Design and Evaluation of Large Language Model-Powered Digital Avatar System for Human–Computer Interaction in Adaptive Knowledge System

In plain English

Fifty participants were randomly split between learning from an LLM-driven digital avatar and searching the internet on their own; the avatar group scored significantly higher on knowledge acquisition, F(1,44) = 4.66, p = 0.036, with a modest effect size of about 10 percent of variance explained. The engineering conclusions are the more durable part: the authors identify pipeline orchestration, latency management and safety alignment as the design decisions that determine whether such a system is usable at all. This is a short conference proceedings paper with a small sample, not a large trial.