Benchmarking Bengali Dialectal Bias: A Multi-Stage Framework Integrating RAG-Based Translation and Human-Augmented RLAIF
This study tested how well large language models handle Bengali dialects using a two-phase framework. Researchers created a gold-labeled dataset by translating standard questions into nine regional dialects and then evaluated 19 open-weight LLMs. The results show severe performance drops in highly divergent dialects, like Chittagong, confirming that model bias is tied to linguistic distance and data exposure.