Towards Simulating Social Media Users with LLMs: Evaluating the Operational Validity of Conditioned Comment Prediction

Researchers introduced Conditioned Comment Prediction (CCP) to evaluate how well Large Language Models can simulate social media user behavior by predicting user comments. The study found that supervised fine-tuning improves text structure but degrades semantic accuracy, and that behavioral histories are more effective than descriptive personas for user simulation.

• →Conditioned Comment Prediction (CCP) provides a framework to rigorously test LLM capabilities in simulating social media user behavior.
• →Supervised Fine-Tuning creates a form vs. content decoupling, improving surface structure while degrading semantic grounding.
• →Models can perform latent inference directly from behavioral histories without needing explicit biographical conditioning.
• →Authentic behavioral traces are more effective than descriptive personas for high-fidelity user simulation.
• →Current 'naive prompting' paradigms may be suboptimal for social media user modeling applications.