{"id":"6885b522-ae54-4bc1-b035-a627a320bd5c","arxiv_id":"2608.10838","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A hybrid VLM-plus-solver pipeline produces more plausible furniture layouts than existing generators on standard rooms and extends to non-rectangular floor plans with doors and windows.","lead":"PolyLayout is a three-stage system that generates furniture layouts for real homes, using a vision-language model to position functional groups inside any floor plan shape. It reports better perceptual quality than two existing layout generators on rectangular rooms and shows it can handle doors, windows, and L/T/U-shaped rooms that other systems do not support.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stage 2 semantic placement is validated with only one VLM (Gemini 2.5 Flash); if reading the color-coded polygon is model-specific, the VLM-guided irregular-room claim lacks support.","rationale":"The reader's conditional verdict is appropriate. The paper has real strengths: a blind human study (Sec. 4.4) independently confirms the plausibility ranking on rectangular rooms, the deterministic solver gives hard geometric guarantees (Sec. 3.1, Eq. 1), and the limitations section honestly acknowledges the eleven-room basis and the lack of comparators. The softest point is the un-isolated, single-backbone dependency of Stage 2. All semantic content - zone placement, wall/corner selection, opening-relative positioning - passes through one Gemini 2.5 Flash model, and the downstream solver can repair geometry but not semantics. The supplementary prompts show some rule-based repairs (App. A.4), but these cover only a few known error classes (e.g., wall-with-door center), not arbitrary misreadings of a color-coded polygon. Because the paper's novel claim is exactly 'VLM-guided macro-routing' beyond rectangles, this is the load-bearing assumption. A cross-backbone rerun is the direct test: if the results hold with a second model family, the concern is resolved and the verdict can move toward accept; if they collapse, the claim must be scoped to the specific backbone. We therefore see no reason to change the reader's CONDITIONAL verdict.","tokens_in":17933,"tokens_out":10647,"duration_ms":113633,"concrete_test":"Run the full benchmark (44 inventories x 3 rectangular and x 5 non-rectangular rooms, two runs each) with Stage 2's semantic suggestion and structured extraction calls served by a different VLM family (e.g., GPT-4o or Claude 3.5 Sonnet) instead of Gemini 2.5 Flash, keeping Stage 1, Stage 3, all prompts, and the solver identical. Report plausibility, in-bounds, placement ratio, and success rate from Tables 2 and 4 for both backbones. If plausibility drops by more than about 0.2 on either setting, or if in-bounds or completeness degrades on non-rectangular plans, the central claim is backbone-specific rather than a property of PolyLayout.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the central pipeline, Stage 2 routes all semantic decisions through four Gemini 2.5 Flash calls: placement ordering, RAG context plan, semantic suggestion, and structured extraction (Sec. 3.1 and App. A). The semantic suggestion receives a top-down rendering with color-coded wall segments and door/window markers; the extraction stage can repair rule violations like placing at a wall center that contains a door (App. A.4 rule 10), but it cannot detect a syntactically valid but semantically wrong reference, e.g., anchoring the sofa to the wrong wall or misreading a corner on a non-rectangular plan. Stage 3's deterministic solver (Sec. 3.1) enforces only geometric feasibility (in-bounds, opening clearance, no overlap); it cannot re-anchor a semantically misplaced bundle. The paper evaluates exactly one backbone with no cross-model check and reports no per-call failure analysis for Stage 2. On the rectangular benchmark the shared Gemini backbone across all three methods controls for VLM capability, but the headline 'extends to irregular boundaries' rests entirely on this single VLM reading the polygon correctly. If the color/label reading is sensitive to the particular model or prompt, the central claim of a general VLM-guided macro-routing method is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"PolyLayout is a three-stage hybrid layout-generation pipeline aimed at retail furniture planning. Stage 1 parses a glTF room into a floor polygon, extracts walls and openings, and deterministically clusters a catalog inventory into functional bundles using a hand-authored mapping of 212 product types to 43 roles and six area templates. Stage 2 invokes a single VLM (Gemini 2.5 Flash) four times — placement ordering, RAG-based context planning, semantic suggestion, and structured extraction — to anchor each bundle or standalone item to a room corner, wall, door, window, or previously placed object. Stage 3 deterministically converts those anchors into metric poses and accepts a pose only when Eq. (1) is satisfied (in-bounds, opening clearance, zero overlap); otherwise it searches discrete offsets (Eq. (3)) or leaves the item unplaced. On three rectangular rooms without openings, the method reports plausibility 2.74 vs 2.63 (Holodeck) and 2.49 (LayoutVLM), 100% success and in-bounds, placement ratio 0.973, and roughly 13× lower latency than LayoutVLM; a blind five-rater human study prefers PolyLayout in 61% of votes. A descriptive capability study on rectangular rooms with openings and five non-rectangular plans reports in-bounds=1.000 and completeness ≥0.95, with plausibility declining from 2.73 to 2.48 as geometry hardens.","tokens_in":18152,"tokens_out":8945,"duration_ms":97307,"significance":"The central geometric claim is sound: because the solver in Stage 3 accepts only poses satisfying Eq. (1), in-bounds and collision-free placement are guaranteed by construction, and unplaced items are not forced into collisions. The rectangular-room comparison is also reasonably controlled: all methods share the Gemini 2.5 Flash backbone, the plausibility gain over LayoutVLM is significant, and the human study independently supports the ranking. The paper is unusually candid about its own limitations — the same-family VLM judge (§4.2), the descriptive rather than comparative nature of the irregular-plan results (§4.5), and the eleven-room basis of the geometric conclusions (Conclusion). If the single-VLM dependency is addressed, this is a credible systems contribution for retail-scale layout generation. The principal unresolved risk is that Stage 2 is validated with exactly one VLM and without a direct measure of semantic reference correctness, so the 'beyond rectangular rooms' capability has not yet been shown to be a property of the method rather than of one model.","major_comments":[{"comment":"Stage 2 routes all semantic decisions through four Gemini 2.5 Flash calls (App. A.1–A.4), and the paper provides no cross-model check and no per-call failure analysis. The extraction rules (App. A.4, rules 10–12) can repair rule violations such as a 'center' placement on a door-bearing wall, but they cannot detect a syntactically valid but semantically wrong reference, e.g., anchoring a bundle to the wrong wall or misreading a corner or wall color on an L/T/U plan. Stage 3's feasibility condition Φ in Eq. (1) verifies only geometric containment, opening clearance, and overlap; it cannot re-anchor a semantically misplaced bundle. Since the capability study in Sec. 4.5 rests entirely on this one VLM reading the color-coded polygon correctly, the headline claim that the method 'extends to irregular boundaries' is vulnerable to model-specific color/label reading. I would ask for at least one additional VLM backbone, or a direct semantic-anchor accuracy metric on the non-rectangular plans, or an ablation that removes the color-coded rendering, before treating the irregular-room claim as established.","section":"Sec. 3.1, App. A, Sec. 4.5"},{"comment":"The non-rectangular results are descriptive only: there is no baseline comparator, and the inherent room variety is five hand-picked plans (L, T, U, beveled, cut), with condition counts scaling by inventories and repeated runs rather than by room topology. The paper's own Conclusion states that 'all geometric conclusions rest on eleven rooms (six rectangular, five non-rectangular).' This honestly bounds the claim, but it also means the central 'beyond rectangular rooms' contribution is currently supported by a small, non-random sample. I would not require a full benchmark, but a per-plan breakdown of the Table 4 metrics (or an expanded set of irregular plans) would materially strengthen the generalization claim.","section":"Sec. 4.5 and Conclusion"}],"minor_comments":[{"comment":"Ablation A2 (w/o RAG plan) is not statistically significant on any metric after Bonferroni correction, yet the text says RAG lets the retailer's design practice shape the layout. This practical value is plausible but not directly tested; the wording should distinguish an architectural design choice from an empirically measured benefit.","section":"Sec. 4.6, Table 5"},{"comment":"The judge prompt (App. B) instructs the model to bias toward in-bounds and to pick the higher score when torn; this may compress score differences and inflate absolute plausibility. The human study mitigates the ranking claim, but the absolute 2.7x values should be interpreted with care.","section":"Sec. 4.2"},{"comment":"The Holodeck plausibility gain is only marginal (p≈0.05); the paper's wording ('marginal') is accurate, but the abstract's 'attains the highest perceptual plausibility' should perhaps note that the advantage over Holodeck is small in the VLM-judge score and is carried mainly by the human preference.","section":"Sec. 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within the scope of the journal and is honestly written. The main risk is not correctness but generalizability: the single-VLM validation of Stage 2. If the authors supply a cross-model run or a semantic-error audit, I would be comfortable with acceptance; as it stands, a major revision is warranted. The RAG component's empirical contribution is weak, but this is not central to the headline claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nPolyLayout is worth a look if you care about applied layout generation. The genuinely new piece is the decoupling: rule-based functional clustering, VLM macro-routing over annotated polygon renderings, and a deterministic solver that accepts only poses satisfying in-bounds, opening clearance, and zero overlap. The solver gives a real guarantee, and unplaced items are reported rather than forced into collisions. That is a meaningful step beyond LayoutVLM and Holodeck, which treat non-rectangular boundaries and fixed openings as afterthoughts or not at all.\n\nThe paper is also honest about where it is exploratory. The capability study on L/T/U/beveled/cut rooms is explicitly descriptive, no baseline conditions on fixed openings, and the limitations section admits the eleven-room scale, single catalog, and no cross-model VLM checks. That candor is earned; the ablations actually help. Removing clustering and removing the solver produce significant drops with sensible failure modes, and the RAG component is honestly reported as not statistically significant on its own.\n\nThe weak spots are in proportion. Eleven rooms is a small base, and condition counts scale with inventories and repeated runs rather than room variety. No code or data is released, which limits reproducibility. The plausibility judge is a Gemini model, same family as the generator backbone; the blind human study mitigates that, but the two are not entirely independent assessments. And the single-VLM dependency in Stage 2 is real: if reading the color-coded polygon is Gemini-specific, the irregular-room generalization claim is weaker. Still, the paper does not overclaim there; it says the pipeline extends, not that any VLM works.\n\nWho is this for? People building practical furniture planning tools, and researchers in hybrid LLM/optimization methods. It does not reshape the field, but it is a solid, well-engineered contribution with a clear separation of semantic and geometric concerns. It deserves a serious referee, primarily for the applied community. I would not desk-reject it.\n\nRecommendation: send it to review, with requests for code/data release and ideally a second VLM backbone check on the non-rectangular plans. The central argument holds up; the gaps are in scope and generality, not in the load-bearing geometry.","headline":"Hybrid VLM-plus-deterministic-solver pipeline is a genuinely useful integration for applied layout generation, with an honest though narrow evaluation.","tokens_in":18738,"tokens_out":1920,"would_cite":true,"duration_ms":20717,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PolyLayout claims that decoupling semantic placement from geometric solving yields room layouts that are both more plausible and perfectly in-bounds, even on irregular floor plans.","keywords":["Spatial Planning","Hybrid Layout Generation","VLM","3D room layout generation","non-rectangular floor plans","furniture placement","constraint satisfaction","retail catalog"],"falsifier":"Re-run the full 264-condition rectangular benchmark and the 440-condition non-rectangular benchmark with a different vision-language model in Stage 2, keeping the prompts, taxonomy, and solver unchanged; if success or in-bounds falls below 100 percent and 1.000, or the plausibility advantage over the two comparison systems disappears, the central result is specific to the chosen backbone rather than to the hybrid design.","tokens_in":17697,"feed_emoji":"🛋️","tokens_out":11882,"duration_ms":119583,"temperature":0.7,"pith_summary":"PolyLayout claims that plausible, constraint-respecting room layouts can be generated without end-to-end black-box optimization by splitting the problem into three stages: rule-based functional clustering into furniture bundles, vision-language-model (VLM) macro-routing of those bundles over an annotated top-down floor plan, and deterministic geometric solving that converts coarse anchors into exact collision-free poses. On rectangular rooms without openings, the reported system achieves the highest perceptual plausibility of the three compared methods (mean 2.74 on a 1–4 scale, versus 2.63 and 2.49) while keeping success and in-bounds placement at 100 percent, and a blind human panel prefers its layouts in 61 percent of votes. The same pipeline, with no modification, is then run on rectangular rooms with doors and windows and on five non-rectangular plans (L, T, U, beveled, cut); in all settings in-bounds stays at 1.000 and average placement completeness stays at or above 0.95. If the claim is right, automatically generated layouts can serve as trustworthy starting points for real homes, handling the irregular boundaries and fixed openings that prior systems natively reject.","feed_headline":"Room layout model wins plausibility, keeps items in bounds","feed_subtitle":"By splitting semantic choices from geometric solving, it also handles L-, T-, and U-shaped plans with doors and windows.","key_machinery":"The load-bearing mechanism is the three-stage decoupling, anchored by two concrete objects: the functional bundle and the Stage-3 dynamic-placement solver. A functional bundle is a group of catalog items arranged by a deterministic template around a mandatory anchor piece (a bed, desk, table, sofa, seating, or bookcase), so the VLM reasons about a few composite objects rather than dozens of individual items. The solver receives each bundle's placement as a place_object call—one reference object plus one relative position from the constrained vocabulary—maps it to a candidate pose with a calculator, then tests it with the feasibility predicate $\\Phi(o)$: the footprint must lie inside the floor polygon, obstruct no door or window, and overlap no placed object. If the candidate fails, the solver searches discrete offsets along the wall, re-anchors to nearby walls, and only then accepts a minimum-overlap pose or reports the item unplaced; this is the component that makes irregular boundaries and fixed openings behave as hard constraints.","core_discovery":"The central claim is that the completeness–validity trade-off seen in existing generators is not intrinsic: semantic quality and hard geometric compliance can be obtained together by decoupling them. Concretely, the paper argues that a fixed taxonomy can group a retail catalog into functional bundles with hand-authored internal geometry, a VLM can then choose each bundle's coarse anchor from a small closed vocabulary (wall center, corner, next to a door or window, in front of or beside a previously placed object), and a deterministic dependency-ordered solver can turn those anchors into metric poses while enforcing a feasibility predicate that requires every footprint to lie inside the floor polygon, every opening to stay clear, and zero overlap with already placed items. The reported outcomes are perfect in-bounds placement and 100-percent success on all tested settings, the best plausibility score in the controlled rectangular comparison, and graceful rather than catastrophic degradation as plans acquire openings and irregular boundaries. The paper reads the dominant remaining flaw—uncovered product types placed standalone and drifting from their functional groups—as a limitation of its hand-authored taxonomy, not of the geometric solver.","pith_inferences":["A direct test of the hybrid split: rerun Stage 2 with a different vision-language backbone and the same solver; if in-bounds stays 1.000 and success stays 100 percent, the geometric guarantee is backend-independent, while plausibility differences would quantify how much of the quality comes from the language model rather than the pipeline.","The same closed-vocabulary interface (reference object plus relative position) is a natural interchange format for interactive refinement: a user's edit can be expressed as a new place_object call, and the retrieval-augmented channel is already positioned to inject retailer-specific design rules.","The hand-authored template set is the scaling bottleneck; learning bundle templates from catalog data or from human-preference feedback could extend the method to kitchens and other appliance-driven rooms without changing the solver.","The decoupling principle could transfer to other spatial-planning domains—retail shelf planograms, event floor plans, or warehouse staging—where coarse semantic zoning and hard geometric constraints coexist."],"forward_implications":["No generated layout exits the room polygon, blocks a door or window, or overlaps an already placed object, on any tested floor-plan shape.","The plausibility gain is not an artifact of the judging model: a blind human panel prefers the new layouts 61 percent of the time, and on 72 percent of conditions the panel's collective choice matches the judge's top score.","The pipeline runs on CPU at roughly one-seventh the latency of one comparison system and about thirteen times faster than the other, making interactive retail use plausible.","The bottleneck for further quality is the hand-authored product-type taxonomy, not the solver: items whose product type maps to no bundle role are placed standalone and can drift from their natural group."],"supporting_citations":[{"why":"Supplies the strongest comparison baseline: a system that generates its own floor plan and openings but assumes rectangular boundaries, against which the plausibility and completeness gains are measured.","marker":"[23]"},{"why":"Supplies the VLM-grounded placement baseline that packs items densely inside a rectangular extent, motivating the hard-constraint handling and the latency comparison.","marker":"[19]"},{"why":"Establishes the prompting pattern for turning a large language model into a layout generator, which Stage 2 extends to visual context and a constrained vocabulary.","marker":"[5]"},{"why":"Represents the learned sequential-placement paradigm that the hybrid pipeline is contrasted against for its lack of hard-constraint enforcement.","marker":"[21]"}],"fun_headline_variants":["VLM-guided layout handles L-shaped rooms with ease","Decoupling semantics and geometry yields realistic room plans","Room layout AI nails in-bounds placement, even for odd shapes","Hierarchical planning beats rectangle-only layout generators","VLM-guided layout now fits L, T, and U-shaped rooms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All semantic routing rests on one vision-language model correctly reading a color-coded top-down rendering of the room and replying only in the allowed placement vocabulary; if it misreads a wall color, a corner, or an opening marker on an irregular plan, the deterministic solver can repair the pose but not the semantic error, and the paper tests only this single backbone.","fun_headline_variants_meta":{"raw":{"variants":["VLM-guided layout handles L-shaped rooms with ease","Decoupling semantics and geometry yields realistic room plans","Room layout AI nails in-bounds placement, even for odd shapes","Hierarchical planning beats rectangle-only layout generators","VLM-guided layout now fits L, T, and U-shaped rooms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000791,"raw_usage":{"total_tokens":3497,"prompt_tokens":971,"completion_tokens":2526,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":2460}},"tokens_in":587,"tokens_out":2526,"duration_ms":17347,"temperature":1.0,"reasoning_tokens":2460,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:20:43.712692+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the full 264-condition rectangular benchmark and the 440-condition non-rectangular benchmark with a different vision-language model in Stage 2, keeping the prompts, taxonomy, and solver unchanged; if success or in-bounds falls below 100 percent and 1.000, or the plausibility advantage over the two comparison systems disappears, the central result is specific to the chosen backbone rather than to the hybrid design.","supporting_citations":[{"cited_title":"In: CVPR","cited_arxiv_id":null,"evidence_quote":"Supplies the strongest comparison baseline: a system that generates its own floor plan and openings but assumes rectangular boundaries, against which the plausibility and completeness gains are measured."},{"cited_title":"In: CVPR","cited_arxiv_id":null,"evidence_quote":"Supplies the VLM-grounded placement baseline that packs items densely inside a rectangular extent, motivating the hard-constraint handling and the latency comparison."},{"cited_title":"In: NeurIPS","cited_arxiv_id":null,"evidence_quote":"Establishes the prompting pattern for turning a large language model into a layout generator, which Stage 2 extends to visual context and a constrained vocabulary."},{"cited_title":"ACM TOG37(4), 70:1–70:14 (2018)","cited_arxiv_id":null,"evidence_quote":"Represents the learned sequential-placement paradigm that the hybrid pipeline is contrasted against for its lack of hard-constraint enforcement."}],"review_version":1}