acceptodds
Withdrawn from ICLR 2027

Klein Flow Matching in Hyperbolic Latent Space

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generating realistic graphs requires capturing both local connectivity and large-scale organization, including hierarchical and tree-like patterns. Hyperbolic space is naturally suited to representing these patterns, making it a promising latent geometry for graph generation. We introduce Klein Flow Matching (KFM), a two-stage latent graph generation model that exploits a simple geometric property: hyperbolic geodesics appear as straight line segments in Klein coordinates. Specifically, KFM represents each conditional velocity using a fixed direction and a time-dependent scalar factor. This gives explicit positions and velocity targets for simulation-free flow-matching training. A graph autoencoder provides deterministic latent targets and condition codes; a conditional transformer learns their transport from a wrapped Gaussian base using the Klein metric. Sampling combines Riemannian Heun integration with graph decoding, including structural constraints. Across five main benchmarks, KFM shows competitive results in degree and clustering statistics. On additional benchmarks for generating planar graphs and trees, constrained decoding and candidate selection yield 100% jointly valid, unique, and novel outputs, with the lowest aggregate structural-discrepancy ratio among the compared models on trees. These results demonstrate the practical promise of KFM for generating graphs across diverse structural regimes.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Flow Matching, Hyperbolic space, Klein model

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_JBdROpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Klein Flow Matching (KFM), a two-stage graph generator combining a autoencoder with conditional flow matching in hyperbolic latent space. Its main idea is to express Klein geodesic velocities as a fixed chord direction multiplied by a time-dependent scalar, alongside Klein-metric regression and Riemannian Heun sampling. The structural benchmarks additionally use training-derived conditions, constrained decoding, and candidate selection.

Critical strengths

None

Critical weaknesses
  1. GGBall already combines hyperbolic latent representations with Riemannian flow matching, and KFM’s closed-form Klein geodesics also derive from prior work. Moreover, GGBall’s performance is stronger than KFM on the Community-small and Ego-small in Table 11.

  2. The comparison is unfair. KFM generates 8x candidates, ranks candidates by their distance from the training profile, and favors non-isomorphic candidates, whereas the baselines retain their original protocols. Some of the selection criteria, particularly the degree-related statistics, are closely related to the reported distributional metrics. Furthermore, constrained construction guarantees validity, and even the profile-only control achieves 100% V.U.N. Hence, the reported results do not support better generative performance in a fair way.

  3. Grid, Planar, and Tree use an additional 40D structural profile, with sampled node/edge counts replacing decoder predictions and the degree profile further guiding construction. However, no profile-free ablation is reported, so the contribution of KFM versus the supplied structural information is unclear.

Questions for authors
  1. Does hyperbolic latent space provide a clear advantage for molecular graph generation, especially given that many of the compared baselines have already been evaluated on molecular graphs? Could the authors evaluate KFM on QM9, ZINC250k, MOSES, and GuacaMol?

  2. Why is GGBall, the closest prior method, not included directly in the main results table? In Table 11, GGBall shows better performance than KFM.

Reviewer_GJ1UOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Klein Flow Matching, a latent graph generation method that performs Riemannian flow matching in Klein coordinates. It exploits the straight-chord representation of hyperbolic geodesics to obtain explicit velocity targets with a fixed direction and time-dependent speed. Experiments on several graph benchmarks show competitive performance.

Critical strengths
  1. The paper provides an explicit formulation of hyperbolic flow matching in Klein coordinates, including closed-form geodesic positions and velocity targets.
  2. Within the selected graph benchmarks, the paper provides comprehensive experimental analysis.
Critical weaknesses
  1. The hyperbolic flow matching for graph generation has already been studied. Since the Klein and Poincaré models represent the same intrinsic hyperbolic geometry, the main contribution appears to be a coordinate reparameterization rather than a fundamentally new generative method. The novelty is incremental.
  2. The experiments mainly involve relatively small and unlabeled graph topologies. The evaluation does not demonstrate generation of realistic molecular graphs or other complex attributed structures, especially graphs that show hierarchical features.
  3. The paper motivates Klein coordinates as providing simpler velocity targets, but does not report clear improvements in computational cost or sampling efficiency compared with other Riemannian flow matching methods.
Questions for authors
  1. Can the method be evaluated on more complex attributed graphs, such as molecular graphs with atom and bond types?
  2. Could the authors further clarify how the Klein formulation distinguishes KFM from existing Riemannian and hyperbolic flow-matching methods?
  3. Does the simpler velocity formulation translate into practical improvements in training efficiency, sampling speed, or stability?
Reviewer_9TdkOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In this work, the authors propose to a new flow-based graph generation approach which uses the Klein formulation of hyperbolic space to represent the latents. The choice of hyperbolic space for the latent embedding is motivated by the observation that many hierarchical networks embed more naturally in hyperbolic space because of the ``tree-like'' growth patterns. In [Bu, 2026], a similar approach is taken which uses a Poincar'e representation of the hyperbolic space instead of the Klein representation. In appendices A and B, they provide a brief comparison of the three different representations of hyperbolic space, Poincar'e, Klein, and the hyperboloid model as well as a through description of the flow procedure in the Klein model. In order to evaluate their KFM model, they compare it to the results for other models as reported in [Fu, 2024] and [Luo, 2026].

Critical strengths
  1. There is a thorough documentation of three models of hyperbolic geometry included in the appendix as well as derivation of the flow model necessary for flow on the Klein representation of hyperbolic geometry. However, while useful, I do not believe this discussion is accessible to a general reader of ICLR.

  2. The proposed approach is well motivated in terms of the needs of graph generation and the preliminary results seem promising

Critical weaknesses
  1. The datasets are not adequately described nor are proper citations provided. The two citations provided [You, 2018] and [Fu, 2024] have differing descriptions of the Community and Ego networks and no citation information is provided for the MUTAG nor IMDB-B datasets. Additionally, given the computational limitations of the proposed approach, I have difficulty believing that Grid dataset aligns with the even the 100-400 vertex dataset of [You, 2018] let alone the 1296-2025 vertex data also referred to as Grid. Additionally, rough statistics on these data sets should be provided at least in the appendices.

  2. The authors state that ``Planar and Tree test constrained generation,'' however in the description of the generation method the constraint is enforced exogenously in the decoding step. Given that the number of non-isomorphic unrooted trees is exponential in the number of vertices [Otter, 1948] and even the number of 3-connected planar maps seems to exhibit exponential growth in terms of the number of edges (for instance, see [Walsh, 1982]), for reasonable the resulting graphs will almost certainly be unique and novel. This essentially makes the 100% VUN measure reported vacuous.

  3. As stated by the authors, the primary motivation for a hyperbolic latent space is the ability capture tree-like or hierarchical features of source data. However, the ability of the Klein Flow matching to generate hierarchical or tree-like data is not explicitly evaluated in this paper. There is a possibility that the Ego or IMDB-D datasets could provide some confirmation of this, however insufficient data is provided about these data sets to evaluate their potential hierarchical structure.

  4. As noted by the authors, the GGBall architecture [Bu, 2026] also used hyperbolic representation of the latent space and seems to be the likely closest competitor. While the relatively novelty of the GGBall architecture excuses the limited comparisons (taken directly from [Bu, 2026]), there is also no meaningful discussions about why to prefer the Klein representation over the Poincar'e.

Questions for authors

No questions.

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

Related papers

Discussion (0)

Sign in to comment.