acceptodds
Under review as a conference paper at ICLR 2027

T-GROW: Robust and Scalable Tree-based Gromov-Wasserstein

Abstract

Comparing structured and multi-modal data across different representation spaces remains a core challenge in machine learning, as metric distances between individual samples are not directly defined across domains. The Gromov-Wasserstein (GW) distance addresses this by seeking a geometry-preserving alignment between probability measures supported on different metric spaces. However, its underlying non-convex quadratic formulation becomes computationally prohibitive at scale. To alleviate this burden, scalable tree-based methods replace full pairwise geometry with simpler summaries of hierarchical structure. While relying solely on root-to-node path lengths enables fast alignment, this structural reduction discards rich pairwise relational geometry within the tree hierarchy. In this work, we introduce T-GROW, a robust and scalable tree-based GW discrepancy built on co-depth, defined as the length of the root path shared by two samples before their paths diverge. For any fixed hierarchy, we prove that solving GW over its leaf clusters under co-depth is exactly equivalent to solving GW over the original samples under the induced tree metric. This theoretical equivalence guarantees that T-GROW retains the complete pairwise geometry of the hierarchy without approximation loss, while shifting the quadratic optimization to a significantly compressed leaf resolution. Experimental evaluations show that T-GROW consistently improves structural matching over scalable GW baselines, exhibits strong robustness to structural noise, and scales efficiently to tens of millions of samples.

Primary area:
optimization
Keywords:
Gromow-Wasserstein, Tree-based, Optimal Transport

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_zchDOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

T-GROW propose a way to compute Gromow-Wasserstein (GW) that rely on trees. Each of the input measures is clustered with randomized hierarchies which are then used to compute an average of the discrepancy (eq. 13 in the paper). They use the Gromov product (or co-depth) to encode the geometry between the leaves (which corresponds to the distance between the root and the lowest common ancestor to those points). The paper gives theoretical insights on the method by showing for instance that GW over the leaf clusters is equal to GW over the samples under the pulled-back tree metric so that the reduction to the computation to the leaves add no optimization gap (Theorem 1). Experiments show case the method over several datasets that have up to ∼107\sim 10^7 points which corresponds to real-world scenarios.

Critical strengths
  • The idea is well motivated: FlowAlign reduces the tree to a weighted list of root depth and solve a 1d problem so any two trees with the same depth distribution cannot be distinguished with GW. The paper propose to restore the co-depth term in the tree distance as they identify this as a substential loss of information.
  • Cost and empirical gains: Each cloud point is clustered into a tree and this operation costs O(SHκNp)O(SH\kappa Np) (SS number of random trees, HH the depth, κ\kappa the branching factor and pp the feature dimension), this is linear in NN which allows it to be applied to large scale datasets. After building the tree, the cloud point are reduced to the LL leaves and GW is solved on those leaves. The numerical experiments highlight improvement in terms of accuracy on the downstream tasks over FlowAlign which relates to the claim that restoring the co-depth helps.
Critical weaknesses
  • The theory is a bit weaker than what the abstract suggest:
    • Theorem 1 only states that solving GW whose distances are constant on cluster of points can be solved on the clusters and not the whole points and this hold for any partition. The "zero optimization gap" applies only to the global optimum and not to the local optimum that will be reached by gradient descent on this non-convex problem. Theorem 2 bounds the gap to the true value to GW by a quantity that the paper does not control.
    • In the experiments, no experiment isolates the contribution of using a tree metric. FlowAlign and DepthAlign change both the geometry of the data and the solver of the problem. A control experiment that runs the same pipeline with the original anchor distances instead of the tree distance is required to show that co-depth drives the numerical gains of the method. The unary-only variant does not answer this as it only compares co-depth with root depth.
Questions for authors
  • Does the separation demonstrated in Proposition 2 extend to DepthAlign?
  • Since the problem is non-convex, how sensitive is T-GROW to the initialization of the conditional-gradient solver?
  • How does T-GROW perform on a large scale datasets when (κ,H)(\kappa, H) differ from the generating hierarchy? on a benchmark whose hierarchy is not a regular tree?
  • Does T-GROW converge to GW in any regime?
6 comments
AuthorsOpenReview ↗

Response to Reviewer zchD Part 1

We thank the reviewer for the constructive comments. We answer each weakness (WiW_i) and question (QjQ_j) below. For the sake of clarity, we divide weaknesses into W1 and W2, corresponding to the 2 bullet points.

W1: "The theory is a bit weaker than what the abstract suggest: Theorem 1 only state … the paper does not control."

Answer

Theorem 1 and the role of the tree. The reduction holds for any partition. The co-depth hierarchy supplies a nested family of partitions: at depth ℓ\ell, the represented point-level relation is d~ℓ=ZℓRℓZℓ⊤\widetilde d_\ell=Z_\ell R_\ell Z_\ell^\top, where ZℓZ_\ell is the membership matrix and RℓR_\ell is a fixed node relation. Theorem 1 therefore applies at every level, supporting coarse-to-fine matching. The partition also affects the represented geometry. In a separate partition control with the relation estimator fixed, SNARE-seq FOSCTTM is 0.14980.1498 for tree partitions, 0.15400.1540 for random Voronoi cells and 0.15250.1525 for flat farthest-point cells, at the same cell budget. On CIFAR-100, replacing the geometry-aware hierarchy with random balanced splits lowers accuracy from 0.6530.653 to 0.3680.368. The CIFAR-100 results below use the same 50 instances: 10 seeds × 5 target permutations, with shuffled targets and independent trees. Results are mean ± sample standard deviation unless a range is stated.

Local optima. The proof of Theorem 1 gives more than the statement about global optima. For every feasible point coupling π\pi, E(π)=F(P),P=agg⁡(π),E(\pi)=F(P), P=\operatorname{agg}(\pi), and every feasible node coupling has a lift with the same cost. So the cost of any coupling computed at node resolution, optimal or not, equals the cost of its lift. Stationary points are preserved as well. Let gE(π)=max⁡γ∈Π(a,b)⟨∇E(π),π−γ⟩g_E(\pi)=\max_{\gamma\in\Pi(a,b)}\langle\nabla E(\pi),\pi-\gamma\rangle be the Frank-Wolfe gap, the stopping criterion of conditional gradient. Since E=F∘agg⁡E=F\circ\operatorname{agg}, we have ∇E=agg⁡∗∇F\nabla E=\operatorname{agg}^{*}\nabla F, and since aggregation maps Π(a,b)\Pi(a,b) onto Π(α,β)\Pi(\alpha,\beta), gE(π)=max⁡Q∈Π(α,β)⟨∇F(P),P−Q⟩=gF(P).g_E(\pi)=\max_{Q\in\Pi(\alpha,\beta)}\langle\nabla F(P),P-Q\rangle=g_F(P). Thus, a first-order stationary node coupling lifts to a first-order stationary coupling of the represented point-level problem, with exactly the same objective value. The reduction does not make the problem convex; it guarantees that nothing is lost by working at node resolution, at any point of the optimization. As a numerical check on the same SNARE-seq relations, point-level and leaf-level CG agree to numerical precision in both objective values and aggregated couplings across six matched starts, including runs ending at different objective values.

Theorem 2: the distortion is controlled. For one domain with input distances dijd_{ij}, point masses aia_i, node masses αp\alpha_p, node-pair means BpqB_{pq}, and q(i)q(i) the node containing ii, the distortion of a node relation RR satisfies δ2(R)=∑i,jaiaj[dij−Rq(i),q(j)]2=δ2(B)+∑p,qαpαq(Bpq−Rpq)2.\delta^2(R)=\sum_{i,j}a_i a_j\big[d_{ij}-R_{q(i),q(j)}\big]^2=\delta^2(B)+\sum_{p,q}\alpha_p\alpha_q(B_{pq}-R_{pq})^2 .

With the input distances at a fixed scale and the within-node means kept on the diagonal, the mean relation R=BR=B therefore minimizes the distortion over all node-pair relations on a given partition. T-GROW keeps the co-depth hierarchy and estimates this mean relation between its nodes from the input distances; we call it co-depth (ours). Under nested refinement this minimum cannot increase.

The optimal represented GW values converge to the original GW value when the distortion in Theorem 2 vanishes. For the mean relation at a common fixed input scale, finite partitions refined to singletons recover the original distances exactly. More generally, on compact metric spaces, let the largest cell diameter εℓ\varepsilon_\ell tend to zero. For points i,i′i,i' in one cell and j,j′j,j' in another, ∣d(i,j)−d(i′,j′)∣≤d(i,i′)+d(j,j′)≤2εℓ|d(i,j)-d(i',j')|\le d(i,i')+d(j,j')\le2\varepsilon_\ell. The mean member-pair relation satisfies ∥d−d~ℓ∥∞≤2εℓ⟶0\lVert d-\widetilde{d}_{\ell}\rVert_{\infty}\le 2\varepsilon_{\ell}\longrightarrow 0 and Theorem 2 gives convergence of optimal GW values. The same bound permits zero diagonals because within-cell distances are at most εℓ\varepsilon_\ell. For the tree path distance, convergence requires its distortion to vanish, as when the weighted tree exactly represents the input metric; nested refinement alone is insufficient for arbitrary metrics. These guarantees concern optimal values, not elimination of CG optimization error.

The distortion is also measurable. On SNARE-seq, averaged over ten tree pairs, δX+δY\delta_X+\delta_Y is 0.3630.363, 0.2560.256, 0.2100.210 and 0.1690.169 for L=4,16,64,256L=4,16,64,256 leaves. The separate refinement study measures values between 0.090.09 and 0.150.15 at nominal budget L=1024L=1024. The point-level CG solver returns a value of about 0.7460.746, and computed reduced and point-level CG values differ by less than 0.0130.013 across the tested resolutions. These numerical diagnostics are distinct from the optimal-value convergence guarantee.

AuthorsOpenReview ↗

Response to Reviewer zchD Part 2

W2: "In the experiments, no experiment isolates … co-depth with root depth."

Answer

We ran the relation control on three tasks. Within each comparison, partitions, masses, slices, solver settings and readout are fixed. For nodes u,vu,v with member sets Iu,IvI_u,I_v, masses aia_i and input distances d(i,j)d(i,j), we compare:

  • Representative distance: d(cu,cv)d(c_u,c_v) between node representatives, centroids on S3DIS and SNARE-seq and medoids on CIFAR-100.
  • Tree path distance: DT(u,v)D_T(u,v), with edge lengths given by the distances between parent and child representatives. This is the tree distance of Eq. (9).
  • Co-depth (ours): the mean relation estimated from the input distances on the same hierarchy, R(u,v)=∑i∈Iu∑j∈Ivaiaj d(i,j)a(Iu)a(Iv),a(Iu)=∑i∈Iuai.R(u,v)=\frac{\sum_{i\in I_u}\sum_{j\in I_v}a_i a_j\,d(i,j)}{a(I_u)a(I_v)},\qquad a(I_u)=\sum_{i\in I_u}a_i.

Co-depth is the relation used in the experiments below. Its definition will be explicit in Sec. 2. On S3DIS we use the RMS over the same member pairs, RRMS(u,v)=∥cu−cv∥2+tr⁡Σu+tr⁡ΣvR_{\mathrm{RMS}}(u,v)=\sqrt{\|c_u-c_v\|^2+\operatorname{tr}\Sigma_u+\operatorname{tr}\Sigma_v}, computed from cell means and covariance traces at linear cost. Each relation receives the same normalization rule: S3DIS matrices have zero diagonals and are divided by their maxima, SNARE-seq matrices retain their original scale, CIFAR-100 uses off-diagonal standardization at every level.

CIFAR-100 blind matching, 10 seeds × 5 permutations, the relation is used at every level of the solver:

Relation at each levelAccuracy
Tree path distance0.358±0.2880.358\pm0.288
Representative distance0.509±0.2740.509\pm0.274
Co-depth (ours)0.631±0.1060.631\pm0.106

S3DIS, three seeds; the relation is used at the leaves:

Relation on the same tree partitionLobby accuracyPair C accuracy
Tree path distance0.421±0.0500.421\pm0.0500.181±0.0150.181\pm0.015
Representative distance0.452±0.0420.452\pm0.0420.394±0.0090.394\pm0.009
Co-depth (ours)0.464±0.0380.464\pm0.0380.404±0.0250.404\pm0.025

SNARE-seq, ten seeds, shuffled targets and independent trees, relations are used at the leaves. At each leaf budget, partitions are identical across relations.

LeavesTree path distanceRepresentative distanceCo-depth (ours)
40.5070±0.17800.5070\pm0.17800.4310±0.18500.4310\pm0.18500.1920±0.03000.1920\pm0.0300
80.5680±0.05900.5680\pm0.05900.2580±0.17200.2580\pm0.17200.1550±0.00400.1550\pm0.0040
160.1511±0.00080.1511\pm0.00080.1505±0.00040.1505\pm0.00040.1504±0.00040.1504\pm0.0004
320.2420±0.18500.2420\pm0.18500.1505±0.00040.1505\pm0.00040.1499±0.00020.1499\pm0.0002
640.1940±0.13100.1940\pm0.13100.1503±0.00040.1503\pm0.00040.1498±0.00020.1498\pm0.0002

Co-depth has the highest mean alignment quality in these controls. On CIFAR-100 it matches or exceeds representative distances on 45 of the 50 instances and outperforms tree path distances on 40. The relation control uses common preprocessing across variants, explaining its 0.6310.631 accuracy versus 0.6530.653 in the main configuration. On SNARE-seq, the advantage over representative distances is largest with four and eight leaves, where it wins on nine of the ten seeds in each setting. At finer resolutions the two relations perform similarly. These results support retaining within-node information that a single representative can omit.

The hierarchy itself also contributes: with the relation estimator fixed, S3DIS Pair C accuracy is 0.4040.404 with tree partitions versus 0.2690.269 with random Voronoi cells. W1 gives the SNARE-seq and CIFAR-100 partition controls. The unary ablation uses the same trees and search and gives 0.2470.247 versus 0.6530.653 for T-GROW. These controls separately assess root-depth information, partitioning and relation estimation.

Q1: "Does the separation demonstrated in Proposition 2 extend to DepthAlign?"

Answer

Proposition 2 applies to FlowAlign's root-depth summary. DepthAlign recursively aligns child-depth measures and retains more information, so the same indistinguishability argument does not establish a separation for DepthAlign. We will make this scope explicit. On identical trees, T-GROW has higher mean label-transfer accuracy than DepthAlign in every tested configuration of Q3.

Q2: "Since the problem is ... of the conditional-gradient solver?"

Answer

On SNARE-seq, for each of five hierarchy seeds we fixed the trees, relation matrices, solver settings and ensemble procedure. We compared one product-coupling start with three random feasible starts per seed, giving five and fifteen runs, respectively:

CG initializationFOSCTTM ↓
Product coupling0.1498±0.00030.1498\pm0.0003
Random feasible couplings0.1502±0.00030.1502\pm0.0003

The two T-GROW means differ by 0.00040.0004, with a worst reported FOSCTTM of approximately 0.15100.1510. This indicates limited sensitivity to the tested initializations under the fixed ensemble procedure, which retains ten of one hundred hierarchy candidates by their leaf objective. W1 separately establishes objective and first-order stationarity identities for aggregation and lifting.

AuthorsOpenReview ↗

Response to Reviewer zchD Part 3

Q3: "How does T-GROW perform on a large scale datasets when (κ,H)(\kappa, H) differ from the generating hierarchy? on a benchmark whose hierarchy is not a regular tree?"

Answer

We built a large benchmark with an irregular generating hierarchy, fixed before any method was run. Each internal node has between two and five children at random; leaves occur at different depths, up to four, with random masses; two independent nonlinear maps produce observations in 16 and 24 dimensions; each domain has 2⋅1052\cdot10^5 points. Encoder shapes are fixed in advance, and all tree methods use the same trees. Label-transfer accuracy over ten seeds:

Encoder (κ,H)(\kappa,H)Tree constructionT-GROWDepthAlignFlowAlign
(4,3)(4,3)hierarchical k-means0.393±0.1490.393\pm0.1490.184±0.1270.184\pm0.1270.123±0.0720.123\pm0.072
(4,4)(4,4)hierarchical k-means0.470±0.2590.470\pm0.2590.182±0.1250.182\pm0.1250.127±0.0900.127\pm0.090
(4,3)(4,3)farthest-point0.245±0.1080.245\pm0.1080.124±0.0720.124\pm0.0720.097±0.0830.097\pm0.083
(4,4)(4,4)farthest-point0.312±0.1290.312\pm0.1290.119±0.0680.119\pm0.0680.116±0.0920.116\pm0.092

T-GROW has the highest mean accuracy in all four configurations, despite the mismatch between regular encoders and the irregular hierarchy. Paired Wilcoxon tests give p≤0.014p\le0.014 against both tree baselines; with k-means trees, T-GROW exceeds DepthAlign on all ten seeds. QGW with 256 cells reaches 0.1780.178 and SGW 0.0850.085. For the (4,4)(4,4) encoder, total runtime is about 34 s with k-means trees and 17 s with farthest-point trees.

We also tested data generated without any tree. Latent clusters pass through two independent random nonlinear maps, to 64 and 48 dimensions, with independent additive noise. For every tree method the shape (κ,H)(\kappa,H) is chosen without labels, as the one with the lowest matching objective among 13 fixed shapes. Blind-matching accuracy:

NoiseT-GROWQGWEGWGWDepthAlignFlowAlign
0.11.001.000.580.580.390.390.360.360.530.530.170.17
0.250.990.990.410.410.350.350.180.180.380.380.510.51
0.50.820.820.240.240.270.270.140.140.280.280.170.17

On the Fig. 1 generator, a matching encoder gives 0.5780.578 versus at most 0.190.19 for FlowAlign and DepthAlign, a finer (4,4)(4,4) encoder on (4,3)(4,3) data gives 0.3860.386 versus at most 0.1820.182. These tests support using an encoder at least as fine as the generating structure when that structure is known. The irregular and non-tree benchmarks above assess performance without such a match.

Q4: "Does T-GROW converge to GW in any regime?"

Answer

See W1 for convergence of optimal GW values under vanishing relation distortion, including the finite-data and vanishing-cell-diameter regimes and the numerical diagnostics.

Reviewer_A77POpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes T-GROW, a scalable approximation of Gromov-Wasserstein (GW) for aligning point clouds that live in different spaces. Each space is encoded by a random hierarchical clustering tree, and the original distances are replaced by tree path distances between leaf clusters. GW is then solved between the two sets of leaves, and the result is averaged over several random trees. For a fixed tree size the cost is linear in the number of points, and the leaf coupling can be lifted back to the points to transfer labels. On the theory side, the authors show that the leaf problem and the point problem coincide once the points carry the tree distances (Theorem 1). They also show that root depths alone, as used by FlowAlign, miss the pairwise information in the tree (Proposition 2), and they bound the gap to GW by the distortion of the trees (Theorem 2). The experiments cover synthetic data with up to 50 million points, robustness to rotations, and real benchmarks on single-cell data (SNARE-seq), 3D scans (S3DIS) and blind vision-language alignment. T-GROW reports the best results in all of them.

Critical strengths
  1. The paper explains clearly what root-based tree methods such as FlowAlign miss. They only use how far each point is from the root, while the tree distance between two points also depends on how much of the path from the root they share, and Proposition 2 gives a concrete case where this matters. This weighs in my assessment because it is the conceptual core of the paper, and I think it is correct and useful beyond this particular method.

  2. The method is simple and fast at a scale where GW is out of reach. Only the small leaf problem is quadratic, and labels are transferred without building an n by m coupling. Running 50 million points per cloud takes about 12 minutes, and on SNARE-seq T-GROW matches entropic GW (FOSCTTM 0.148 vs 0.150) about 200 times faster. If the accuracy results hold up, this would make GW-type alignment practical for much larger problems.

Critical weaknesses
  1. I find the main blind matching results hard to trust as they stand. T-GROW is meant to approximate GW, but it beats GW by a wide margin (65.6% vs 30.0% on CIFAR-100), and the paper does not explain how. One possible reason is the evaluation protocol. Both trees of a slice are built with the same random seed, so if source and target are stored in the same order, they start from corresponding points. Points that fall in the same leaf also get the same coupling, so the final assignment among them comes down to tie-breaking, which often follows the index order, and in this benchmark the index order is the ground truth unless the classes are shuffled. I am also skeptical of Table 10, where the method still reaches 58.4% with only 8 leaves for 100 classes. This concern is the main reason for my rating, because if the results do not hold with shuffled targets and independent seeds, the main empirical claim of the paper does not stand.
  2. Even if the results hold, it is not clear that the tree geometry is what makes T-GROW work. T-GROW combines several random trees while GW seems to be run once, which matters for a non-convex problem (on S3DIS, GW is far below EGW on the same anchors). The unary-only ablation also already beats every baseline on CIFAR-100. It would help to rerun the same pipeline, with the same partitions, restarts and aggregation, but with the original distances between leaf anchors instead of tree distances, to see whether the tree distances matter at all.

I am also not convinced by the theory and the robustness experiments. Theorem 1 holds by construction, because points in the same leaf are given zero distance, and Theorem 2 bounds the error by a tree distortion that is never measured and does not go to zero as the leaves get finer. The synthetic data come from a tree with the same branching factor and depth as T-GROW, and the perturbations rotate parts of the data around branch centers or around the center of the data, which tree distances are largely blind to, since the distance between two siblings does not depend on the angle between them. Since the claimed contribution is the tree geometry itself, this is the second main reason for my rating.

Questions for authors
  1. Are the target samples shuffled, and do the two trees of a slice use independent seeds? Could the authors rerun blind matching and SNARE-seq under both conditions, and explain how 100 classes can be told apart when the tree has only 8 leaves?
  2. How were the branching factor, depth and number of trees chosen, and were they tuned on the test benchmarks?
  3. Could the authors run the control suggested in weakness 2 and report the GW objective reached by each method? This would show whether T-GROW finds better GW solutions or optimizes something else.
5 comments
AuthorsOpenReview ↗

Response to Reviewer A77P Part 1

We thank the reviewer for the constructive comments. We answer each weakness (WiW_i) and question (QjQ_j) below. All CIFAR-100 blind-matching tables below use shuffled targets, independent trees, and the same 50 instances: 10 seeds × 5 target permutations.

W1 "I find the main blind matching … does not stand."

Answer

Protocol. The two trees of a slice use separate random draws; Algorithm 2 will be clarified accordingly. We reran every method with randomly permuted targets to remove the correspondence implied by the original storage order. Candidate scoring, consensus weighting and final selection use JJ, defined below. The ground-truth correspondence is used only after optimization to report accuracy and normalize the reported cost. On CIFAR-100 with N=100N=100, T-GROW gives 0.6530.653, compared with 0.3120.312 for EGW and 0.1990.199 for GW. On SNARE-seq, we separately varied the tree draws and cell order over five seeds. FOSCTTM remains similar across the four conditions.

Tree seeds, cell ordersame, originalsame, shuffledindependent, originalindependent, shuffled
T-GROW0.14980.14980.14990.14990.15010.15010.15010.1501

Eight leaves for 100 classes. The number of leaves sets the resolution of the coarse matching, not of the final answer. The hierarchy matches groups of classes. Inside a matched pair of leaves, each class is described by 16 quantiles of its distances to all other classes, and classes are assigned by a Hungarian step whose costs are determined by these geometric profiles rather than by class indices. The full permutation is then refined on the original class relations under JJ, and classes can move across the initial leaf matches. With shuffled targets, the 8-leaf configuration gives 0.5780.578, above every baseline.

Why T-GROW beats GW. Every method returns a permutation π\pi, scored by the same objective, Eq. (56):J(π)=∑i,j[Xij−Yπ(i)π(j)]2.J(\pi)=\sum_{i,j}\big[X_{ij}-Y_{\pi(i)\pi(j)}\big]^2. X,YX,Y are the standardized class-distance matrices of Eq. (55), and JJ is their GW cost for a permutation. The table reports final JJ divided by the ground-truth permutation's JJ on the same 50 instances:

T-GROWEGWGWQGWLR-GW
Accuracy0.6530.6530.3120.3120.1990.1990.1600.1600.0980.098
Final JJ / ground-truth JJ0.9230.9231.0321.0321.1091.1091.1361.1361.1601.160

A ratio below 11 is possible because the true correspondence need not minimize the cross-modal GW objective. Over all runs, JJ and accuracy have a Spearman correlation of −0.86-0.86.

T-GROW generates candidates by coarse-to-fine matching and refines them under JJ, whereas GW and EGW optimize their transport objectives at full resolution. Its higher accuracy and lower final JJ are consistent with an optimization benefit of the hierarchical procedure. Theorem 1 guarantees exact reduction for the represented relation, not global optimality of the numerical solver. W2 tests candidate generation under the complete search.

W2 "Even if the results hold, it is not clear … the tree distances matter at all."

Answer

Restarts and search. In W1, each baseline uses 20 restarts, retains its lowest-cost solution and receives the same final refinement as T-GROW. The control below applies the complete T-GROW search with the same settings to different candidate generators. The random-hierarchy variant also uses its random hierarchy for the later tree-based moves.

Candidates given to the same searchAccuracyFinal JJ / ground-truth JJTime (sec)
GW, 20 restarts0.468±0.2790.468\pm0.2790.9760.9763.3
EGW, 20 restarts0.474±0.2210.474\pm0.2210.9520.95258
Random balanced hierarchy0.368±0.2470.368\pm0.2471.0051.0052.0
T-GROW0.653±0.1170.653\pm0.1170.9230.9231.8

Under the same search settings, the T-GROW hierarchy yields the highest mean accuracy and lowest mean JJ. Paired Wilcoxon tests give p<10−3p<10^{-3} against each alternative. The random-hierarchy control further supports the contribution of geometry-aware splits to candidate generation and later tree-based moves.

Number of trees.

Trees124
Accuracy0.432±0.2980.432\pm0.2980.548±0.2390.548\pm0.2390.653±0.1170.653\pm0.117

A single tree exceeds the GW and EGW baselines with 20 restarts; two trees also exceed their candidates under the full search. Additional trees improve mean accuracy and reduce variability. On S3DIS, the GW–EGW gap on identical anchors reflects the solver choice, which we will make explicit.

Unary ablation. With shuffled targets, the same trees and the same search, the unary variant gives 0.2470.247 versus 0.6530.653 for T-GROW. T-GROW is better on 46 of the 50 instances. This is consistent with Proposition 2, root-depth summaries can miss shared-ancestry information.

AuthorsOpenReview ↗

Response to Reviewer A77P Part 2

Following up on the previous comment for W2.

Anchor-distance control. Proposition 1 describes how co-depth encodes shared-path information on the hierarchy. T-GROW keeps this hierarchy and estimates the relation between its nodes from the input distances. For two nodes u,vu,v with member sets Iu,IvI_u,I_v and masses aia_i, the mean relation is R(u,v)=∑i∈Iu∑j∈Ivaiaj d(i,j)a(Iu) a(Iv),a(Iu)=∑i∈Iuai.R(u,v)=\frac{\sum_{i\in I_u}\sum_{j\in I_v}a_i a_j\,d(i,j)}{a(I_u)\,a(I_v)}, a(I_u)=\sum_{i\in I_u}a_i. We call this relation co-depth and report the tree path distance of Eq. (9) separately. The definition will be stated explicitly in Sec. 2. Theorem 1 applies to either relation, since both are fixed relations between node pairs. The requested control replaces R(u,v)R(u,v) by the distance d(cu,cv)d(c_u,c_v) between the medoids of the same nodes. Hierarchies, class relations, preprocessing, restarts, search and aggregation are the same in all rows:

Relation on the same hierarchyAccuracyFinal JJ / ground-truth JJ
Tree path distance0.358±0.2880.358\pm0.2881.0261.026
Anchor distance0.509±0.2740.509\pm0.2740.9580.958
Co-depth (ours)0.631±0.1060.631\pm0.1060.9290.929

Co-depth has the highest mean accuracy, lowest mean J and smallest standard deviation. It matches or exceeds anchor distances on 45 of the 50 instances and outperforms tree path distances on 40. W3 motivates the estimator through distortion; this control evaluates its matching performance. The control uses common preprocessing across relation variants, explaining its 0.6310.631 accuracy versus 0.6530.653 in the main configuration.

On SNARE-seq, we vary the leaf budget over ten seeds. At each budget, partitions, slices, solver and readout are identical across the three relations, and only the leaf relation changes.

LeavesTree path distanceAnchor distanceCo-depth (ours)
40.5070±0.17800.5070\pm0.17800.4310±0.18500.4310\pm0.18500.1920±0.03000.1920\pm0.0300
80.5680±0.05900.5680\pm0.05900.2580±0.17200.2580\pm0.17200.1550±0.00400.1550\pm0.0040
160.1511±0.00080.1511\pm0.00080.1505±0.00040.1505\pm0.00040.1504±0.00040.1504\pm0.0004
320.2420±0.18500.2420\pm0.18500.1505±0.00040.1505\pm0.00040.1499±0.00020.1499\pm0.0002
640.1940±0.13100.1940\pm0.13100.1503±0.00040.1503\pm0.00040.1498±0.00020.1498\pm0.0002

Co-depth has the lowest mean FOSCTTM at each tested budget. The advantage over anchor distances is largest with four and eight leaves, where it wins on nine of the ten seeds in each setting; the two relations perform similarly at finer resolutions.

We refer the last paragraph of the 'Critical Weaknesses' as W3: "I am also not convinced … main reason for my rating."

Answer

Theorem 1 certifies a lossless reduction for the chosen hierarchical representation. For any fixed relation RℓR_\ell between the nodes of level ℓ\ell, the point-level relation is d~ℓ=ZℓRℓZℓ⊤\widetilde d_\ell=Z_\ell R_\ell Z_\ell^\top, with ZℓZ_\ell the membership matrix. The proof gives E(π)=F(agg⁡π)E(\pi)=F(\operatorname{agg}\pi) for every feasible point coupling π\pi, every feasible node coupling has a lift with the same cost. This holds at every level of the hierarchy and does not require zero within-node distances: it includes relations with nonzero within-node means on the diagonal. The cost-preserving correspondence therefore also applies to the non-optimal couplings returned by a numerical solver. The optimization error of that solver is a separate matter and is not covered by the reduction.

As a numerical check on SNARE-seq, point-level and leaf-level CG objectives agree to numerical precision across six matched starts, including runs ending at different objective values. To make the distortion bound in Theorem 2 concrete, we characterize the mean relation and state conditions under which the distortion vanishes. For one domain with input distances dijd_{ij} at a fixed scale, point masses aia_i, node masses αp\alpha_p, node-pair means BpqB_{pq}, and q(i)q(i) the node containing ii, the weighted least-squares identity is δ2(R)=∑i,jaiaj[dij−Rq(i),q(j)]2=δ2(B)+∑p,qαpαq (Bpq−Rpq)2.\delta^2(R)=\sum_{i,j}a_i a_j\big[d_{ij}-R_{q(i),q(j)}\big]^2=\delta^2(B)+\sum_{p,q}\alpha_p\alpha_q\,(B_{pq}-R_{pq})^2. Consequently the mean relation R=BR=B, including its within-node means on the diagonal, minimizes the distortion among all node-pair relations on a fixed partition. Under nested refinement this minimum cannot increase, and it is zero at singleton resolution.

More generally, for distributions on compact metric spaces, if the largest cell diameter εℓ\varepsilon_\ell tends to zero, the mean and RMS member-pair relations satisfy ∥d−d~ℓ∥∞≤2εℓ\lVert d-\widetilde{d}_{\ell}\rVert_{\infty}\le 2\varepsilon_{\ell}. The distortion then vanishes, and Theorem 2 gives convergence of the optimal GW values of the reduced problems to the original one. For the tree path distance, the same criterion applies whenever the weighted tree represents the input metric.

AuthorsOpenReview ↗

Response to Reviewer A77P Part 3

Following up on the previous comment for W3.

As a numerical diagnostic, we measured the distortion on SNARE-seq with the original Euclidean distances. Averaged over ten tree pairs, δX+δY\delta_X+\delta_Y for the mean relation is 0.3630.363, 0.2560.256, 0.2100.210 and 0.1690.169 at L=4,16,64,256L=4,16,64,256 leaves, so it decreases as the leaves get finer. For the tree path distance it increases from about 0.690.69 to 1.691.69. The point-level CG solver returns a value of about 0.7460.746, and the reduced and point-level CG values differ by less than 0.0130.013 at every resolution. These are values computed by a numerical solver, and we report them separately from the guarantee on optimal values.

Synthetic data and rotations. We add two benchmarks whose data are not generated by a tree matching the encoder and contain no rotation; Fig. 1 and the rotation test will move to the appendix with this caveat. In the first, latent clusters pass through two independent random nonlinear maps, to 64 and 48 dimensions, with independent additive noise. The shape (κ,H)(\kappa,H) is chosen by the label-free rule of Q2 from 13 fixed shapes, and FlowAlign and DepthAlign get the same rule. Blind matching, 5 seeds. Including the search over all shapes, T-GROW takes about 26 s per instance, against 35 to 45 s for one EGW run.

NoiseT-GROWQGWEGWGWDepthAlignFlowAlign
0.11.000.580.390.360.530.17
0.250.990.410.350.180.380.51
0.50.820.240.270.140.280.17

In the second, each node has between 2 and 5 children at random, depth and leaf masses are random, and there are 2⋅1052\cdot10^5 points per side. The encoder shapes are fixed in advance, and all tree methods use the same trees. Label-transfer accuracy over ten seeds:

Encoder (κ,H)(\kappa,H)Tree constructionT-GROWDepthAlignFlowAlign
(4,3)(4,3)hierarchical k-means0.393±0.1490.393\pm0.1490.184±0.1270.184\pm0.1270.123±0.0720.123\pm0.072
(4,4)(4,4)hierarchical k-means0.470±0.2590.470\pm0.2590.182±0.1250.182\pm0.1250.127±0.0900.127\pm0.090
(4,3)(4,3)farthest-point0.245±0.1080.245\pm0.1080.124±0.0720.124\pm0.0720.097±0.0830.097\pm0.083
(4,4)(4,4)farthest-point0.312±0.1290.312\pm0.1290.119±0.0680.119\pm0.0680.116±0.0920.116\pm0.092

No regular encoder matches this hierarchy, and T-GROW has the highest mean accuracy in every tested configuration. The differences are significant on paired seeds in all four configurations (Wilcoxon p≤0.014p\le0.014), and with k-means trees T-GROW is better than DepthAlign on all ten seeds. The contribution combines the co-depth hierarchical representation, exact reduction at each resolution and coarse-to-fine matching. Propositions 1 and 2 explain the shared-path information that root-depth summaries can lose. Theorem 1 establishes exact reduction for each represented relation. The distortion analysis motivates relation estimation and gives convergence conditions through Theorem 2.

Q1 "Are the target samples shuffled … when the tree has only 8 leaves?"

See W1 for the shuffled-target and independent-tree checks, and for class-level decoding with eight leaves.

Q2 "How were the branching ... they tuned on the test benchmarks?"

Answer

How the values were chosen. For blind matching, the configuration κ=3\kappa=3, H=4H=4, four trees and the search settings were selected using accuracy and JJ on tuning instances whose seeds are disjoint from those of the reported instances, and were then fixed for evaluation. For SNARE-seq, the shape (4,2)(4,2) was selected using FOSCTTM on the benchmark. We therefore report both the sensitivity across configurations and a fixed label-free selection rule.

Sensitivity. All 13 CIFAR-100 shapes exceed every baseline in mean accuracy; the lowest is 0.3510.351, versus EGW's 0.3120.312. The six SNARE-seq shapes give FOSCTTM from 0.14980.1498 to 0.15090.1509. W2 reports sensitivity to the number of trees.

A label-free rule. We fix the candidate shapes in advance, 13 for blind matching and six for SNARE-seq, run T-GROW with each, and keep the result with the lowest value of a common objective: JJ in blind matching, and the GW objective of the final coupling on the input distances on SNARE-seq. No label or correspondence is used.

TaskLabel-free ruleTime (sec)
CIFAR-100, accuracy0.7240±0.04000.7240\pm0.040030.2; EGW baseline: 54.9
SNARE-seq, FOSCTTM ↓0.1501±0.00020.1501\pm0.000214.6; EGW baseline: 142

Label-free selection improves mean CIFAR-100 accuracy and reduces variability; on SNARE-seq it stays within the narrow range across shapes. These results include the cost of evaluating every candidate shape.

Reviewer_jjTROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces T-GROW. This is a method for making Gromov–Wasserstein matching more practical on large datasets. It groups samples into a tree hierarchy and uses their shared paths from the root to capture relationships between clusters. Matching is then performed at the cluster level, with results combined across multiple randomized trees. The authors show that this reduction preserves the optimal GW value for a fixed tree representation and bound the error introduced by approximating the original geometry. Experiments examine matching accuracy, robustness, and runtime on synthetic data, single-cell data, and vision–language tasks.

Critical strengths
  1. Co-depth retains shared ancestry, while exact leaf aggregation preserves the GW optimum under the fixed tree geometry.
  2. Fixed-resolution scaling is practical; SNARE-seq matches EGW quality with approximately 200× lower reported runtime.
Critical weaknesses
  1. The blind-matching procedure is not fully specified. Appendix A.1.5 explains soft-coupling lifting, but A.4.2 does not specify how each slice produces a class permutation; its Hungarian step combines existing candidates. Classes sharing a leaf are indistinguishable in that slice's geometry (Eq. (11)), making this step essential. Without its specification and the candidate-selection objective, the CIFAR-100 gains might be difficult to attribute to leaf-level GW.

  2. The benefit of tree distances remains unclear. The unary ablation (A.6) supports pairwise information, but does not compare tree distances with original distances between identical representatives. On S3DIS, T-GROW and GW/QGW also use different partitions (A.5.2), so better partitioning could explain some gains. Since Eq. (10) is standard GW on compressed supports, a controlled comparison is needed to establish the value of tree geometry itself.

Questions for authors
  1. How does each slice produce a class permutation? Please provide pseudocode showing how classes sharing a leaf are distinguished, whether original geometry is used for refinement, and which objectives govern Eq. (57) and final selection. If refinement is used, how much does it improve CIFAR-100 accuracy with the trees and search budget fixed?

  2. Can you isolate the effect of tree distances? On S3DIS or SNARE-seq, replace tree distances with original distances between the same representatives, keeping partitions, masses, slices, solver settings, and lifting/aggregation fixed. Report alignment quality and runtime under a consistent normalization convention to assess the contribution of tree geometry.

3 comments
AuthorsOpenReview ↗

Response to Reviewer jjTR Part 1

We thank the reviewer for the constructive comments. We answer each weakness (WiW_i) and question (QjQ_j) below.

W1: "The blind-matching procedure is not fully specified… attribute to leaf-level GW."

Answer

How a slice produces a class permutation. Each slice builds independent trees for the two modalities. Node relations are mass-weighted means of the original class relations, with zero diagonals. The solver enumerates coarse assignments, retains the best candidates and extends them at each finer level by Hungarian matching of child descriptors, followed by refinement of the level's matching objective. Leaf assignments are decoded into class permutations as follows. Q1 gives the pseudocode.

How classes sharing a leaf are distinguished. Each class is described by 16 quantiles of its distances to all other classes in its own modality. Inside a matched pair of leaves, classes are assigned by a Hungarian step on these profiles, so this step uses the original geometry. The full permutation is then refined on the original class relations, and classes may move across the initial leaf matches. The Hungarian step of Eq. (58) is a different one: it combines the candidates of all slices into one consensus candidate.

Objective. Candidate scores, the costs csc_s in Eq. (57), the acceptance of the consensus candidate and the final selection all use Eq. (56), J(π)=∑i,j[Xij−Yπ(i)π(j)]2J(\pi)=\sum_{i,j}\big[X_{ij}-Y_{\pi(i)\pi(j)}\big]^2 with X,YX,Y the standardized class-distance matrices of Eq. (55). The ground-truth correspondence is used only after optimization, to report accuracy and to normalize the reported cost, it is not used for candidate scoring or selection.

What the gains are attributable to. We keep the complete search settings, including up to 6 polished candidates, cost-weighted consensus and 480 kicks, and compare candidate generators. EGW and GW each supply 20 restarts, 5 per slice. The random-hierarchy variant also uses its hierarchy for later tree-based moves. All CIFAR-100 tables in this response use the same 50 instances with shuffled targets: 10 seeds × 5 target permutations. Under the same search settings, T-GROW hierarchy gives the highest mean accuracy and lowest mean JJ. So the gains reflect the combination of geometry-aware hierarchical matching and full-resolution search, the latter alone does not explain the result. (See answer to Q1 for the stage ablation which measures the contribution of full-resolution search.)

Candidates given to the same searchAccuracyFinal JJ / ground-truth JJTime per run (sec)
T-GROW hierarchy0.653±0.1170.653\pm0.1170.9230.9231.8
EGW candidates, 20 restarts0.474±0.2210.474\pm0.2210.9520.95258
GW candidates, 20 restarts0.468±0.2790.468\pm0.2790.9760.9763.3
Random balanced hierarchy0.368±0.2470.368\pm0.2471.0051.0052.0

W2: "The benefit of tree … the value of tree geometry itself."

Answer

The three relations. Let u,vu,v be two nodes at the same depth, with member sets Iu,IvI_u,I_v, masses aia_i, and d(i,j)d(i,j) the original distance between members.

  • Representative distance: d(cu,cv)d(c_u,c_v), with cuc_u the centroid of node uu on S3DIS and SNARE-seq, or its medoid on CIFAR-100. This is the requested control.
  • Tree path distance: DT(u,v)D_T(u,v), each edge length being the distance between the representatives of a parent and its child.
  • Co-depth (ours): the mass-weighted mean of the original distance over all member pairs, R(u,v)=1a(Iu) a(Iv)∑i∈Iu∑j∈Ivaiaj d(i,j),a(Iu)=∑i∈Iuai.R(u,v)=\frac{1}{a(I_u)\,a(I_v)}\sum_{i\in I_u}\sum_{j\in I_v}a_i a_j\, d(i,j),\qquad a(I_u)=\sum_{i\in I_u}a_i.

The last two are defined on the co-depth hierarchy. Proposition 1 describes how co-depth encodes shared-path information on the hierarchy, and the tree path distance of Eq. (9) is built from it with edge lengths. Co-depth keeps the same hierarchy and estimates the relation between its nodes from the input distances. Theorem 1 applies to either relation, since both are fixed relations between node pairs. Co-depth is the relation T-GROW uses in the experiments reported here. On S3DIS we use the root mean square over the same member pairs, computed from the mean and covariance trace of each cell at linear cost.

Controlled relation comparisons. Within each relation control, partitions, masses, slices, solver settings and lifting are identical. The same normalization rule is applied to each relation: S3DIS matrices have zero diagonals and are divided by their maxima. SNARE-seq matrices retain their original scale, CIFAR-100 uses off-diagonal standardization at every level, On S3DIS, we compare the three relations at the leaves over three seeds.

Relation on the same tree partitionLobby accuracyPair C accuracyTime (sec): Lobby / C
Tree path distance0.421±0.0500.421\pm0.0500.181±0.0150.181\pm0.01513.63 / 16.36
Representative distance0.452±0.0420.452\pm0.0420.394±0.0090.394\pm0.00913.62 / 14.12
Co-depth (ours)0.464±0.0380.464\pm0.0380.404±0.0250.404\pm0.02513.64 / 13.51
AuthorsOpenReview ↗

Response to Reviewer jjTR Part 2

Following up on the previous comment for W2.

For SNARE-seq, we vary the leaf budget to examine the effect of relation choice at different resolutions. The leaf-level comparisons use ten seeds, shuffled targets and independent trees, with identical partitions across relations at each budget. We report FOSCTTM, where lower values are better. At 16 leaves, total runtimes for tree path distances, representative distances and co-depth are 0.28 s, 0.33 s and 0.38 s.

LeavesTree path distanceRepresentative distanceCo-depth (ours)
40.5070±0.17800.5070\pm0.17800.4310±0.18500.4310\pm0.18500.1920±0.03000.1920\pm0.0300
80.5680±0.05900.5680\pm0.05900.2580±0.17200.2580\pm0.17200.1550±0.00400.1550\pm0.0040
160.1511±0.00080.1511\pm0.00080.1505±0.00040.1505\pm0.00040.1504±0.00040.1504\pm0.0004
320.2420±0.18500.2420\pm0.18500.1505±0.00040.1505\pm0.00040.1499±0.00020.1499\pm0.0002
640.1940±0.13100.1940\pm0.13100.1503±0.00040.1503\pm0.00040.1498±0.00020.1498\pm0.0002

On CIFAR-100, we compare the relations at every level of the solver using the same 50 shuffled-target instances as in W1:

Relation at each levelAccuracyTotal time (s)
Tree path distance0.358±0.2880.358\pm0.2881.75
Representative distance0.509±0.2740.509\pm0.2741.89
Co-depth (ours)0.631±0.1060.631\pm0.1061.81

Co-depth has the highest mean alignment quality in these controls. It matches or exceeds representative distances on 45 of the 50 CIFAR-100 instances. On SNARE-seq, its advantage is largest with four and eight leaves, where it wins on nine of the ten seeds in each setting; the relations perform similarly at finer resolutions. The CIFAR-100 control uses common preprocessing across variants, explaining its 0.6310.631 accuracy versus 0.6530.653 in W1. Theorem 2 motivates the mean estimator: at a fixed input scale, with within-node means kept on the diagonal, it minimizes distortion among all node-pair relations on the partition. The controls above evaluate its effect on matching.

Partitions. In a separate partition control, we fix the co-depth (ours) relation estimator, solver and readout, and replace the tree partition with Voronoi cells around random anchors, using the same number of cells:

Fixed relation and taskTree partitionRandom-anchor Voronoi
S3DIS Lobby accuracy0.46400.46400.42800.4280
S3DIS Pair C accuracy0.40400.40400.26900.2690
SNARE-seq FOSCTTM ↓0.14980.14980.15400.1540

The value of tree geometry beyond GW on compressed supports. At one level, Eq. (10) is GW on compressed supports. The controls identify three contributions of the hierarchy:

  • A geometry-aware partition: tree partitions outperform random ones with the relation estimator fixed, and geometry-aware hierarchies outperform random splits under the same search settings.
  • A nested family of exact problems: Theorem 1 applies to each level's represented relation, enabling coarse-to-fine matching. Its candidates outperform full-resolution GW and EGW candidates under the same search settings.
  • A data-estimated relation on the hierarchy: co-depth (ours) outperforms representative and tree path distances in the relation controls.

The central idea of the paper is unchanged: keep the pairwise shared-ancestry information of the hierarchy and match at controlled resolutions with an exact reduction. The relation estimator and the class-level decoder make explicit how this is carried out.

Q1 "How does each slice produce a class permutation? Please provide pseudocode … with the trees and search budget fixed?"

Answer

Building on W1, the pseudocode makes the original-geometry decoding and refinement explicit. Candidate scoring, the costs in Eq. (57) and final selection all use JJ of Eq. (56).

Input: class representations, standardized relations X, Y, search settings.

Candidates = {}
For each slice:
  Trees = BuildIndependentTrees(L2-normalized class representations)
  Relations = NodePairMeans(X, Y; Trees), with zero diagonals
  Beam = BestDistinct(EnumerateCoarseAssignments(Trees); coarse QAP cost)

  For each finer level:
    Starts = HungarianProlong(Beam; child descriptors)
    Starts += PerturbedStarts(Starts)
    Beam = BestDistinct(
        RefineLevelQAP(Starts; current-level Relations);
        current-level QAP cost)

  Profiles = QuantileProfiles(original class distances)
  Candidates += WithinLeafHungarian(Beam; Profiles),
                followed by global assignment of unmatched classes

Candidates = BestDistinct(Candidates; J)
Candidates = RefineAndSubtreeSwap(Candidates; J)
Costs = {J(pi) for pi in Candidates}
Consensus = Refine(WeightedConsensus(Candidates, Costs); J)
pi = Best(Candidates + {Consensus}; J)
pi = SubtreeSearch(pi; tree pool, local repair, polishing; J)
return CommonFinalRefinement(pi; J)
AuthorsOpenReview ↗

Response to Reviewer jjTR Part 3

Following up on the previous comment for Q1.

How much refinement improves accuracy. On the same 50 CIFAR-100 instances, we fix the trees and candidate-generation settings and enable later stages progressively:

Stages enabledAccuracyTime per run (sec)
Coarse-to-fine candidates and consensus only0.140±0.0830.140\pm0.0830.58
common final refinement0.264±0.1860.264\pm0.1860.69
polishing of the best six candidates and subtree swaps0.465±0.1650.465\pm0.1650.74
the full method0.653±0.1170.653\pm0.1171.80

The later stages substantially improve accuracy. This stage ablation changes computation as stages are enabled. W1 separately holds the full search settings fixed when comparing candidate generators.

Q2 "Can you isolate the effect of tree distances? … to assess the contribution of tree geometry."

Answer:

See W2 for the controlled relation comparisons, including alignment quality, runtimes and normalization conventions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.