Verified65% confidenceTradeoffExact time
RLHF 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.
3
Sources
65%
Confidence
Long-term
Relevance
7/15/2026
First Seen
Sources
AI大模型入门：核心概念、发展历程与国产模型格局
Related Claims
Unverified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79% similarVerifiedIn RLHF, human evaluators tend to give higher scores to answers that are structurally complete and politely worded when making subjective creative judgments; this bias is referred to as a variant of reward hacking.78% similarVerifiedTraditional RLHF relies on preference judgments from human annotators, which is costly and subject to annotation quality bottlenecks.77% similarUnverifiedRLHF relies on large amounts of human annotation to construct reward signals, which is more costly but better at capturing implicit human preferences.75% similarUnverifiedMeans of mitigating reward hacking include reward shaping, multi-objective constraints, and introducing human judgment in reinforcement learning from human feedback (RLHF), but completely eliminating reward hacking remains an open problem.75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/515616API
curl https://kongchang.com/api/v1/knowledge/claims/515616MCP
get_claim(id=515616)