DocAtlas: Long-Document Understanding as Mutable-State Interaction

DocAtlas helps an AI answer questions about long documents by maintaining an evolving record of what it has found. Its document environment supplies search and reading tools, an organized document tree, and notes that change as the agent gathers evidence. That lets the agent revisit useful material while keeping the information presented to the model within a fixed limit. In the reported experiment, a four-billion-parameter Qwen model trained through reinforcement learning in this environment scored 63.7% on the MMLongBench-Doc benchmark, compared with 54.4% for direct document input. The comparison supports the combined training-and-environment approach; it does not isolate how much of the gain came from the notes or any other individual component.