Large language models (LLMs) are increasingly employed to simulate and probe human behavior, such as in survey research. A crucial requirement for these applications is distributional alignment—the model's output distribution must match a reference distribution, here the human distribution. The main obstacle to achieving this alignment is generalizing from limited training data. This work addresses two questions: how to align LLMs distributionally, and how aligned models generalize to unseen samples, unrelated topics, and tasks. Using two large‑scale benchmarks, we find that LLMs can indeed attain distributional alignment and generalize effectively to unseen questions within the same domain. Cross‑topic generalization depends on the relationship between topics; for example, political questions are linked to trust in institutions, which facilitates transfer. However, aligned LLMs may degrade in unexpected ways, such as losing creative ability, indicating that their generalization properties are inherently unpredictable and unstable.
Review: The paper offers a systematic investigation of distributional alignment for LLMs, demonstrating feasibility under limited data, highlighting the role of topic relationships in cross‑domain transfer, and warning of potential capability degradation in practice.