跳转至

DocMemo：基于概率记忆引导检索的多模态文档理解动态证据发现

文章背景与核心概要

长文档理解（Long-document understanding）一直是人工智能领域的难点之一，因为这通常要求系统在数百页的文档中精准定位稀疏且异构的证据。传统的检索系统往往受制于静态检索的局限性以及脆弱的跨轮记忆，一旦在早期检索阶段出现错误，系统很难进行自我纠正。为了解决这一痛点，本文提出了 DocMemo 这一先进的记忆引导框架，将长文档推理建模为一个动态的证据探索过程。

DocMemo 核心创新在于引入了三级检索状态（tri-level retrieval state）：文档架构记忆（捕捉结构先验）、页面信念记忆（处理动态相关性估计）以及问题情境记忆（追踪特定查询的推理轨迹）。通过采用结合汤普森采样（Thompson sampling）的贝叶斯页面信念更新、空间邻近度传播机制，以及结构感知自适应粒度的证据访问方式，DocMemo 能够持续优化页面选择，并无缝融合细粒度视觉区域。在三个基准测试上的实验结果表明，该方法达到了最先进的性能，充分验证了结构化记忆与动态信念更新机制的有效性。

摘要 (Summary)

DocMemo 是一种先进的记忆引导框架，旨在解决长文档推理中的挑战。长文档推理通常需要在数百页中定位稀疏且异构的证据。传统系统受到静态检索局限性和脆弱的跨轮记忆的困扰，经常无法从早期的检索错误中恢复。DocMemo 将长文档推理构想为一个动态的证据探索过程，使用了三级检索状态： 1. 文档架构记忆 (Document Schema Memory)：捕获结构先验。 2. 页面信念记忆 (Page Belief Memory)：处理动态相关性估计。 3. 问题情境记忆 (Question Episodic Memory)：追踪特定查询的推理轨迹。

通过采用结合汤普森采样的贝叶斯页面信念更新、空间邻近度传播以及结构感知的自适应粒度证据访问，DocMemo 持续优化页面选择并结合了细粒度视觉区域。在三个基准测试上的经验评估展示了最先进的性能，证实了结构化记忆和动态信念更新的有效性。

DocMemo is an advanced memory-guided framework designed to solve the challenges of long-document reasoning, which often requires locating sparse and heterogeneous evidence across hundreds of pages. Traditional systems suffer from static retrieval limitations and fragile cross-round memory, frequently failing to recover from early retrieval errors. DocMemo formulates long-document reasoning as a dynamic evidence exploration process using a tri-level retrieval state: 1. Document Schema Memory: Captures structural priors. 2. Page Belief Memory: Handles dynamic relevance estimation. 3. Question Episodic Memory: Tracks query-specific reasoning trajectories.

By employing Bayesian page belief updating with Thompson sampling, spatial proximity propagation, and structure-aware adaptive-granularity evidence access, DocMemo continuously refines page selection and incorporates fine-grained visual regions. Empirical evaluations across three benchmarks demonstrate state-of-the-art performance, confirming the effectiveness of structured memory and dynamic belief updates.

论文元数据 (Paper Metadata)

  • arXiv ID: arXiv:2608.07067 [cs.AI]
  • 提交时间: 2026年8月7日
  • 作者: Hanshu Yao, Janfeng Zhong, Niu Lian, Jinpeng Wang
  • 主学科: 人工智能 (cs.AI)
  • 其他学科: 计算与语言 (cs.CL)、信息检索 (cs.IR)、多媒体 (cs.MM)
  • 许可证: 知识共享署名 4.0 国际 许可证图标
  • arXiv ID: arXiv:2608.07067 [cs.AI]
  • Submitted On: August 7, 2026
  • Authors: Hanshu Yao, Janfeng Zhong, Niu Lian, Jinpeng Wang
  • Primary Subject: Artificial Intelligence (cs.AI)
  • Other Subjects: Computation and Language (cs.CL), Information Retrieval (cs.IR), Multimedia (cs.MM)
  • License: Creative Commons Attribution 4.0 International license icon

Abstract (摘要原文)

长文档理解需要定位数百页中稀疏且异构的证据，然而现有系统仍然受到静态检索和脆弱的跨轮记忆的限制。主流的单轮方法在开始时就确定一个固定的 top-\(k\) 页面集，难以从早期的检索错误中恢复；最近的迭代方法允许进行多轮证据获取，但它们没有研究跨轮状态的传播机制，使得追踪页面相关性的动态变化变得困难。

为了解决这些局限性，我们提出了 DocMemo，这是一个记忆引导的框架，将长文档推理构建为动态证据探索。DocMemo 维护了一个三级检索状态，由文档架构记忆、页面信念记忆和问题情境记忆组成，它们分别捕获结构先验、动态相关性估计和查询特定的推理轨迹。在推理过程中，DocMemo 通过结合汤普森采样的贝叶斯页面信念更新、空间邻近度传播和结构感知的自适应粒度证据访问，不断精炼跨轮页面选择，同时用细粒度视觉区域补充页面级证据。在 3 个基准测试上的实验表明，DocMemo 实现了最先进的性能，并验证了结构化记忆和动态页面信念更新的功效。

Long-document understanding requires locating sparse and heterogeneous evidence across hundreds of pages, yet existing systems remain limited by static retrieval and fragile cross-round memory. Mainstream single-round methods commit to a fixed top-\(k\) page set at the outset and struggle to recover from early retrieval errors; recent iterative approaches allow multi-round evidence acquisition, but they do not investigate the propagation mechanism of cross-round states, making it difficult to track the dynamic changes in page relevance.

To address these limitations, we propose DocMemo, a memory-guided framework that formulates long-document reasoning as dynamic evidence exploration. DocMemo maintains a tri-level retrieval state consisting of Document Schema Memory, Page Belief Memory, and Question Episodic Memory, which respectively capture structural priors, dynamic relevance estimation, and query-specific reasoning trajectories. During reasoning, DocMemo continuously refines cross-round page selection through Bayesian page belief updating with Thompson sampling, spatial proximity propagation, and structure-aware adaptive-granularity evidence access, while supplementing page-level evidence with fine-grained visual regions. Experiments on 3 benchmarks show that DocMemo achieves state-of-the-art performance and validate the efficacy of structured memory and dynamic page belief updating.