DR Pixel
An end-to-end multimodal deep research agent that uses rendered document page images directly as the sole unit for retrieval and reasoning, skipping OCR text extraction and preserving information from charts, formulas, and layout structure.
Timeline (last 90 days)
DR Pixel is the first end-to-end multimodal deep research agent that uses page images as the sole retrieval and reasoning unit.
DR Pixel runs on a corpus containing 1.06 million pages.
On frontier models such as GPT-5.4, Gemini-3.5-Flash, and Qwen3.6-27B, the page-image-based approach comprehensively outperforms OCR-based baselines and requires fewer tool calls.
After training, DR Pixel improves retrieval recall on DRPixelBench by about 15 percentage points and final accuracy by more than 10 percentage points.
On ViDoRe-V3, the gap between DR Pixel and the oracle that is "directly given the gold page" is only 1.8 percentage points.
The study proposes a counterfactual credit reweighting method that weights each reasoning and action step according to its counterfactual contribution to the final answer, in order to handle noisy teacher trajectories.
The pipeline of extracting text via OCR and then retrieving is neither necessary nor optimal for complex documents.
All Facts (7)
DR Pixel is the first end-to-end multimodal deep research agent that uses page images as the sole retrieval and reasoning unit.
50%UnverifiedDR Pixel runs on a corpus containing 1.06 million pages.
50%UnverifiedOn frontier models such as GPT-5.4, Gemini-3.5-Flash, and Qwen3.6-27B, the page-image-based approach comprehensively outperforms OCR-based baselines and requires fewer tool calls.
50%UnverifiedAfter training, DR Pixel improves retrieval recall on DRPixelBench by about 15 percentage points and final accuracy by more than 10 percentage points.
50%UnverifiedOn ViDoRe-V3, the gap between DR Pixel and the oracle that is "directly given the gold page" is only 1.8 percentage points.
50%UnverifiedThe study proposes a counterfactual credit reweighting method that weights each reasoning and action step according to its counterfactual contribution to the final answer, in order to handle noisy teacher trajectories.
50%UnverifiedThe pipeline of extracting text via OCR and then retrieving is neither necessary nor optimal for complex documents.
50%