outer alignment
The problem or property of specifying a training objective that adequately captures the designers' intended goal or broader alignment target, assuming the learned system then robustly pursues that specification. The training objective may itself be implemented through a loss, reward-derived return, preference feedback, or another criterion. The inner/outer distinction is useful but its exact boundary varies across authors and settings.
Metadata
- aliases
- en
- outer alignment problem
- reward misspecification problem
- concept_id
- outer_alignment
- preferred
- en
- outer alignment
- ja
- アウターアラインメント
- relations
- relations
- broader
- prerequisite
- related
- status
- active