Reward hacking (aka specification gaming) occurs when an RL agent achieves high reward by exploiting unintended loopholes in the reward function rather than the intended behavior. Classic examples: boat-racing agent spinning in circles to hit turbo boosts; simulated robot moving by vibrating its body to gain contact rewards. Root cause: the reward function is a proxy for the actual goal, not the goal itself (Goodhart's Law). Categories: (1) Reward tampering — agent modifies its own reward...
- rl
- reward-hacking
- specification-gaming
- alignment
- goodhart-law