning across text, figures, tables and equations, it outperformed OCR-based baselines even without training. A training method that weights reasoning and action steps by their contribution to the answer improved retrieval recall by about 15 points and final accuracy by more than 10 points. On the separate ViDoRe-V3 benchmark, it finished within 1.8 points of an oracle given the correct pages directly.
HF Daily Papers' note
score 5