Ibrahim, L., Hafner, F. S., Rocher, L. (2026). Training language models to be warm can reduce accuracy and increase sycophancy. Nature. https://doi.org/10.1038/s41586-026-10410-0