DocAtlas beats human experts on long-document AI with mutable-state search
71.4% on MMLongBench-Doc beats human 65.8% benchmark—here's how
Deep Dive
DocAtlas reimagines long-document understanding as a mutable-state interaction—an external harness that decides what gets searched, read, stored, reviewed, and shown at each step. It combines self-improving retrieval, selective evidence access, and active working memory under a fixed context budget. With GPT-5.4, it scores 71.4% on MMLongBench-Doc, beating the 65.8% human-expert reference. A compact Qwen3.5-4B agent trained with end-to-end RL in DocAtlas jumps to 63.7% from a 54.4% direct-input baseline.
Key Points
- DocAtlas hits 71.4% on MMLongBench-Doc with GPT-5.4, exceeding the 65.8% human-expert baseline
- A compact Qwen3.5-4B VLM trained with end-to-end RL in DocAtlas reaches 63.7%, up from 54.4% direct-input
- Uses mutable-state harness with search, reading, note-taking, and review tools plus a hierarchical evidence tree
Why It Matters
DocAtlas shows interactive, mutable-state retrieval can beat static RAG and even human experts on complex documents.