Tag
This paper proposes using Bayesian Truth Serum as a reward in reinforcement learning fine-tuning to mitigate sycophancy in large language models, showing improved accuracy and reduced answer-flip rates without labeled data.