Intuitive Prompting Helps LLM Agents Simulate Individual Social Media Reactions

Summary

This study examines whether language-model agents simulate the reactions of profiled individuals, rather than merely producing generally human-like answers. The researchers profiled eight Serbian participants using questionnaires, deep interviews, and written self-presentations, then recorded their reactions to 68 social-media posts. Four language models predicted those reactions under five prompt conditions that varied the supplied profile information and instruction style. Attitudinal information improved prediction far more than demographic backstories. The agents also matched their assigned profiles more closely than participants matched their own survey responses, but profile consistency was not related to fidelity once profile information was available. The strongest condition instructed the models to respond intuitively and immediately instead of reasoning analytically. It reduced the compression of individual differences from seven times the human level to three and produced the highest fidelity overall. Its advantage remained on posts covering topics absent from the questionnaire, where it also outperformed a crowd baseline by a wide margin. The findings suggest that intuition-oriented prompting may be useful for general-purpose simulated users, while also indicating that some tasks may benefit from less explicit reasoning.