{
  "id": 5933,
  "url": "https://arxiv.org/abs/2604.12710v2",
  "title": "LASA: Language-Agnostic Semantic Alignment at the Semantic Bottleneck for LLM Safety",
  "summary": "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 representa",
  "authors": "Junxiao Yang, Haoran Liu, Jinzhe Tu, Jiale Cheng, Zhexin Zhang, Shiyao Cui et al.",
  "category": "research",
  "topics": "bias-fairness,safety-alignment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-04-13T15:59:50.000Z",
  "fetched_at": "2026-07-14T16:32:06.469Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/5933",
  "original_url": "https://arxiv.org/abs/2604.12710v2",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}