AI RESEARCH
LASA: Language-Agnostic Semantic Alignment at the Semantic Bottleneck for LLM Safety
arXiv CS.AI
•
ArXi:2604.12710v1 Announce Type: cross Large language models (LLMs) often nstrate strong safety performance in high-resource languages, yet exhibit severe vulnerabilities when queried in low-resource languages. We attribute this gap to a mismatch between language-agnostic semantic understanding ability and language-dominant safety alignment biased toward high-resource languages.