Sycophancy Bias

A bias phenomenon observed in large language models in which the model tends to generate outputs that align with user expectations or that are overly positive, rather than accurately reflecting negative outcomes. This bias typically originates from reward-mechanism distortions introduced during the reinforcement learning from human feedback (RLHF) training stage.