Should we align models based on human preference or human behaviour? We explore Alignment Makes Language Models Normative, Not Descriptive (Shapira, et al., 2026) to find out.

https://doi.org/10.48550/arXiv.2603.17218