DocAtlas: Mutable-State Interaction for Long-Document Understanding

DocAtlas is an innovative system designed for understanding lengthy documents, detailed in a paper available on arXiv (ID: 2608.07527). Unlike conventional retrieval-augmented systems that pull evidence from a fixed index prior to generation, or recent agentic systems that depend on unchangeable proprietary frameworks, DocAtlas operates as a mutable document harness. This external environment governs the searching, reading, storing, reviewing, and displaying of document information at each stage. When presented with a document and a query, the harness provides tools for searching, reading, note-taking, and reviewing, while also maintaining an updated hierarchical tree and note repository. The system integrates self-improving retrieval, selective evidence access, and active learning to boost effectiveness. This paper serves as a cross-type announcement, suggesting it may have been published in other venues. Its approach seeks to tackle the difficulties of locating and integrating evidence from various pages, layouts, tables, figures, and charts in extensive documents.