The Current State of AI for Conversion of Mathematics to and from Braille
Moritz Groß, Duncan Soiffer, Neil Soiffer.
Accepted at ICCHP 2026.
There are many different braille codes for mathematics around the world. Most have limited or no support for reading online mathematical content or for entry and conversion from those codes to math notation. Meanwhile, AI has had tremendous success with language translation. We investigate the use of AI to translate to and from two of the most widely used braille codes, testing several different techniques across leading commercial and open-weight large language models. In our best configuration, accuracy reached 92% for both Nemeth-to-MathML and MathML-to-Nemeth conversion using thousands of examples. By contrast, results for translating to and from UEB were lower. Although there are some caveats to these numbers, it appears that translation of technical braille is nearing viability.
(The project grew out of my work on MathCAT and my university side job at Math4VIP)