The paper "The Multilingual Alignment Prism: Aligning Global and Local Preferences to Reduce Harm" addresses the critical issue of aligning AI systems with global and local preferences while minimizing both global and local harms. The authors, Aakanksha, explore the viability of different alignment approaches in a multilingual setting, focusing on balancing dual objectives: addressing and optimizing for a non-homogeneous set of languages and cultural preferences. They collect the first set of human-annotated red-teaming prompts in multiple languages, distinguishing between global and local harm, to understand the reliability of alignment techniques in non-stationary preference distributions across geographies and languages. Key findings include: 1. **New Dataset**: The authors release the first multilingual red-teaming dataset, *Aya Red-teaming*, which includes human-annotated harmful prompts in eight languages, covering a wide range of harm categories. 2. **Evaluation Methods**: They evaluate Direct Preference Optimization (DPO) and Supervised Fine-tuning (SFT) for multilingual safety alignment, demonstrating that DPO outperforms SFT in balancing safety and general performance. 3. **Cross-Harm Transfer**: The study shows that training schemes based on "local" harms can aid in mitigating "global" harms, and vice versa, with a significant reduction in harmful model generations. The paper highlights the importance of language-specific datasets and alignment techniques in achieving effective multilingual safety alignment, providing insights into cross-lingual transfer and novel optimization approaches. The findings underscore the need for a balanced approach to safety and general performance, showing that it is possible to have both in large language models (LLMs) with the right alignment methods and datasets.The paper "The Multilingual Alignment Prism: Aligning Global and Local Preferences to Reduce Harm" addresses the critical issue of aligning AI systems with global and local preferences while minimizing both global and local harms. The authors, Aakanksha, explore the viability of different alignment approaches in a multilingual setting, focusing on balancing dual objectives: addressing and optimizing for a non-homogeneous set of languages and cultural preferences. They collect the first set of human-annotated red-teaming prompts in multiple languages, distinguishing between global and local harm, to understand the reliability of alignment techniques in non-stationary preference distributions across geographies and languages. Key findings include: 1. **New Dataset**: The authors release the first multilingual red-teaming dataset, *Aya Red-teaming*, which includes human-annotated harmful prompts in eight languages, covering a wide range of harm categories. 2. **Evaluation Methods**: They evaluate Direct Preference Optimization (DPO) and Supervised Fine-tuning (SFT) for multilingual safety alignment, demonstrating that DPO outperforms SFT in balancing safety and general performance. 3. **Cross-Harm Transfer**: The study shows that training schemes based on "local" harms can aid in mitigating "global" harms, and vice versa, with a significant reduction in harmful model generations. The paper highlights the importance of language-specific datasets and alignment techniques in achieving effective multilingual safety alignment, providing insights into cross-lingual transfer and novel optimization approaches. The findings underscore the need for a balanced approach to safety and general performance, showing that it is possible to have both in large language models (LLMs) with the right alignment methods and datasets.