DocMemo Enhances Multi-Modal Long Document Understanding

Understanding long, multi-modal documents, especially those spanning hundreds of pages, presents a significant challenge for AI systems. Existing methods often struggle with static retrieval, committing to a fixed set of pages early on, or with iterative approaches that lack robust mechanisms for tracking dynamic changes in page relevance. This can lead to errors that are difficult to recover from.Researchers have introduced DocMemo, a novel framework that redefines long-document reasoning as a dynamic exploration of evidence. DocMemo employs a sophisticated tri-level retrieval state, comprising Document Schema Memory, Page Belief Memory, and Question Episodic Memory. These components respectively capture structural priors, dynamically estimate page relevance, and track query-specific reasoning paths.During the reasoning process, DocMemo continuously refines its cross-round page selection. It achieves this through Bayesian page belief updating, Thompson sampling, spatial proximity propagation, and structure-aware adaptive-granularity evidence access. Furthermore, it augments page-level evidence with fine-grained visual regions, crucial for multi-modal understanding. Experiments across three benchmarks demonstrate DocMemo's state-of-the-art performance, validating the effectiveness of its structured memory and dynamic belief updating mechanisms.