Stabilizing language models under continual learning via condition-anchored distillation

Continual adaptation of language models can change their output distribution on prompts learned earlier, while retaining every old prompt-answer pair may be undesirable or impossible.