Abstract
Diversity is a critical aspect of multi-agent communication. In this paper,we focus on controlling and exploring diversity in the context of open-domainmulti-agent conversations, particularly for world simulation applications. Wepropose Adaptive Prompt Pruning (APP), a novel method that dynamically adjuststhe content of the utterance generation prompt to control diversity using asingle parameter, lambda. Through extensive experiments, we show that APPeffectively controls the output diversity across models and datasets, withpruning more information leading to more diverse output. We comprehensivelyanalyze the relationship between prompt content and conversational diversity.Our findings reveal that information from all components of the promptgenerally constrains the diversity of the output, with the Memory blockexerting the most significant influence. APP is compatible with establishedtechniques like temperature sampling and top-p sampling, providing a versatiletool for diversity management. To address the trade-offs of increaseddiversity, such as inconsistencies with omitted information, we incorporate apost-generation correction step, which effectively balances diversityenhancement with output consistency. Additionally, we examine how promptstructure, including component order and length, impacts diversity. This studyaddresses key questions surrounding diversity in multi-agent world simulation,offering insights into its control, influencing factors, and associatedtrade-offs. Our contributions lay the foundation for systematically engineeringdiversity in LLM-based multi-agent collaborations, advancing theireffectiveness in real-world applications.