Reward Hacking

Advanced

Maximizing reward without fulfilling real goal.

Full Definition

Maximizing reward without fulfilling real goal.

Keywords

proxy exploitation

Domains

AI Safety & Alignment

Related Terms

Instrumental Convergence related to

Tendency for agents to pursue resources regardless of final goal.

 Value Misalignment related to

Model optimizes objectives misaligned with human values.

 Outer Alignment related to

Correctly specifying goals.

 Inner Alignment related to

Ensuring learned behavior matches intended objective.

 Alignment Problem related to

Ensuring AI systems pursue intended human goals.

 Specification Gaming related to

Model exploits poorly specified objectives.

Concept Map

See how Reward Hacking connects to other concepts.

Open Knowledge Graph