Bayesian Chronicle Agents Add Controllable Beliefs to LLM Social Simulation
Bayesian Belief Layer for Controllable Opinion Dynamics in LLM Agents
Unlike implicit in-context opinion shifts, BCA lets social-simulation builders verify a stubbornness setting numerically, matching Friedkin–Johnsen fixed points at R² 0.93–0.99 and recovering prescribed parameters perfectly across four LLMs.
The paper introduces Bayesian Chronicle Agents (BCA), a belief layer for LLM agents in which each stance is a probability updated by one Bayesian step per utterance heard. A single prior-strength parameter κ encodes stubbornness, modeled on its role in Friedkin–Johnsen (FJ) opinion dynamics. Sweeping κ yields three regimes on demand — consensus, persistent disagreement and committed-minority influence — with persistent disagreement matching FJ closed-form fixed points at R² of 0.93 to 0.99. The prescribed κ remains recoverable after the language round-trip, with perfect rank-order recovery across all four models tested. Explicit belief numbers also make simulations auditable, exposing systematic per-model stance biases across the four models.
Bayesian Belief Layer for Controllable Opinion Dynamics in LLM Agents
LLM agents in social simulation revise their opinions implicitly, in context: how open an agent is to persuasion can neither be specified nor verified, and collective outcomes inherit the model's training prior. We introduce Bayesian Chronicle Agents (BCA), a minimal belief layer separating \emph{what} an agent believes from \emph{how} it speaks. Each stance is a probability, updated by one Bayesian step per utterance heard. A single prior-strength parameter $κ$ encodes stubbornness, modeled after its role in Friedkin--Johnsen (FJ) opinion dynamics. We then sweep this parameter to yield three canonical regimes of opinion dynamics on demand (consensus, persistent disagreement, committed-minority influence), with persistent disagreement matching the FJ closed-form fixed points at $R^2\!=\!0.93$--$0.99$. We further show that prescribed $κ$ remains recoverable after the language round-trip, with perfect rank-order recovery across all four models. Explicit belief also makes simulation auditable: the layer surfaces systematic per-model stance biases that end-to-end simulation would silently absorb.