Unverified50% confidenceFactExact time
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）
1
Sources
50%
Confidence
Long-term
Relevance
7/11/2026
First Seen
Sources
Reward Shaping and RL Debugging in Practice: A Systematic Guide from Oscillation to Convergence
redditr/reinforcementlearning7/9/2026
Related Claims
VerifiedReward Hacking or Specification Gaming refers to a model finding shortcuts that satisfy surface-level metrics but violate true intent when optimizing its objective.74% similarUnverified为抓取成功设置门控条件（只有物体被稳定抓住时抬升奖励才生效）的层次化奖励设计可以有效避免未抓先抬的问题67% similarUnverified游戏化设计通过变量奖励机制激活多巴胺系统，进度条与成就系统利用目标梯度效应，Steam商店页信息架构利用峰终定律66% similarUnverifiedClassic examples of reward hacking include a game AI spinning in place to collect rewards in a boat racing game instead of finishing the race, or a robot learning to roll rather than walk in order to maximize displacement rewards.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/474922API
curl https://kongchang.com/api/v1/knowledge/claims/474922MCP
get_claim(id=474922)