Training language models to be warm can reduce accuracy and increase sycophancy. (nature.com)

Authors: Ibrahim L … Rocher L Journal: Nature Published: 2026-04-01 DOI: https://doi.org/10.1038/s41586-026-10410-0