{"id":"9df403d1-4db1-45bb-b3ae-be9919f3f501","arxiv_id":"2411.11752","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper presents an MR system that uses LLMs to interpret spatial commands and generative AI to create 3D objects, with a small user study reporting positive perceived usability.","lead":"sMoRe is a mixed-reality application that turns voice or typed commands into generated 3D objects placed in the user's physical room using LLMs and text-to-3D AI. A six-person lab study reports positive perceived usability and enjoyment, though the evidence is exploratory and lacks a baseline comparison.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Placement accuracy is the unmeasured load-bearing link: the claimed usability benefit cannot be attributed to sMoRe's locate-and-place pipeline unless LLM dimensions and sampled positions are shown to match user intent.","rationale":"The paper is an implemented prototype with a plausible modular design and unusually detailed prompts (Appendix A); those engineering contributions deserve credit. The reader's conditional verdict is appropriate. My stress-test did not find a reason to reject. The most load-bearing condition for the central claim is not the LLM's linguistic parsing or mesh quality, but whether the physical placement actually follows user intent. That condition is never measured. The system's own Section 5.2 announces that the matching between predicted and actual locations should be evaluated in future work, which is an explicit self-acknowledged missing support for the core mechanism. The Location Optimizer's static validation rules are too coarse to guarantee 'On' or 'near' semantics, and random sampling with a large maximum attempt count plus silent floor fallback could fail in ways invisible to a six-participant subjective study. A concrete logging-and-reference-marker test would settle whether the pipeline places objects correctly or merely often displays plausible scenes. If the test shows low error, the conditional verdict can be upgraded; if it shows high error or frequent fallback, the paper's effectiveness claim should be re-scoped to 'voice-driven object generation with approximate placement,' which is still a useful system contribution. Thus my recommendation is UNCHANGED: keep the conditional verdict.","tokens_in":22452,"tokens_out":4512,"duration_ms":48017,"concrete_test":"Re-run or instrument the existing system so that every prompt logs: Object Preparer dimensions, every sampled p_c with its validation result, number of attempts, and final candidate or floor-fallback. Before each placement, the participant places a movable reference marker at their intended location. After the object appears, compute the 3D distance between the reference and the final object pose, normalized by the object's bounding-box diagonal. Collect at least five placements per participant. If the median normalized error exceeds, say, 0.25-0.5, or if any fallback occurs in a non-trivial fraction of trials, the central claim of accurate, intention-following placement is not supported. Publish the logs and the failure counts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The framework's core contribution is creating and locating virtual objects, and the user study is the only support for this claim. The Location Optimizer (Algorithm 2) generates random surface points and validates them only through Table 1's coarse static rules; e.g., 'On' is accepted when the candidate is inside the target object's x/z boundaries, with no explicit check that the object rests on the surface or that the generated mesh will not interpenetrate it. Object Preparer dimensions come from LLM estimates (Appendix A.3), and those dimensions define both the placeholder collider and the validation geometry. If the LLM underestimates or overestimates real-world scale, the 'precise placement' in DG2 is wrong even when the random sampling succeeds. Additionally, if no candidate passes validation after up to 10,000 attempts, the system silently falls back to the floor in front of the user; the paper does not report how often this happens. The paper itself concedes in Section 5.2 that 'Future studies should evaluate how accurate the respective matching between the predicted and the actual locations is.' Because no objective placement measurement exists, the positive SUS/TLX and qualitative responses cannot be causally attributed to placement accuracy; they could reflect the novelty of voice-to-3D generation or the 2.5D map. This is a load-bearing gap, not a cosmetic one, since the claimed contribution is precisely 'create and locate virtual objects.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents sMoRe, a mixed-reality system that combines an LLM-based pipeline (Planner, Analyzer, Object Preparer, Location Optimizer) with a text-to-3D generator (Genie) to let users create and place virtual objects using voice or typed text. The system also provides a 2.5D layout map and hand/controller interaction. The authors evaluate sMoRe in a lab study with six participants using NASA TLX, SUS, Likert items, and open-ended questions, reporting generally positive descriptive scores and qualitative feedback. The paper claims that sMoRe enhances user comprehension, interaction, and organization in MR environments.","tokens_in":22716,"tokens_out":4136,"duration_ms":38252,"significance":"If the framework works as described, sMoRe is a useful integration of LLM-driven spatial reasoning with generative 3D asset creation in MR, with a clearly described architecture, algorithm pseudocode, and an appendix of module prompts that would allow replication. The 2.5D layout map and the voice-to-3D workflow are interesting interaction ideas. The main limitation is that the evaluation does not provide objective or comparative evidence for placement accuracy or for the claimed usability improvements; therefore the significance is conditional on a strengthened evaluation. The paper is honest about several limitations in Section 5.2, but the headline claims go beyond what the data support.","major_comments":[{"comment":"The conclusion that sMoRe \"significantly enhances user interaction\" and the abstract's claim that the user study \"demonstrates the effectiveness\" of the system are not supported by the reported evidence. The study has six participants, no control condition, no inferential statistics, and no objective task-performance measures; only descriptive means and standard deviations are reported. Because Contribution 3 claims a \"comprehensive user study to validate the usability and effectiveness,\" this is a load-bearing gap rather than a cosmetic one.","section":"§4.3 and §6"},{"comment":"Placement accuracy, which is central to DG2 and to the first contribution (creating and locating virtual objects), is never measured. Algorithm 2 validates candidate locations only against the coarse static criteria in Table 1; for example, \"On\" checks only whether the candidate is inside the target object's x/z boundaries, with no explicit check that the object rests on the surface or that the generated mesh does not interpenetrate the environment. The Object Preparer's dimensions come from LLM estimates (Appendix A.3), and these dimensions define the placeholder collider and the validation geometry. The paper itself concedes in Section 5.2 that \"Future studies should evaluate how accurate the respective matching between the predicted and the actual locations is.\" Without any objective measurement, the reported SUS/TLX and Likert responses cannot be causally attributed to the locate-and-place pipeline.","section":"§5.2 and Algorithm 2 / Table 1"},{"comment":"The fallback behavior of the Location Optimizer is not instrumented. When no candidate point passes validation after up to 10,000 attempts, the system silently places the object on the floor in front of the user. The paper does not report how often this fallback occurs, in which usage scenarios it occurs, or whether any participant's objects were placed by fallback. Since fallback placement overrides the user-specified location, its frequency and consequences are directly relevant to the usability claims and should be measured.","section":"§3.3.4 and Algorithm 2"},{"comment":"The user-study procedure does not include predefined placement targets or a baseline condition. Participants freely explored generating and placing objects for ten minutes, so the paper cannot report a task success rate or compare sMoRe against an alternative interaction method (e.g., controller-based placement). A within-subjects comparison with a baseline, or at least a task-based accuracy metric with predefined target locations, would substantially strengthen the efficacy claim that the current descriptive Likert scores are meant to support.","section":"§4.2 and §4.3"}],"minor_comments":[{"comment":"The NASA TLX and SUS instruments are cited with \"[?]\" placeholders; full references are needed.","section":"§4.2"},{"comment":"The Object Preparer example contains \"chariity\" where the object name appears to be misspelled (likely \"charity\" or the intended word); this should be corrected.","section":"Appendix A.3 and Figure 5"},{"comment":"The coordinate-mapping formulas are typeset without clear multiplication symbols and with ambiguous subscripts; defining P_virtual, P_layout, and S_virtual as single variables and presenting the equations with standard math notation would improve readability.","section":"§3.4.2"},{"comment":"The page footer and ACM reference format display the year 2018 and a placeholder DOI, which are inconsistent with the manuscript's actual submission year (2024); these metadata should be fixed.","section":"Title page and ACM Reference Format"},{"comment":"The text refers to \"Pseudo-code 1\" and \"Pseudo-code 2\" while the displayed items are labeled \"Algorithm 1\" and \"Algorithm 2\"; the terminology should be made consistent. Additionally, the sentence introducing the Planner's prompt processing is incomplete and should read as a full sentence.","section":"§3.3"},{"comment":"The conclusion contains the duplicated phrase \"significantly enhances user interaction and interaction within MR environments\"; one occurrence of \"interaction\" should be removed.","section":"§6"}],"recommendation":"major_revision","confidential_remarks":"In its current form, the paper is better suited as a systems/demonstration contribution than as a full archival IMWUT paper given the small, non-comparative user study. The clear system description and prompt-level implementation details are strengths worth preserving. I would encourage the authors to add an objective placement-accuracy assessment and a within-subjects comparison to a baseline interaction technique before resubmission; these additions would speak directly to the load-bearing gaps identified in the major comments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: sMoRe is a genuine, working integration of LLM-driven command parsing, text-to-3D generation, and MR scene placement, with a nice 2.5D layout map. The paper's architecture is clear, the prompts in the appendix are detailed enough to reproduce, and the limitations section is unusually honest. But the evaluation does not support the central claim. Six self-selected participants, no baseline, no statistical tests, and no objective measurement of placement accuracy. The authors say 'demonstrates the effectiveness' and 'significantly enhances' in the abstract and conclusion; that is overreach.\n\nWhat is actually new: the specific pipeline (Planner, Analyzer, Object Preparer, Location Optimizer, Mesh Generator) and the bidirectional 2.5D map that syncs layout objects to virtual objects. The conceptual move is incremental—LLMR, OCTOPUS, and voice-prompted VR scene creation already do LLM-driven placement. The contribution is a concrete systems integration, not a new interaction paradigm.\n\nWhere it is soft: the load-bearing link is placement accuracy. The Location Optimizer validates candidates with coarse static rules from Table 1 (e.g., 'On' just checks x/z boundaries; no penetration check), dimensions come from LLM guesses, and if no candidate passes after 10,000 tries it silently falls back to the floor in front of the user. The paper never reports fallback frequency or measures whether the final object is where the user intended. Section 5.2 concedes exactly this. So positive SUS/TLX and qualitative responses could be driven by the novelty of voice-to-3D rather than by precise placement—the thing the system claims to provide.\n\nAlso, no code or raw data released. For a systems paper that is fixable; the appendix prompts make replication plausible.\n\nMy recommendation: send it to peer review, but the authors need to reframe as a usability exploration, add objective placement metrics (or at least report fallback rate and placement error), and tone down the claims. The prototype is real and the limitations are acknowledged; that deserves a serious referee, not a desk reject.","headline":"A real and well-documented MR prototype whose central placement-accuracy claim is not actually measured; worth reviewing but needs a major reframing and objective evaluation.","tokens_in":23271,"tokens_out":1878,"would_cite":false,"duration_ms":17511,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes sMoRe, a mixed-reality system that turns voice or text commands into newly generated 3D objects placed at user-specified physical locations.","keywords":["mixed reality","large language models","generative AI","text-to-3D generation","natural language interfaces","object placement","spatial mapping","user study"],"falsifier":"Run a placement-accuracy study in which a fixed set of commands with known intended locations is repeated, measuring the distance between the placeholder position and the intended position and checking whether objects stay within the identified surface boundaries; systematic error would show that the reported usability gains come from the interaction style rather than reliable placement.","tokens_in":22240,"feed_emoji":"🥽","tokens_out":6215,"duration_ms":55210,"temperature":0.7,"pith_summary":"The paper proposes sMoRe, a mixed-reality application that turns voice or typed commands into newly generated 3D objects placed in the user's physical space. Its central claim is that an LLM-driven pipeline paired with text-to-3D generative AI can replace asset libraries, controllers, and 3D-modeling skill as the main way users create and organize virtual objects. In a six-participant user study, the authors report that users found the system easy to learn, low in physical and mental demand, and effective for achieving what they had in mind. The paper argues this points toward a more intuitive, accessible way to interact with mixed-reality environments.","feed_headline":"Speak a command, and a headset builds virtual objects","feed_subtitle":"A five-stage LLM pipeline places text-to-3D creations on scanned real surfaces by voice.","key_machinery":"The carrying mechanism is the five-module pipeline (Planner, Analyzer, Object Preparer, Location Optimizer, Mesh Generator), with the Location Optimizer doing the placement work by generating random points on LLM-chosen surface types and accepting only candidates that satisfy the twelve spatial-relation criteria in Table 1. The Object Preparer's LLM-produced scale determines collider boundaries, so the entire placement chain depends on that estimate. The 2.5D layout map synchronizes positions with the virtual scene through a linear scaling equation and gives users an overview and an indirect manipulation handle.","core_discovery":"sMoRe establishes a framework in which a natural-language request is decomposed by a Planner LLM into an object name, attributes, and target locations; an Analyzer LLM links those locations to objects in a scanned scene; an Object Preparer LLM assigns real-world dimensions, shape, and gravity; a Location Optimizer samples candidate surface points and validates them against pre-defined spatial-relation rules; and a Mesh Generator calls a text-to-3D service and attaches physics and interaction components. The resulting virtual object appears at the chosen location, persists as the user moves, and is mirrored on a 2.5D layout map. The reported user study supports that this improves comprehension, interaction, and organization of the MR space.","pith_inferences":["The paper does not measure whether the predicted placement matches the user's intended location; a controlled accuracy study would test whether the positive usability ratings come from reliable spatial reasoning or from the novelty of voice-generated objects.","Because the Location Optimizer falls back to the floor in front of the user when no valid point is found, failures could silently place objects in unintended spots; making that fallback visible in the interface would let users correct placement immediately.","The one-object-per-prompt constraint suggests the Planner would need extension before the system can handle compound organization tasks such as moving several objects into a new arrangement at once.","The 2.5D map's real-time synchronization implies the map could function as an external spatial memory aid, but whether it actually improves recall over headset-only views is untested."],"forward_implications":["Users can generate bespoke 3D objects from voice or text descriptions and have them appear on specified physical surfaces without asset libraries or 3D modeling experience.","The 2.5D layout map provides a persistent bird's-eye overview and lets users reposition virtual objects indirectly by moving their low-poly representations.","Virtual objects respond to gravity and collide with physical and virtual objects during hand or controller manipulation.","The same framework can be applied to visual reminders, interior design previews, and creative prototyping in physical spaces.","If the usability results generalize, speech-driven generation can lower the cognitive and physical load of manipulating mixed-reality content."],"supporting_citations":[{"why":"LLMR, a prior system for real-time prompting of interactive worlds with LLMs; the main baseline for combining LLM interactions with MR object creation.","marker":"[10]"},{"why":"DreamFusion, a text-to-3D method using 2D diffusion; the conceptual basis for generating 3D assets from text descriptions.","marker":"[31]"},{"why":"Roberts et al., voice prompts to 'speak the world into existence'; the prior work this paper extends toward dynamic MR object creation.","marker":"[37]"},{"why":"Aghel Manesh et al., voice prompts for modifying virtual scenes and proxemic variables; supports the natural-language spatial manipulation claim.","marker":"[1]"},{"why":"SituationAdapt, which uses LLM reasoning for contextual UI optimization in MR; supports using LLMs for scene-aware placement decisions.","marker":"[23]"},{"why":"DreamCodeVR, speech-driven programming and behavior generation in VR; grounds the idea of using natural language to control virtual world content.","marker":"[14]"}],"fun_headline_variants":["Voice commands spawn 3D objects in your space","LLM pipeline places text-to-3D objects by voice","Talk to MR, see objects appear right where you want","sMoRe: Your voice dictates where 3D objects land","Natural language directs virtual objects in mixed reality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system assumes that the dimensions the LLM assigns to an object and the random surface points the Location Optimizer samples will place virtual objects where users mean them to go, and the user study never measures placement accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Voice commands spawn 3D objects in your space","LLM pipeline places text-to-3D objects by voice","Talk to MR, see objects appear right where you want","sMoRe: Your voice dictates where 3D objects land","Natural language directs virtual objects in mixed reality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000186,"raw_usage":{"total_tokens":1270,"prompt_tokens":837,"completion_tokens":433,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":353}},"tokens_in":453,"tokens_out":433,"duration_ms":4555,"temperature":1.0,"reasoning_tokens":353,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:10:19.163282+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a placement-accuracy study in which a fixed set of commands with known intended locations is repeated, measuring the distance between the placeholder position and the intended position and checking whether objects stay within the identified surface boundaries; systematic error would show that the reported usability gains come from the interaction style rather than reliable placement.","supporting_citations":[{"cited_title":"SituationAdapt: Contextual UI Optimization in Mixed Reality with Situation Awareness via LLM Reasoning","cited_arxiv_id":"2409.12836","evidence_quote":"SituationAdapt, which uses LLM reasoning for contextual UI optimization in MR; supports using LLMs for scene-aware placement decisions."}],"review_version":1}