LLMs Fail Braille Translation, Highlighting Accessibility Gaps.

This study investigates the performance of leading Large Language Models (LLMs) in translating between Korean text and Braille, a critical accessibility modality. Despite expectations that these multilingual, instruction-tuned models could generalize to Braille through text representations, the research found consistently poor and unstable outputs. The findings indicate a lack of Braille-aware tokenization and weak alignment between Korean and Braille patterns within current LLMs. This systematic limitation suggests that current general-purpose LLMs struggle with the unique structural constraints of Braille. In contrast, the research demonstrated that supervised fine-tuning of a smaller model, T5-small, on the same dataset resulted in substantial and stable performance gains over zero-shot and prompted LLM baselines. This highlights the effectiveness of targeted, task-specific supervision for improving accessibility features in AI.