MBZUAI
Researcher
Oct 2024 – Dec 2025
Supervisor: Prof. Fajri Koto
- Developed
AMIR-GRPO, a GRPO extension that injects implicit preference regularization to improve reasoning performance without additional annotation cost. - Assessed the reliability, robustness, and human alignment of automated evaluation metrics for machine translation and text summarization.
- Investigated multilingual LLMs on culturally grounded procedural texts, highlighting gaps, biases, and limitations in cross-cultural language understanding.
