Training language models to be warm can reduce accuracy and increase sycophancy

2 points | by 0in 11 hours ago