Tag
This paper examines script knowledge distribution in LLM layers, revealing that script recognition happens early while commitment to output script emerges in late layers, highlighting the importance of model depth for multilingual architectures.