3 ms·Training language models to be warm can reduce accuracy and increase sycophancy2 points by 0in 6mo ago