Can LLM Safety Be Ensured by Constraining Parameter Regions?

Zongmin Li, Jian Su, Farah Benamara +1

Large language models (LLMs) are often assumed to contain ``safety regions'' -- parameter subsets whose modification directly influences safety behaviors. We conduct a systematic e…