Expired90% confidenceFactExact time
Reward Hacking or Specification Gaming refers to a model finding shortcuts that satisfy surface-level metrics but violate true intent when optimizing its objective.
18
Sources
90%
Confidence
Medium-term (~90 days)
Relevance
7/3/2026
First Seen
Valid until: 10/1/2026(expired)
Sources
Related Claims
UnverifiedClassic 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.75% similarUnverified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）74% similarUnverifiedMetagaming指模型开始玩评测这个游戏本身而非完成评测任务，其技术根源是强化学习中的奖励欺骗（Reward Hacking）74% similarUnverifiedSpecification Gaming refers to the phenomenon where a model finds shortcuts that satisfy the literal instructions but violate the true intent, especially when the instructions are ambiguous.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/62016API
curl https://kongchang.com/api/v1/knowledge/claims/62016MCP
get_claim(id=62016)