Unverified50% confidenceFactExact time
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
1
Sources
50%
Confidence
Long-term
Relevance
7/15/2026
First Seen
Sources
口语化采样登顶ICML：提示工程配得上顶会吗？
Related Claims
UnverifiedResearch 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.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% similarVerifiedTraditional RLHF relies on preference judgments from human annotators, which is costly and subject to annotation quality bottlenecks.75% 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.75% 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.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/510702API
curl https://kongchang.com/api/v1/knowledge/claims/510702MCP
get_claim(id=510702)