Personalization of Large Language Models for Predictive Typing

We study improving predictive typing via per‑user personalization of a large language model (LLM). We evaluate personalization in a time-ordered setting, using a user's earlier messages to adapt the LLM to improve predictions on later messages. We compare a static n-gram and a static LLM against an LLM adapted via full-parameter fine-tuning, one-time low-rank adaptation (LoRA), and multiple progressive LoRA updates. In a simulated predictive keyboard, the static LLM increased keystroke savings by 13\% compared to the n-gram model. Adapting the parameters of the LLM only offered a small additional increase of 0.5\%. We found this was due to the strong personalization provided by conditioning the LLM on the 512 previous characters of a user's sent emails. Without this conditioning, perplexity increased sharply by 20\% relative.