We present new results from a real-user evaluation of a data-driven approach to learning user-adaptive referring expression generation (REG) policies for spoken dialogue systems. Referring expressions can be difficult to understand in technical domains where users may not know the technical ‘jargon’ names of the domain entities. In such cases, dialogue systems must be able to model the user’s (lexical)domain knowledge and use appropriate referring expressions. We present a reinforcement learning (RL) framework in which the system learns REG policies which can adapt to unknown users online. For real users of such a system, we show that in comparison to an adaptive hand-coded baseline policy, the learned policy performs significantly better, with a 20.8% average increase in adaptation accuracy, 12.6% decrease in time taken, and a 15.1% increase in task completion rate. The learned policy also has a significantly better subjective rating from users. This is because the learned policies adapt online to changing evidence about the user’s domain expertise. We also discuss the issue of evaluation in simulation versus evaluation with real users.