Title: Think-Before-Speak: Bridging Internal Assessment and Public Discourse in Multi-Agent Social Simulations

Abstract:

LLM-driven multi-agent simulations present a compelling avenue for investigating collective opinion dynamics, deliberation, and social interaction. Despite this potential, current dialogue simulation models primarily focus on observable turn-taking or aggregated outputs, often obscuring the underlying internal evaluative mechanisms. Consequently, the processes governing speaking intentions and public expression remain difficult to analyze. To address this gap, we present TBS (Think-Before-Speak), a novel interval-based simulation framework that distinguishes between an agent’s private reasoning and its public utterances.

During each simulation interval, agents update structured internal states by integrating shared dialogue history with their individual memories. These internal states comprise several key dimensions: dissonance-related appraisal, perceived opinion climate, perceived isolation risk, response strategy, and willingness to speak. Following this internal update, an orchestrator manages competing speaking intentions, selecting a single utterance to add to the public dialogue. This design allows internal evaluations and public interactions to co-evolve dynamically.

We validated TBS through simulated town hall discussions focused on a climate-related policy issue. Our analysis reveals that TBS generates coherent internal-state traces that shift systematically in response to variations in turn-allocation, silence, and memory conditions. Specifically, we found that dissonance-related appraisal enhances an agent’s willingness to speak, while silence-pressure appraisal diminishes it. However, once a speaking intention is established, the actual public expression is predominantly determined by the turn-allocation rules. These results indicate that TBS facilitates mechanism-sensitive social simulation by rendering the trajectory from internal evaluation to public expression both observable and analyzable.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC