{"id":"0125ee34-ad67-48ce-81c0-6f90665b84d8","arxiv_id":"2505.12880","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Point clouds are lifted to Anti-de Sitter space and message passing uses the AdS proper distance, producing a network approximately equivariant under the full conformal group.","lead":"AdS-GNN is a graph neural network whose connections are built from distances in Anti-de Sitter space, making the network react to conformal transformations of the input points in a controlled way. It combines a known physics idea, the AdS/CFT correspondence, with message-passing machine learning to handle scale changes and angle-preserving deformations in point clouds.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Full conformal equivariance is not delivered by the lift; the decisive SCT part rests on one MNIST augmentation test, leaving the central claim empirically unverified in the Ising setting.","rationale":"The reader's weakest_assumption identifies two issues: the unproved z-limit and SCT breaking. I agree that SCT breaking is the load-bearing one; the z-limit is not the bottleneck. For distinct points, the Galperin construction has a well-behaved limit, as Eq. (24) shows for N=2, and the averaging formula in Appendix A.2 gives finite limits for generic configurations; a missing proof is addressable and does not threaten the central claim. The real gap is that the headline claim asserts full conformal equivariance, while the model delivers equivariance only under the similarity subgroup, with SCT equivariance approximate and unquantified. The paper is candid about this limitation, and the geometric machinery in Appendix D is careful and self-contained; the MNIST test is suggestive but insufficient because the central physics experiment (Ising) never applies an SCT. If the concrete test above shows small error, conditional acceptance is justified and the paper makes a useful contribution. If the test shows large error, the contribution should be reframed as a scale/rotation/translation-equivariant graph network with an AdS structure rather than a conformally equivariant one. I therefore keep the reader's CONDITIONAL verdict unchanged, while requiring the quantitative SCT test and a stated z0 as conditions for acceptance.","tokens_in":33415,"tokens_out":11645,"duration_ms":127647,"concrete_test":"Measure the equivariance error on the Ising task: for a fixed trained AdS-GNN, generate SCT parameters b of varying norm and direction, compute the relative L2 error between Pred(g·{x_i}) and the exactly transformed ground-truth correlation function (Eqs. 29–30), and repeat for z0 ∈ {1e-2, 1e-4, 1e-6} and for N=4 and N=16. If the error grows with ||b|| or exceeds ~5–10% for any b in the sampled range, the 'mild breaking' assumption fails and the central claim should be downgraded to similarity equivariance. Also report the same curve for SuperPixel MNIST with k_lift and z0 explicitly stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the architecture is 'equivariant under general conformal transformations' (Abstract; Section 1) is not what is constructed. The lift in Algorithm 1 maps {x_i} to {(x_i, ẑ_i)} via Eq. (11); this map commutes with translations and rotations, and with scalings only in the z0→0 limit, and it does not commute with special conformal transformations (SCTs). Section 4.1 states that the breaking is 'gentle' and Section 6.1 concedes it. The empirical basis for 'mild' is a single MNIST experiment (Fig. 4) with one SCT family b=(0,b2), and no stated z0, k_lift, or error magnitude. The Ising experiments (Table 2, Figs. 5–6), which are the strongest empirical evidence, do not exercise SCTs at all: they test scale covariance, which is already enforced by the readout Eq. (17) and the limiting z_i proportional to nearest-neighbour separation. Hence the unique new capability of the model (SCT equivariance) is untested in the task where the physics claims are made. If the SCT breaking is not actually small there, the model reduces to a similarity-equivariant GNN, and the advertised conformal equivariance is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces AdS-GNN, a graph neural network for point clouds that lifts input points from Euclidean space R^d to Anti de Sitter space AdS_{d+1} using a per-point scale coordinate derived from the local neighborhood via an AdS center-of-mass computation (Algorithm 1). Message passing is then performed on the lifted points using the AdS proper distance, which is invariant under the isometry group PO(d+1,1) of AdS, the same group as the global conformal group of R^d. The authors state that the resulting network is equivariant under general conformal transformations, with exact invariance under translations, rotations, and scalings, and only mild breaking under special conformal transformations (SCTs) introduced by the lift. They validate the model on superpixel MNIST, shape segmentation, PascalVOC, and on predicting N-point correlation functions of the 2D Ising CFT, where the learned conformal dimensions are close to the exact values Delta_sigma=1/8 and Delta_epsilon=1. A substantial appendix develops the conformal geometry background and details of the AdS construction.","tokens_in":33668,"tokens_out":8481,"duration_ms":93499,"significance":"If the approximate conformal equivariance claim is accepted, the paper makes a useful contribution: it provides a simple, computationally cheap message-passing architecture that is exactly isometry-equivariant on AdS and that demonstrably generalizes under scaling and across system sizes in the Ising task. The clean two-point-function argument and the interpretable extraction of conformal dimensions are attractive and go beyond standard benchmark reporting. The mathematical appendix is self-contained and grounds the construction in standard results on the conformal group and AdS geometry. However, the central novelty advertised in the abstract and introduction—equivariance under special conformal transformations—is not as strongly supported as the rest of the paper. The Ising experiments, which are the most physically meaningful evidence, test scale covariance rather than SCT equivariance, and the only SCT experiment is a single MNIST augmentation test with incomplete reporting. The paper is therefore of genuine interest but needs substantial qualification and additional evidence before the central claim can be accepted at face value.","major_comments":[{"comment":"The paper's headline claim of equivariance under general conformal transformations is stronger than what is actually constructed. Algorithm 1 and the readout in Eqs. (12), (15), and (17) are exactly equivariant only under translations, rotations, and scalings; the lift is explicitly stated to break special conformal transformations, and Section 6.1 concedes this. The abstract and the sentences in Sections 1 and 4.2 that describe a 'conformal group equivariant GNN' should therefore be qualified as approximate equivariance under the full conformal group. This is not merely a wording issue: SCT equivariance is the unique new capability of the model relative to scale-equivariant networks, so the present formulation overstates the contribution.","section":"Abstract, Section 1, Section 4.1, Section 6.1"},{"comment":"The empirical evidence that the SCT breaking is 'mild' is insufficient for the weight placed on it. Figure 4 reports a single MNIST experiment with one one-parameter family b=(0,b2) of special conformal transformations, and the text does not state the values of b used, the error magnitude, or the hyperparameters z0, k_lift, and k_con for that run. More importantly, the Ising experiments in Section 5.2, where the physics claims are made, never apply SCTs to the point sets; they test scale covariance, which is already enforced by the readout structure of Eq. (17). Please add quantitative SCT equivariance error measurements on the point-cloud tasks (including the Ising correlation task), or explicitly restrict the claims to approximate scale-and-rotation equivariance with an empirical statement of the SCT error.","section":"Section 5.1, Figure 4, Section 5.2"},{"comment":"The claim that the AdS center-of-mass coordinate \\hat z_i has a finite, scale-covariant limit as z0->0 is asserted without proof and depends on the KNN convention in Algorithm 1. For the value k_lift=1 used in all Ising experiments (Section B.3), the behavior is ambiguous: if KNN includes the point itself, then \\hat z_i = z0 and the z-coordinate does not encode any local scale; if KNN excludes the point itself, this convention must be stated. Please specify the KNN convention, provide a derivation or argument for the finite limit for general k_lift, and report the resulting range of \\hat z_i values. This matters because the scale-covariance of the readout in Eqs. (12), (15), and (17) relies on \\hat z_i transforming as z -> lambda z under a global scaling.","section":"Section 4.1, Algorithm 1, Section B.3"}],"minor_comments":[{"comment":"The text refers to 'Figure ??' when describing the two connected components of the hyperboloid; this is a missing reference and should be fixed.","section":"Section 3.2"},{"comment":"There is a typo in 'how a a model trained on a given number of nodes' — the duplicated article should be removed.","section":"Section B.2"},{"comment":"The figure would be much more informative with labeled axes, the range of the SCT parameter b, and error bars or the number of trials; please add these details.","section":"Figure 4"},{"comment":"The sentence 'which yields an efficient conformal group equivariant GNN' should be adjusted to 'approximately conformally equivariant GNN' to match the acknowledged SCT breaking in the lift.","section":"Section 4.2"},{"comment":"The expression log(\\hat z_i) uses a dimensionful quantity; if a reference scale or normalization is intended, it should be stated explicitly, especially because the logarithmic term is central to the learned conformal dimensions.","section":"Eq. (17)"}],"recommendation":"major_revision","confidential_remarks":"The main reason for major revision is the mismatch between the advertised full conformal equivariance and the actual approximate construction, combined with weak empirical support for the SCT part. The underlying idea is solid and the Ising results are interesting, so I do not see grounds for rejection; the issues should be fixable by rephrasing the claims and adding targeted SCT experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the AdS lift: point clouds get a per-point z from a local center-of-mass computation, and message passing is conditioned on the AdS proper distance. That construction is clean, the geometry in the appendices is standard and worked through carefully, and the copy of isometry-equivariant message passing onto AdS is legitimate. The Ising experiments are the strongest part: the learned conformal dimensions come out close to 1/8 and 1, and the generalization across N and outside the training domain is real evidence that the model has picked up the scale structure. For that part of the paper I have no serious objection.\n\nThe soft spots are concentrated in the claim hierarchy. The abstract says equivariant under general conformal transformations, but the lift itself is equivariant only under translations, rotations, and scaling; special conformal transformations are broken by construction, as the authors concede in Sections 4.1 and 6.1. The GNN on the lifted cloud is exactly isometry-equivariant on AdS, which is a real property, but it is not the advertised property. The paper is candid about this, which I credit, but the abstract still overstates what is delivered.\n\nMy main empirical quarrel is that the only SCT test is one MNIST augmentation experiment with a single one-parameter family, and no numerical error magnitude is reported. The Ising tasks, where the physics claims are made, do not exercise SCTs at all; they test scale behavior, which is already enforced by the readout and the z-lift scaling. So the unique selling point of the architecture is currently supported by one experiment, and the rest of the evidence would still be consistent with a similarity-equivariant GNN. That is a real gap, but it is addressable: run SCT-augmented Ising tests or report SCT-breaking magnitudes on the tasks where conformal claims matter.\n\nMinor but worth noting: the claimed finite limit of z_i as z0 tends to 0 is asserted without proof, z0 is never stated, no code is released, and the learned Delta values are fitted parameters, which is fine but should not be oversold. The baseline comparisons are fair enough and the PascalVOC null result is honestly reported.\n\nWho this is for: people building scale- and rotation-equivariant point-cloud models, and anyone wanting a bridge between AdS/CFT kinematics and geometric deep learning. I would bring it to a reading group, and I would cite it. It deserves a serious referee: the core construction is sound enough for conditional acceptance, with the equivariance claims tightened and SCT validation strengthened.","headline":"Novel AdS lift plus proper-distance message passing gives an interesting scale- and rotation-equivariant GNN, but the headline full conformal equivariance is not delivered and the new SCT capability is tested only once on MNIST.","tokens_in":34190,"tokens_out":1289,"would_cite":true,"duration_ms":12440,"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":"A graph neural network that lifts point clouds into Anti-de Sitter space achieves approximate equivariance under the full conformal group, and on the 2D Ising model it recovers the conformal dimensions $\\Delta_\\sigma=1/8$ and…","keywords":["conformal equivariance","graph neural networks","Anti-de Sitter space","point clouds","scale invariance","Ising model","conformal field theory","conformal dimensions"],"falsifier":"Test the full pipeline under a family of special conformal transformations on point clouds with sharply varying local density, or inspect Algorithm 1 directly on a pair of coincident points: if the equivariance error grows with density contrast, or if the computed $\\hat{z}_i$ diverges or vanishes as $z_0\\to0$, the central claim of mild, controlled symmetry breaking would be falsified.","tokens_in":33194,"feed_emoji":"📐","tokens_out":8128,"duration_ms":77649,"temperature":0.7,"pith_summary":"This paper tries to build a graph neural network whose predictions are unchanged when the input point cloud is subjected to any conformal (angle-preserving) transformation of $\\mathbb{R}^d$, including scalings and special conformal transformations that other equivariant architectures do not handle. The route is to lift each point cloud into $\\mathrm{AdS}_{d+1}$, where conformal transformations of the boundary act as ordinary isometries, and then run message passing keyed on the AdS proper distance. The paper argues that this makes the network exactly equivariant once the data is lifted, with only a mild symmetry breaking coming from the lift itself. On the 2D Ising conformal field theory, the trained network predicts $N$-point correlation functions and its readout exponents reproduce the known scaling dimensions $\\Delta_\\sigma=1/8$ and $\\Delta_\\epsilon=1$.","feed_headline":"Lift point clouds into AdS space for conformally equivariant GNNs","feed_subtitle":"The network learns Ising correlation functions and recovers known scaling dimensions straight from data.","key_machinery":"The load-bearing object is the AdS lift: Algorithm 1 embeds every point at boundary scale $z_0$, computes the AdS center of mass of its $k_{\\mathrm{lift}}$ nearest neighbours using the Galperin formula, and re-embeds the point at the resulting height $\\hat{z}_i$, so that $X_i=(x_i,\\hat{z}_i)$ lives in $\\mathrm{AdS}_{d+1}$. This turns each conformal transformation of the boundary into an isometry of the bulk, and the invariant proper distance $\\cosh D(X_i,X_j)=(z_i^2+z_j^2+\\|x_i-x_j\\|^2)/(2z_iz_j)$ becomes the only geometric input to the message function. The same construction supplies the feature lift $h_i^{\\mathrm{lifted}}=\\hat{z}_i^{\\Delta} h_i^{\\mathrm{input}}$ and the conjugate readout $O(x_i)=\\hat{z}_i^{-\\Delta} h_i^{\\mathrm{final}}$, which is what lets the network learn and expose conformal dimensions.","core_discovery":"The central discovery is a construction that reduces conformal equivariance for point-cloud networks to isometric equivariance: each point $x_i$ is embedded at a scale-dependent height $z_i$ in $\\mathrm{AdS}_{d+1}$, computed with a nearest-neighbour center of mass so that $z_i$ encodes the local length scale of the point cloud. Once the lifted points $X_i=(x_i,z_i)$ are fixed, the message-passing layer conditioned on the AdS proper distance $D(X_i,X_j)$ is exactly invariant under the isometry group of $\\mathrm{AdS}_{d+1}$, which is the same group as the global conformal group $\\mathrm{PO}(d+1,1)$ of $\\mathbb{R}^d$. The paper reports that the residual symmetry breaking from the lift is small, and that on the 2D Ising model the readout $\\log \\mathrm{Pred}_a = \\mathrm{AdSGNN}_a - \\Delta_a \\sum_i \\log \\hat{z}_i$ turns the trained network into a calculator for conformal dimensions, recovering $\\Delta_\\sigma=1/8$ and $\\Delta_\\epsilon=1$.","pith_inferences":["Beyond the paper, the same readout trick could be aimed at the three-point coefficients $c_{abc}$, which together with the conformal dimensions determine all higher correlation functions in a conformal field theory; the paper raises this question but does not answer it.","The tested special-conformal perturbation is a single one-parameter family; a natural stress test would apply a wider range of such transformations to point clouds with strongly nonuniform density, where the center-of-mass lift's symmetry breaking should be largest.","The scalar-feature restriction points to a plausible extension: combining the AdS lift with orientation-carrying feature spaces would give vector-valued conformal primaries and could close the expressivity gap the authors observe on rotation-sensitive tasks.","If the claimed finite limit of the lift as $z_0\\to0$ can be established rigorously, the construction would apply to degenerate configurations with coincident or nearly coincident points, where the current empirical check is weakest."],"forward_implications":["A network built this way is exactly invariant under the isometry group of $\\mathrm{AdS}_{d+1}$, so once a point cloud is lifted its output is conformally invariant by construction, independent of how the input was discretised.","Because the height $\\hat{z}_i$ tracks local density, the model should be scale-invariant and should keep its predictions accurate on inputs whose spatial extent lies far outside the training range.","The conformal-dimension readout in the Ising experiments gives a direct, interpretable interface between a trained network and the universal data of a conformal field theory.","The same architecture transfers across system sizes: a network trained on correlation functions of $N=8$ points can predict $N=16$ correlation functions better than the flat-space baselines, suggesting it has learned the physics rather than a lookup table."],"supporting_citations":[{"why":"Supplies the AdS/hyperbolic center-of-mass formula used in Algorithm 1 to assign each point its height $z_i$.","marker":"[34]"},{"why":"Establishes that the global conformal group of $\\mathbb{R}^d$ is the projective orthogonal group $\\mathrm{PO}(d+1,1)$, the isomorphism that makes the AdS lift work.","marker":"[24]"},{"why":"Gives the theorem that the isometry group of $\\mathrm{AdS}_{d+1}$ is $\\mathrm{PO}(d+1,1)$.","marker":"[30]"},{"why":"Provides the 2D Ising CFT correlation functions and the conformal dimensions $\\Delta_\\sigma=1/8$, $\\Delta_\\epsilon=1$ used as ground truth.","marker":"[4]"},{"why":"Defines the message-passing layer the paper adapts by substituting the AdS proper distance for the Euclidean distance.","marker":"[36]"},{"why":"Supplies the bulk-to-boundary intuition that boundary fields couple to bulk data through factors of $z^\\Delta$, motivating the feature lift and readout.","marker":"[33]"}],"fun_headline_variants":["Lift to AdS, get conformally equivariant GNN","AdS-GNN: Equivariant under all conformal maps","From flat to AdS: GNNs respect conformal symmetry","Conformal GNN via AdS isometry, learns scaling dims","AdS lift stabilizes GNN conformal equivariance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The construction stands on the assumption that the nearest-neighbour center-of-mass lift assigns every point a finite, non-degenerate height $z_i$ in the $z_0\\to0$ limit and that the resulting breaking of special conformal invariance is small enough not to destroy the equivariance benefit.","fun_headline_variants_meta":{"raw":{"variants":["Lift to AdS, get conformally equivariant GNN","AdS-GNN: Equivariant under all conformal maps","From flat to AdS: GNNs respect conformal symmetry","Conformal GNN via AdS isometry, learns scaling dims","AdS lift stabilizes GNN conformal equivariance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000614,"raw_usage":{"total_tokens":2850,"prompt_tokens":940,"completion_tokens":1910,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":1820}},"tokens_in":556,"tokens_out":1910,"duration_ms":14286,"temperature":1.0,"reasoning_tokens":1820,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:23:45.816402+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Test the full pipeline under a family of special conformal transformations on point clouds with sharply varying local density, or inspect Algorithm 1 directly on a pair of coincident points: if the equivariance error grows with density contrast, or if the computed $\\hat{z}_i$ diverges or vanishes as $z_0\\to0$, the central claim of mild, controlled symmetry breaking would be falsified.","supporting_citations":[{"cited_title":"A concept of the mass center of a system of material points in the constant curvature spaces,","cited_arxiv_id":null,"evidence_quote":"Supplies the AdS/hyperbolic center-of-mass formula used in Algorithm 1 to assign each point its height $z_i$."},{"cited_title":"Schottenloher,A Mathematical Introduction to Conformal Field Theory","cited_arxiv_id":null,"evidence_quote":"Establishes that the global conformal group of $\\mathbb{R}^d$ is the projective orthogonal group $\\mathrm{PO}(d+1,1)$, the isomorphism that makes the AdS lift work."},{"cited_title":"An introduction to Cartan geometries","cited_arxiv_id":"2302.14457","evidence_quote":"Gives the theorem that the isometry group of $\\mathrm{AdS}_{d+1}$ is $\\mathrm{PO}(d+1,1)$."},{"cited_title":"E(n) equivariant graph neural networks,","cited_arxiv_id":null,"evidence_quote":"Defines the message-passing layer the paper adapts by substituting the AdS proper distance for the Euclidean distance."}],"review_version":1}