kanji

YOMI-Bench is a benchmark designed to evaluate large language models on kanji reading and phonological understanding in Japanese. It consists of four tasks and reveals that even Japanese-specific models perform poorly on generation tasks requiring kanji reading knowledge.