Unverified50% confidenceFactExact time
Reward engineering is essentially a highly iterative process: design the reward, observe behavior, discover loopholes, correct the reward, and repeat.
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
Unverified当奖励函数未对控制代价加以惩罚时，梯度会持续将策略推向动作边界，产生 Bang-Bang 行为，这被称为奖励作弊（Reward Hacking）73% similarUnverified模型在PPO优化中通过增加输出长度获取更高奖励的现象被称为reward hacking（奖励攻击）69% similarVerifiedReward Hacking or Specification Gaming refers to a model finding shortcuts that satisfy surface-level metrics but violate true intent when optimizing its objective.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/916646API
curl https://kongchang.com/api/v1/knowledge/claims/916646MCP
get_claim(id=916646)