The method focuses on how agents make incremental decisions, not just whether a final output looks good. That is important for simulations where believable behavior emerges from many small actions over time.
Better preference learning could make generative agents more useful for testing social dynamics, product behavior and policy scenarios before real-world deployment.