If an AI just asks you good questions, does that keep a task on track as well as showing you a preview too?

This explores whether asking clarifying questions alone can make a task progress as steadily as asking questions and also showing a preview of the likely result. The corpus can't answer that head-to-head. The one study on this pairing, a 24-person VR geometry-editing experiment, found that Do spatially-anchored previews stabilize VR geometry editing tasks? questions combined with 3D previews anchored in the scene reduced variability in task progress and needed fewer conversation rounds. But it reports that result against no disambiguation, so it never shows what questions alone would have done.

What the study does show is the kind of gain at stake. Peak performance was the same across all conditions. The combination didn't make the best runs better; it made the unlucky runs less unlucky. So 'matching' here means matching on consistency, not on the best case, and a small study of this size can't say much about how the two approaches compare on that.

There's a plausible reason previews could add something that questions can't, though this is my inference, not a tested finding. Which clarifying questions actually improve user satisfaction? finds that questions work best when they target a concrete gap ('What type of monitor?') and when users can foresee how answering will improve the results. A preview is that foresight made literal. The user sees a candidate result and corrects it, instead of putting a spatial intention into words and hoping the question captured it.

'Questions alone' also isn't one fixed condition, because question quality varies a lot. Can models learn to ask genuinely useful clarifying questions? breaks quality into attributes like clarity, relevance and specificity. Models trained on those attributes ask better questions, and the gain is largest in clinical reasoning, where the right question changes the decision. How can models select the most informative question to ask? goes further and simulates possible futures, then picks the question whose answers would cut uncertainty the most. That is a preview done internally: the model imagines the outcomes so the user doesn't have to see them.

So a well-chosen question may close part of the gap, especially where words describe the task well. Previews probably matter most where the thing being described is visual or spatial and hard to say out loud. The experiment the corpus lacks is a questions-only arm run next to the questions-plus-previews arm.