DR Pixel: open-source page-image deep research agent, 27B model matches GPT-5.4
sewon__min · x · 2026-11-04
Berkeley NLP and rtnldynamics released DR Pixel (arXiv 2611.03097), the first open-source end-to-end multimodal deep research agent that retrieves and reasons directly over rendered document page images—no OCR.
- Motivation: scientific papers hold key info in tables, figures, and equations; existing DR agents flatten documents into OCR text.
- Results: on a 1.06M-page corpus, untrained page-native DR already beats OCR-based baselines across GPT-5.4, Gemini-3.5-Flash, and Qwen3.6-27B, with fewer tool calls; after training, their 27B model matches GPT-5.4.
- Training: introduces counterfactual credit reweighting—weighting individual reasoning/action steps by their counterfactual contribution to the answer—lifting retrieval recall 15 points and accuracy over 10 points.
- Also releases DRPixelBench, a benchmark requiring reasoning over text, figures, tables, and equations.
Related event: Berkeley-Led Team Open-Sources DR Pixel, a Pixel-Based Deep Research Agent(2 posts)→
More from coding & agent
- Model Vendors Post-Train for Tool Search, Squeezing Out MCP-Layer Workarounds — pvncher · 2026-11-04
- Sentry CEO: LLMs Struggle With Large Complex Systems, Humans Must Guide Design — zeeg · 2026-11-04
- Engineer builds agent that clears years-old code backlog without any mandate — BradPorter_ · 2026-11-04
- Sentry CEO: AI Code Verification Is About Design Correctness, Not Compilation — trq212 · 2026-11-04
- AnnotateKit launches: agent-readable in-app feedback for iOS that drives surgical AI code edits — julianweisser · 2026-11-04
- AIOps platform cuts agent token costs 30%+ with Pydantic's open-source sandbox Monty — samuelcolvin · 2026-11-04