The More Popular, The Harder to Forget: Adaptive Popularity for LLM Unlearning

Summary

Popular facts are memorized 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 AdaPop (Adaptive Popularity), 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. Internal metrics support our analysis: 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.