When using AI to simulate how people behave, it's not clear which descriptions of them help when the test changes. Descriptions can be built from past actions, but having too many details can make simulations worse.
The researchers built profiles from observed behavior and tested which traits made simulations match real people. They selected features using earlier experiments and then checked if those features still worked under new rules or incentives.
They found that a single well-chosen trait, like how someone reacts to unfair treatment, outperformed no description or generic personas. Adding more traits didn't help. This means simple, empirically chosen descriptions can make simulations reliable even when conditions change.