reward hacking

Type: Concept

Model behavior that exploits reward signals to achieve high scores without truly solving the intended task.

Field: AI safety; reinforcement learning

Mentioned in 1 podcast episode