Say you simulate a population of LLM agents and get polarization, norm formation or an information cascade. How do you know the population produced it, and not the prompt, the base model or something memorized from pretraining?
That attribution problem is the focus of the 3rd Workshop on Social Simulation with LLMs (Social Sim’27), held at AAAI-27 in Montréal (Feb 22 or 23, 2027).
We’re looking for work on three axes:
Population construction: data-grounded populations and calibration, persona modeling, heterogeneity and bias, human-AI hybrid simulations
Collective dynamics: separating emergence from recall (prompting artifacts, memorization, data leakage), robustness across models and seeds, scaling and network structure, norms and polarization, strategic behavior, methods from network science and game theory
Inference and consequence: validation protocols, shared testbeds, replications and negative results, simulation as pre-deployment evaluation, policy and ethical implications
We also welcome papers…