AI Safety and Alignment
← Researchresearch

AI Safety and Alignment

We researches methods to make large language models safer, more reliable, and better aligned with human intentions.



This research focuses on improving model behavior in Burmese-language interactions. Key risks include harmful responses, misinformation, bias, and the disclosure of personal information. The approach considers local language, cultural context, and the needs of Burmese-speaking communities. Instruction tuning and preference optimization are explored to encourage helpful and appropriate responses.

The evaluation approach combines automated checks with human assessment. Testing examines instruction following, appropriate handling of sensitive requests, and acknowledgement of uncertainty. Human evaluators assess response quality, contextual relevance, and potential harm. Evaluation findings guide further training and identify limitations requiring additional research. This work aims to support the responsible development and use of Burmese language models.