Title: A Training-Free Mixture-of-Agents Framework for Multi-Document Summarization using LLMs and Knowledge Graphs

Abstract:

Multi-Document Summarization (MDS) is essential for extracting key insights from large bodies of text. However, current methods frequently face challenges in capturing intricate inter-document dynamics, depend significantly on extensive labeled datasets for supervised learning, and often show poor generalization across different languages and domains. To overcome these hurdles, we introduce a training-free mixture-of-agents framework for MDS that combines the unique advantages of large language models (LLMs) and knowledge graphs. This methodology breaks down the summarization process into distinct agent roles: extractive selection, knowledge-aware abstraction, and iterative refinement. Notably, each of these specialized tasks is performed without the need for task-specific fine-tuning. We integrate the results from these agents through a multi-perspective consistency mechanism, which is orchestrated by LLMs. Our evaluation across four datasets in both English and Vietnamese reveals state-of-the-art or highly competitive results, confirming the robustness and flexibility of our modular architecture.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC