Go to ACL ARR 2025 May homepageKG-MASD: Knowledge Graph-guided Multi-Agent System Distillation
Abstract: Recent research has focused on optimizing multi-agent LLMs for complex reasoning tasks, revealing that such architectures can significantly enhance reasoning abilities. Nevertheless, there are critical challenges, such as the uncontrollable hallucinations caused by the multi-model and multi-round iteration mechanism. The current paradigm also fails to effectively distill the collaborative reasoning power of distributed multi-agent systems into a single deployable model, which limits reasoning efficiency in practical application scenarios. To address these issues, we propose a solution that combines knowledge graphs with multi-agent systems. Focusing on industrial-field intelligent QA systems, we design a Knowledge Graph-guidedMulti-Agent System Distillation (KG-MASD) . The framework makes three main contributions. First, it constructs an industrial field knowledge graph to provide prior information. Second, it establishes a collaborative reasoning mechanism for a multi-teacher model. Third, it develops a multi-agent distillation methodology. To verify its effectiveness, this study introduces the first standard industrial production instruction dataset. It comprises approximately 52k domain-specific question-and-answer pairs and an industrial knowledge graph encompassing around 36k entities and 131k relationships. Experimental results indicate that the KG-MASD framework may offer a potential domain adaptation advantage over existing single-model and multi-agent distillation frameworks, with a possible improvement ranging from 2.4% and 20.1% in domain adaptation.
Paper Type: Long
Research Area: NLP Applications
Research Area Keywords: Multi-Agent System, Knowledge Distillation, Knowledge Graph
Contribution Types: NLP engineering experiment, Data resources
Languages Studied: English, Chinese
Submission Number: 1506
Loading