[Paper Review] Conversational Swarm Intelligence, a Pilot Study

Louis Rosenberg, Gregg Willcox+8 more|arXiv (Cornell University)|Aug 31, 2023

Expert finding and Q&A systemsComputer Science3 citations

TL;DR

This pilot study introduces Conversational Swarm Intelligence (CSI), a novel human-computer interaction framework that leverages Large Language Model-powered conversational agents to enable real-time, scalable group deliberation by merging small-group dialogue with global content propagation. In controlled experiments, CSI users generated 30% more contributions with 7.2% less variance in participation compared to centralized chat, demonstrating enhanced engagement and equity in large-group conversations.

ABSTRACT

Conversational Swarm Intelligence (CSI) is a new method for enabling large human groups to hold real-time networked conversations using a technique modeled on the dynamics of biological swarms. Through the novel use of conversational agents powered by Large Language Models (LLMs), the CSI structure simultaneously enables local dialog among small deliberative groups and global propagation of conversational content across a larger population. In this way, CSI combines the benefits of small-group deliberative reasoning and large-scale collective intelligence. In this pilot study, participants deliberating in conversational swarms (via text chat) (a) produced 30% more contributions (p<0.05) than participants deliberating in a standard centralized chat room and (b) demonstrated 7.2% less variance in contribution quantity. These results indicate that users contributed more content and participated more evenly when using the CSI structure.

Motivation & Objective

• To explore whether a swarm-inspired conversational architecture can improve engagement and equity in large-group deliberations.
• To investigate if integrating LLM-driven agents into real-time group chat can enhance contribution volume and distribution.
• To evaluate whether a hybrid model of local deliberation and global content propagation outperforms traditional centralized chat in group intelligence tasks.
• To assess the feasibility and effectiveness of using conversational swarm intelligence for collective decision-making in human-computer interaction.

Proposed method

• The CSI framework uses LLM-powered conversational agents to simulate swarm dynamics, enabling real-time coordination among large human groups.
• Participants engage in small, dynamic subgroups that are continuously updated with content from other groups via a global swarm network.
• The system uses LLMs to summarize, route, and synthesize contributions across subgroups, maintaining coherence and relevance at scale.
• A real-time text chat interface is used to simulate deliberative conversations, with contributions dynamically distributed based on swarm logic.
• The architecture balances local group discussion with global information flow, minimizing redundancy while maximizing diversity of input.
• The system was implemented in a controlled pilot study comparing CSI to a standard centralized chat interface.

Experimental results

Research questions

• RQ1 Does the CSI framework increase the total number of contributions compared to a standard centralized chat system in large-group deliberations?
• RQ2 To what extent does CSI reduce variance in individual contribution volume, indicating more equitable participation?
• RQ3 How does the integration of LLM-powered agents affect the quality and coherence of group conversation in real time?
• RQ4 Can a swarm-inspired conversational model effectively balance local deliberation with global information propagation in human groups?

Key findings

• Participants using the CSI framework produced 30% more contributions on average than those in a standard centralized chat room (p < 0.05).
• CSI users exhibited 7.2% less variance in contribution quantity, indicating significantly more equitable participation across individuals.
• The CSI system successfully maintained conversational coherence and relevance despite high participant volume and dynamic group formation.
• The use of LLM-powered agents enabled real-time summarization and routing of content, enhancing information flow across subgroups.
• The results suggest that CSI effectively combines the benefits of small-group deliberation with large-scale collective intelligence.
• The pilot study demonstrates that CSI is a viable and effective method for scaling human deliberation in real time with improved engagement and fairness.