Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications

Evidence card

Claim

Safety-relevant behavior can be brittle under sparse parameter or low-rank changes in studied settings.

Evidence level

Experimentally observed

Source

https://arxiv.org/abs/2402.05162

Publication date

2024-02-07

Authors or institution

Boyi Wei, Kaixuan Huang, Yangsibo Huang, Tinghao Xie, Xiangyu Qi, Mengzhou Xia, Prateek Mittal, Mengdi Wang, Peter Henderson

System tested

Safety-critical regions and low-rank/pruning modifications in tested LLMs.

Limitations

Model families, tasks, and safety metrics define the scope.

What the evidence does show

Safety-relevant behavior can be brittle under sparse parameter or low-rank changes in studied settings.

What the evidence does not show

That every compression or low-rank change destroys alignment.

Date last reviewed in UTC

2026-06-26T00:00:00Z