Cooperative Multi-Agent Architectures for Sentiment Classification: Round-Robin and Society of Mind Approaches
Year: 2026
Type: conference-paper
Abstract: This study presents a comparative investigation of three multi-agent system architectures for sentiment analysis, leveraging an ensemble of local large language models (LLMs) deployed via the Ollama platform. Specifically, three architectures are developed and evaluated: (1) a conventional Round-Robin Group Chat, (2) an optimized Round-Robin variant with improved efficiency, and (3) a three-layer hierarchical Society of Mind architecture. All approaches employ a committee of four LLMs (llama3.2, qwen2, gemma3:4b, and qwen3) that collaborate through the AutoGen framework. In the sentiment analysis of 1,500 news articles spanning the Business, Entertainment, and Technology categories, the experimental results show that Approach 1 (Original) achieves the best performance, with an accuracy of 41.53% and a macro F1-score of 0.3659, while Approaches 2 and 3 attain accuracies of 35.67% and 34.67%, respectively. A more detailed analysis reveals that all three approaches are relatively strong in classifying the POSITIVE class but struggle with the NEGATIVE and NEUTRAL classes, reflecting the inherent complexity and difficulty of sentiment analysis in this setting.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203661087