ExPerT: New AI framework personalizes LLM responses using keystroke dynamics

Large language models (LLMs) are becoming ubiquitous, but personalization remains limited to static profiles or text-only signals that fail to adapt to a user's changing expertise across different queries. Researchers from KAIST have introduced ExPerT (Expertise Personalization via keystroke and semantic cues), a query-wise personalization framework that dynamically adapts LLM responses based on the user's domain expertise for each specific question. ExPerT uses two components: a semantic-behavioral inference module that interprets both the query text and real-time keystroke dynamics (typing speed, pauses, corrections) via in-context LLM prompting, and an expertise-conditioned response generator that adjusts detail, terminology, and complexity accordingly.

In a user study with 40 participants generating 1,270 queries across multiple domains, ExPerT reduced expertise inference error by 65.7% compared to the strongest baseline (mean absolute error of 0.398 vs. 1.162). Response satisfaction improved by 17.52%, jumping from 3.71 to 4.36 on a 5-point Likert scale. The system effectively distinguishes between a user who types slowly and pauses frequently—indicating lower expertise—versus a fast, fluent typist with domain-specific abbreviations. Accepted as a long paper at ACL 2026, ExPerT opens the door for privacy-sensitive, real-time personalization that doesn't require explicit user profiles or history.