Unverified50% confidenceOpinionExact time
RLHF标注员群体在地域、文化背景、教育程度上分布不均，导致模型习得的偏好是特定人群价值观的统计偏态
1
Sources
50%
Confidence
Long-term
Relevance
7/15/2026
First Seen
Sources
Related Claims
UnverifiedRLHF对齐的方向本质上反映的是参与标注的人群的价值观，而非某种客观标准75% similarVerifiedTraditional RLHF relies on preference judgments from human annotators, which is costly and subject to annotation quality bottlenecks.66% 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.65% similarVerifiedRLHF has limitations: human annotators have biases and inconsistencies, reward models may be over-optimized (reward hacking), and the process is costly and difficult to scale.65% similarUnverified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65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/520775API
curl https://kongchang.com/api/v1/knowledge/claims/520775MCP
get_claim(id=520775)