Unverified50% confidenceFactExact time
The study uses two automated reward dimensions: generating understandable surprise and predicted audience amusement.
1
Sources
50%
Confidence
Long-term
Relevance
10/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe amusement reward that predicts audience laughter can be exploited: adding explicit laughter cues to a message inflates the predicted score.75% similarUnverifiedThe surprise reward based on word vector embeddings has a vulnerability: nonsensical responses with randomly shuffled words can score just as high as genuinely clever witticisms.70% similarUnverifiedHuman annotators exhibit a cognitive bias toward more information being better when evaluating answers, causing reward models to implicitly link length with quality, which is one of the causes of model verbosity.68% similarUnverifiedGao et al.（2023）的研究发现随着策略模型对奖励模型的优化程度增加，真实人类偏好评分先升后降呈现倒U形曲线66% similarUnverifiedThe greatest risk of self-improvement comes from the reward signal itself: if the reward is improperly defined, the agent will optimize evaluation metrics rather than the true objective (the reinforcement learning version of Goodhart's Law).64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/966760API
curl https://kongchang.com/api/v1/knowledge/claims/966760MCP
get_claim(id=966760)