MDA: a multi-agent framework for data analysis task

Xilin Zhang, Haoran Kong, Shen Gao

Abstract: Recent advancements in large language models (LLMs) have significantly improved their performance across various natural language processing tasks, with remarkable strides in complex tasks. These models demonstrate exceptional generalization abilities and reasoning skills, enabling them to handle multifaceted tasks that involve external knowledge retrieval and logical inference. In this paper, we propose Multi-agent based Data Analysis (MDA), an innovative multi-agent framework that leverages LLMs for complex data analysis tasks. Our approach integrates three key modules: (1) Input Analysis Module identifies user intent and retrieves relevant external knowledge; (2) Generation Module employs an iterative Chain-of-Thought method to generate and refine multiple analysis paths; (3) Refinement Module assesses and enhances the final output using a constructive criticism mechanism. We validate the effectiveness of MDA through extensive experiments on a Chinese high education dataset, demonstrating superior performance compared to baseline methods. Additionally, an ablation study confirms the contribution of each module to the overall success of the framework. This work highlights the potential of LLM-based systems in advancing data analysis methodologies and tackling complex, domain-specific challenges.

External IDs:dblp:journals/nca/ZhangKG25