Quantization-Robust LLM Unlearning via Low-Rank Adaptation

João Vitor Boer Abitante 1,2 , Joana Meneguzzo Pasquali 1 , Luan Fonseca Garcia 2 , Ewerton de Oliveira 3 , Thomas da Silva Paula 3 , Rodrigo C. Barros 1,4 , Lucas S. Kupssinskü 1 · 2026

Read Quantization-Robust LLM Unlearning via Low-Rank Adaptation by João Vitor Boer Abitante 1,2 , Joana Meneguzzo Pasquali 1 , Luan Fonseca Garcia 2 , Ewerton de Oliveira 3 , Thomas da Silva Paula 3 , Rodrigo C. Barros 1,4 , Lucas S. Kupssinskü 1 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-13151.txt.

Description

Removing knowledge from a language model is hard enough, and quantising the model for deployment can undo the removal, reviving forgotten behaviour. João Vitor Boer Abitante and six co-authors show full-parameter fine-tuning usually makes changes too small to survive 4-bit quantisation. Their fix concentrates unlearning into low-rank adapters while freezing the base model, so the effective update persists after quantisation.

On Llama-2-7B with the MUSE dataset, low-rank unlearning lifts 4-bit utility by up to 7.93 points on BOOKS and improves NEWS utility too, while cutting privacy leakage sharply: for one method on BOOKS, PrivLeak moves from minus 25.68 to minus 5.86, near the ideal zero, with forgetting measures staying near zero. For deployed unlearning, the lesson is to unlearn where quantisation cannot erase it.

Catalog record

CREATORJoão Vitor Boer Abitante 1,2 , Joana Meneguzzo Pasquali 1 , Luan Fonseca Garcia 2 , Ewerton de Oliveira 3 , Thomas da Silva Paula 3 , Rodrigo C. Barros 1,4 , Lucas S. Kupssinskü 1
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE32 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-13151

Browse the full library · About the Tower of Babel