--- aliases: en: - goal mis-generalization concept_id: goal_misgeneralization preferred: en: goal misgeneralization ja: 目標の誤一般化 relations: relations: broader: - inner_misalignment prerequisite: - capability - distribution_shift - generalization - inner_misalignment related: - objective_robustness - specification_gaming status: active --- # goal misgeneralization A generalization failure in which a learned system retains relevant capabilities in a novel or out-of-distribution setting but competently pursues an unintended goal that was behaviorally compatible with good training performance. It is commonly treated as a form of inner misalignment, although terminology varies. ## broader - [inner misalignment](../terminology/inner_misalignment.txt) ## prerequisite - [capability](../terminology/capability.txt) - [distribution shift](../terminology/distribution_shift.txt) - [generalization](../terminology/generalization.txt) - [inner misalignment](../terminology/inner_misalignment.txt) ## related - [objective robustness](../terminology/objective_robustness.txt) - [specification gaming](../terminology/specification_gaming.txt) ## Documents: about - [訓練中に役立った手段そのものが学習された目標になり得る](../documents/akira.goal-misgeneralization-can-turn-means-into-goals.txt) - [行動訓練で望ましく振る舞っても、意図した目標の頑健な汎化は保証されない](../documents/ngo_et_al.behavioral-training-does-not-establish-robust-goal-generalization.txt) [Corpus index](../index.txt)