LASA: Language-Agnostic Semantic Alignment at the Semantic Bottleneck for LLM Safety

TL;DR

LASA enhances LLM safety across languages by anchoring safety alignment in the model's language-agnostic semantic bottleneck, significantly reducing attack success rates.

Contribution

The paper introduces LASA, a novel method that improves multilingual safety in LLMs by aligning safety understanding within the semantic bottleneck layer.

Findings

LASA reduces attack success rate from 24.7% to 2.8% on LLaMA-3.1-8B-Instruct.

LASA maintains low attack success rates (~3-4%) across multiple models.

Semantic bottleneck is key to language-agnostic safety alignment.

Abstract

Large language models (LLMs) often demonstrate 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. Consistent with this hypothesis, we empirically identify the semantic bottleneck in LLMs, an intermediate layer in which the geometry of model representations is governed primarily by shared semantic content rather than language identity. Building on this observation, we propose Language-Agnostic Semantic Alignment (LASA), which anchors safety alignment directly in semantic bottlenecks. Experiments show that LASA substantially improves safety across all languages: average attack success rate (ASR) drops from 24.7% to 2.8% on LLaMA-3.1-8B-Instruct and remains around 3-4% across Qwen2.5 and Qwen3 Instruct models (7B-32B). Together, our analysis and method offer a representation-level perspective on LLM safety, suggesting that safety alignment requires anchoring safety understanding not in surface text, but in the model's language-agnostic semantic space.