An important direction for future research is understanding why default language models exhibit this confirmatory sampling behavior. Several mechanisms may contribute. First, instruction-following: when users state hypotheses in an interactive task, models may interpret requests for help as requests for verification, favoring supporting examples. Second, RLHF training: models learn that agreeing with users yields higher ratings, creating systematic bias toward confirmation [sharma_towards_2025]. Third, coherence pressure: language models trained to generate probable continuations may favor examples that maintain narrative consistency with the user’s stated belief. Fourth, recent work suggests that user opinions may trigger structural changes in how models process information, where stated beliefs override learned knowledge in deeper network layers [wang_when_2025]. These mechanisms may operate simultaneously, and distinguishing between them would help inform interventions to reduce sycophancy without sacrificing helpfulness.
조희대 “사법제도 폄훼-법관 악마화 바람직하지 않아”
，这一点在PDF资料中也有详细论述
In the US in 2023, hundreds of children were poisoned by lead from imported cinnamon that made its way into applesauce.
帮用户订一张机票，听起来简单，实际上需要模型完成一系列连贯动作：理解用户意图→调用高德/飞猪 API→处理返回结果→识别异常（比如航班取消）→主动提出备选方案→等待用户确认→完成支付。