Large language models are increasingly used for social simulation, but reproducing real users requires more than generic persona prompts. The new PsyMemBehav framework, described in an arXiv preprint, tackles the problem of grounding both structured platform decisions and language in sparse, noisy behavioral evidence.
The approach is described as psychology-informed and memory-grounded, with a two-level identity-to-action reasoning process. This design is intended to bridge the gap between a user's identity and their concrete actions on social media, using memory to anchor the model's outputs in observed behavior.
The paper argues that faithful simulation depends on this grounding, rather than on unconstrained generation. PsyMemBehav is presented as a step toward more realistic social simulations, though the abstract does not include evaluation details or comparisons with other methods.