DocAtlas：将长文档理解视为可变状态交互
原标题：DocAtlas: Long-Document Understanding as Mutable-State Interaction
AI 摘要
DocAtlas 是一个将长文档理解视为可变状态交互过程的系统，通过外部环境管理文档信息的搜索、读取、存储和展示。该系统结合了自改进检索、选择性证据访问和主动工作记忆，在固定上下文预算下运行。使用 GPT-5.4 时，DocAtlas 在 MMLongBench-Doc 上达到 71.4% 的准确率，超过人类专家参考水平；用 Qwen3.5-4B VLM 进行端到端强化学习训练后达到 63.7%，显著优于直接输入基线。
正文节选
Computer Science > Computation and Language Title:DocAtlas: Long-Document Understanding as Mutable-State Interaction View PDF HTML (experimental) Abstract:Long-document understanding requires models to find and combine evidence across many pages, layouts, tables, figures, and charts. Existing retrieval-augmented systems usually select evidence from a static index before generation, while recent agentic systems add multi-turn tool use but often rely on frozen proprietary backbones who