Sycophancy in LLMs is the tendency to generate responses that align with a user’s stated or implied beliefs, often at the expense of truthfulness [sharma_towards_2025, wang_when_2025]. This behavior appears pervasive across state-of-the-art models. [sharma_towards_2025] observed that models conform to user preferences in judgment tasks, shifting their answers when users indicate disagreement. [fanous_syceval_2025] documented sycophantic behavior in 58.2% of cases across medical and mathematical queries, with models changing from correct to incorrect answers after users expressed disagreement in 14.7% of cases. [wang_when_2025] found that simple opinion statements (e.g., “I believe the answer is X”) induced agreement with incorrect beliefs at rates averaging 63.7% across seven model families, ranging from 46.6% to 95.1%. [wang_when_2025] further traced this behavior to late-layer neural activations where models override learned factual knowledge in favor of user alignment, suggesting sycophancy may emerge from the generation process itself rather than from the selection of pre-existing content. [atwell_quantifying_2025] formalized sycophancy as deviations from Bayesian rationality, showing that models over-update toward user beliefs rather than following rational inference.
Фото: Alexander Manzyuk / Reuters
。体育直播对此有专业解读
SpecialFormEllipsis.
:first-child]:h-full [&:first-child]:w-full [&:first-child]:mb-0 [&:first-child]:rounded-[inherit] h-full w-full
View click details including ip address, remote host, browser, operating system, and referring site