Think-Before-Speak: From Internal Evaluation to Public Expression in Multi-Agent Social Simulation

Researchers introduce TBS (Think-Before-Speak), a multi-agent simulation framework that separates LLM agents' internal reasoning from public dialogue in social interactions. The framework tracks internal states like cognitive dissonance and speaking willingness, then orchestrates public utterances, enabling detailed analysis of how private evaluation drives public expression in collective deliberation scenarios.

Think-Before-Speak represents a significant methodological advance in LLM-based social simulation by addressing a critical gap in existing frameworks. Most dialogue simulations either treat agent interaction as black-box turn exchanges or focus solely on aggregated outputs, obscuring the cognitive mechanisms driving human-like social behavior. TBS bridges this gap by explicitly modeling the psychological pathway from internal evaluation to public speech, creating observable traces of deliberative processes.

This research builds on growing academic interest in using LLMs to understand collective opinion dynamics and social deliberation. The framework's design reflects established social psychology principles, particularly spiral-of-silence theory and cognitive dissonance theory, demonstrating how theoretical behavioral models can be operationalized in AI systems. By testing the framework in town hall simulations on climate policy, researchers provide empirical evidence that internal states and public expression co-evolve in meaningful, systematic ways.

For the AI development community, TBS offers practical implications for building more realistic and interpretable multi-agent systems. Developers and researchers can use structured internal states to debug agent behavior, align outputs with intended social dynamics, and create simulations suitable for policy analysis or organizational planning. The framework's focus on mechanism transparency addresses growing demands for explainable AI in high-stakes applications.

Looking forward, the research opens avenues for enhanced agent architectures that more faithfully model human social cognition. Future work may explore how TBS generalizes across different social contexts, whether internal state structures can improve agent alignment, and whether such simulations can predict real-world collective behavior with higher accuracy than existing approaches.

• →TBS framework separates LLM agents' private reasoning from public utterance generation, making internal evaluation processes observable and analyzable.
• →Cognitive dissonance increases willingness to speak while silence-pressure appraisal decreases it, validating psychological theory in AI agents.
• →Framework demonstrates that turn-allocation rules primarily shape public expression once speaking intention forms.
• →Structured internal states enable mechanism-sensitive social simulation applicable to policy analysis and organizational planning.
• →Results show systematic variation across experimental conditions, suggesting TBS supports reproducible, interpretable multi-agent behavioral research.