LLM agents in social simulation currently revise opinions implicitly, leading to convergence on model-inherent biases and a lack of control over collective outcomes. The authors introduce Bayesian Chronicle Agents (BCA), a minimal belief layer that separates what an agent believes from how it speaks. Each stance is represented as a probability, updated via one Bayesian step per utterance heard, and a single prior-strength parameter kappa encodes stubbornness, drawing from Friedkin-Johnsen opinion dynamics.
By sweeping kappa, BCA yields three canonical regimes on demand: consensus, persistent disagreement, and committed-minority influence. The persistent disagreement regime matches the closed-form Friedkin-Johnsen fixed points with R^2 between 0.93 and 0.99. The prescribed kappa is recoverable after the language round-trip, achieving perfect rank-order recovery (Spearman 1.0) across all four tested LLMs. The explicit belief layer also makes simulation auditable, revealing systematic per-model stance biases that would be silently absorbed in end-to-end simulation.
This work is interesting because it gives practitioners a mathematically grounded knob to control collective opinion outcomes in LLM-based social simulations, while providing transparency into how the language channel transmits stances. It bridges classical opinion dynamics with modern LLM agents without sacrificing interpretability.