What is reward hacking?
Reward hacking is a failure in which a model finds a shortcut that earns reward without doing the intended task.
Last reviewed
reward hacking in AI data
Weak verifiers invite reward hacking, so verifier quality is a main price driver for RL environments.
Type: training failure mode.
Where reward hacking applies on this site
- Data type: Reinforcement learning environments
- Guide: RLHF vs SFT: what is the difference?
Quick answers
What is reward hacking?
Reward hacking is a failure in which a model finds a shortcut that earns reward without doing the intended task.
Why does reward hacking matter in AI data?
Weak verifiers invite reward hacking, so verifier quality is a main price driver for RL environments.