Long Recent Context Drives Personalization Gains in Predictive Typing

We study how to improve predictive typing by personalizing a pretrained large language model for each user. Using time-ordered email data from 44 users, we compare two statistical language-model baselines (a 12-gram model and a mixture of that model with a user-specific Prediction by Partial Matching model) with the 13-billion-parameter Open Pre-trained Transformer model (OPT-13B) using long recent left context, one-time parameter adaptation, and progressive adaptation. We evaluated different methods: a non-adapted model, full-parameter fine-tuning, one-time low-rank adaptation (LoRA), and progressive LoRA that updates every 128 evaluation phrases and base-initialized progressive LoRA, which starts from a non-personalized OPT-13B without any adaptation from the seed set. Our results showed that increasing the left-context limit from 0 to 512 characters reduced mean per-character perplexity from 2.46 to 2.05 for non-adapted OPT-13B. Also, Non-adapted OPT-13B outperformed both statistical baselines, and Progressive LoRA achieved the lowest mean perplexity among all methods. Parameter adaptation produced only small improvements in keystroke savings, showing that lower perplexity does not always lead to a comparable reduction in typing effort.