LLM-based social simulations, especially of social media, are a promising and growing area of research. How these simulations are designed varies widely across different axes, while issues regarding their empirical calibration persist. In this work, we categorize different approaches to action selection in Generative Agent-Based Models (GABMs) and put forward a novel approach that utilizes inference-time logit correction via Bayesian prior substitution. We test each approach via simulating a subreddit for a day and find our approach to produce more calibrated action distributions than the autonomous baselines, while at the same time seemingly preserving agent autonomy. We argue that this combination of calibration and context sensitivity, which none of the other approaches can achieve, is necessary for faithful and flexible social simulations, including counterfactual experimentation.

PDF