A new research paper investigates the universality of refusal directions in AI models across low-resourced languages. The study, which extended the PolyRefuse framework to Belarusian, Bashkir, and Tajik, found that while directions from aligned languages transfer broadly, those from misaligned languages show limited effectiveness. This suggests that safety guarantees may be weaker for users of the least-resourced and most misaligned languages. AI
IMPACT Suggests potential safety gaps for users of low-resourced languages with misaligned AI models.
RANK_REASON Research paper on AI safety and cross-lingual transfer of refusal directions.
Read on llama.cpp — Releases →
- Anastasiia Ivanova
- Bashkir
- Belarusian
- PolyRefuse
- Qwen2.5-14B-Instruct
- Tajik
- Wang et al., 2025
- Yoruba
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →