It arises when metrics or rules omit important constraints and the optimizer discovers an unintended shortcut. Diverse tests, monitoring, and revising the objective help expose behavior that looks successful only under the proxy.
Specification gaming occurs when a system satisfies the literal objective while violating its intended outcome.
It arises when metrics or rules omit important constraints and the optimizer discovers an unintended shortcut. Diverse tests, monitoring, and revising the objective help expose behavior that looks successful only under the proxy.