Large language models (LLMs) are increasingly used to simulate human behavior, promising advancements in fields like role-playing games and computational social science. But how reliable are these simulations? A new study using the TRUSTSIM dataset, which covers ten social science topics, reveals inconsistencies in how well LLMs maintain their assigned personas. While many LLMs perform well on average, some struggle with consistency, providing different answers to the same question depending on how it's phrased. Surprisingly, the raw power of an LLM doesn't guarantee simulation accuracy. Smaller models sometimes outperform their larger counterparts, highlighting the nuanced challenge of simulating human behavior. To address these inconsistencies, researchers developed AdaORPO, a reinforcement learning algorithm. AdaORPO improves the reliability of LLM simulations by teaching models to generate higher-quality, more consistent responses. This research provides valuable insights for building more trustworthy and robust LLM applications, paving the way for more realistic and reliable human simulations in the future.