3 ms·Training language models to be warm can reduce accuracy and increase sycophancy4 points by Anon84 6mo ago