Tag
This paper investigates how dialectal biases persist and accumulate throughout the entire language modeling pipeline, from tokenization to inference, showing that disparities are encoded at every step.