DocMemo: Dynamic Evidence Discovery via Probabilistic Memory-Guided Retrieval for Multi-Modal Document Understanding
The hosts discuss the paper 'DocMemo,' which introduces a memory-guided retrieval system for multi-modal document understanding. It uses three memory layers—schema, belief, and episodic—to dynamically discover evidence across long documents. The system outperforms baselines on benchmarks like MMLongBench-Doc and PaperTab, with lower retrieval costs and better handling of unanswerable questions.