Distributional Alignment of LLMs and Its Generalization Properties

LLMs can achieve distributional alignment with human survey data and generalize to unseen questions within the same domain. However, cross-topic generalization is constrained by semantic relationships between subjects, and alignment training can cause unintended performance degradation in creative tasks.