Technical Summary: Think-Before-Speak (TBS)

Problem Statement

Existing Large Language Model (LLM)-based multi-agent dialogue frameworks primarily model social interaction as a sequence of observable turn exchanges or aggregated outputs. While effective for generating public transcripts, these approaches fail to capture the internal evaluative processes that occur between turns. In real-world public deliberation (e.g., town halls), silence does not imply cognitive inactivity; participants continuously interpret remarks, assess opinion climates, experience psychological pressure, and decide whether speaking is warranted. Current paradigms—specifically hierarchical aggregation (where agents reason in parallel but ignore same-round contributions) and sequential turn-taking (where agents only reason when their turn arrives)—treat public utterances as direct products of the simulation protocol rather than outcomes of evolving internal states. This limitation obscures the pathway from internal evaluation to speaking intention and public expression, making it difficult to study mechanisms like cognitive dissonance, the spiral of silence, and self-censorship.

Methodology: The TBS Framework

The authors propose Think-Before-Speak (TBS), a discrete-time, interval-based multi-agent simulation framework designed to separate private reasoning from public utterance generation.

Core Architecture

1. Interval-Based Interaction: Time is abstracted into fine-grained intervals ($v$). Unlike sequential models where only one agent acts per round, TBS allows all agents to reason continuously at every interval based on shared dialogue history and their own memory.
2. Dual-Module Agent Design:
   • Internal Reasoning Module: Updates structured internal states ($s$) at every interval. These states include:
     • Dissonance-related appraisal: Perceived inconsistency, dissonance tension, and motivation to reduce tension.
     • Silence-pressure appraisal: Perceived opinion climate and perceived isolation risk.
     • Response strategy: Selection from a pool (defend, qualify, converge, challenge, bridge, withhold).
     • Time cost: An estimated latency for reasoning, used to resolve conflicts.
   • Speaking Module: Generates the actual utterance ($u$) only when the agent is selected to speak.
3. Orchestrator & Conflict Resolution:
   • The orchestrator manages the pipeline. At each interval, all agents submit speaking intentions accompanied by their estimated time_cost.
   • If multiple agents intend to speak, the orchestrator selects the agent with the smallest time cost (fastest reasoning) as the speaker for that interval.
   • Non-speaking agents update their internal states but do not generate public text, reducing token overhead.
4. Memory Management: Agents utilize a three-part memory schema (Group, Monologue, Self) with dynamic retrieval mechanisms:
   • Latest: Retains only the most recent $n$ rounds.
   • Sampling: Randomly samples history with a long-tail distribution favoring recent events.
   • Summary: Iteratively compresses dialogue and internal states into summaries, modeling memory accumulation and loss of detail.

Experimental Setup

The framework was evaluated using town hall discussions on a climate policy topic ("solar photovoltaic mandate").

• Agents: Utilized personas from the "Six Americas" (six distinct climate attitudes) and "Balanced Stakeholders" (supportive vs. opposing pairs) taxonomies.
• Backbone Models: Gemini-2.5-Flash-Lite and Gemini-2.5-Flash.
• Variables:
  • Turn Allocation: Willing Mode (agents self-select to speak based on internal state) vs. Rotation Mode (fixed sequential order).
  • Silence Constraints: Force Speak (orchestrator forces a speaker if no one volunteers) vs. No Force Speak (silence is a valid outcome).
  • Memory Modes: Latest, Random, and Summary.

Key Contributions

1. Discrete-Time Simulation Protocol: A novel protocol bridging continuous-time reasoning and discrete communication through interval-based interaction and conflict resolution.
2. Unified Parallel-Competitive Protocol: Integration of parallel reasoning (all agents think at every step) with competitive turn-taking (only one speaks), improving both realism and efficiency.
3. Structured Interpretability: A framework that tracks agent cognition (internal states) and interaction dynamics explicitly, enabling fine-grained analysis of social behavior and decision processes.
4. Empirical Validation: Preliminary experiments demonstrating the framework's ability to generate coherent internal-state traces that vary systematically across different simulation conditions.

Results

The study analyzed structured process indicators (dissonance index, silence-pressure index, willingness to speak, and public expression) using linear mixed-effects models.

• Internal State Coherence: The framework produced internally consistent process traces. The dissonance index ($\alpha = .89$) and silence-pressure index ($\alpha = .92$) showed strong reliability.
• Impact of Conditions on Internal States:
  • Dissonance: Increased modestly over time. It was significantly higher in the Willing Mode and Force Speak conditions compared to Rotation/No-Force settings. The "Six Americas" persona ecology showed lower average dissonance than Balanced Stakeholders.
  • Silence Pressure: Did not show a significant linear trend over time but was significantly higher in the Willing Mode, Force Speak, and Summary Memory conditions. Summary memory increased both the average level and growth rate of silence pressure.
• From Evaluation to Expression (RQ2):
  • Internal Evaluation predicts Willingness: Higher dissonance significantly increased the probability of wanting to speak (odds ratio $\approx 4.34$ per 0.1 unit increase), while higher silence pressure significantly decreased it (odds ratio $\approx 0.24$).
  • Role of Turn Allocation: In the Willing Mode, the strength of the relationship between internal states and speaking intention was weaker than in Rotation mode, suggesting self-selected participation raises the threshold for action, though the directional effects remained.
  • From Intention to Public Expression: Once an agent formed a speaking intention, internal evaluations had only a marginal effect on whether they actually received the floor. Instead, turn-allocation rules were the primary predictor of realized expression. In the Willing Mode, agents who wanted to speak were ~3.34 times more likely to be selected than in Rotation mode.

Significance and Claims

The paper claims that TBS supports mechanism-sensitive social simulation by making the pathway from internal evaluation to public expression observable and analyzable.

• Theoretical Alignment: The results validate that TBS successfully operationalizes communication theories, specifically distinguishing between dissonance-related appraisal (an expressive motivator) and silence-pressure appraisal (an expressive constraint).
• Process vs. Output: The framework demonstrates that public utterances are not direct products of the simulation protocol but emerge from a two-stage process: (1) internal evaluations shaping the willingness to speak, and (2) external rules (turn allocation) determining the access to the public floor.
• Limitations and Future Work: The authors acknowledge the work is preliminary. Limitations include heuristic memory management (not grounded in psychological memory theories), the lack of content-based retrieval (e.g., RAG), and limited experimental scale regarding topics and agent numbers. Future work aims to expand these scales and refine memory mechanisms to better balance temporal decay with content-based reactivation.

The paper concludes that TBS provides a realistic, interpretable, and experimentally controllable testbed for studying latent communicative processes that are difficult to observe directly in human interactions.