Pith. sign in

REVIEW 3 major objections 7 minor 55 references

The paper argues that combining clarifying questions with spatially anchored graphical previews makes LLM-assisted VR geometry editing steadier, shorter in conversation, and more positively rated than clarification alone or no disambiguatio

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Combining clarification questions with in-VR graphical previews in an LLM-assisted geometry editor reduces conversation rounds and steadies task progress compared with clarification alone.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Solid, transparent VR/LLM user study whose headline efficiency claim is weakened by a conversation-round counting confound; still worth refereeing. the 3 major comments →

arxiv 2607.23201 v1 pith:UCOUX56C submitted 2026-07-25 cs.HC

Beyond Conversations: Spatially-Anchored Previews for Intent Disambiguation in LLM-Assisted Geometry Editing in Virtual Reality

Junlong Chen , Amr Gomaa , Jens Grubert , Per Ola Kristensson This is my paper
classification cs.HC
keywords virtual realitylarge language modelsintent disambiguationclarification questionsgraphical previewsparameter-driven editinggeometry editinguser experience
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that when people edit 3D objects in virtual reality by giving spoken commands to a large language model assistant, the ambiguity that arises—users do not know the numerical parameters defining the object—is best resolved by combining a clarifying question with a spatially anchored graphical preview of what each possible edit would do. In a within-subjects study of 24 participants editing three parameterized geometries, the full hybrid condition produced significantly smoother progress toward the target, significantly fewer conversation rounds, and higher user-experience ratings than the condition with only clarifying questions. Peak closeness to the target was not significantly different across conditions, so the benefit is not that people finish closer, but that they get there with less floundering and less dialogue. The authors take this as empirical evidence for embedding disambiguation directly in the spatial scene rather than relying on chat alone, and they derive design guidelines for future LLM-assisted VR/AR editing systems.

Core claim

On the paper's own terms, the central claim is that in an LLM-assisted, parameter-driven VR editing workflow, augmenting dialogue-based clarification questions with in-place graphical previews yields a smoother, more efficient, and more positively experienced interaction than either no disambiguation or clarification questions alone. Quantitatively, the hybrid condition had significantly lower moment-to-moment variability in the normalized closeness score, significantly fewer conversation rounds (about 8.2 versus 12.3 with no support and 14.3 with clarification only), and significantly higher overall user-experience ratings than clarification alone. Maximum closeness scores did not differ si

What carries the argument

The central mechanism is DisambVR, a VR system in which a large language model receives the user's spoken command, the current parameter values, and the edit history, then decides whether the command is ambiguous. If it is, the system emits a clarification question plus candidate parameter edits, each shown with a 2D preview image and an in-situ semi-transparent 3D overlay on the object when hovered. These spatially anchored previews let users evaluate the consequence of an edit without knowing the geometry's parameter names. The quantitative workhorse is the normalized closeness score—one minus the mean normalized parameter error across the four parameters—with mean-square successive differ

Load-bearing premise

The load-bearing premise, stated in the paper's limitations, is that interactions in immersive editing can be represented as changes in latent numerical parameters; if real editing intent is not parameter-shaped, the smoother-progress and reduced-dialogue benefits of the hybrid preview condition may not transfer beyond the four-parameter task geometry.

What would settle it

Run the same three conditions on a free-form mesh-sculpting task that cannot be reduced to four parameters: if the hybrid condition no longer yields significantly fewer conversation rounds and lower closeness-score variability, the parameter-driven premise is doing the work. A cheaper check is the paper's own powered replication—an equivalence study with roughly 56–72 participants that fails to reproduce the variability and round-count differences, or that shows peak closeness differs by more than the pre-specified ±0.03 bound.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • LLM-assisted VR/AR editors should include spatially anchored previews of candidate edits alongside clarifying text, because the hybrid cut conversational rounds from about 12.3 (no support) and 14.3 (clarification only) to 8.2.
  • Evaluations of such systems should track progression variability and dialogue cost, not just final accuracy; peak closeness was not significantly different, so endpoint-only metrics would miss the measured benefit.
  • The design guidelines directly follow: support coarse-to-fine edit flows, make disambiguation depth adaptive to user expertise, balance guidance with autonomous exploration, and phrase LLM responses in accessible language to avoid introducing new ambiguity.
  • Qualitative feedback indicates previews mainly helped because users could see the consequences of an edit before committing, reducing back-tracking, though some users found too much guidance reduced their sense of agency.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • An extension the paper leaves open: a preview-only condition might match or beat the hybrid for fluent users, since the hybrid still adds text and cognitive load; the additive design cannot rule this out.
  • If the latent-parameter premise transfers to object layout, lighting, or material editing, spatially anchored previews could become a general pattern for LLM control of any parameterized scene, but the paper only demonstrates this on four-parameter shapes.
  • The qualitative finding that some users felt reduced agency suggests an adaptive disambiguation depth—fewer previews for experts, more for novices—could improve on the fixed hybrid, but such an adaptive system is not tested here.
  • Because the study manufactures ambiguity by hiding parameter names, an alternative path is to teach users the parameter vocabulary directly, which might reduce the need for previews at lower implementation cost than building in-situ overlays.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. This paper presents DisambVR, a VR system in which GPT-4o interprets spoken commands to edit parameterized 3D geometries. A within-subjects study (N=24) compared three conditions: no disambiguation (NONE), clarification questions (CQ), and clarification questions plus spatially-anchored graphical previews (CQGP). The main reported findings are that CQGP reduces variability in closeness score over time (MSSD) relative to NONE, reduces the number of conversation rounds relative to both NONE and CQ, and improves UEQ-S user experience ratings relative to CQ, while peak closeness does not differ across conditions. The paper also presents qualitative themes and derives design recommendations for LLM-assisted immersive editing.

Significance. If the findings hold, this is a timely and useful empirical contribution to HCI/XR+LLM research. The study is carefully constructed: within-subjects, counterbalanced, with a concrete task, logged interaction data, and a priori equivalence testing with a post-hoc sensitivity analysis. The authors are also appropriately cautious about null peak-performance results and explicitly acknowledge several limitations. However, the central efficiency claim about conversation rounds is currently compromised by an undefined, modality-confounded metric. Since this metric is load-bearing for contribution C1 and for the paper's framing, the contribution is not yet fully substantiated. The missing GP-only condition is acknowledged and is a design-scope limitation rather than an internal error.

major comments (3)
  1. [Section 4.2 vs Sections 3.1/3.4] The headline 'fewer conversation rounds' result is not interpretable as reported. The paper never defines what counts as a 'conversation round.' In the CQ condition, Section 3.1 states users 'respond verbally' to clarification questions; in the CQGP condition, Section 3.4 states users can 'double-click on the option to confirm the edit.' If a round is counted as a speech-to-response exchange, CQGP will produce fewer rounds whenever a user clicks instead of speaking, independent of disambiguation quality. The reported means (NONE 12.3, CQ 14.3, CQGP 8.23) therefore cannot support the efficiency claim without knowing whether click confirmations were counted as rounds. Please define the metric, report how many confirmations were click-based versus speech-based, and re-analyze with click confirmations counted equivalently or restrict the comparison to speech-initiated rounds. The UEQ-S advan
  2. [Section 4.2, Eq. (3)] The MSSD calculation is internally inconsistent. The text says xi is recorded 'each time a change in closeness score occurs' and n is 'the total number of different closeness scores in one trial'; the next sentence says MSSD values are 'obtained at a same sampling rate of 5 Hz.' These two descriptions are incompatible. Additionally, since participants could end a trial before the 4-minute limit, it is unclear whether MSSD is computed on fixed-length resampled series or on variable-length event-based series. Because MSSD is one of the main quantitative claims (lower variability in CQGP), the exact sequence construction and the role of the 5 Hz rate need clarification to confirm that the comparison is not affected by condition-dependent numbers of edits or interpolation artifacts.
  3. [Section 4.2 and Section 5.2] The interpretation of the MSSD result as evidence for the hybrid condition should be sharpened. The pairwise comparison shows significantly lower MSSD for CQ and CQGP relative to NONE, but the CQ-versus-CQGP difference is not significant (T=1.28, p=.21). Thus the data support the claim that disambiguation support stabilizes progression, but they do not support the stronger reading that graphical previews specifically add stabilization beyond clarification questions alone. Please align the wording in C1 and in Section 5.2 with this statistical outcome.
minor comments (7)
  1. [Abstract] The phrase 'VR simulate scenarios' should be 'VR-simulated scenarios' or 'VR simulating scenarios.'
  2. [Figure 5] The legend uses 'GP' for Gaussian Process while 'GP' is also the paper's abbreviation for graphical previews. Rename the legend (e.g., 'Gaussian Process mean') to avoid ambiguity.
  3. [Section 3.4] The statement that condition and geometry sequences were 'fully counterbalanced' across 24 participants is imprecise: with 3 conditions and 3 geometries there are 6 condition orders and 6 geometry orders, so not all 36 pairings can be fully balanced with N=24. Please describe the exact counterbalancing scheme.
  4. [Section 4.1] The TOST equivalence bound of ±0.03 is stated to be fixed a priori, but no justification is given for this value on the closeness-score scale. A brief rationale or a reference to a convention would help readers assess the sensitivity analysis.
  5. [Section 4.4] The thematic analysis does not report the number of coders, inter-rater reliability, or whether the analysis was conducted by one or multiple researchers. Please add these details.
  6. [Throughout] There are formatting/spacing issues such as 'theCQGPcondition' and 'theCQcondition'; also 'two one sided t-tests' should be 'two one-sided t-tests.'
  7. [Section 5.6] The latent-parameter assumption is acknowledged as a limitation, which is good. Consider adding a sentence on which classes of VR editing tasks (e.g., layout, lighting) plausibly fit this assumption, to help readers gauge generalizability.

Circularity Check

0 steps flagged

No circular derivation: the claims are empirical measurements; only minor non-load-bearing self-citations.

full rationale

The paper's contributions (C1) are empirical comparisons of logged interaction data (MSSD, conversation rounds, closeness scores) and questionnaire ratings across three conditions. The closeness metric (Eqs. 1-2) is a transparent normalization adapted from external prior work [23,24], not a parameter fitted to the study's outcomes and then re-derived as a prediction. MSSD (Eq. 3) is a standard successive-difference statistic computed directly from logged closeness series. The Gaussian-process curves in Fig. 5 are descriptive smoothing, not the basis of any significance test; the inferential claims use Friedman/Conover tests on observed MSSD and round counts. The only self-citations ([7], [8]) appear in the introduction and related work to motivate the research gap and to scope the work; they do not supply the measured outcomes, and no uniqueness theorem or ansatz is imported from them. The stated limitations in Section 5.6, such as "interactions in immersive environments can be represented as a change in latent parameters" and the absence of a GP-only condition, are external-validity caveats rather than definitional loops. The skeptic's concern that CQGP's double-click confirmation may lower the conversation-round count relative to CQ is a plausible measurement confound, but the paper never defines "conversation round" in a way that makes that result true by construction, so it does not meet the circularity threshold. At most, there is a minor non-load-bearing self-citation, which is why the score is 2 rather than 0.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 1 invented entities

The central quantitative result is an empirical measurement, so there are no fitted model parameters in the physics sense. The free parameters above are methodological choices that affect the analysis. The main load-bearing assumptions are the parameter-shaped view of scene editing and the reliability of the LLM's ambiguity detection. DisambVR is an invented artifact but with no public code or data, hence no independent evidence.

free parameters (3)
  • TOST equivalence bound = ±0.03
    Chosen by hand and fixed before data collection (Section 4.1); used as the smallest meaningful effect for equivalence testing, not fitted to the observed data.
  • Closeness metric parameter weights = 1/4 equal weights across four normalized parameter errors
    Equation (2) aggregates four Δp values with equal weight; this weighting is a design choice, not derived from data or from any independent benchmark.
  • MSSD sampling rate = 5 Hz
    Section 4.2 states MSSD values were obtained at a 5 Hz sampling rate; the choice affects the variability metric and is not derived from a theoretical requirement.
axioms (5)
  • domain assumption Interactions in immersive environments can be represented as changes in latent parameters
    The entire task and system design depend on this; explicitly acknowledged in Section 5.6 ('This work also hypothesizes that interactions in immersive environments can be represented as a change in latent parameters').
  • domain assumption GPT-4o reliably detects ambiguity and emits structured, parseable outputs, with hallucination guards sufficient for the study
    Section 3.1 relies on the LLM to classify input as ambiguous or not and to output structured segments; no independent accuracy benchmark of the LLM behavior is reported.
  • domain assumption Counterbalancing and practice trials adequately control learning, order, and geometry-difficulty effects
    Section 3.4 describes counterbalancing and practice; this is a standard methodological assumption, though inherent difficulty differences across geometries are acknowledged rather than eliminated.
  • domain assumption NASA-TLX and UEQ-S are valid instruments for perceived workload and UX in this VR context
    Used in Sections 3.4 and 4.3; standard instruments but their measurement properties in immersive LLM-assisted editing are not independently validated here.
  • standard math Statistical test assumptions (normality where claimed, sphericity correction, distribution-free tests) are appropriate
    The authors check Shapiro-Wilk, Mauchly's test, and use Greenhouse-Geisser/Friedman tests where required; the assumptions are standard and partially verified.
invented entities (1)
  • DisambVR no independent evidence
    purpose: Testbed system for comparing NONE, CQ, and CQGP disambiguation conditions in LLM-assisted parameter-driven VR editing.
    DisambVR is an original system artifact, but no code, parameter files, prompts, or commit hash are publicly released; its behavior is only evidenced by the in-study measurements, so it lacks an independent falsifiable handle outside the paper.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Conversations: Spatially-Anchored Previews for Intent Disambiguation in LLM-Assisted Geometry Editing in Virtual Reality." pith.science (2026). https://pith.science/paper/UCOUX56C

@misc{pith2026260723201,
  author       = {Pith},
  title        = {Pith review of: Beyond Conversations: Spatially-Anchored Previews for Intent Disambiguation in LLM-Assisted Geometry Editing in Virtual Reality},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UCOUX56C}},
  note         = {Machine review of arXiv:2607.23201}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

User intent disambiguation remains a key challenge in intelligent interactive systems. While they have been widely studied in dialogue systems in 2D interfaces, research on how intent disambiguation could be incorporated within Large Language Model (LLM) assisted editing workflows in immersive environments remains limited. Recent advances in LLMs create opportunities to leverage the immersive nature of virtual and augmented reality (VR/AR) environments to provide better disambiguation support. In this paper, we evaluate how traditional dialogue-based disambiguation can be augmented with spatially-anchored graphical previews to resolve ambiguous user commands in LLM-assisted parameter-driven editing workflows. A within-subjects study in which 24 participants completed complex geometry editing tasks in VR simulate scenarios where VR scenes are controlled by numerical parameters. Compared with the condition where disambiguation is not available, quantitative metrics and qualitative feedback indicate that a hybrid approach which combines clarification questions and graphical previews can support better interaction stability with fewer conversation rounds while improving user experience. These findings provide empirical evidence on the effectiveness of disambiguation methods in LLM-assisted editing of parameter-driven immersive scenes and inform design guidelines for future integration of LLMs in advanced VR/AR systems.

Figures

Figures reproduced from arXiv: 2607.23201 by Amr Gomaa, Jens Grubert, Junlong Chen, Per Ola Kristensson.

Figure 1
Figure 1. Figure 1: Examples of a user editing a polyhedral torus without disambiguation support (left), the user editing a superellipsoid with the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Workflow of DisambVR. The system feeds the system instructions prompt, historical input messages, the currently-recognized speech input, parameter edit history, and current parameters to GPT-4o. If the LLM decides that the input is ambiguous and can only be resolved with a low confidence, the clarification question and information for the graphical preview options are extracted. Otherwise if the input is n… view at source ↗
Figure 3
Figure 3. Figure 3: An example of the DisambVR system UI. The user toggles the speech recognition button on and asks the system to make the object bigger. The system detects ambiguity in the input and asks the user to clarify which parameter to increase, followed by three options with graphical previews, a heading, and a brief description of the edit to be made. The user is not satisfied with either of the options, and asks h… view at source ↗
Figure 4
Figure 4. Figure 4: Examples of the polyhedral torus (left), superellipsoid (middle), and supertoroid (right). The values of the four parameters [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Progression lines showing the closeness score with respect to elapsed time in each trial conducted by each participant under [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Violin plots for quantitative measures in the study where significant differences were found. Red dots indicate the arithmetic [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

55 extracted references · 1 canonical work pages

  1. [1]

    Andrea Alfieri, Ralf Wolter, and Seyyed Hadi Hashemi. 2022. Intent Disambiguation for Task-oriented Dialogue Systems. InProceedings of the 31st ACM International Conference on Information & Knowledge Management. 5079–5080

    2022

  2. [2]

    Gaurav Arora, Shreya Jain, and Srujana Merugu. 2024. Intent detection in the age of LLMs. InProceedings of the 2024 conference on empirical methods in natural language processing: Industry track. 1559–1570

    2024

  3. [3]

    Jinheon Baek, Nirupama Chandrasekaran, Silviu Cucerzan, Allen Herring, and Sujay Kumar Jauhar. 2024. Knowledge-Augmented Large Language Models for Personalized Contextual Query Suggestion. InProceedings of the ACM Web Conference 2024. 3355–3366

    2024

  4. [4]

    Put-that-there

    Richard A Bolt. 1980. “Put-that-there” Voice and gesture at the graphics interface. InProceedings of the 7th annual conference on Computer graphics and interactive techniques. 262–270

    1980

  5. [5]

    Chun-Peng Chang, Alain Pagani, and Didier Stricker. 2024. 3D Spatial Understanding in MLLMs: Disambiguation and Evaluation.arXiv preprint arXiv:2412.06613(2024)

    Pith/arXiv arXiv 2024

  6. [6]

    Di Laura Chen, Ravin Balakrishnan, and Tovi Grossman. 2020. Disambiguation Techniques for Freehand Object Manipulations in Virtual Reality. In 2020 IEEE Conference on Virtual Reality and 3D User Interfaces (VR). IEEE, 285–292

    2020

  7. [7]

    Junlong Chen, Jens Grubert, and Per Ola Kristensson. 2025. A Comparative Study of Speech-and-Pointing and Disocclusion Mini-Map Techniques for Object Selection in Virtual Reality. InProceedings of the 2025 ACM Symposium on Spatial User Interaction. 1–12

    2025

  8. [8]

    Junlong Chen, Jens Grubert, and Per Ola Kristensson. 2025. Analyzing Multimodal Interaction Strategies for LLM-Assisted Manipulation of 3D Scenes. In2025 IEEE Conference Virtual Reality and 3D User Interfaces (VR). IEEE, 206–216

    2025

  9. [9]

    Jiangong Chen, Xiaoyi Wu, Tian Lan, and Bin Li. 2025. LLMER: Crafting interactive extended reality worlds with json data generated by large language models.IEEE Transactions on Visualization and Computer Graphics(2025)

    2025

  10. [10]

    Xing Yu Chen and Geetha Kanaparan. 2023. ARID—An Augmented Reality Mobile Application for Interior Design. InResilient and Responsible Smart Cities: The Path to Future Resiliency. Springer, 3–17

    2023

  11. [11]

    Yizhou Chi, Jessy Lin, Kevin Lin, and Dan Klein. 2024. CLARINET: Augmenting Language Models to Ask Clarification Questions for Retrieval. arXiv preprint arXiv:2405.15784(2024)

    Pith/arXiv arXiv 2024

  12. [12]

    Fernanda De La Torre, Cathy Mengying Fang, Han Huang, Andrzej Banburski-Fahey, Judith Amores Fernandez, and Jaron Lanier. 2024. LLMR: Real-time Prompting of Interactive Worlds using Large Language Models. InProceedings of the CHI Conference on Human Factors in Computing Systems. 1–22

    2024

  13. [13]

    Kaustubh D Dhole. 2020. Resolving Intent Ambiguities by Retrieving Discriminative Clarifying Questions.arXiv preprint arXiv:2008.07559(2020)

    Pith/arXiv arXiv 2020

  14. [14]

    Siamak Farshidi, Kiyan Rezaee, Sara Mazaheri, Amir Hossein Rahimi, Ali Dadashzadeh, Morteza Ziabakhsh, Sadegh Eskandari, and Slinger Jansen

  15. [15]

    Zhiying Fu, Rui Xu, Shiqing Xin, Shuangmin Chen, Changhe Tu, Chenglei Yang, and Lin Lu. 2022. EasyVRModeling: Easily Create 3D Models by an Immersive VR System.Proceedings of the ACM on Computer Graphics and Interactive Techniques5, 1 (2022), 1–14

    2022

  16. [16]

    Runze Gan, Jiaming Liang, Bashar I Ahmad, and Simon Godsill. 2020. Modeling intent and destination prediction within a Bayesian framework: Predictive touch as a usecase.Data-Centric Engineering1 (2020), e12

    2020

  17. [17]

    Kenny Jones, and Daniel Ritchie

    Aditya Ganeshan, Ryan Huang, Xianghao Xu, R. Kenny Jones, and Daniel Ritchie. 2024. ParSEL: Parameterized Shape Editing with Language.ACM Trans. Graph.43, 6, Article 197 (Nov. 2024), 14 pages. doi:10.1145/3687922

    work page doi:10.1145/3687922 2024

  18. [18]

    Daniele Giunchi, Nels Numan, Elia Gatti, and Anthony Steed. 2024. DreamCodeVR: Towards Democratizing Behavior Design in Virtual Reality with Speech-Driven Programming. In2024 IEEE Conference Virtual Reality and 3D User Interfaces (VR). IEEE, 579–589

    2024

  19. [19]

    Google. 2023. Gemini with AI personalisation – Get help made just for you. https://gemini.google/overview/personalization/?hl=en-GB. Online, accessed 09-April-2025

    2023

  20. [20]

    Gregory Guest, Kathleen M MacQueen, and Emily E Namey. 2012. Introduction to Applied Thematic Analysis.Applied Thematic Analysis3, 20 (2012), 1–21

    2012

  21. [21]

    Hart and Lowell E

    Sandra G. Hart and Lowell E. Staveland. 1988. Development of NASA-TLX (Task Load Index): Results of Empirical and Theoretical Research. InHuman Mental Workload, Peter A. Hancock and Najmedin Meshkati (Eds.). Advances in Psychology, Vol. 52. North-Holland, 139–183. doi:10.1016/S0166- 4115(08)62386-9

    doi:10.1016/s0166- 1988

  22. [22]

    Eric Horvitz. 1999. Principles of mixed-initiative user interfaces. InProceedings of the SIGCHI conference on Human Factors in Computing Systems. 159–166

    1999

  23. [23]

    Ishtiaque Hossain, I-Chao Shen, Takeo Igarashi, and Oliver van Kaick. 2023. Data-guided Authoring of Procedural Models of Shapes. InComputer Graphics Forum, Vol. 42. Wiley Online Library, e14935

    2023

  24. [24]

    Ishtiaque Hossain, I-Chao Shen, and Oliver van Kaick. 2025. Approximating Procedural Models of 3D Shapes with Neural Networks. InComputer Graphics Forum. Wiley Online Library, e70024

    2025

  25. [25]

    Xiang Hu, Zujie Wen, Yafang Wang, Xiaolong Li, and Gerard De Melo. 2020. Interactive Question Clarification in Dialogue via Reinforcement Learning. InProceedings of the 28th International Conference on Computational Linguistics: Industry Track. 78–89

    2020

  26. [26]

    Yongquan ‘Owen’ Hu, Jingyu Tang, Xinya Gong, Zhongyi Zhou, Shuning Zhang, Don Samitha Elvitigala, Florian ‘Floyd’ Mueller, Wen Hu, and Aaron J Quigley. 2025. Vision-Based Multimodal Interfaces: A Survey and Taxonomy for Enhanced Context-Aware System Design. InProceedings of Manuscript submitted to ACM Beyond Conversations: Spatially-Anchored Previews for ...

    2025

  27. [27]

    Ulas Berk Karli and Tesca Fitzgerald. 2023. Extended Abstract: Resolving Ambiguities in LLM-enabled Human-Robot Collaboration. In2nd Workshop on Language and Robot Learning: Language as Grounding. https://openreview.net/forum?id=LtwuJx83Rc

    2023

  28. [28]

    Daniël Lakens, Anne M Scheel, and Peder M Isager. 2018. Equivalence Testing for Psychological Research: A Tutorial.Advances in Methods and Practices in Psychological Science1, 2 (2018), 259–269

    2018

  29. [29]

    Jaewook Lee, Jun Wang, Elizabeth Brown, Liam Chu, Sebastian S Rodriguez, and Jon E Froehlich. 2024. GazePointAR: A context-aware multimodal voice assistant for pronoun disambiguation in wearable augmented reality. InProceedings of the 2024 CHI Conference on Human Factors in Computing Systems. 1–20

    2024

  30. [30]

    Toby Jia-Jun Li, Jingya Chen, Haijun Xia, Tom M Mitchell, and Brad A Myers. 2020. Multi-Modal Repairs of Conversational Breakdowns in Task-Oriented Dialogs. InProceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology. 1094–1107

    2020

  31. [31]

    Chen Liang, Yuxuan Liu, Martez Mott, and Anhong Guo. 2025. HandProxy: Expanding the Affordances of Speech Interfaces in Immersive Environments with a Virtual Proxy Hand.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies9, 3 (2025), 1–30

    2025

  32. [32]

    Florian Mathis, Brad A Myers, Ben Lafreniere, Michael Glueck, and David PS Marques. 2024. MR-Driven Near-Future Realities: Previewing Everyday Life Real-World Experiences Using Mixed Reality. InProceedings of the 26th International Conference on Multimodal Interaction. 76–85

    2024

  33. [33]

    Daniel Mendes, Daniel Medeiros, Maurício Sousa, Ricardo Ferreira, Alberto Raposo, Alfredo Ferreira, and Joaquim Jorge. 2017. Mid-air modeling with Boolean operations in VR. In2017 IEEE Symposium on 3D User Interfaces (3DUI). IEEE, 154–157

    2017

  34. [34]

    Meta. 2025. Interaction SDK Overview | Meta Horizon OS Developers. https://developers.meta.com/horizon/documentation/unity/unity-isdk- interaction-sdk-overview/. Online, accessed 09-April-2025

    2025

  35. [35]

    Lin Ning, Luyang Liu, Jiaxing Wu, Neo Wu, Devora Berlowitz, Sushant Prakash, Bradley Green, Shawn O’Banion, and Jun Xie. 2025. User-LLM: Efficient LLM Contextualization with User Embeddings. InCompanion Proceedings of the ACM on Web Conference 2025. 1219–1223

    2025

  36. [36]

    Sharon Oviatt. 2003. User-centered modeling and evaluation of multimodal interfaces.Proc. IEEE91, 9 (2003), 1457–1468

    2003

  37. [37]

    Sundar Pichai. 2024. Introducing Gemini 2.0: Our new AI model for the agentic era. https://blog.google/technology/google-deepmind/google- gemini-ai-update-december-2024/#ceo-message. Online, accessed 09-April-2025

    2024

  38. [38]

    Carl Edward Rasmussen and Christopher K. I. Williams. 2005.Gaussian Processes for Machine Learning. The MIT Press. doi:10.7551/mitpress/3206. 001.0001

    doi:10.7551/mitpress/3206 2005

  39. [39]

    Martin Schrepp, Andreas Hinderks, and Jörg Thomaschewski. 2017. Design and Evaluation of a Short Version of the User Experience Questionnaire (UEQ-S).International Journal of Interactive Multimedia and Artificial Intelligence, 4 (6), 103-108.(2017)

    2017

  40. [40]

    Svetlana Stoyanchev, Alex Liu, and Julia Hirschberg. 2014. Towards Natural Clarification Questions in Dialogue Systems. InAISB Symposium on Questions, Discourse and Dialogue, Vol. 20

    2014

  41. [41]

    Yiliu Tang, Jason Situ, Andrea Yaoyun Cui, Mengke Wu, and Yun Huang. 2025. LLM Integration in Extended Reality: A Comprehensive Review of Current Trends, Challenges, and Future Perspectives. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–24

    2025

  42. [42]

    Zhao Tian, Junjie Chen, and Xiangyu Zhang. 2025. Fixing Large Language Models’ Specification Misunderstanding for Better Code Generation. In 2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE). IEEE Computer Society, 645–645

    2025

  43. [43]

    Cyrus Vachha, Yixiao Kang, Zach Dive, Ashwat Chidambaram, Anik Gupta, Eunice Jun, and Björn Hartmann. 2025. Dreamcrafter: Immersive Editing of 3D Radiance Fields Through Flexible, Generative Inputs and Outputs. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–13

    2025

  44. [44]

    John Von Neumann, Robert H Kent, HR Bellinson, and BI Hart. 1941. The Mean Square Successive Difference.The Annals of Mathematical Statistics 12, 2 (1941), 153–162

    1941

  45. [45]

    Zhan Wang, Lin-Ping Yuan, Liangwei Wang, Bingchuan Jiang, and Wei Zeng. 2024. VirtuWander: Enhancing Multi-modal Interaction for Virtual Tour Guidance through Large Language Models. InProceedings of the 2024 CHI Conference on Human Factors in Computing Systems. 1–20

    2024

  46. [46]

    Nelson Wong and Carl Gutwin. 2010. Where are you pointing? The accuracy of deictic pointing in CVEs. InProceedings of the SIGCHI conference on human factors in computing systems. 1029–1038

    2010

  47. [47]

    Yixuan Yang, Junru Lu, Zixiang Zhao, Zhen Luo, Wanxi Dong, Victor Sanchez, and Feng Zheng. 2025. LLplace: Embodied 3D Indoor Layout Synthesis Framework with Large Language Model. In2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 20685–20691

    2025

  48. [48]

    Hamed Zamani, Susan Dumais, Nick Craswell, Paul Bennett, and Gord Lueck. 2020. Generating Clarifying Questions for Information Retrieval. In Proceedings of the Web Conference 2020. 418–428

    2020

  49. [49]

    Chenwei Zhang, Wei Fan, Nan Du, and Philip S Yu. 2016. Mining User Intentions from Medical Queries: A Neural Network Based Heterogeneous Jointly Modeling Approach. InProceedings of the 25th International Conference on World Wide Web. 1373–1384

    2016

  50. [50]

    Dongsong Zhang, Boonlit Adipat, and Yaser Mowafi. 2009. User-Centered Context-Aware Mobile Applications — The Next Generation of Personal Mobile Computing.Communications of the Association for Information Systems24, 1 (2009), 3

    2009

  51. [51]

    Lei Zhang, Ashutosh Agrawal, Steve Oney, and Anhong Guo. 2023. VRGit: A Version Control System for Collaborative Content Creation in Virtual Reality. InProceedings of the 2023 CHI Conference on Human Factors in Computing Systems. 1–14

    2023

  52. [52]

    Lei Zhang, Jin Pan, Jacob Gettig, Steve Oney, and Anhong Guo. 2024. VRCopilot: Authoring 3D Layouts with Generative AI Models in VR. In Proceedings of the 37th Annual ACM Symposium on User Interface Software and Technology. 1–13

    2024

  53. [53]

    Michael JQ Zhang and Eunsol Choi. 2025. Clarify When Necessary: Resolving Ambiguity Through Interaction with LMs. InFindings of the Association for Computational Linguistics: NAACL 2025. 5526–5543. Manuscript submitted to ACM 20 Chen et al

    2025

  54. [54]

    Shunpeng Zou and Xiaohui Zou. 2017. Understanding: How to Resolve Ambiguity. InIntelligence Science I: Second IFIP TC 12 International Conference, ICIS 2017, Shanghai, China, October 25-28, 2017, Proceedings 2. Springer, 333–343. Manuscript submitted to ACM

    2017

  55. [2024]

    Understanding User Intent Modeling for Conversational Recommender Systems: A Systematic Literature Review.User Modeling and User-Adapted Interaction(2024), 1–64

    2024

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.