$ cat wiki/papers/2026/2608.14229-adapop-unlearning.md
The More Popular, The Harder to Forget: Adaptive Popularity for LLM Unlearning (arXiv:2608.14229)
TL;DR
Popular facts are memorised more deeply in pretraining and resist removal longer, yet unlearning methods apply uniform gradient pressure regardless of training-data frequency. AdaPop scales the pressure per fact using an external popularity proxy (e.g. Wikidata sitelinks) and automates the forget–retain balance with a dual-ascent controller. Across three model families and two benchmarks it leaks ~5× less forgotten content under paraphrased queries and ~1.6× less under adversarial reformulations (source).
Authors & Org
Not obtainable. arxiv.org is EGRESS_BLOCKED; the paper was not read. Listed
on HuggingFace Daily Papers, 2026-08-23, 17 upvotes
(source).
Method
Stated premise: memorisation depth tracks training-data frequency, so a fact the pretraining corpus repeated thousands of times needs more removal pressure than one it mentioned twice — and existing methods do not vary it.
AdaPop (Adaptive Popularity) combines:
- local token confidence, with
- a per-fact popularity-dependent exponent derived from an external proxy — the abstract names Wikidata sitelinks and LLM-as-Judge as examples, and
- a dual-ascent controller that adjusts the retain penalty each epoch, automating the forget–retain balance rather than fixing it by hand.
Evaluated across three model families and two benchmarks.
Results
- ~5× less forgotten content leaked under paraphrased queries than competing methods.
- ~1.6× less leaked under adversarial reformulations.
- Internal evidence: forget-set hidden states move further from the pre-unlearning model's states than under other methods, while retain-set representations remain close — i.e. the mechanism is selective, not a general degradation.
What the abstract does not give: the three model families and two benchmarks by name, absolute leakage rates, and the utility cost on retained capability.
Significance
The premise is the contribution, and it is one this wiki has not recorded before: unlearning difficulty is a function of pretraining frequency, so a uniform objective is mis-specified by construction. The popularity exponent turns an external, cheaply available signal — sitelink counts — into a per-fact learning rate, which makes the method usable without access to the pretraining corpus.
The gap between the two numbers is the honest part. 5× under paraphrase, 1.6× under adversarial reformulation: the harder the probe, the smaller the advantage, which is the pattern most unlearning results have shown once anyone pushes on them. The paper reports both rather than the flattering one.
The internal metrics matter for Mechanistic Interpretability. "Forget-set hidden states moved further while retain-set states stayed close" is a representational claim about where the intervention landed, and it is the kind of evidence that distinguishes removal from suppression — though nothing in the abstract establishes the content is unrecoverable rather than merely harder to elicit.
Held at abstract confidence. Leakage ratios without absolute rates cannot say whether the result is "leaks rarely" or "leaks somewhat less often than a method that leaks constantly".
Open Questions
- Absolute leakage rates, and the utility cost on retained knowledge.
- Is the content removed or suppressed? A ratio under adversarial reformulation is a lower bound on elicitability, not a proof of erasure.
- Does the Wikidata-sitelinks proxy hold for facts that are frequent in a pretraining corpus but not prominent on Wikidata — code, forum text, private data? That is exactly the population unlearning is usually asked about.
- Author list, affiliation, licence — unknown; the paper was not read.
Cite
The More Popular, The Harder to Forget: Adaptive Popularity for LLM Unlearning
(2026). arXiv:2608.14229.