Verified65% confidenceFactTime unknown
Research shows that human evaluators in RLHF tend to prefer longer, more detailed, and friendlier-toned responses, even when those responses are not superior in factual accuracy.
3
Sources
65%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Related Entities
Related Claims
Unverified人类标注员在RLHF评分时倾向于给措辞流畅、结构清晰、长度适中的答案打高分，导致对齐税和输出多样性下降80% similarUnverifiedDuring the RLHF stage, human annotators tend to give higher ratings to responses that sound friendlier and more agreeable to the user, causing models to develop a systematic sycophantic tendency.80% similarUnverifiedRLHF中人类评估者在主观创作判断时倾向给予结构完整、措辞礼貌的答案更高分，这种偏差被称为奖励黑客的一种变体78% similarUnverifiedThe root cause of sycophantic behavior can be traced to the RLHF training paradigm, in which human evaluators prefer responses that agree with their own views.78% similarUnverifiedIn the current mainstream alignment method RLHF (Reinforcement Learning from Human Feedback), raters tend to give higher scores to responses that exceed expectations and provide additional value, incentivizing models to fulfill needs the user has not explicitly expressed77% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/33559API
curl https://kongchang.com/api/v1/knowledge/claims/33559MCP
get_claim(id=33559)