Metraipolis › archive › dc_bf3087dd877b
The More Popular, The Harder to Forget: Adaptive Popularity for LLM Unlearning — arXiv:2608.14229
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2608.14229
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-19T02:27:56.656Z
- SHA-256 of the archived text
4dfb45b3cbc3936a7fbe1b040ac5e363067de18d5a07c7c45eefef9415529fd2- Stable URL of this page
https://metraipolis.com/documento/dc_bf3087dd877b
How to cite
The More Popular, The Harder to Forget: Adaptive Popularity for LLM Unlearning — arXiv:2608.14229 — arXiv (en), https://arxiv.org/abs/2608.14229 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_bf3087dd877b (Metraipolis archive dc_bf3087dd877b, sha256 4dfb45b3cbc3936a7fbe1b040ac5e363067de18d5a07c7c45eefef9415529fd2)
BibTeX
@misc{dcbf3087dd877b,
title = {The More Popular, The Harder to Forget: Adaptive Popularity for LLM Unlearning — arXiv:2608.14229},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2608.14229},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_bf3087dd877b; sha256 4dfb45b3cbc3936a7fbe1b040ac5e363067de18d5a07c7c45eefef9415529fd2},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.14229. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
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. Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-08-17 in cs.CL). Categorie arXiv: Computation and Language (cs.CL). Pagina del lavoro: https://arxiv.org/abs/2608.14229 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-19T02:20:31.908Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-08-17?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).