large language models (LLMs) have demonstrated remarkable performance in
various tasks and gained significant attention. LLMs are also used for local
sequence transduction tasks, including grammatical error correction (GEC) and
formality style transfer, where most tokens in a source te