Saved in:
Bibliographic Details
Main Authors: Li, Zongmin, Su, Jian, Benamara, Farah, Sun, Aixin
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2602.17696
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917283712991232
author Li, Zongmin
Su, Jian
Benamara, Farah
Sun, Aixin
author_facet Li, Zongmin
Su, Jian
Benamara, Farah
Sun, Aixin
contents Large language models (LLMs) are often assumed to contain ``safety regions'' -- parameter subsets whose modification directly influences safety behaviors. We conduct a systematic evaluation of four safety region identification methods spanning different parameter granularities, from individual weights to entire Transformer layers, across four families of backbone LLMs with varying sizes. Using ten safety identification datasets, we find that the identified safety regions exhibit only low to moderate overlap, as measured by IoU. The overlap drops significantly when the safety regions are further refined using utility datasets (\ie non-harmful queries). These results suggest that current techniques fail to reliably identify a stable, dataset-agnostic safety region.
format Preprint
id arxiv_https___arxiv_org_abs_2602_17696
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Can LLM Safety Be Ensured by Constraining Parameter Regions?
Li, Zongmin
Su, Jian
Benamara, Farah
Sun, Aixin
Machine Learning
Artificial Intelligence
Large language models (LLMs) are often assumed to contain ``safety regions'' -- parameter subsets whose modification directly influences safety behaviors. We conduct a systematic evaluation of four safety region identification methods spanning different parameter granularities, from individual weights to entire Transformer layers, across four families of backbone LLMs with varying sizes. Using ten safety identification datasets, we find that the identified safety regions exhibit only low to moderate overlap, as measured by IoU. The overlap drops significantly when the safety regions are further refined using utility datasets (\ie non-harmful queries). These results suggest that current techniques fail to reliably identify a stable, dataset-agnostic safety region.
title Can LLM Safety Be Ensured by Constraining Parameter Regions?
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2602.17696