Do spatially-anchored previews stabilize VR geometry editing tasks?

The paper asks whether dialogue-based disambiguation can be "augmented with spatially-anchored graphical previews" when a user gives an ambiguous command to an LLM that edits a VR scene through numerical parameters. In a within-subjects study with 24 participants doing complex geometry editing tasks, the abstract reports that a hybrid of clarification questions and graphical previews "can support better interaction stability with fewer conversation rounds while improving user experience," relative to a condition with no disambiguation. The discussion narrows this: clarification questions alone (CQ) and the combined condition (CQGP) both significantly reduced task progression variability, measured by MSSD, and CQGP "required significantly fewer conversation rounds to complete the task."

The stability result is the one that matters, because the performance result is flat. The maximum closeness score "did not differ significantly across conditions," so disambiguation did not raise the best state users reached. What changed was the path: less erratic progression toward it, and in the hybrid condition fewer conversational turns spent getting there. The paper frames this as text and visual feedback guiding users "towards a smoother task progression trajectory," which it calls novel against 2D interfaces, where disambiguation improved task performance. Its motivation is the "limited user agency due to ambiguous user input" in LLM-assisted scene editing. The previews are a hybrid of 2D UI image previews and spatially-anchored 3D overlays, so the user sees candidate interpretations in the scene rather than reading them in a question.

Against the vault, this is a spatial variant of a text-side result. Which clarifying questions actually improve user satisfaction? concerns what a good clarifying question looks like in text, and here CQ alone already improves stability, which is consistent with clarification helping. The added claim is that a preview anchored in the workspace lets the hybrid condition resolve the ambiguity in fewer rounds. Why do language models fail in gradually revealed conversations? describes the failure that ambiguity produces in text-only multi-turn use, and this study looks at a related setting with a different remedy: show the user the candidate outcome. The paper also states its scope: it "does not study disambiguation on the input side," unlike the speech-and-pointing work on referential disambiguation. That separates it from Can reference resolution work as a language modeling problem?, which works on the input side.

The excerpt leaves several things open. It names three conditions (no disambiguation, CQ, CQGP), so the contribution of previews alone cannot be isolated, and it does not say whether CQGP differs from CQ on stability. It does not state the baseline for the "fewer conversation rounds" comparison, effect sizes, task details or the qualitative themes behind "improving user experience." It gives no evidence on why previews help, and it does not test whether users' intent matured, the concern in How do users actually form intent when prompting AI systems?. A cautious reading is that, in this one task family and sample, showing a candidate result in the scene reduced conversational effort and erratic progression without lifting the best outcome.