The More Popular, The Harder to Forget: Adaptive Popularity for LLM Unlearning
TL;DR
The AdaPop (Adaptive Popularity) method is proposed, which combines local token confidence with a per-fact popularity-dependent exponent derived from an external proxy, and automates the forget-retain balance via a dual-ascent controller that adjusts the retain penalty each epoch.
Abstract
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.