acceptodds
Under review as a conference paper at ICLR 2027

Autoregressive Frontier Expansion: Growing Trees with Graph Machine Learning

Abstract

Tree-like branching structures are common in nature, from botanical trees to neurons, blood vessels and respiratory trees. Their branching shape often reflects function, making structural modelling central to understanding how these systems work. Because acquiring real-world 3D data is often expensive or infeasible, realistic generative models are valuable for simulation and data augmentation. Existing morphology-specific models either constrain how topology is generated or rely on hand-tuned, mechanistic procedures. Generic 3D graph generators, by contrast, do not exploit or enforce the structure of trees. We propose Autoregressive Frontier Expansion, a generative framework that constructs trees through an iterative expansion process, simulating the biological growth of real trees. At each step, a flow-matching model parameterised by an SO(2)-equivariant GNN expands the frontier by predicting whether each active branch bifurcates or terminates. We evaluate our method on cortical neurons and botanical trees in unconditional, class-conditioned, and morphology-guided generation. Across both domains, the generated morphologies agree closely with the reference distributions and, in conditional experiments, with the specified targets.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
tree, graph generation, iterative expansion, graph neural network, persistent homology

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_gpo9OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a method to generate a realistic tree in 3D using Graph Neural network. Unlike other previous works, this work predicts topology and geometry at the same time. This work uses QSM reconstructions from BioDiv-3DTrees dataset to evaluate the performance. From unconditional neuronal morphonolgy generation, this methos shows 100% valid tree generations, and have four best scores out of five, in distributional agreement, compared to SemlaFlow and MorphoGen methods. Also, this method shows an unconditional generation performance as well. This also compared with SemlaFlow, MorphoGen and the proposed method has at least 14 times fewer parameters which use only 1.57 M parameters.

Critical strengths
  • Introducing a new iterative 3D tree generation method that simulates the biological growth of real trees.
  • This approach predicts topology and geometry at the same time unlike other previous works that are introduced in this paper.
Critical weaknesses
  • Missing core tree related works including Space Colonization, Tree-D Fusion.
  • The title and abstract show the major contribution of this paper is generating a realistic tree. But why it shows molecular generation for the validation?
  • Line #18-#20, the authors claim that “simulating the biological growth of real trees”, yet the method is compared with QSM, which is non-deterministic reconstruction algorithm and this paper does not show any tree related metrics and validation such as DBH, the number of branches, branching angles, etc.
  • From my view, the trees that are visualized in Figure 5, does not realistic as the paper claims in Line #18-#20. Also, in Figure 6, the trees in d_max=20, in the middle shows flat branches which look unrealistic.
Questions for authors
  • The title says “Growing Trees” and other parts showing some neurons and trees. Which one is the main contribution of this paper?
  • There are no tree specific evaluations in this paper. Why those are not included and why authors say Line#18-#20 that “simulating the biological growth of real trees”?
  • How do you know the generated trees in Figure 5 and Figure 6 are realistic?
  • One of the baselines, SemlaFlow, uses QM9 and GEOM dataset. Any reasons authors did not make comparisons using those datasets?
Reviewer_9WiEOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method for generating binary trees. Trees are iteratively grown, achieved by gathering all leaves at an intermediate generation stage, and predicting their 3D positions, as well as whether the node should bifurcate or terminate as a leaf. An SO(2)-equivariant GNN is used to predict a velocity field for both positions and (continuously-relaxed) bifurcation indicators, optimized via flow matching. Conditional generation is further supported to ensure trees satisfy topological features, while still exhibiting geometric variability.

The main contribution of the paper is a method of binary tree generation which ensures topological validity, while still adhering closely to features in the data distribution. The method is evaluated over neurons and botanical trees, and compares favorably against existing methods for molecule generation and more general tree generation.

Critical strengths

The main strength of the paper is the combination of two main ideas: tree-conditioned generation of 3D positions at the tree's frontier, and a method of iterative tree expansion by predicting whether or not to bifurcate frontier nodes. The approach is simple, yet effective, as demonstrated by its comparison against other graph generation methods across a variety of metrics, and in unconditional and conditional generation settings. Although limited to binary trees, it seems plausible to extend the approach to more general trees, with more significant modifications to support graphs; as such, the approach could serve to inspire future work in graph generation.

Critical weaknesses

The main weakness of the paper is its restriction to binary trees. This impacts the datasets used for evaluation, where n-ary trees must be converted to binary trees prior to training and testing. Moreover the conversion can lose information, where for nodes of degree > 3 "... we keep the two thickest children and drop the rest" (it is further unclear what "thickest" means). A consequence is that the results shown for MICrONS neurons and BioDiv-3DTrees trees correspond to modifications of the original data. In contrast, methods that generate generic trees do not face the same limitations, yet for consistent comparison must be run on the same modified data. Moreover, it would be ideal to compare against methods that generate binary trees. However, SemlaFlow is focused on graph generation, while MorphoGen generates n-ary trees. It is thus somewhat expected that the proposed method, which generates a more restrictive class of graphs, would be an improvement over prior works that cannot make simplifying assumptions based on binary trees.

Questions for authors

To what extent can the improvement over prior works be attributed to the assumption of a principal axis? It would be helpful to ablate this choice to more clearly show the benefits of the proposed work.

As the approach is focused on iterative expansion of trees, how would the approach compare to other expansion-like generation methods (see [1] below), as well as tree generation schemes (see [2] below)? A discussion on how these methods relate to the prior work, if not experimental comparisons, would strengthen the paper.

[1] Rose, Daniel, Roxane Axel Jacob, Johannes Kirchmair, and Thierry Langer. "NEAT: Neighborhood-Guided, Efficient, Autoregressive Set Transformer for 3D Molecular Generation." arXiv preprint arXiv:2512.05844 (2025).

[2] Wang, Hanxiao, Biao Zhang, Jonathan Klein, Dominik L. Michels, Dong-Ming Yan, and Peter Wonka. "Autoregressive generation of static and growing trees." In Proceedings of the SIGGRAPH Asia 2025 Conference Papers, pp. 1-12. 2025.

Reviewer_q8Q9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes autoregressive frontier expansion, a generative model that grows a tree one level at a time by predicting, for every node on the current frontier, whether it bifurcates and where its children are placed. Both quantities come from a single flow-matching model parameterized by an SO(2)-equivariant GNN, so topology and geometry are generated together and every sample is a valid tree by construction. The experimental results show the effectiveness of the proposed method.

Critical strengths

S1. This paper proposes autoregressive frontier expansion, an iterative procedure that conditions future branching on the intermediate morphology.

S2. It is interesting to jointly predict both topology and geometry.

S3. Both 26,469 MICrONS neurons and 3,386 botanical trees are used in the experiments.

Critical weaknesses

W1. This paper is not well motivated. For example, the paper only focuses on the generation of independent single trees/neurons, without considering the environmental constraints in real biological scenarios. However, in real situations, neuron growth is restricted by the spatial constraints of surrounding neural circuits, and tree growth may be affected by light/nutrient competition from adjacent vegetation. The potential application value of the proposed method may be limited in real scenarios.

W2. The structural simplification assumptions should be carefully evaluated. This paper contracts all path nodes with degree 2, forcibly converts multifurcations of non-root nodes into binary structures, and only retains branch points and leaf nodes, which is equivalent to abandoning the modeling of local branch smooth morphology and natural multi-furcation structures. Also, the paper does not demonstrate the impact of this simplification on the generation effect, nor does it explain its applicable scenarios.

W3. The downstream task validation is missing. This paper mentions that the model can be used for data augmentation, however all experiments only verify the matching degree of statistical distributions between generated samples and real samples, and do not verify the actual effect of generated samples for downstream tasks (such as neuron reconstruction, tree biomass estimation). It seems that the application value of the proposed method lacks empirical support.

W4. The writing of the paper could be improved. For example, the contributions of this work should be clearly summarized and presented at the end of the first section. For another example, there is just one numbered subsection (i.e., 1.1 RELATED WORK) in Section 1 (i.e., 1 INTRODUCTION).

W5. Some SOTA baselines should be considered, such as [1][2][3][4].

[1] SynCoGen: Synthesizable 3D Molecule Generation via Joint Reaction and Coordinate Modeling. ICLR 2026

[2] NEAT: Neighborhood-Guided, Efficient, Autoregressive Set Transformer for 3D Molecular Generation. arxiv 2025

[3] Toward Better Geometric Representations for Molecule Generative Models. arxiv 2026

[4] Learning Canonical Representations for Unified 3D Molecular Modeling. 2026

Questions for authors

W1-W5

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of revisions and responses to reviewer concerns

We thank the reviewers for their constructive feedback. In response, we have substantially revised the manuscript with new experiments, analyses, and clarifications to explain the method and scope more clearly, and better illustrate our contribution.

Strengths

  • Reviewers recognised the core contribution as an interesting, simple, and effective way to jointly generate tree topology and 3D geometry.
  • They highlighted that the model conditions future branching on the intermediate morphology, produces valid trees by construction, and performs well across unconditional and conditional settings.
  • They noted the large-scale evaluation on 26,469 neurons and 3,386 botanical trees, the breadth of metrics and comparisons, and the method’s potential to inform future work on tree and graph generation.

Changes in the revised paper

The main revisions are listed below:

  • New baselines. We added NEAT and TheMol on neurons and botanical trees, including metrics, runtimes, validity checks, and samples.
  • Local branch geometry. We added and evaluated a post-hoc model that restores fine neuron paths.
  • Binarisation analysis. We measured node removal and a normalised W1W_1 shift of 0.020, which is below real–real variation.
  • New visual evidence. We added reference-versus-generated botanical trees and cross-domain galleries for both new baselines.
  • Related work. We added Space Colonization, Tree-D Fusion, and Wang et al.’s autoregressive tree generator.
  • Presentation. We added contribution bullets, a standalone Related Work section, a clearer method description, and more precise captions.

Reviewer q8Q9

  • W1. Environmental context - scope clarified. The current scope is individual morphologies, with environment-conditioned generation left for future work.
  • W2. Structural simplification - added and evaluated. We added local-path reconstruction, quantified binarisation effects, and used the same processed reference for every method.
  • W3. Downstream validation - scope clarified. Data augmentation is presented as motivation, while downstream-task validation remains separate future work.
  • W4. Presentation - revised. We added contribution bullets and promoted Related Work to its own section.
  • W5. Baselines - added and discussed. We added NEAT and TheMol and explained the applicability or availability limits of the other suggested methods.

Reviewer 9WiE

  • W1. Binary-tree restriction - quantified and clarified. We measured the effect of binarisation and explained how the comparison remains fair.
  • Q1. Principal-axis bias - clarified. We report gains on rotation-invariant metrics against E(3)-equivariant and non-equivariant baselines.
  • Q2. Related generators - added and discussed. We evaluated NEAT and added Wang et al. to related work; their implementation is unavailable.

Reviewer gpo9

  • Missing related work - added. We added Space Colonization and Tree-D Fusion and explained how their assumptions differ from ours.
  • Scope of “trees” - clarified. We clarified that no molecules are generated and that both evaluation domains are tree-shaped branching structures.
  • Biological-growth claim - revised. We changed “simulating biological growth” to “inspired by biological growth” because the datasets are static snapshots.
  • Tree-specific evaluation - clarified. We report branch length, angles, branch order, path length, and asymmetry; DBH is undefined without predicted radii.
  • Visual realism - revised and extended. We clarified that Fig. 5 shows neurons, added a botanical reference comparison, and reconstructed fine paths in the neuron figures.
  • Molecular datasets - scope clarified. We clarified that QM9 and GEOM contain cyclic molecules and are outside the scope of tree generation.

As the discussion period is nearing its end, we would appreciate feedback on whether the changes and clarifications resolved the points raised, or which points remain outstanding.

Thank you again for your thoughtful reviews. We have responded in detail to each point raised and revised the manuscript accordingly. With the discussion period closing soon, we would greatly appreciate a brief indication of whether any concerns remain.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.