Unverified50% confidenceFactExact time
标准RLHF存在奖励黑客问题，模型可能通过迎合评分者偏好而非真正提升质量来最大化奖励，是Goodhart定律的体现
1
Sources
50%
Confidence
Long-term
Relevance
7/8/2026
First Seen
Sources
元认知反馈：用强化学习让大模型准确表达不确定性
hackernewshackernews7/7/2026
Related Claims
VerifiedRLHF 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.79% 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77% similarUnverifiedThe phenomenon of models increasing their score by adding length is referred to in academia as 'reward hacking' and is a manifestation of Goodhart's Law.76% similarUnverifiedRLHF（基于人类反馈的强化学习）会放大规避风险效应，因为人类评分者倾向于给保守、无争议的输出打高分68% similarUnverifiedRLHF relies on large amounts of human annotation to construct reward signals, which is more costly but better at capturing implicit human preferences.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/285633API
curl https://kongchang.com/api/v1/knowledge/claims/285633MCP
get_claim(id=285633)