Abstract
Multi-Document Summarization (MDS) play a critical role in distilling essential information from collections of textual data. Existing approaches often struggle to capture complex inter-document relationships, rely heavily on large amounts of labeled data for supervised training, or exhibit limited generalization across domains and languages. To address these limitations, we introduce the Mixture of Agents (MoA) framework, a novel, training-free and modular paradigm for MDS. MoA orchestrates three specialized agents operating in parallel across refinement layers: (1) an Extractor Agent that selects key sentences, (2) a KGSum Agent that constructs and summarizes knowledge graphs, and (3) an Abstractor Agent that generates coherent abstractive summaries. Each agent leverages pre-trained Large Language Models (LLMs), enabling MoA to operate without task-specific supervised training. This work introduces a multi-agent architecture for MDS and provides comprehensive evaluations on both English and Vietnamese datasets. Experiments on four benchmarks—Multi-News, Multi-XScience (English), VN-MDS, and ViMs (Vietnamese)—demonstrate that MoA achieves state-of-the-art ROUGE scores on Multi-News and yields competitive or superior results on the remaining datasets. Our findings highlight MoA as a robust, generalizable, and data-efficient approach to MDS.
Abstract
Multi-Document Summarization (MDS) play a critical role in distilling essential information from collections of textual data. Existing approaches often struggle to capture complex inter-document relationships, rely heavily on large amounts of labeled data for supervised training, or exhibit limited generalization across domains and languages. To address these limitations, we present a training-free mixture-of-agents framework for MDS that leverages the complementary strengths of large language models (LLMs) and knowledge graphs. Our approach decomposes summarization into specialized agent tasks: extractive selection, knowledge-aware abstraction, and iterative refinement, each operating without task-specific fine-tuning. We unify their outputs using a multi-perspective consistency mechanism guided by LLMs.} Experiments across four datasets in English and Vietnamese demonstrate state-of-the-art or competitive performance, validating the effectiveness and adaptability of our modular design.