{"id":"9f78e7c5-aab5-442e-90f0-6cc86acf9e3e","arxiv_id":"2505.13539","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"EuLearn is a benchmark of 3,300 embedded surfaces with uniformly distributed genus (0 to 10), plus a graph-sampling method and adjacency-aware PointNet and Transformer variants that reach about 0.8 genus-classification accuracy.","lead":"The authors built a 3D dataset of 3,300 surfaces where the number of holes is evenly balanced from zero to ten, and showed that standard AI models often fail to count the holes. A new sampling method that keeps the connections between points helps the models reach about 80% accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Genus labels are asserted, not verified, and the stated construction cannot yield genus 0; an Euler characteristic audit of all released meshes would settle it.","rationale":"Good-faith summary: the paper would be a useful benchmark if the labels are correct; the adjacency-aware accuracy gains are plausible because mesh connectivity encodes cycle structure, and the public code and data are positive evidence. My concern is not with novelty or the ML comparison but with the unverified and internally inconsistent genus guarantee. The proof of Theorem 3 is circular ('By construction' with no reported χ counts), the explicit formula in Section 2.1 excludes genus 0, and coarse Marching Cubes can change topology in principle. These are checkable: the dataset is public and the audit is cheap. I do not think this warrants rejection; a condition that the authors release the verification script and results and clarify genus-0 generation is appropriate. The reader's weakest assumption already named this gap, and I agree with that assessment, so the existing CONDITIONAL verdict stands.","tokens_in":24824,"tokens_out":10757,"duration_ms":122665,"concrete_test":"Download the released HuggingFace repository and run an independent audit: for every .stl mesh compute V, E, F (trimesh or equivalent), verify orientability and χ=V−E+F=2−2g against the label; also reproduce one genus-0 instance from its listed parameters or identify an explicit sphere-generation code path. If all 3,300 meshes pass and a genus-0 construction exists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—uniform genus labels from 0 to 10—depends on Theorem 3, whose proof is 'By construction'. That support is internally inconsistent: Section 2.1 states 'n self-intersections result in a surface with genus n+1', so any closed-curve thickening has genus at least 1; the paper never describes how the 300 genus-0 instances are made (Appendix B lists 15 'singular knot types' per genus including genus 0). The intermediate operations (distance-field thickening, 100^3 discretization, Marching Cubes) are asserted to preserve genus, but no Euler characteristic counts are reported for the released meshes, and the datasheet's 'no errors' claim (Appendix J, item 9) is not backed by a verification step. If any mesh has χ ≠ 2−2g, the uniform distribution is mislabeled and Table 1's accuracy numbers lose their target labels.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces EuLearn, a synthetic dataset of 3,300 embedded triangulated surfaces, with 300 instances claimed for each genus from 0 to 10. Each surface is built by thickening a closed, possibly self-intersecting curve, discretizing the resulting scalar field on a 100^3 grid, and extracting a mesh via Marching Cubes; each instance is shipped as a mesh, a smoothed mesh, and a scalar field. The authors also propose a graph-based point sampling scheme and two adjacency-informed architectures (GS PointNet and GS Attention), and report that these reach about 0.79-0.81 accuracy on an 11-class genus classification task, compared with 0.10-0.63 for vanilla baselines. The central theoretical claim is Theorem 3, that the genus distribution in EuLearn is uniform, with a proof stated as 'by construction.'","tokens_in":24961,"tokens_out":6852,"duration_ms":73456,"significance":"If the uniform-genus guarantee were actually verified, EuLearn would fill a real gap: existing 3D datasets are topologically unbalanced or have ill-defined genus, and a controlled benchmark for topological classification could be useful for geometric deep learning. The paper also contributes a public dataset, a reproducible generation workflow, and a non-Euclidean sampling idea that demonstrably changes model behavior. However, the significance is currently conditional. The proof of the main theorem is not a proof, and the construction as stated cannot produce genus-0 surfaces. The experimental comparison is also presented without repeated-run statistics, and the adjacency matrix input makes the classification task partly a graph-topology task rather than a raw point-cloud task. These issues are fixable within the manuscript's scope, but they need to be addressed before the dataset guarantee can be accepted.","major_comments":[{"comment":"The stated construction rule, 'n self-intersections result in a surface with genus n+1' (Section 2.1), is inconsistent with the presence of 300 genus-0 instances. A thickened simple closed curve has a torus boundary, hence genus 1, so the family of thickened closed curves described in the paper cannot produce genus-0 surfaces. Appendix B still lists 15 singular knot types per genus, including genus 0, but no construction for the sphere-like instances is given. Please describe explicitly how the genus-0 instances are generated, or revise the rule linking self-intersections to genus; this is load-bearing for the uniform-distribution claim.","section":"§2.1, Appendix B"},{"comment":"Theorem 3 is not proved. The proof says only 'By construction' and states that the Euler characteristic 'we may verify,' but no actual Euler characteristic counts, per-instance audit, or verification script are reported anywhere in the manuscript. The datasheet (Appendix J, item 9) asserts 'None that we are aware of' without an audit, and item 17 admits that some parameterizations were discarded because the genus was altered when the radius exceeded the reach. Please provide a complete per-instance audit of all 3,300 meshes, reporting chi = 2 - 2g, and release the verification code; otherwise the uniform genus labels are asserted rather than established.","section":"Appendix A.3, Appendix J"},{"comment":"The graph sampling procedure is claimed to 'preserve the topological structure of the surface' and is described as 'a fine grained coarsening of the graph itself,' but no check is reported that the 3,000-vertex sampled graph has the same Euler characteristic as the original mesh. Since all model inputs are sampled graphs, any topology-changing sample would invalidate the classification labels. Please report Euler characteristic or genus before and after sampling across the dataset, and add a safeguard that rejects samples that change the genus.","section":"§3.1, Algorithm 2"},{"comment":"All reported accuracies, precisions, recalls, and F1 scores come from a single run, with no standard deviations or confidence intervals. The sensitivity analysis in Appendix I.2 reports that the random seed is the most influential hyperparameter (relative importance 0.45), so the observed gaps between models could be sensitive to initialization. Please report repeated-seed means and standard deviations for all methods. In addition, the per-class tables (Tables 6 and 7) show F1 scores as low as 0.24 (genus 8, GS Attention) and recall as low as 0.30 (genus 5, GS PointNet), so the average accuracy claim overstates the degree to which the models 'recognize' all genera.","section":"§3.3, Table 1, Appendix I.2"},{"comment":"The comparison between vanilla models and the GS models conflates two different inputs: the GS models receive the mesh adjacency matrix, which directly encodes the combinatorial structure of the triangulation (edge count, and with coordinates, much of the face structure). Since the target is the topological genus, a large part of the performance gain may come from the input containing the answer in combinatorial form rather than from the models learning geometry. This does not invalidate the experiments, but the paper should reframe the task as a graph-topology-informed classification benchmark and should include a control, for example k-NN graphs or permuted adjacency matrices, to quantify how much of the gain is due to genuine mesh connectivity rather than the raw embedding coordinates.","section":"§3.2, §D.4, §D.6"}],"minor_comments":[{"comment":"The text says the 20 sinusoidal variations 'by construction, have the same isotopy type,' while the datasheet says some variations were discarded 'because the genus was altered due to the contact of the surface with itself when the radius exceeded the reach.' Please reconcile these two statements.","section":"Appendix B.4 vs Appendix J item 17"},{"comment":"The proofs of Theorems 1 and 2 are described as 'By computation,' but the computations, the version of Trimesh, and the exact counting procedures are not fully specified; please include a reproducibility script or more detail, since these theorems are used to motivate the dataset.","section":"Appendix A.1, A.2"},{"comment":"There are numerous typos and minor errors, including 'hi+i' and 'comutative' in the FNO and edge-convolution definitions, 'PoitNet++' in Appendix D.3, 'PointNett++' in Section 3.1, and 'presernts' in Appendix I.2; a careful proofreading pass is needed.","section":"Throughout"},{"comment":"The definitions are numbered as Definition 3.1 through 3.7 even though they appear across Sections 2 and 3 and in the appendices; please renumber them consistently with the section in which they appear.","section":"Appendix D, Definitions"},{"comment":"The datasheet says the dataset 'will be made available upon submission of the paper,' while the abstract and datasheet item 27 give a HuggingFace repository URL; please clarify the current release status.","section":"Appendix J, item 33"}],"recommendation":"major_revision","confidential_remarks":"The paper's core dataset guarantee is currently unsupported by the stated construction and proof, and the experiments lack repeated-run statistics. However, the identified issues are addressable within the manuscript's scope: adding a per-instance Euler characteristic audit, clarifying the genus-0 construction, and adding multi-seed results would substantially strengthen the paper. I would not reject outright, because the proposed dataset and the adjacency-preserving sampling idea have genuine value for the geometry-processing and topological-machine-learning communities."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my take on EuLearn. The core idea is genuinely useful: a 3D surface dataset with exactly 300 instances per genus from 0 to 10, all watertight meshes, publicly released. That fills a real gap: FAUST is all spheres; ShapeNet, Thingi10K, ABC, SimJEB are all skewed or topologically messy. The graph-sampling algorithm and the adjacency-aware PointNet/attention variants are reasonable contributions too, and the experimental contrast (0.49–0.63 for vanilla methods vs ~0.8 for graph-sampled ones) is a useful data point for people building topology-aware models.\n\nBut the paper as written does not support its main theorem. The construction section says 'n self-intersections result in a surface with genus n+1,' which implies a minimum genus of 1. Yet Appendix B lists 15 singular knot types for genus 0 and Figure 1 shows spheres. The paper never explains how a genus-0 surface is produced by thickening a closed curve. Thickening a simple closed curve gives a torus, not a sphere. This is not a minor omission; it is an internal contradiction in the central construction. Theorem 3, whose proof is 'By construction,' therefore rests on an inconsistency.\n\nRelatedly, there is no verification step. For a dataset whose labels are topological genera, the paper should report an Euler characteristic audit—at minimum, χ = V − E + F for every mesh, or for a random sample with confidence bounds. The datasheet says 'no errors' but nothing in the pipeline checks that Marching Cubes at 100^3 preserves the genus of the thickened curve. The stress-test note is right: a simple script would settle it.\n\nTwo experimental soft spots. First, the train/test split is not described. If the 20 sinusoidal variations of the same seed knot are split randomly across train and test, that is leakage, because those surfaces share the same isotopy type and similar geometry. The split should be by seed knot. Second, the main table reports a single run with no error bars; the sensitivity analysis mentions seed variance but doesn't translate into confidence intervals for Table 1.\n\nI also want to flag the circularity: for a connected orientable triangulated surface, the adjacency matrix determines the Euler characteristic, so GS models have the answer baked into the input. That's fine for a benchmark, but the paper should not oversell it as evidence that these architectures 'understand' topology.\n\nBottom line: the dataset might be exactly what the geometric deep learning community needs, and the flaws are fixable. Send this to referees, with fingers crossed that a careful one will ask for the Euler audit and a split-by-knot protocol. As it stands, I wouldn't cite it for the uniform-genus claim until the genus-0 construction is explained and verified.","headline":"A genuinely useful uniform-genus dataset, but the construction as written cannot produce genus 0 and the labels are never checked; fixable but not ready to cite.","tokens_in":25539,"tokens_out":5868,"would_cite":false,"duration_ms":58142,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["57M25","68T07","65D18"],"pacs":[],"model":"deepseek-v4-flash","headline":"EuLearn, a new dataset of 3,300 watertight surfaces with uniformly distributed genus from 0 to 10, shows that feeding adjacency information to PointNet and attention layers lifts genus-classification accuracy from at most 0.63 to about…","keywords":["Euler characteristic","genus classification","singular knots","3D surface dataset","graph sampling","PointNet","transformer attention","Marching Cubes"],"falsifier":"Compute $V - E + F$ on the 3,000-point sampled graphs of EuLearn and compare with the mesh genus; any mismatch falsifies the coarsening claim. Also check whether 300 distinct genus-0 meshes exist in the released files and whether their construction is documented, since the 'n self-intersections → genus n+1' rule allows no genus-0 case.","tokens_in":24616,"feed_emoji":"🪢","tokens_out":6614,"duration_ms":57846,"temperature":0.7,"pith_summary":"This paper introduces EuLearn, a dataset of 3,300 watertight, randomly embedded surfaces in 3D space with exactly 300 examples for each genus from 0 to 10, so that topological type is uniformly distributed for the first time in a surface dataset. The surfaces are built by thickening self-intersecting Fourier knots (Lissajous and Fibonacci families) into tubular neighborhoods and discretizing them with Marching Cubes; the number of self-intersections determines the genus. The authors report that standard 3D neural architectures—PointNet, PointNet++, FNO, DGCNN, and a transformer—classify genus poorly on this data, reaching at most 0.63 accuracy, whereas their adjacency-preserving graph-sampling method combined with adjacency-informed PointNet and attention layers reaches about 0.79–0.81 accuracy. The point is to expose and help close a blind spot: current deep learning tools do not reliably recognize topological structure, and a controlled, topologically balanced dataset makes that measurable.","feed_headline":"Mesh adjacency lifts hole-count accuracy to 0.81","feed_subtitle":"The EuLearn dataset gives every genus from 0 to 10 equal weight, exposing what vanilla point-cloud networks miss.","key_machinery":"The load-bearing construction is a pipeline: parameterize a singular knot (Lissajous or Fibonacci), detect its self-intersections by minimizing pair distances with differential evolution, thicken the curve into a tubular neighborhood whose radius is kept below the surface reach so distinct regions do not merge, discretize the signed-distance scalar field on a $100^3$ grid, and extract the surface with Marching Cubes. The identity that carries the genus control is the Euler characteristic $\\chi = V - E + F = 2 - 2g$, with $n$ self-intersections producing genus $n+1$. For learning, the central device is graph sampling (Algorithm 2): starting from the mesh triangulation, it iteratively removes a random subset of neighbors of each traversed node and reconnects the retained node to the removed nodes' neighbors, so a 3,000-point sample inherits the adjacency structure of the original mesh. The adapted layers, GS PointNet and GS attention, replace Euclidean-neighbor aggregation with aggregation over these adjacency-defined neighbors, and the attention variant drops the learned linear projections, using a masked softmax over raw point dot products.","core_discovery":"EuLearn is claimed to be the first 3D surface dataset whose topological types are both well-defined and uniformly distributed: by construction, every instance is a single watertight, compact, orientable triangulated surface, and the genus is uniform across the dataset (Theorem 3). Each surface is the boundary of a tubular neighborhood of a closed, self-intersecting curve, with each self-intersection adding one handle; the genus is then computed from the Euler characteristic $\\chi = V - E + F = 2 - 2g$. The paper further claims that when the point cloud is sampled by a graph-coarsening procedure that preserves adjacency (Algorithm 2), and when PointNet and attention layers aggregate over mesh neighbors rather than over a Euclidean ball, genus classification accuracy rises from 0.10–0.63 for vanilla baselines to 0.79–0.81, with attention reaching 0.81. This is presented as evidence that topological information in the adjacency structure is what current architectures lack.","pith_inferences":["If genus is truly preserved by the coarsening, the same pipeline could generate larger datasets with arbitrary genus ranges (including $g \\geq 11$) by adding more Fourier-knot parameter families; the paper's data-generation methodology is designed to be reused.","The strong role of adjacency suggests a natural experiment the paper does not run: feeding the same 3,000-point samples to a model that uses only geometric features (e.g., persistent-homology descriptors) should change accuracy in a way that isolates what topological information adjacency actually encodes.","The stated rule 'n self-intersections result in a surface with genus n+1' implies a minimum genus of 1; if genus-0 surfaces come from a separate construction (e.g., non-singular knots), the paper's description leaves that unspecified, and a reader should check the released dataset metadata for how the 300 genus-0 instances were generated.","The sensitivity analysis finding that random seed is the most influential hyperparameter suggests the reported 0.79–0.81 accuracies may have a few-point spread across seeds; re-running with a fixed seed list would give a tighter performance estimate."],"forward_implications":["If the construction works as claimed, researchers gain a benchmark where genus is a controlled label, so failures in topology recognition become attributable to the model rather than to dataset imbalance.","The adjacency-preserving sampling method is a standalone contribution: any point-cloud network can consume EuLearn's 3,000-point graphs, and the reported gains suggest that retaining mesh edges matters more than the specific architecture.","The dataset's scalar-field files (100³ grids per surface) enable testing of FNO-style and other field-based models on the same topology classification task.","The reported accuracy of about 0.8 on 11 balanced classes is a concrete target: any future model can be compared against it as a baseline on EuLearn."],"supporting_citations":[{"why":"Provides the Marching Cubes algorithm used to discretize the tubular neighborhood into a watertight mesh.","marker":"[44]"},{"why":"Supplies the definition of reach, the bound that keeps the thickening radius from merging separate regions and changing the genus.","marker":"[12]"},{"why":"Defines the Fourier-knot family from which the Lissajous and Fibonacci curve families are taken.","marker":"[23]"},{"why":"Defines Lissajous knots, one of the two curve families used to generate the dataset.","marker":"[5]"},{"why":"Gives the phase restrictions that let Lissajous knots avoid unwanted self-intersections, so the number of self-intersections controls the genus.","marker":"[26]"},{"why":"Supplies the PointNet architecture whose adjacency-informed variant reaches 0.79 accuracy and serves as a vanilla baseline at 0.49.","marker":"[34]"},{"why":"Provides the strongest vanilla baseline (0.63 accuracy) and the ball-sampling idea adapted by the graph-sampling algorithm.","marker":"[35]"},{"why":"Supplies the transformer/attention baseline architecture and the Noam optimizer used in all experiments.","marker":"[40]"},{"why":"Provides the DGCNN dynamic graph convolution baseline with 0.16 accuracy.","marker":"[43]"},{"why":"Supplies the Fourier Neural Operator baseline with 0.20 accuracy.","marker":"[29]"}],"fun_headline_variants":["Topology-aware nets jump genus accuracy to 0.81","Knot-built surfaces trip up vanilla point cloud nets","Adjacency graph boosts hole counting in 3D nets","Graph-aware PointNet masters genus, hits 0.81"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline preserves the genus of every instance: thickening the curve, discretizing it on a 100³ grid, and graph-sampling it down to 3,000 points never changes the Euler characteristic, and the genus-0 class is actually produced by the stated construction.","fun_headline_variants_meta":{"raw":{"variants":["Topology-aware nets jump genus accuracy to 0.81","Knot-built surfaces trip up vanilla point cloud nets","Adjacency graph boosts hole counting in 3D nets","Graph-aware PointNet masters genus, hits 0.81"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1273,"prompt_tokens":889,"completion_tokens":384,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":316}},"tokens_in":505,"tokens_out":384,"duration_ms":4596,"temperature":1.0,"reasoning_tokens":316,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:32:38.447167+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $V - E + F$ on the 3,000-point sampled graphs of EuLearn and compare with the mesh genus; any mismatch falsifies the coarsening claim. Also check whether 300 distinct genus-0 meshes exist in the released files and whether their construction is documented, since the 'n self-intersections → genus n+1' rule allows no genus-0 case.","supporting_citations":[{"cited_title":"CRC Press, 2013","cited_arxiv_id":null,"evidence_quote":"Provides the Marching Cubes algorithm used to discretize the tubular neighborhood into a watertight mesh."},{"cited_title":"Kauffman","cited_arxiv_id":null,"evidence_quote":"Defines the Fourier-knot family from which the Lissajous and Fibonacci curve families are taken."},{"cited_title":"There are infinitely many Lissajous knots.manuscripta mathematica, 93:29–37, 1997","cited_arxiv_id":null,"evidence_quote":"Gives the phase restrictions that let Lissajous knots avoid unwanted self-intersections, so the number of self-intersections controls the genus."},{"cited_title":"Attention is all you need","cited_arxiv_id":null,"evidence_quote":"Supplies the transformer/attention baseline architecture and the Noam optimizer used in all experiments."},{"cited_title":"Sarma, Michael M","cited_arxiv_id":null,"evidence_quote":"Provides the DGCNN dynamic graph convolution baseline with 0.16 accuracy."}],"review_version":1}