Can Language Models Handle a Non-Gregorian Calendar? The Case of the Japanese wareki

The study on language models' ability to handle non-Gregorian calendars, particularly the Japanese wareki, underscores a significant gap in current AI capabilities. While some models, including GPT-4o and Deepseek V3, can manage basic calendar conversions, they falter in more complex wareki date calculations. This limitation is attributed to a lack of corpus frequency for Japanese calendar expressions and a prevailing bias towards the Gregorian calendar in the models' training data. The findings are crucial as they point to the necessity of enhancing language models to effectively engage with culturally specific temporal systems, thereby fostering a more inclusive AI landscape.