{"id":"9cb901f6-e603-4b1a-97db-9958bea5cfa6","arxiv_id":"2511.13431","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A flow-matching representation of shape maps: per-shape flows to a Gaussian anchor compose into a zero-shot map between any two shapes, working across meshes, point clouds, SDFs, and volumes.","lead":"FUSE trains one flow-matching model per 3D shape, mapping each shape's pointwise feature distribution onto a shared Gaussian, and composes the inverse of one shape's flow with the forward flow of another to transfer points directly. The result is a representation-agnostic way to establish correspondences across meshes, point clouds, and signed distance fields without per-pair optimization.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4.2's anchor-alignment premise is unproven: independent CFM flows only match each shape's embedding distribution to the Gaussian, so nothing in Eq. (1) constrains semantically corresponding points to share latent codes; the zero-shot claim is supported only by benchmark averages.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing gap: the anchor-alignment premise. I agree. The paper itself flags (§4.3, Table 1) that FUSE helps only for partially aligned embeddings and can corrupt already-aligned FMNet features, which limits the universality claim. Independent support includes the simple, reproducible construction and consistent benchmark improvements over KNN, but there is no formal argument or direct measurement of anchor-space alignment. The proposed residual test would either substantiate or refute the premise. If the separation is strong, conditional acceptance could be upgraded; if not, the central claim would be overstated. Thus the reader's CONDITIONAL verdict remains appropriate.","tokens_in":13411,"tokens_out":10234,"duration_ms":112649,"concrete_test":"On a held-out FAUST pair, train flows independently with the paper's geodesic embeddings, then compute anchor-space residuals r(x,y) = ||ψ1^{-1}(f1(x)) − ψ2^{-1}(f2(y))|| for all ground-truth correspondence pairs, plus the same distances for random non-corresponding pairs. Report the separation (e.g., mean/quantile ratio and ROC AUC). If the true-correspondence residual is not substantially below the random-pair distribution (AUC ≈ 0.5), Eq. (2) is not aligning semantics, and the reported accuracy must instead be attributed to the target-space nearest-neighbor stage operating on nearly identical distributions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the assertion in §4.2 (after Eq. 3): \"Given the nature of the FM solutions and the selection of a shared Gaussian distribution as anchor, points with similar embeddings from different shapes are mapped to nearby regions in the anchor distribution.\" This is not a theorem. The CFM objective Eq. (1) is per-shape: it only constrains the pushforward of p0=N(0,1)^d to equal the empirical embedding distribution p_l. It does not constrain the joint law of (ψ1^{-1}(f1(x)), ψ2^{-1}(f2(y))) for corresponding x,y. Nothing in training rewards two independent flows for using the same latent coordinate for corresponding points. FM solutions are not unique — many velocity fields generate the same marginal path — and the selected solution depends on the arbitrary stochastic pairing of x0~p0 with x1~p_l during training. Consequently, Eq. (2) can in principle reduce to an arbitrary measure-preserving rearrangement between p1 and p2, equivalent to random matching. The empirical gains over KNN are consistent with the anchor-alignment premise but do not establish it for unseen pairs/modalities; indeed Table 1 shows FUSE degrades FMNet embeddings (0.00116 vs 0.00006), contradicting the universality claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"FUSE proposes a neural representation for maps between 3D shapes based on flow matching. For each shape, a flow model is trained independently to transport the shape's pointwise feature embedding distribution to a shared Gaussian anchor. The map between two shapes is then defined as the composition of the source inverse flow and the target forward flow (Eq. 2). The method claims zero-shot, representation-agnostic matching across meshes, point clouds, SDFs, and volumes, and reports competitive accuracy on FAUST, SMAL, SHREC20, Kinect, and additional tasks including UV mapping, human scan registration, and volumetric correspondence. The central premise is that independently trained flows map semantically corresponding points to nearby locations in the Gaussian anchor, enabling the composed map to align embeddings without any cross-shape supervision.","tokens_in":13783,"tokens_out":4352,"duration_ms":49256,"significance":"If the central premise holds, FUSE offers an elegant and novel probabilistic view of shape maps: a per-shape flow representation that supports zero-shot composition for arbitrary pairs and representations. The idea of using flow matching over feature embeddings rather than raw coordinates is a meaningful extension of GeomDist, and the empirical breadth—spanning several benchmarks and modalities—is commendable. The method is simple, computationally lightweight per shape, and yields high coverage by construction if the flows are truly invertible. However, the paper's main claim rests on an unproven alignment property of independent flow solutions, and at least one experimental result (FMNet embeddings in Table 1) directly contradicts the stated universality. The paper would benefit from a precise theoretical or empirical characterization of when the anchor-alignment property holds, as well as a more careful framing of its claims.","major_comments":[{"comment":"The load-bearing assumption that independently trained flows map corresponding points to nearby anchor locations is not established. The CFM objective (Eq. 1) only constrains the per-shape pushforward p_0 to p_l; it does not constrain the joint law of (ψ_1^{-1}(f_1(x)), ψ_2^{-1}(f_2(y))) for corresponding x,y. FM solutions are non-unique, and the stochastic pairing in training can produce arbitrary measure-preserving rearrangements. The sentence after Eq. (3) merely asserts the property. Please provide either a theoretical condition (e.g., uniqueness of the optimal transport map) or direct empirical evidence: for known correspondences, measure the distance between latent codes, compare the composition against a random bijection, and test stability across training seeds. Without this, Eq. (2) could in principle be equivalent to random matching.","section":"§4.2, Eq. (2)"},{"comment":"The claim that FUSE is a general alignment mechanism is contradicted by the FMNet row: FUSE increases the correspondence error from 0.000062 (KNN) to 0.001158, a degradation of nearly 20x. The text acknowledges this ('may even slightly corrupt them'), but such a caveat undermines the abstract's claim of a universal representation. The authors should either restrict the claim to embeddings that are not already well aligned, or provide a quantitative criterion—e.g., distributional similarity or a learned classifier—that predicts when FUSE will help. As is, the method is not universally beneficial across embeddings, which is a central advertised property.","section":"§4.3, Table 1"},{"comment":"The paper states that flow matching 'guarantees that the inverse flow deterministically maps any shape back to the exact Gaussian.' In practice, the learned velocity field is a neural network and the ODE integration is numerical; the inverse map is only approximate. Table 2 measures KL/JS divergence after inversion for geodesic embeddings, but this is a single embedding and the analysis does not quantify how inversion error propagates into the composition Φ_{12}. Please report inversion accuracy across embeddings and shapes, and discuss the effect of integration tolerance on the final correspondence accuracy. The distinction from GeomDist's approximate inversion is quantitative, not qualitative, and needs a controlled comparison.","section":"§4.1, §4.4"},{"comment":"The 'Coverage' metric is never formally defined, yet it is used to argue that FUSE achieves better coverage than baselines. If coverage is simply the size of the image of the discrete map, then any surjective map (or a map defined as ψ_2∘ψ_1^{-1} applied to a dense enough sample) trivially achieves high coverage, making the comparison with functional-map methods that produce soft or low-rank maps unfair. Please provide the exact definition, how it is computed from the map matrix, and what property it is intended to measure. The statement that FUSE achieves better coverage 'due to the nature of the flows' is circular unless coverage is defined independently of the map's construction.","section":"§5.1, metrics"},{"comment":"The geodesic embedding relies on a small set of landmarks (5 for FAUST/SMAL, 6 for SHREC20) selected 'at the protrusions' or 'common landmarks.' The procedure for choosing these landmarks is not specified: Are they placed manually, automatically, or via a semantic prior? If landmarks are chosen to correspond across shapes, then the embedding itself already encodes correspondence information, and the 'no learned priors' claim is weakened. The sensitivity of FUSE to the number and choice of landmarks is not analyzed, yet the landmark embedding is the primary unsupervised descriptor used in the main experiments. Please clarify the selection protocol and provide an ablation over landmark sets.","section":"§5.1, landmark selection"}],"minor_comments":[{"comment":"Typo: 'architechtures' should be 'architectures'.","section":"Related Work"},{"comment":"Typo: 'gedesics' should be 'geodesics'.","section":"§5.2, Inter-representation"},{"comment":"The Kinect dataset description says 'intoduced' instead of 'introduced'.","section":"§5.1, Datasets"},{"comment":"'V olumetric' has a stray space.","section":"Table 5"},{"comment":"The phrase 'we consider all the pairs of the 5 test set' is unclear. Please specify the test set size and the exact pairing protocol.","section":"§5.1, SHREC20"},{"comment":"The KNN-IN-GAUSS baseline is not defined precisely. Does it invert the flow using the same ODE solver as FUSE? If not, the comparison may conflate inversion error with alignment quality.","section":"§5.1, Baselines"},{"comment":"No code or supplement link is provided. Given the number of implementation details (flow architecture, training hyperparameters, ODE solver settings), the paper would benefit from releasing code or a detailed supplementary to ensure reproducibility.","section":"General"},{"comment":"The reference to [16] as 'The Fourth Blogpost Track at ICLR 2025' is informal; consider citing the formal equivalence paper (e.g., Lipman et al. or follow-up work).","section":"§4.4"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a simple and intriguing idea, but the central alignment premise is currently an article of faith. The Table 1 FMNet degradation is a direct counterexample to the universality claim and should be addressed head-on. I see the issue as fixable: the authors could add a theoretical condition, a direct latent-space evaluation, and a sensitivity analysis over embeddings and landmarks. The landmark selection protocol also deserves scrutiny, as manual selection of protrusions may inject correspondence supervision. If those points are addressed, the paper could be a solid contribution; in the current form, the overclaiming and missing evidence prevent acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"FUSE is a genuinely new idea: instead of learning a map between shapes directly, you train one flow-matching model per shape from a shared Gaussian anchor to that shape's pointwise feature distribution, then get correspondences by composing inverse and forward flows. That is simple, representation-agnostic, and it actually works on the benchmarks they run. The results are broad — FAUST, SMAL, SHREC20, Kinect point clouds, SDFs, volumes — and FUSE is competitive with or better than classic baselines, particularly with geodesic embeddings, and it beats KNN and OT in coverage. The UV-mapping and scan-fitting applications are a nice bonus.\n\nThe soft spot is the one the stress-test note identifies: the anchor-alignment premise in Section 4.2 is asserted, not proven. The CFM objective is per-shape; nothing forces independently trained flows to map corresponding points to the same latent coordinates. The velocity field is not unique, and the stochastic pairing in training can yield arbitrary measure-preserving rearrangements. So in principle the composition could be a random bijection. That is a real gap. But I don't think it is fatal: the empirical evidence across many benchmarks and modalities shows consistent gains over KNN and KNN-IN-GAUSS, which would not happen if the composition were garbage. The authors need to say why alignment emerges — perhaps from smoothness of the embeddings or properties of flow matching — but they don't need to prove a theorem for this paper to be useful.\n\nOther soft spots: Table 1 shows FUSE degrades FMNet and xyz embeddings, so the \"universal\" framing in the abstract is too strong. The authors acknowledge this in the text, but the paper should be clearer that FUSE helps when embeddings are moderately aligned, not always. Reproducibility is a real issue — no code, no architecture details, no hyperparameters, and no error bars. The SDF benchmark excludes topology-failed shapes, which is a mild selection bias. These are fixable.\n\nThe stress-test note is right that the alignment claim is unproven, but it overstates the risk: it says the method \"can in principle reduce to arbitrary measure-preserving rearrangement,\" and while that is logically true, the benchmarks suggest it doesn't happen in practice. That tension is exactly what the authors should investigate in revision.\n\nThis paper deserves a serious referee. The idea is novel, the evaluation is solid enough to warrant the community's attention, and the gap between the theoretical premise and the empirical behavior is interesting in its own right. I would accept it for peer review, and conditional on acceptance I'd ask for code, error bars, and an analysis of the alignment assumption — perhaps a toy experiment where they measure correspondence between latent codes for known pairs.","headline":"FUSE is a genuinely new zero-shot representation for shape maps built on per-shape flow matching, with broad and mostly solid empirical results; the anchor-alignment premise is the real soft spot, but the empirics carry it.","tokens_in":14210,"tokens_out":4228,"would_cite":true,"duration_ms":44427,"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":"FUSE represents the map between any two 3D shapes as the composition of two independently trained flow-matching models through a shared Gaussian anchor, producing bijective correspondences without per-pair optimization.","keywords":["shape matching","flow matching","shape correspondence","cross-representation matching","geodesic embeddings","invertible flows","zero-shot correspondence","probabilistic alignment"],"falsifier":"For a benchmark with dense ground-truth correspondences, compare the Euclidean distances in the Gaussian anchor between ψ₁⁻¹(E1(x)) and ψ₂⁻¹(E2(y)) for true corresponding pairs (x,y) against the same distances for random point pairs; the central claim collapses if these two distance distributions are statistically indistinguishable, since the anchor then carries no correspondence information.","tokens_in":13318,"feed_emoji":"🗺️","tokens_out":7528,"duration_ms":66259,"temperature":0.7,"pith_summary":"The paper introduces FUSE, a neural representation for correspondences between 3D shapes that treats each shape's feature distribution as something to be transported to a common Gaussian anchor by an invertible flow. Once each shape has its own flow, matching any pair becomes the composition of the source's inverse flow and the target's forward flow — a zero-shot operation that requires no further training or optimization. The authors argue that this probabilistic alignment improves correspondence accuracy whenever the underlying features are not already perfectly aligned, and they demonstrate the construction across meshes, point clouds, signed distance fields, and volumetric data. The significance is that shape matching becomes a one-time-per-shape investment that then spans representations and deformations, with bijectivity and coverage built in.","feed_headline":"Compose two flows to map between any 3D shapes","feed_subtitle":"Aligning every shape's features to one shared distribution gives instant, representation-agnostic shape maps.","key_machinery":"The central object is the flow-matching model ψ: [0,1]×R^d→R^d, an invertible diffeomorphism trained via Conditional Flow Matching to transport samples from a fixed Gaussian anchor p0 = N(0,1)^d to a shape's pointwise embedding distribution pℓ. The map between shapes is the composition Φ12 = ψ₂ ∘ (ψ₁)⁻¹: the inverse flow pulls source embeddings back to the shared Gaussian, and the forward flow pushes them into the target's embedding space. The shared Gaussian is what makes the construction representation-agnostic and zero-shot, since any two flows from the same anchor can be composed without additional optimization. The geodesic feature embedding (distances to a handful of landmarks) is the","core_discovery":"The central claim is that, given two trained flows ψ₁ and ψ₂ from a shared Gaussian anchor to the embedding distributions of S₁ and S₂, the composition Φ12 = ψ₂ ∘ (ψ₁)⁻¹ is an accurate, bijective, high-coverage pointwise map from S₁ to S₂. Because each flow is trained independently on its own shape with no cross-shape constraint, the method is a purely generative representation: corresponding points end up near each other in the anchor because the flow-matching objective makes distributional alignment coincide with semantic alignment. The paper argues that this holds for any embedding that encodes shape semantics, and that the flow composition acts as a probabilistic alignment mechanism — im","pith_inferences":["The same anchored-flow construction suggests a universal indexing scheme for shape collections: if every shape maps to the shared Gaussian, then nearest-neighbor search in the anchor space could support retrieval, interpolation, and shape analogy without ad hoc matching — a direction the paper leaves implicit.","Because the benefit is tied to distributional misalignment, a testable extension is to characterize the regime of feature alignment in which the flow bridge is net positive, and to intentionally use features that are order-invariant but distributionally misaligned so the flow does the alignment work.","If the core locality conjecture holds, flow-matching over feature distributions is a general alignment primitive that could port to other domains where semantically linked points live in different distributions across instances (e.g., audio or text embedding alignment), not just 3D geometry.","The anchored Gaussian also makes the representation transitive: composing ψ₃∘(ψ₁)⁻¹ can be viewed as a chain through intermediate shapes, suggesting a compositional structure that functional maps lack, though the paper does not explore it."],"forward_implications":["Once each shape has its flow, any pair of shapes becomes matchable by composing flows, with no per-pair optimization or learned correspondence priors.","The representation is modality-agnostic: because only surface sampling and embedding computation are required, the same flows support mesh-to-mesh, point-cloud-to-mesh, SDF-to-mesh, and volumetric matching.","Bijectivity and high coverage are inherited from the invertibility of flows, improving on functional-map and optimal-transport baselines in coverage while achieving competitive accuracy.","The flow acts as a probabilistic aligner: it consistently improves accuracy and distributional similarity for moderately aligned embeddings (e.g., geodesic features), and it can initialize refinement methods like ZoomOut and Neural Adjoint Maps to state-of-the-art results.","Beyond correspondence, the construction serves as a backbone for UV parametrization, human-scan reposing, and volume matching."],"fun_headline_variants":["Flow composition gives instant shape maps","Invertible flows map any shapes directly","Flow-matching maps shapes without training","One flow per shape, compose to match","From anchor to shape: flows align geometry"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that independently trained flows map semantically corresponding points to nearby locations in the shared Gaussian anchor; if the flows from different shapes do not cluster corresponding points in the anchor, the composition Φ12 degenerates to a random bijection between distributions and gains nothing over direct nearest-neighbor search.","fun_headline_variants_meta":{"raw":{"variants":["Flow composition gives instant shape maps","Invertible flows map any shapes directly","Flow-matching maps shapes without training","One flow per shape, compose to match","From anchor to shape: flows align geometry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1075,"prompt_tokens":690,"completion_tokens":385,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":434,"completion_tokens_details":{"reasoning_tokens":322}},"tokens_in":434,"tokens_out":385,"duration_ms":4017,"temperature":1.0,"reasoning_tokens":322,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T21:47:35.266915+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a benchmark with dense ground-truth correspondences, compare the Euclidean distances in the Gaussian anchor between ψ₁⁻¹(E1(x)) and ψ₂⁻¹(E2(y)) for true corresponding pairs (x,y) against the same distances for random point pairs; the central claim collapses if these two distance distributions are statistically indistinguishable, since the anchor then carries no correspondence information.","supporting_citations":[],"review_version":1}