Anatomy of Unlearning: The Dual Impact of Fact Salience and Model Fine-Tuning
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866914620072001536 |
|---|---|
| author | Borisiuk, Anna Savchenko, Andrey Panchenko, Alexander Tutubalina, Elena |
| author_facet | Borisiuk, Anna Savchenko, Andrey Panchenko, Alexander Tutubalina, Elena |
| contents | Machine Unlearning (MU) enables Large Language Models (LLMs) to remove unsafe or outdated information. However, existing work assumes that all facts are equally forgettable and largely ignores whether the forgotten knowledge originates from pretraining or supervised fine-tuning (SFT). In this paper, we introduce DUET (Dual Unlearning Evaluation across Training Stages), a benchmark of 28.6k Wikidata-derived triplets annotated with fact popularity using Wikipedia link counts and LLM-based salience scores. Our experiments show that pretrained and SFT models respond differently to unlearning. An SFT step on the forget data yields smoother forgetting, more stable tuning, and 10-50% higher retention, while direct unlearning on pretrained models remains unstable and prone to relearning or catastrophic forgetting. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2602_19612 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Anatomy of Unlearning: The Dual Impact of Fact Salience and Model Fine-Tuning Borisiuk, Anna Savchenko, Andrey Panchenko, Alexander Tutubalina, Elena Computation and Language Machine Unlearning (MU) enables Large Language Models (LLMs) to remove unsafe or outdated information. However, existing work assumes that all facts are equally forgettable and largely ignores whether the forgotten knowledge originates from pretraining or supervised fine-tuning (SFT). In this paper, we introduce DUET (Dual Unlearning Evaluation across Training Stages), a benchmark of 28.6k Wikidata-derived triplets annotated with fact popularity using Wikipedia link counts and LLM-based salience scores. Our experiments show that pretrained and SFT models respond differently to unlearning. An SFT step on the forget data yields smoother forgetting, more stable tuning, and 10-50% higher retention, while direct unlearning on pretrained models remains unstable and prone to relearning or catastrophic forgetting. |
| title | Anatomy of Unlearning: The Dual Impact of Fact Salience and Model Fine-Tuning |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2602.19612 |