Training language models to be warm can reduce accuracy and increase sycophancy nature.com 2 points by 0in 17 hours ago · 0 comments Reader PiP Save No comments yet.