Late-Layer Script Commitment in Multilingual LLMs

Logit-lens and probing analysis show that LLMs encode input and output scripts in early layers but only commit to the target script in the final layers. Intermediate representations frequently default to Latin text.