Conventional unlearning methods typically adopt a direct-training approach, which optimizes a forget loss on undesired content and a retain loss on preserved content. However, due to the unbounded loss functions and insufficient retain set coverage, they often suffer from catastrophic forgetting and unstable optimization. This chapter introduces an alternative framework based on external-output construction, where unlearned predictions are first generated through auxiliary methods and then distilled into the target model. We detail two representative strategies: (1) Unlearning from Logit Difference, which constructs unlearned outputs by subtracting an assistant model’s logits-trained to memorize forget content and forget retain content-from the target model’s logits; and (2) Who’s Harry Potter, which constructs unlearned predictions via entity name replacement in the input. Both methods enable controllable and effective unlearning without directly optimizing the model for the forget and retain losses. Empirical results on benchmark datasets, including TOFU for synthetic knowledge and WPU for real-world knowledge, demonstrate that these approaches achieve a better balance between unlearning and preserving general knowledge, while also offering more stable training dynamics.