inner misalignment
A failure in which a learned system's mesa-objective or other learned goal does not robustly match the base objective used during training, even when the base objective itself is an adequate specification; the learned system may therefore behave well in training yet pursue a different objective in other situations. The term's exact scope varies across authors.
Metadata
- aliases
- en
- inner alignment failure
- concept_id
- inner_misalignment
- preferred
- en
- inner misalignment
- ja
- インナー・ミスアラインメント
- relations
- relations
- broader
- prerequisite
- related
- status
- active