{"id":"31f0dae9-aa07-40ad-afef-70a5c33e3135","arxiv_id":"2509.03737","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"LayoutGKN uses a differentiable path-based graph kernel over learned room embeddings to rank floor plans, matching LayoutGMN's accuracy at about 20x lower inference cost.","lead":"LayoutGKN is a floor-plan retrieval model that compares floor plans by running a graph kernel on learned room embeddings, avoiding the expensive cross-graph attention of graph matching networks while matching their accuracy. Because room embeddings can be computed once per floor plan, querying a database is about 20 times faster than the matching-network baseline, which matters for real-time search.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported P@k scores are computed inside each model's own top-50 candidate set, so the headline accuracy comparison may not reflect true retrieval quality; full-corpus P@k should be measured.","rationale":"The reader's weakest_assumption was the validity of sGED as a proxy for human-perceived similarity. That is an important task-definition concern, but it applies equally to both methods and is partially addressed by the user study. The P@k protocol is more immediately load-bearing for the empirical comparison: if the protocol is biased, the specific numerical claims of 'comparable or better' and the zero-shot advantage could be artifacts of the evaluation, even under the chosen sGED ground truth. The speed claim is well-supported by the precomputation argument and FLOPs analysis, so the method retains value regardless. The paper is transparent about the protocol in the supplementary, which makes the concern checkable rather than fatal. A conditional verdict remains appropriate, contingent on re-evaluating P@k on the full corpus or with a fixed candidate set.","tokens_in":11968,"tokens_out":14269,"duration_ms":151817,"concrete_test":"Recompute P@5 and P@10 against the full RPLAN and MSD galleries, with relevance defined by a fixed sGED threshold (or by the sGED top-k in the full ranking). Also run a controlled candidate-set version using the same candidate set for all methods (e.g., the union of all methods' top-50, or the sGED top-50). If GKN's P@k advantage over GMN narrows or reverses, the central accuracy claim is not supported as stated; if the gap persists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The main accuracy claim (Table 1: GKN P@5 0.623 vs GMN 0.616 on RPLAN; zero-shot MSD P@5 0.674 vs 0.585) rests on a non-standard P@k protocol. The supplementary 'Precision' section states: 'we take the top-50 results returned by the model for each query. We then re-rank them according to the ground-truth similarity values. Finally, we compute P@k between this ground-truth order and the model's original ranking.' This is not standard top-k retrieval precision over the full database. It measures how well the model orders a candidate set that the model itself selected. If a model's top-50 omits relevant floor plans, those omissions never count against it. A method with good intra-top-50 ordering but poor recall can receive a high P@k, while a method that retrieves all relevant items but ranks them slightly outside the top k can be penalized. Because the top-50 boundary is method-specific, the comparison is not apples-to-apples. This directly threatens the 'comparably or better than LayoutGMN' part of the central claim; the speed contribution is unaffected.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LayoutGKN, a graph-kernel network for floor-plan similarity retrieval. It replaces the cross-graph node-interaction layers of graph matching networks (GMNs) with a differentiable GraphHopper kernel computed on independently learned node embeddings, so that node embeddings can be precomputed offline. Training uses triplets mined with a normalized graph-edit-distance target (sGED), and evaluation is on RPLAN with zero-shot transfer to MSD, comparing triplet accuracy, P@5/P@10, and inference time against LayoutGMN, LayoutGEN, and a non-learning graph kernel. The paper reports comparable or better ranking accuracy than LayoutGMN (RPLAN P@5 0.623 vs 0.616, P@10 0.683 vs 0.675) with roughly a 20x speedup per 10K pairs, and stronger zero-shot P@k on MSD.","tokens_in":12303,"tokens_out":2804,"duration_ms":32369,"significance":"The architectural insight is clear and plausible: cross-graph interactions can be postponed to the similarity function without sacrificing accuracy, enabling precomputation and much faster retrieval. The method is simple, the equations are internally consistent, the loss is differentiable, and the authors provide open code and data. If the ranking evaluation is repaired, this would be a useful contribution to floor-plan retrieval and to graph-similarity learning more broadly. The speed advantage over GMNs is real and well motivated, since GMN embeddings cannot be precomputed. The paper also includes a user study connecting sGED to human judgment, although its scale is limited.","major_comments":[{"comment":"The definition of sGED assumes an exact GED value, but GED is NP-hard and the paper does not state how it is computed (exact algorithm, approximation, or upper bound) or give the maximum graph size. Since sGED is used both for triplet mining and for ground-truth evaluation, an uncontrolled approximation in Eq. (9) could affect all results. Please specify the GED solver and any limits on graph size.","section":"Eq. (9) and GED computation"}],"minor_comments":[{"comment":"The text says 'P@5 and 20' but Table 1 reports P@5 and P@10. Please correct the typo.","section":"Sec. 5.1 / Table 1"},{"comment":"The term 'triplet accuracy' is used without a definition. Please state whether it is the fraction of triplets where the anchor-positive distance is smaller than anchor-negative, or some other criterion.","section":"Sec. 5.1"},{"comment":"The triplet-mining ranges (positive sGED in (0.6,0.9), negative ratio in (0.7,0.9)) are introduced as fixed heuristics. A sentence on the sensitivity of results to these thresholds, or a reference to prior use, would help.","section":"Sec. 4.2"},{"comment":"The shape feature vector is written as s = [cx; cy; w; h; sqrt(a); p/4] in Sec. 3.1 but in the supplementary it is [c(u)_x, c(u)_y, w, h, sqrt(a), p/4]^T. The text in Sec. 3.1 also uses 'lx' and 'ly' inconsistently. Please align notation.","section":"Sec. 3.1"},{"comment":"The qualitative example is helpful, but the caption should specify whether the shown rankings are from the top-50 protocol or from full-corpus ranking. This matters for interpreting the gray-background items.","section":"Fig. 4 caption / Sec. 5.2"}],"recommendation":"major_revision","confidential_remarks":"The core speed contribution is convincing and the method is well-motivated. The ranking evaluation protocol is the main obstacle: the top-50 re-ranking procedure makes the reported P@k numbers non-standard and potentially misleading for the central comparative claim. This is fixable by reporting standard corpus-level metrics. I would also encourage the editor to require the GED computation details, since sGED underpins both training and evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: LayoutGKN's speedup is real and its central architectural idea — moving cross-graph interactions from the GNN encoder to a differentiable graph kernel over learned node embeddings — is sensible. But the paper's ranking evaluation is built on a non-standard P@k protocol that makes the \"comparable or better than LayoutGMN\" claim shaky. The triplet accuracy is the cleaner number and it holds.\n\nWhat's good: the paper is honest and well-engineered. Precomputable embeddings are a genuine practical win for floor-plan retrieval, and 1.8s vs 35.6s per 10k pairs is a meaningful speedup. They include a no-learning graph kernel baseline, ablate hidden dimension and depth, and release code/data. The zero-shot MSD result is reported with a plausible but clearly post-hoc explanation — they acknowledge it.\n\nThe soft spot is the evaluation. As the supplementary explains, for each query they take the model's top-50, re-rank those 50 by ground-truth sGED, and then compute P@k. That is not retrieval precision over the database. It measures how well the model orders the candidate set it chose. If a model's top-50 happens to omit relevant floor plans, those omissions never count against it, and because each model has its own top-50, the comparison isn't apples-to-apples. The published P@5/P@10 numbers could be inflated, and the GKN-vs-GMN ranking gap could change under a standard protocol. The triplet accuracy (97.78 vs 97.74) doesn't have this problem and supports the \"comparable\" claim; the speed claim is unaffected.\n\nThe sGED ground truth is a reasonable choice, but the user study backing it is only 50 queries with acknowledged disagreement — thin, but not a fatal issue. A comparison to existing deep graph kernel baselines (e.g., SimGNN-style methods) would also help to substantiate the novelty, since this is conceptually a deep graph kernel.\n\nOverall: this is a solid, applied paper with one significant methodological flaw in the headline ranking metric. It deserves a serious referee, and the authors should fix that metric before publication. Worth reading for the efficiency idea and as a cautionary example of evaluation design.","headline":"Real speedup, shaky P@k protocol: the ranking claim needs a proper evaluation, but the triplet accuracy and efficiency hold up.","tokens_in":12758,"tokens_out":3630,"would_cite":true,"duration_ms":37051,"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":"LayoutGKN computes floor-plan similarity by learning independent node embeddings and scoring them with a differentiable graph kernel, matching or beating the slower graph matching network while cutting inference time by roughly 20x.","keywords":["floor plan retrieval","graph similarity learning","graph matching networks","graph kernel","graph edit distance","joint embedding architecture","zero-shot generalization","differentiable kernel"],"falsifier":"Conduct a user study with a few hundred floor-plan queries, asking architects to rank near-duplicate plans; if rankings correlate more strongly with mean intersection-over-union than with normalized graph edit distance, the sGED training target that LayoutGKN optimizes would be invalidated, and the reported precision numbers would no longer reflect human-perceived similarity.","tokens_in":11850,"feed_emoji":"⚡","tokens_out":7815,"duration_ms":72558,"temperature":0.7,"pith_summary":"Floor-plan similarity is usually computed with graph matching networks, which interleave cross-graph node interactions throughout the encoder and therefore cannot precompute embeddings, making them slow. LayoutGKN instead learns node embeddings independently and moves all cross-graph node-level interactions into a differentiable path-based graph kernel applied at the end of the pipeline. This preserves the accuracy of node-level correspondence while letting gallery embeddings be precomputed offline. On the RPLAN and MSD benchmarks, the method matches or beats the LayoutGMN baseline on retrieval precision while reducing inference time by about an order of magnitude (roughly 20x per 10K pairs), and it reports stronger zero-shot transfer to MSD.","feed_headline":"Floor-plan search gets 20x faster without losing accuracy","feed_subtitle":"Moving cross-graph matching into a differentiable kernel keeps retrieval quality and boosts zero-shot transfer.","key_machinery":"The load-bearing component is the differentiable GraphHopper path-based graph kernel applied over learned node embeddings. Each node is endowed with a shortest-path histogram matrix M(u) that counts how often the node appears at each position along shortest paths of each length; the kernel then computes kG(G1, G2) = Σ_{u∈V1} Σ_{v∈V2} ⟨M(u), M(v)⟩ · k_node(h_u, h_v), where k_node is a Gaussian kernel on the learned embeddings. This placement lets the encoder process each graph in isolation while the similarity function still performs a form of node-level matching. A normalized version of the kernel is used in a triplet loss, and because both the kernel and the embeddings are differentiable, t","core_discovery":"The central claim is that explicit cross-graph node-level interactions inside graph encoders are not necessary for effective graph similarity; they can be postponed to a differentiable kernel that scores the final independently learned node embeddings. LayoutGKN embeds each floor plan with a shared message-passing network, then computes similarity with a normalized GraphHopper kernel over the node embeddings. The kernel weighs node pairs by the overlap of their shortest-path histograms, so topological correspondences are still captured at scoring time, but the two graphs never interact during encoding. Because embeddings are computed independently, they can be precomputed for a gallery, whic","pith_inferences":["The same postpone-and-kernel strategy could generalize to other graph comparison tasks, such as molecular similarity or program syntax trees, where cross-graph attention is the computational bottleneck.","Because gallery embeddings are precomputed, LayoutGKN could be combined with approximate nearest-neighbor indexing to scale to millions of floor plans without altering the training objective.","A testable extension is replacing the fixed shortest-path histogram with a learned topology descriptor, which might adapt the kernel to datasets with different edge semantics.","The method makes incremental database updates cheap: adding a new floor plan only requires computing its own embedding and kernel self-normalizer, not pairwise interactions with every existing plan."],"forward_implications":["Real-time retrieval over tens of thousands of floor plans becomes practical: per-query scoring costs about 1–2 seconds per 10K pairs because gallery embeddings are precomputed.","The node embeddings can be much smaller without a large accuracy drop, since the kernel, rather than the embedding dimension, carries the topological information.","The zero-shot result on MSD indicates that the kernel imposes a structural prior that transfers across floor-plan datasets, although the paper notes part of the gain may come from dataset composition.","Training and evaluation under normalized graph edit distance, rather than IoU, aligns better with human ranking, at least in the reported user study."],"supporting_citations":[{"why":"Defines the graph matching network architecture that LayoutGKN compares against and whose cross-graph interactions the method removes.","marker":"[15]"},{"why":"Supplies the floor-plan-specific LayoutGMN baseline and the joint-embedding setup that the paper must match or beat.","marker":"[21]"},{"why":"Introduces the GraphHopper path-based graph kernel used as the differentiable similarity function.","marker":"[6]"},{"why":"Defines graph edit distance, the basis of the sGED training and evaluation target.","marker":"[25]"},{"why":"Provides the RPLAN dataset used for training and testing.","marker":"[35]"},{"why":"Provides the MSD dataset used for zero-shot evaluation.","marker":"[32]"},{"why":"Supplies the triplet network loss formulation used for training.","marker":"[11]"}],"fun_headline_variants":["Postponing graph matching speeds floor-plan search 20x","No cross-graph encoding? Floor-plan search still accurate, 20x faster","LayoutGKN: differentiable kernel cuts floor-plan matching cost","Faster floor-plan retrieval by deferring graph interactions","Graph similarity without early interaction: 20x speedup"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The whole evaluation rests on the premise that the normalized graph edit distance (sGED) used as ground truth agrees with how humans perceive floor-plan spatial similarity; if that link is wrong, the reported ranking quality and zero-shot gains lose their meaning, and the user study backing it up is small (50 queries) with acknowledged disagreement.","fun_headline_variants_meta":{"raw":{"variants":["Postponing graph matching speeds floor-plan search 20x","No cross-graph encoding? Floor-plan search still accurate, 20x faster","LayoutGKN: differentiable kernel cuts floor-plan matching cost","Faster floor-plan retrieval by deferring graph interactions","Graph similarity without early interaction: 20x speedup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1162,"prompt_tokens":643,"completion_tokens":519,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":387,"completion_tokens_details":{"reasoning_tokens":433}},"tokens_in":387,"tokens_out":519,"duration_ms":5040,"temperature":1.0,"reasoning_tokens":433,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:43:17.310322+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Conduct a user study with a few hundred floor-plan queries, asking architects to rank near-duplicate plans; if rankings correlate more strongly with mean intersection-over-union than with normalized graph edit distance, the sGED training target that LayoutGKN optimizes would be invalidated, and the reported precision numbers would no longer reflect human-perceived similarity.","supporting_citations":[{"cited_title":"LayoutGMN: Neural Graph Match- ing for Structural Layout Similarity","cited_arxiv_id":null,"evidence_quote":"Supplies the floor-plan-specific LayoutGMN baseline and the joint-embedding setup that the paper must match or beat."},{"cited_title":"Scalable Kernels for Graphs with Continuous Attributes","cited_arxiv_id":null,"evidence_quote":"Introduces the GraphHopper path-based graph kernel used as the differentiable similarity function."},{"cited_title":"A Distance Measure Be- tween Attributed Relational Graphs for Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Defines graph edit distance, the basis of the sGED training and evaluation target."},{"cited_title":"Data-driven Interior Plan Genera- tion for Residential Buildings","cited_arxiv_id":null,"evidence_quote":"Provides the RPLAN dataset used for training and testing."},{"cited_title":"MSD: A Benchmark Dataset for Floor Plan Generation of Building Complexes","cited_arxiv_id":null,"evidence_quote":"Provides the MSD dataset used for zero-shot evaluation."},{"cited_title":"Deep Metric Learning Using Triplet Network","cited_arxiv_id":null,"evidence_quote":"Supplies the triplet network loss formulation used for training."}],"review_version":1}