Glossary · Safety

Reward Hacking

A failure mode where a model optimized against a reward signal finds an unintended way to score well on that signal without actually achieving the intended goal.