A new approach to multi-document summarization sidesteps the need for expensive training data by deploying specialized AI agents that work together to extract, analyze, and refine information from multiple sources. The method, called Mixture of Agents (MoA), combines large language models (LLMs) with knowledge graphs to handle the complex task of distilling essential information from collections of documents, and it works equally well in English and Vietnamese without requiring task-specific fine-tuning.

Why Is Multi-Document Summarization So Difficult?

As digital information expands at an exponential rate, the ability to synthesize knowledge from multiple sources has become critical. Traditional summarization methods struggle with a fundamental problem: they either produce outputs that lack semantic coherence, or they require massive amounts of labeled training data that is expensive and time-consuming to create. Most existing supervised models demand large task-specific labeled datasets, which limits their usefulness in low-resource domains or for languages like Vietnamese that have fewer available training examples.

Large language models have opened new possibilities through their zero-shot and few-shot capabilities, meaning they can perform tasks without extensive task-specific training. However, they face a critical bottleneck: context length limitations. When you feed an LLM multiple long documents at once, it struggles to learn the complex relationships between them and often fails to reconcile conflicting information across sources.

How Does the New Mixture of Agents Framework Work?

Rather than relying on a single AI model to handle the entire summarization task, the MoA framework decomposes the problem into three specialized agents, each handling a different aspect of the work:

• Extractor Agent: Identifies the most important factual sentences from the source documents using a robust scoring model that recognizes salient information.
• KGSum Agent: Constructs a knowledge graph that explicitly models entities and their relationships, then generates summaries by identifying thematic communities and detecting contradictory information within the graph structure.
• Abstractor Agent: Produces a fluent, coherent summary directly from the source documents without relying on extracted sentences or structured representations.

The outputs from these three parallel agents are then orchestrated by a novel mechanism called Adaptive Multi-Perspective Fusion (AMF). This mechanism assesses metadata generated by each agent and dynamically selects the optimal strategy for synthesizing a final summary, essentially allowing the system to exploit the strengths of each agent based on the specific characteristics of the documents being summarized.

What Makes Knowledge Graphs Essential for Understanding Multiple Documents?

The KGSum agent represents the most sophisticated component of the framework. Knowledge graphs are structured representations that explicitly model how entities relate to one another. By constructing a knowledge graph from multiple documents, the system can identify not just what information exists, but how that information connects across different sources. This is particularly valuable when documents contain contradictory claims or when understanding the relationships between entities is crucial to producing an accurate summary. The knowledge graph approach allows the system to systematically learn intricate connections between multiple documents in a way that traditional text-based approaches cannot.

How to Implement Multi-Perspective Summarization in Your Organization

• Leverage Existing Models: The framework requires no task-specific fine-tuning, meaning organizations can deploy it using pre-trained LLMs they already have access to, reducing implementation costs and time-to-deployment.
• Test Across Languages: The system has been validated on both English and Vietnamese datasets, suggesting it can generalize across different languages without requiring separate training for each language.
• Combine Multiple Perspectives: Rather than relying on a single summarization approach, use the framework's ability to generate extractive, knowledge-graph-based, and abstractive summaries, then intelligently fuse them based on document characteristics.

The research team tested the MoA framework on four different datasets spanning English and Vietnamese, and the results demonstrated state-of-the-art or competitive performance across all benchmarks. This validation is significant because it shows the framework's effectiveness and adaptability across diverse linguistic contexts and document types.

The modular design of the framework offers practical advantages for real-world deployment. Because each agent operates independently without task-specific fine-tuning, organizations can adapt the system to new domains or languages without collecting and labeling thousands of training examples. This training-free approach addresses a persistent pain point in natural language processing (NLP), where the cost and effort of creating labeled datasets often exceeds the cost of the computational infrastructure itself.

The framework's acceptance by Neural Computing and Applications, a peer-reviewed journal, indicates that the research has undergone rigorous evaluation and represents a meaningful advance in how AI systems can handle the increasingly complex task of synthesizing information from multiple sources. As organizations continue to grapple with information overload, tools that can automatically distill essential insights from collections of documents without requiring extensive training data will become increasingly valuable.