Reproduce “Zero-Shot Document Understanding using Pseudo Table of Contents-Guided Retrieval-Augmented Generation” on OpenResearch

About this paper

Hyeon Seong Jeong, Sangwoo Jo, Byeong Hyun Yoon, Yoonseok Heo, Haedong Jeong, Taehoon Kim

DocsRay, a training-free system from Sogang University, enhances complex multimodal document understanding by generating pseudo-Tables of Contents and employing hierarchical retrieval-augmented generation. It achieved an accuracy of 64.7% on MMLongBench-Doc, outperforming GPT-4.1 by 15.0 percentage points, and reduced query latency by 45.4%.

Key insights

• The synergistic orchestration of pseudo-TOC generation, multimodal content conversion, and hierarchical RAG effectively leverages latent LLM capabilities for complex document understanding.
• Concatenating embeddings from models with complementary strengths (e.g., keyword focus and semantic understanding) significantly improves retrieval accuracy over individual or additive approaches.
• Hierarchical retrieval guided by a pseudo-TOC provides substantial efficiency gains (45.4% reduced query latency) with only a marginal impact on accuracy, offering a practical trade-off for real-world applications.

Summary by alphaXiv.