{"id":"153c14d5-cc5b-4392-b78b-6939319ff083","arxiv_id":"2608.11793","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PolarSym adds decoupled directional and radial geometric attention to a CAD floorplan parsing Transformer, improving PQ by 1.73% and mIoU by 4.31% over a reproduced SymPoint V2 baseline.","lead":"This paper presents PolarSym, a Transformer attention module that models direction and distance separately in polar-style geometry to parse CAD floor plans. It reports small gains in parsing quality and faster training over a reproduced SymPoint V2 baseline.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 lists different batch sizes for baseline (2/8) and PolarSym (2/6), contradicting the 'identical training settings' on which the headline gains rest.","rationale":"The reader's weakest assumption focused on whether the reproduced SymPoint V2 baseline is faithful and comparably strong. My concern is closely related but more specific and more damaging: the paper's own Table 2 indicates a batch-size mismatch (2/8 vs. 2/6), which directly undercuts the 'identical training settings' premise of the central comparison. This is a concrete, checkable inconsistency rather than an assumption about implementation quality. It does not prove the method is ineffective, but it means the headline numbers are not currently interpretable as a controlled comparison. The reader's conditional verdict is appropriate; the condition should explicitly include rerunning with equal batch size and clarifying the GPU/batch notation. I therefore keep the verdict at CONDITIONAL, with the added specificity that the batch-size discrepancy must be resolved before the empirical claim can be accepted.","tokens_in":11526,"tokens_out":3851,"duration_ms":37087,"concrete_test":"Rerun the PolarSym experiment with the same total batch size as the reproduced SymPoint V2 baseline (e.g., 8, if Table 2 means total batch) under the same 50-epoch, 2-GPU schedule, and report PQ, RQ, SQ, and mIoU. If the gains over 85.42/89.92/94.98/66.56 shrink or vanish, the headline claim fails. Alternatively, if 'GPU / Batch' is not total batch size, provide the exact per-GPU batch sizes for both runs and confirm they are identical.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim is that PolarSym outperforms the reproduced SymPoint V2 baseline by +1.73% PQ, +1.54% RQ, +0.30% SQ, and +4.31% mIoU 'under identical training settings.' However, Table 2 reports GPU/batch as 2/8 for SymPoint V2 (Reproduction) and 2/6 for PolarSym. If these numbers denote the number of GPUs and total batch size, the comparison is not controlled: batch size affects effective learning rate, gradient noise, and optimization trajectory, so the reported gains may partly reflect a batch-size difference rather than the proposed geometric attention. This is not a speculative concern about baseline quality; it is a direct contradiction between the paper's stated protocol and its own table. The abstract, Section 4.2, and Table 2 all use 'identical' language, so the discrepancy is load-bearing. If the 'GPU / Batch' column has a different intended meaning (e.g., per-GPU batch size vs. total batch), that must be clarified; as written, the central comparison is not valid.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PolarSym, a polar-coordinate geometry-aware attention framework for CAD floorplan parsing. It decouples geometric relationships into a direction branch (SF-RoPE) and a distance branch (RDB), fused through a dynamic gating mechanism that modulates the semantic self-attention scores. The authors claim that, under identical training settings, PolarSym outperforms a reproduced SymPoint V2 baseline by +1.73% PQ, +1.54% RQ, +0.30% SQ, and +4.31% mIoU on a public benchmark, while also converging faster. Ablation experiments are presented to show the complementary contributions of the direction and distance branches.","tokens_in":11696,"tokens_out":3518,"duration_ms":34196,"significance":"If the empirical claims hold, PolarSym offers a conceptually interesting and computationally light way to inject explicit geometric priors into Transformer attention for CAD parsing, and the decoupling of direction and distance is a sensible design choice for architectural symmetry. However, the paper currently does not provide code, pre-trained models, or machine-checkable artifacts, and the core empirical validation rests on a single controlled comparison whose protocol is internally inconsistent. The central contribution is therefore plausible but not yet convincingly established.","major_comments":[{"comment":"The abstract, §4.2, and Table 2 all state that PolarSym and the reproduced SymPoint V2 baseline are compared 'under identical training settings,' but Table 2 lists the GPU/Batch column as 2/8 for SymPoint V2 (Reproduction) and 2/6 for PolarSym. If these numbers denote the number of GPUs and total batch size, then the comparison is not controlled: total batch size affects the effective learning rate, gradient noise, and optimization trajectory, so the reported gains may be partly attributable to the batch-size difference rather than to the proposed geometric attention. If 'GPU / Batch' has a different intended meaning, this must be clarified explicitly. As written, the headline improvement of +1.73% PQ and the related claims of 'identical training settings' are invalidated, which is load-bearing for the paper's central assertion.","section":"§4.2, Table 2"},{"comment":"The directional branch is underspecified. Equation (2) defines the rotation angle as θ_k = λ φ(p_i), but the function φ is never defined anywhere in the text. The only description is that 'different attention heads alternately adopt horizontal and vertical coordinates to encode directional information along the two primary axes.' This leaves the actual mapping from a point's coordinates to θ_k ambiguous, making the method non-reproducible. Please specify φ precisely (e.g., φ(p_i) = x_i for some heads and φ(p_i) = y_i for others) and describe exactly how the head alternation is implemented.","section":"§3.2.1, Eq. (2)"},{"comment":"All reported improvements are based on single-run comparisons without any error bars, confidence intervals, or significance tests. The gain in SQ is only +0.30 percentage points, which is within typical run-to-run variance for such models. Without multiple seeds or a significance analysis, the claim that PolarSym consistently outperforms the baseline across all metrics is not statistically supported. Please provide repeated runs (at least three seeds) or otherwise justify the robustness of the reported numeric differences.","section":"§4.3, Tables 2 and 3"},{"comment":"Table 4 and its surrounding text are confusing and appear internally inconsistent. The baseline PQ in Table 3 is 85.420, while the 'Strong Baseline' baseline PQ in Table 4 is 85.21; moreover, the 'Weak Baseline' uses a different geometry implementation ('Standard Axial + Linear Dist') rather than merely a different initialization, yet the text says the comparison involves 'suboptimal initialization.' The relationship between the weak/strong baselines and the earlier ablation settings is unclear, and the reported final PQ values (83.57 vs. 87.15) are compared without explaining how these baselines relate to the SymPoint V2 reproduction in Table 2. Please clarify the experimental protocol for Table 4 or remove it if it does not support a well-defined claim.","section":"§4.3, Table 4"}],"minor_comments":[{"comment":"The paper claims that the gating mechanism adds 'negligible extra computation' but does not report FLOPs, parameter counts, or runtime comparisons. Please provide a quantitative complexity or efficiency analysis.","section":"§3.2.3"},{"comment":"The acronym RDB is inconsistently expanded: Section 2.3 calls it 'Resonance Distance Deviation,' while §3.2.2 calls it 'Resonant Distance Bias.' Please use a single consistent name and abbreviation throughout.","section":"Section 2.3 and §3.2.2"},{"comment":"There is a typo in the section heading 'V ector-Based Methods'; the space after 'V' should be removed.","section":"Section 2.1.2"},{"comment":"The text says 'LFEM' is used but does not define the abbreviation; define it at first use, even if it refers to the SymPoint-V2 Layer Feature Enhancement Module.","section":"§3.1"}],"recommendation":"major_revision","confidential_remarks":"The batch-size discrepancy in Table 2 is the most serious issue because it directly undermines the paper's central 'identical training settings' claim. It is, however, potentially fixable by rerunning experiments at matched batch sizes or by clearly redefining the column. The underspecification of Eq. (2) is also a reproducibility blocker. The paper would benefit from code release and multi-seed evaluation. I would not recommend acceptance in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is actually reasonable: instead of folding all geometry into one positional encoding, PolarSym splits it into direction and distance branches, with split-feature RoPE for orientation and a Fourier-encoded distance bias for long-range radial correspondences. The intuition that symmetric floorplan elements violate the usual monotonic distance-relevance assumption is well-taken, and the gating mechanism to balance the two branches is clean. For a constrained-compute setting, this is a plausible way to add geometric awareness without touching the backbone.\n\nBut the central empirical claim does not hold up as written. The paper says PolarSym beats the reproduced SymPoint V2 baseline under identical training settings, yet Table 2 lists GPU/batch as 2/8 for the baseline and 2/6 for PolarSym. If those are total batch sizes, the comparison is not controlled. That is not a nitpick about baseline quality; it is a direct contradiction between the stated protocol and the paper's own table. The +1.73 PQ delta could be partly an optimization artifact. This needs to be fixed before the results mean anything.\n\nOther soft spots are real but less severe. There is a single run per configuration, no error bars or significance tests, so the 0.30 SQ gain is within noise. Eq. 2 defines theta_k = lambda * phi(p_i) but never specifies phi, so the directional branch is underspecified. The dataset is not named in Section 4.1, which makes reproduction harder. The paper also releases no code. None of these prove the method is wrong, but they make verification impossible.\n\nWhat the paper does well is the literature review and the problem framing. The distinction between direction consistency and radial correspondence is useful, and the authors are honest that their gains are modest. Table 1's comparison to the official SymPoint V2 (250 epochs, 8 GPUs) is also informative, even if not controlled.\n\nOverall: this deserves a serious referee, not a desk reject. The idea is coherent and the work is honest, but the training protocol must be fixed, the baseline re-run at matched batch size, and ideally with multiple seeds. I would not cite it in its current form, but I would read a revised version. If I were the editor, I would send it out with a clear request to address the batch-size discrepancy and to release code.","headline":"The geometric attention idea is sensible and worth a referee, but the paper's headline gains rest on a flawed comparison: Table 2 uses different batch sizes for baseline and PolarSym, contradicting the 'identical training settings' claim.","tokens_in":750,"tokens_out":946,"would_cite":false,"duration_ms":24356,"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":"PolarSym claims that explicitly splitting CAD floorplan geometry into direction and distance—modeled by SF-RoPE and a resonant distance bias inside attention—improves long-range symmetric correspondence and beats a reproduced SymPoint V2…","keywords":["CAD floorplan parsing","geometry-aware attention","polar coordinates","rotary position embedding","distance bias","Transformer","panoptic segmentation","Building Information Modeling"],"falsifier":"Train PolarSym and the reproduced SymPoint V2 baseline under the official SymPoint V2 protocol (250 epochs, eight GPUs, batch size 16) and compare on the same test split; if PolarSym fails to match the official 90.1% PQ or the margin over the baseline shrinks to near zero, the claimed advantage would be shown to depend on the constrained training budget rather than on the geometric attention itself.","tokens_in":11287,"feed_emoji":"📐","tokens_out":9291,"duration_ms":81453,"temperature":0.7,"pith_summary":"PolarSym sets out to show that a Transformer can parse CAD floorplans more accurately if its attention mechanism is given, separately and explicitly, two geometric facts about each pair of points: their direction relative to a building axis and their distance from each other. The paper argues that ordinary semantic attention plus Cartesian positional encoding conflates these factors and tends to restrict attention to local regions, missing long-range correspondences between repeated, mirrored, or symmetric building elements. By adding a directional branch called SF-RoPE and a distance branch called RDB to the attention score, with learned gates balancing them, the method reports gains of +1.73% PQ, +1.54% RQ, +0.30% SQ, and +4.31% mIoU over a reproduced SymPoint V2 baseline under identical 50-epoch, two-GPU training. The authors also report faster convergence—about 74.6% PQ versus 58.1% at epoch 20—and interpret this as showing that explicit geometric modeling improves optimization efficiency, not just model capacity. If correct, the contribution is a low-cost, architecture-preserving way to make Transformers geometry-aware for structured drawings.","feed_headline":"Two-branch polar attention beats CAD floorplan baseline by 1.73 PQ","feed_subtitle":"Splitting direction from distance in attention yields faster, steadier convergence on the same 50-epoch budget.","key_machinery":"The load-bearing mechanism is the polar decomposition of pairwise geometry into a direction term and a distance term, injected additively into attention rather than into feature vectors. SF-RoPE (Split-Feature Rotary Position Embedding) partitions feature channels into two halves and applies cross-half rotation matrices with angle $\\theta_k = \\lambda \\phi(p_i)$, alternating horizontal and vertical coordinate axes across attention heads, producing the directional increment $\\Delta_{\\mathrm{dir}}$. RDB (Resonant Distance Bias) maps Euclidean distance $d_{ij}$ through Fourier features $\\gamma(d) = (d, \\sin(2^0\\pi d), \\cos(2^0\\pi d), \\ldots, \\sin(2^{L-1}\\pi d), \\cos(2^{L-1}\\pi d))$ and a lightweight MLP, allowing several distance peaks so that distant symmetric pairs can receive high bias; this forms the radial branch $B_{\\mathrm{rad}}$. Learnable gates $\\alpha = \\tanh(g_\\alpha)$ and $\\beta = \\tanh(g_\\beta)$ fuse the two into the final score $S = S_{\\mathrm{sem}} + \\alpha\\Delta_{\\mathrm{dir}} + \\beta B_{\\mathrm{rad}}$, preserving the vanilla Transformer architecture while changing only how attention scores are computed.","core_discovery":"The paper's central claim is that replacing unified positional encodings with two independent geometric attention modifiers yields more accurate and faster-converging CAD floorplan parsing. Directional consistency is captured by SF-RoPE, which splits a query or key feature into two halves and rotates them across halves with angles derived from physical coordinates; the radial branch learns a non-monotonic 'resonant' distance bias by Fourier-encoding Euclidean distance and passing it through a small MLP. A dynamic gating mechanism adds these two terms to the semantic attention score through $S = S_{\\mathrm{sem}} + \\alpha \\Delta_{\\mathrm{dir}} + \\beta B_{\\mathrm{rad}}$, with $\\alpha$ and $\\beta$ initialized so direction acts early and distance enters later. The result, under identical constrained training settings, is a consistent improvement over the reproduced SymPoint V2 baseline on PQ, RQ, SQ, and especially mIoU, together with faster early convergence, which the paper interprets as evidence that the geometric prior helps the network establish global structural correspondences early in training.","pith_inferences":["A testable extension is to run PolarSym under the official SymPoint V2 protocol (250 epochs, eight GPUs, batch size 16); if the official 90.1% PQ remains higher than PolarSym's 87.15% under that setting, the paper's claim reduces to an efficiency advantage under constrained budgets, not an accuracy advantage over the official model.","The RDB's Fourier encoding is a generic periodic-basis trick; one could replace it with alternative bases such as learned periodic functions or random Fourier features to test whether the resonant non-monotonic form or simply the extra expressiveness drives the distance-branch gains.","The paper's framing suggests that the gate initializations might be re-tuned per dataset: in mirror-heavy layouts a larger initial $\\alpha$ could be optimal, while in scattered layouts $\\beta$ could be raised earlier; this is a prediction the authors do not test.","Since the method is architecture-agnostic at the attention level, a natural next step would be to pair PolarSym with linear attention or graph-based backbones to see whether the geometric bias compensates for their reduced global receptive fields."],"forward_implications":["Under the same 50-epoch, two-GPU training schedule, applying PolarSym to a reproduced Point Transformer baseline yields higher panoptic and semantic quality, with the largest relative gain in mIoU (+4.31 points).","Because the geometric terms are additive attention modifications, the method can in principle be dropped into any Transformer-based CAD parsing or vector graphics model without retraining the backbone from scratch.","Faster convergence—PolarSym reaches 74.6% PQ at epoch 20 versus 58.1% for the baseline—implies that explicit geometric priors can reduce the epoch budget needed to reach usable parsing quality, addressing a known pain point of point-based Transformers.","Ablation results indicate that direction and distance contribute differently: SF-RoPE alone raises PQ from 85.42% to 86.83% and mIoU by 2.60 points, while adding RDB raises PQ to 87.15% and mIoU to 70.87%, supporting the claim that the two branches are complementary.","If the polar decomposition is the cause of the gains rather than a side effect of extra parameters, the same additive geometric-attention recipe should transfer to other structured domains with repeated symmetric layouts, such as mechanical drawings or indoor scene layouts."],"supporting_citations":[{"why":"Defines SymPoint V2, the official model whose reproduction serves as the primary baseline and whose 250-epoch result anchors the efficiency comparison.","marker":"[LYYZ24]"},{"why":"Introduces SymPoint's point-set representation for panoptic symbol spotting, which PolarSym's Point Transformer backbone extends.","marker":"[LYW+24]"},{"why":"Contributes the FloorplanCAD benchmark and the PanCADNet baseline that the experiments use as the public dataset and one comparison method.","marker":"[FZL+21]"},{"why":"Supplies the rotary position embedding (RoPE) that SF-RoPE modifies by splitting features across halves.","marker":"[SAL+24]"},{"why":"Presents CADTransformer's position-encoding approach, which the paper criticizes as conflating direction and distance and motivates the decoupled design.","marker":"[FCWW22]"},{"why":"Provides CADSpotting, a recent point-based CAD parsing method compared in Table 1 and relevant to the point-representation lineage.","marker":"[YMZ+24]"}],"fun_headline_variants":["Polar attention splits direction and distance to lift CAD parsing by 4.31 mIoU","Faster convergence and 4.31 mIoU gain from polar-aware attention in CAD parsing","Geometry-aware attention boosts CAD floorplan parsing with negligible extra compute","PolarSym: direction and distance attention lifts CAD parsing by 1.73 PQ and faster","Polar attention framework improves CAD parsing PQ, RQ, mIoU with faster convergence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reproduced SymPoint V2 baseline, trained for 50 epochs on two GPUs, is a faithful and reasonably strong implementation of the official SymPoint V2, so the reported gains reflect PolarSym's geometric attention rather than a weakened comparison point.","fun_headline_variants_meta":{"raw":{"variants":["Polar attention splits direction and distance to lift CAD parsing by 4.31 mIoU","Faster convergence and 4.31 mIoU gain from polar-aware attention in CAD parsing","Geometry-aware attention boosts CAD floorplan parsing with negligible extra compute","PolarSym: direction and distance attention lifts CAD parsing by 1.73 PQ and faster","Polar attention framework improves CAD parsing PQ, RQ, mIoU with faster convergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000594,"raw_usage":{"total_tokens":2825,"prompt_tokens":1030,"completion_tokens":1795,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":1682}},"tokens_in":646,"tokens_out":1795,"duration_ms":13370,"temperature":1.0,"reasoning_tokens":1682,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:26:51.351171+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train PolarSym and the reproduced SymPoint V2 baseline under the official SymPoint V2 protocol (250 epochs, eight GPUs, batch size 16) and compare on the same test split; if PolarSym fails to match the official 90.1% PQ or the margin over the baseline shrinks to near zero, the claimed advantage would be shown to depend on the constrained training budget rather than on the geometric attention itself.","supporting_citations":[],"review_version":1}