Cyclic Ablation: Testing Concept Localization against Functional Regeneration in AI

Eduard Kapelko

Safety and controllability are critical for large language models. A central question is whether undesirable behaviors like deception are localized functions that can be removed, o…