An AI research team has built a system that adjusts how a language model explains things by watching not just what you type, but how you type it.
The framework, called ExPerT, combines two signals to estimate a user's expertise on any given query: the semantic content of the question itself, and keystroke dynamics — the timing and rhythm of how the question was typed. Those signals feed an inference module that gauges domain familiarity on a per-query basis, then conditions the model's response on that estimate, tuning the level of detail, terminology, and conceptual depth accordingly. A user study with 40 participants and 1,270 queries found ExPerT cut expertise inference error by 65.7% compared to the best existing baseline (mean absolute error 0.398 versus 1.162) and lifted response satisfaction scores by 17.52%, moving from 3.71 to 4.36 on a five-point scale.
Most personalization systems today rely on static user profiles or text signals alone — they assume your expertise is fixed and uniform, which it isn't. ExPerT's query-wise approach recognizes that the same person can be a novice in one domain and an expert in another, sometimes within the same session. That's a meaningful shift in how personalization could work at the application layer.
The catch is that keystroke data requires instrumentation that most deployed LLM interfaces don't have, which means ExPerT is less a product than a proof of concept — and the 40-person study is small enough to warrant caution before reading too much into the satisfaction numbers.