DocMemo: Probabilistic Memory-Guided Retrieval Framework for Multi-Modal Long Documents

_reachsumit · x · 2026-08-10

Long-document understanding requires locating sparse and heterogeneous evidence across hundreds of pages, but existing systems are limited by static retrieval and fragile cross-round memory. To address this, the paper introduces DocMemo, a framework formulating long-document reasoning as dynamic evidence exploration.

DocMemo maintains a tri-level retrieval state memory: Document Schema Memory, Page Belief Memory, and Question Episodic Memory. During reasoning, it continuously refines cross-round page selection through Bayesian page belief updating with Thompson sampling, spatial proximity propagation, and structure-aware adaptive-granularity evidence access. Experiments show that DocMemo achieves state-of-the-art performance on three benchmarks.

Original post →

More from Research

Research channel →