越火的知识越难忘，AI 遗忘也得看人气
大模型记住的东西，不是均匀分布的——越多人知道的事实，它记得越牢，想擦掉也越难。但现有的「遗忘」方法对所有知识一视同仁地施压，结果就是冷门知识被暴力擦除，热门知识却还在漏。这篇提出 AdaPop:先拿外部数据(比如维基百科的引用数)给每条知识标个「人气指数」，再按人气高低分配不同的遗忘力度——越火的，擦的力度越大；同时用一个自动控制器每轮调整「忘」和「留」的平衡，不让模型为了忘东西而把正常能力也丢了。在三个模型家族上，AdaPop 在换说法提问时泄露的遗忘内容比竞品少约 5 倍，在对抗性改写下少约 1.6 倍。它不是你明天能用上的——但如果你关心 AI 的「被遗忘权」、模型安全或数据合规，这是目前最诚实的解法：承认遗忘不是开关，而是按人气分级的工程。
📄 原文摘要(英文)
Popular facts are memorised more deeply during pretraining and resist removal longer than rare ones, yet existing LLM unlearning methods apply uniform gradient pressure regardless of training-data frequency. We propose the AdaPop (Adaptive Popularity) method, which combines local token confidence with a per-fact popularity-dependent exponent derived from an external proxy (e.g., Wikidata sitelinks, LLM-as-Judge), and automates the forget-retain balance via a dual-ascent controller that adjusts the retain penalty each epoch. Across three model families and two benchmarks, AdaPop leaks ~5x less forgotten content than competing methods under paraphrased queries and ~1.6x less under adversarial reformulations. We support our analysis with internal metrics: under our method, forget-set hidden states move further from the pre-unlearning model's states than under other methods, while retain-set representations remain close.