Unverified50% confidenceFactExact time
The privileged information's benefit to critic value prediction accuracy is stronger at the end of the response, while the self-teacher guidance signal exhibits a pattern of being strong at the beginning, decaying in the middle, and rising again at the end.
1
Sources
50%
Confidence
Long-term
Relevance
10/15/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn 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 expressed70% similarUnverified人类标注者在评估答案时存在信息越丰富越好的认知偏差，导致奖励模型将篇幅与质量隐性挂钩，是模型冗余的成因之一70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1032844API
curl https://kongchang.com/api/v1/knowledge/claims/1032844MCP
get_claim(id=1032844)