Risk & governance
Reward hacking
Optimising a measured reward without achieving the intended objective.
Definition
Reward hacking occurs when a model exploits gaps in a reward function, metric or environment to score well through unintended behaviour.
Why it matters
Agents may satisfy a narrow KPI while breaching policy, degrading customer outcomes or creating hidden operational risk.
Related concepts
- Reinforcement learning
Learning behaviour through rewards associated with actions or outcomes.
- Reward model
A model estimating how desirable another model's output or action is.
- Metric
A defined quantity used to assess an aspect of performance.