🤖 AI Summary
This work addresses the challenge of answering complex real-world queries that require synthesizing information across numerous heterogeneous documents—a task where traditional retrieval-augmented generation (RAG) systems suffer from insufficient evidence coverage and long-context large language models exhibit unreliable reasoning. To overcome these limitations, the authors propose a hierarchical multi-agent framework in which each document is assigned a dedicated sub-agent for focused retrieval, while a coordinator distributes tasks and aggregates local answers. A token-bounded recursive synthesis layer then integrates these partial results. This architecture introduces, for the first time, a document-wise sub-agent design that enables modular and scalable multi-document question answering, achieving both broad coverage and enhanced reasoning accuracy. Evaluated on the LOONG benchmark, the method attains an average score of 58.1, substantially outperforming standard RAG (33.0) and Agentic RAG (32.8), with only 38% of the API cost of full-context baselines.
📝 Abstract
Answering complex, real-world queries often requires synthesizing facts scattered across vast document corpora. In these settings, standard retrieval-augmented generation (RAG) pipelines suffer from incomplete evidence coverage, while long-context large language models (LLMs) struggle to reason reliably over massive inputs. We introduce SPD-RAG, a hierarchical multi-agent framework for exhaustive cross-document question answering that decomposes the problem along the document axis. Each document is processed by a dedicated document-level agent operating only on its own content, enabling focused retrieval, while a coordinator dispatches tasks to relevant agents and aggregates their partial answers. Agent outputs are synthesized by merging partial answers through a token-bounded synthesis layer (which supports recursive map-reduce for massive corpora). This document-level specialization with centralized fusion improves scalability and answer quality in heterogeneous multidocument settings while yielding a modular, extensible retrieval pipeline. On the LOONG benchmark (EMNLP 2024) for long-context multi-document QA, SPD-RAG achieves an Avg Score of 58.1 (GPT-5 evaluation), outperforming Normal RAG (33.0) and Agentic RAG (32.8) while using only 38% of the API cost of a full-context baseline (68.0).