Learning Not To Learn: When Evolving Norms Meet Those That Ought to Be Preserved

Understanding social norms helps large language models (LLMs) provide culturally appropriate responses that can influence human judgments and values. As norms change, LLMs must revise outdated judgments while retaining knowledge that remains valid. Existing continual learning methods do not explicitly distinguish norms that require revision from those that remain valid, risking outdated judgments or unnecessary forgetting. We propose a continual learning framework guided by normative validity, including 1) a mechanism that distinguishes norms requiring revision from those remaining valid, and 2) an update constraint that protects the latter. We construct a dataset spanning cultures and three temporal stages from the International Social Survey Programme (ISSP) and General Social Survey (GSS). Each scenario includes its country and temporal context, together with a reference judgment and explanation. Comparisons across stages identify stable norms for protection while evolving norms and novel norms guide learning. When these objectives conflict, our update rule adjusts the learning direction to reduce interference. Across three stages, relative improvements in final stage macro F1 reach 13.29% on Qwen and 12.94% on Llama. Improvements in retaining previously learned norms that remain valid exceed those of the strongest baselines by 8.16% on Qwen and 23.02% on Llama. Ablations support the contributions of norm status distinctions and update constraints, while gains on unseen scenarios support generalization. By tying knowledge preservation to continued normative validity, our framework supports adaptation to social change while maintaining cultural continuity.