Bayesian Belief Layer for Controllable Opinion Dynamics in LLM Agents
The paper adds an explicit Bayesian belief state so LLM-agent opinions can be set, audited, and tested instead of left inside the prompt.
Each stance is represented as a probability and updated once for every utterance the agent hears. A prior-strength parameter, kappa, controls stubbornness and can be swept to produce consensus, persistent disagreement, or committed-minority influence. The authors report persistent-disagreement results matching Friedkin-Johnsen fixed points at R² 0.93–0.99, and say kappa remains recoverable after generation across four models. The layer also exposed model-specific stance biases that an end-to-end simulation would hide. ArXiv · AI/CL/LG's note
Each stance is represented as a probability and updated once for every utterance the agent hears. A prior-strength parameter, kappa, controls stubbornness and can be swept to produce consensus, persistent disagreement, or committed-minority influence. The authors report persistent-disagreement results matching Friedkin-Johnsen fixed points at R² 0.93–0.99, and say kappa remains recoverable after generation across four models. The layer also exposed model-specific stance biases that an end-to-end simulation would hide. ArXiv · AI/CL/LG's note
score 4