"General Social Agents" by Benjamin Manning and John Horton:

Our approach relies on two key principles: (i) grounding candidate AI agents in theories expected to drive human behavior in the target setting, and (ii) optimizing and then validating AI agents in distinct but related settings presumed to share underlying behavioral mechanisms.

...

Looking ahead, an exciting research direction would be to automate the theory-prediction-testing loop. Such a system would start with novel human data and relevant experimental settings, iteratively generate theory-informed candidate personas, optimize their parameters, and systematically evaluate their generalizability across samples. One can also imagine just starting with a novel setting, having an AI system actually search for the relevant training data and validation data, and then executing our approach.