acceptodds
Under review as a conference paper at ICLR 2027

Differentiable Latent-Tree Causal Discovery via Triplet Branch-Point Objectives and Three-Terminal Root Identification

Abstract

Discovering directed causal structure in the presence of latent variables remains challenging, particularly when identifiability principles must be translated into practical learning objectives. We study latent-tree structural causal models, in which observed variables are leaves and latent variables form the internal structure of a directed tree. We introduce a differentiable framework for recovering both the latent-tree skeleton and its causal orientation from observational data. For skeleton learning, we develop a differentiable triplet branch-point (TBP) formulation for latent-tree completion: using the branch-additivity form of the established median property, simplex-valued triplet-center assignments jointly learn unknown branch-point identities and latent-related distances. For causal orientation, we establish a tree-conditioned three-terminal independence (TTI) characterization of the latent root under linear non-Gaussian LT-SCM assumptions; orienting all edges away from the identified root then recovers the directed latent-tree causal graph. Synthetic experiments across diverse latent-tree structures demonstrate strong structure-recovery performance, while experiments on the White Wine Quality dataset and a VAE-based generative model illustrate the applicability of the proposed framework to real data and end-to-end representation learning.

Primary area:
causal reasoning
Keywords:
Causal Discovery, Differentiable Structure Learning, Latent Variable Models, Latent Tree Models, Causal Identifiability

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_Lj2MOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies causal structure learning for latent-tree structural causal models from observational data. The authors propose a differentiable framework that combines two main components. First, they formulate latent-tree skeleton recovery as a differentiable triplet branch-point (TBP) completion problem, where simplex-valued assignments are used to jointly infer unknown branch-point identities and latent-related distances. Second, they introduce a three-terminal independence (TTI) condition for identifying the unique latent root under linear non-Gaussian assumptions, after which all edges can be oriented away from the recovered root. The paper also provides population-level identifiability results and evaluates the proposed method on a broad collection of synthetic latent-tree structures, the White Wine Quality dataset, and a VAE-based proof-of-concept experiment.

Critical strengths
  • The paper presents a novel differentiable formulation for latent-tree causal discovery. The proposed TBP objective provides an interesting way to jointly recover latent-tree structure and latent branch-point information.
  • The TTI criterion provides a principled mechanism for causal orientation. Under the linear non-Gaussian latent-tree assumptions, the paper establishes a root-identification result that enables recovery of the directed tree.
Critical weaknesses
  • There is a gap between the theoretical guarantees and the practical algorithm. The main recovery results assume that the learned distance matrix is already a tree metric, while the proposed optimization does not explicitly enforce this condition.
  • The empirical validation of some key claims is still limited. In particular, the paper does not directly evaluate root-identification accuracy for TTI.
  • The applicability of the method is limited by its strong reliance on a linear non-Gaussian latent-tree model with a unique latent root.
Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Lj2M

We thank the reviewer for recognizing our proposed differentiable latent-tree learning framework and the theoretical contribution of TTI to root identification, as well as for raising important questions and suggesting additional experiments. We address each point below.

W1: Gap Between the Optimization Objective and the Tree-Metric Guarantee

Theorem 5.1 establishes that a tree metric with zero TBP loss admits an additive-tree realization. In practice, we implement this process in two stages: we first use the differentiable TBP objective to learn information distances involving latent nodes, and then apply MST to extract an explicit tree skeleton from the learned distances. Ideally, when the learned distances satisfy the tree-metric condition and the TBP loss is sufficiently small (ideally zero), the MST-derived skeleton provides the additive-tree realization characterized by Theorem 5.1. We also explored incorporating explicit tree-structure constraints directly into the optimization, but found them less effective in practice. Empirically, we observe that TBP optimization typically produces information-distance matrices that are already close to those induced by the ground-truth tree, allowing the subsequent MST step to recover the skeleton effectively. The strong structural recovery results reported in the main paper further support this two-stage design.

W2: Empirical Evaluation of TTI's Root Identification Capability

We thank the reviewer for the suggestion on additional experiments: To independently evaluate the root-identification capability of TTI while excluding the influence of skeleton recovery errors, we conduct additional oracle-skeleton experiments. Specifically, we directly provide TTI with the ground-truth undirected latent-tree skeleton, bypassing the skeleton recovery stage, and consider all latent nodes as candidate roots.

We conduct experiments on three structures corresponding to Figures 6(a), 3, and 7 in the paper. Figures 6(a) and 3 contain six observed nodes and three and four latent nodes, respectively, while Figure 7 contains 30 observed nodes and four latent nodes. For each structure, we use 50 datasets, each containing 5,000 samples with exponential noise. The root recovery rate is 100% across all three experimental settings. That is, the root selected by TTI agrees with the task-specified ground-truth root mapping in every run. By bypassing the skeleton recovery stage, these experiments directly evaluate the root-selection performance of TTI given the correct skeleton. We will further evaluate this experiment across all structures considered in the paper and include the corresponding results in the revised manuscript.

W3: Model Applicability and Methodological Novelty

We agree with the reviewer's observation that the model class studied in this paper is relatively restrictive, particularly in terms of the tree-structure assumption and the general SCM assumptions. At the same time, we view the tree structure as a deliberate structural inductive bias that captures sparse, hierarchical, and interpretable latent causal relationships. Furthermore, the novelty of our work lies not simply in imposing a tree-structure constraint, but in translating theoretical properties of latent-tree models into differentiable structure-learning criteria and further enabling complete causal structure recovery. By integrating information-distance learning, skeleton recovery, root identification, and causal orientation into a unified framework, our work establishes a connection between theoretical identifiability and practical differentiable learning.

We sincerely thank the reviewer again for their patience and valuable suggestions.

Thank you for the detailed response and additional experiments, which have addressed my concerns. I appreciate the authors' clarifications and will maintain my current positive rating.

Thank you very much for your positive assessment and constructive suggestions. Following your suggestion, we have added an evaluation of TTI's root-identification performance using ground-truth skeletons to Section 6.2 and the corresponding appendix of the revised manuscript.

Reviewer_48cMOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies causal discovery in linear non-Gaussian latent-tree SCMs, where observed variables are leaves and latent variables are internal nodes. It proposes a differentiable TBP objective to jointly learn latent-related distances and triplet-center assignments, followed by a TTI criterion for root identification and edge orientation. Experiments cover synthetic latent trees, the White Wine Quality dataset, and integration with a VAE.

Critical strengths
  1. The proposed framework separates skeleton recovery from orientation, making the roles of the assumptions explicit.

  2. TTI characterizes the root under the stated assumptions, while the TBP formulation translates established branch-point geometry into an optimizable objective.

  3. The benchmark covers 190 topologies, and the near-Gaussian experiments support the different sensitivities of skeleton recovery and orientation.

Critical weaknesses
  1. This paper investigates a restrictive setting where the underlying SCM is a latent variable linear non-Gaussian acyclic model with the causal structure being a tree, each latent variable having degree at least three, and each variable having at most one parent. According to my knowledge, this setting is strictly more restrictive than those studies in [1, 2]. If I'm wrong, please clarify the differences and include their proposed method as baselines in the revised paper.

  2. The recovery guarantees require the learned distances to form a tree metric, but the optimization does not enforce this condition. Consequently, minimizing the TBP objective does not directly establish the premises needed for correct recovery.

  3. Comparisons with classical latent-tree reconstruction methods, including combinations with the same TTI orientation procedure, are needed to isolate the benefit of TBP-based skeleton learning.

Reference

[1] Structural estimation of partially observed linear non-gaussian acyclic model: A practical approach with identifiability. ICLR 2024.
[2] Efficient and trustworthy causal discovery with latent variables and complex relations. ICLR 2025.

Questions for authors

Please see weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 48cM

We thank the reviewer for recognizing the experimental contributions of our work. We also appreciate the reviewer for pointing out the issues concerning model applicability and the need for comparisons with classical latent-tree reconstruction methods. We address each point below.

W1: Relationship to More General Latent-Variable Causal Discovery Methods

We agree with the reviewer that the methods suggested by the reviewer, PO-LiNGAM [1] and ETCD [2], as well as the LaHiCSL baseline originally included in our paper, have broader structural applicability to latent-variable causal discovery. At the same time, our work pursues a different methodological objective: developing a differentiable framework for latent-tree causal discovery based on continuous optimization, integrating skeleton recovery, latent-node identification, and causal orientation into a unified pipeline, with the potential for further integration into end-to-end deep learning models. In particular, these existing methods do not offer the same type of unified, end-to-end differentiable structure-learning framework based on continuous optimization, which is the focus of our work.

Following the reviewer's suggestion, we include PO-LiNGAM and ETCD as additional baselines and compare them on the structure corresponding to Figure 3 in our paper, which contains six observed nodes and four latent nodes. We use exponential noise, 5,000 samples, and 50 independent datasets. The results are as follows:

StructureMethodLCER ↑SER ↑SHD ↓
Figure 3PO-LiNGAM0%0%11.72 ± 3.69
Figure 3ETCD78%78%0.59 ± 1.67
Figure 3Ours100%100%0.00 ± 0.00

The results show that our method achieves better performance on this representative latent-tree structure. We will include these two methods as additional baselines in the revised manuscript and provide comparisons on more structures.

W2: Gap Between the Optimization Objective and the Tree-Metric Guarantee

Theorem 5.1 establishes that a tree metric with zero TBP loss admits an additive-tree realization. In practice, we implement this process in two stages: we first use the differentiable TBP objective to learn information distances involving latent nodes, and then apply MST to extract an explicit tree skeleton from the learned distances. Ideally, when the learned distances satisfy the tree-metric condition and the TBP loss is sufficiently small (ideally zero), the MST-derived skeleton provides the additive-tree realization characterized by Theorem 5.1. We also explored incorporating explicit tree-structure constraints directly into the optimization, but found them less effective in practice. Empirically, we observe that TBP optimization typically produces information-distance matrices that are already close to those induced by the ground-truth tree, allowing the subsequent MST step to recover the skeleton effectively. The strong structural recovery results reported in the main paper further support this two-stage design.

AuthorsOpenReview ↗

Response to Reviewer 48cM (Continued)

W3: Comparison with Classical Latent-Tree Structure Recovery Methods

We thank the reviewer for suggesting additional experiments to evaluate the contribution of TBP to latent-tree skeleton learning.

We conduct additional comparison experiments with three classical methods: Neighbor Joining (NJ) [3], Recursive Grouping (RG) [4], and SLLT [5,6]. These methods are classical algorithmic reconstruction approaches rather than end-to-end differentiable structure-learning frameworks like ours. They primarily recover undirected latent-tree structures. Therefore, for a fair comparison, we feed the undirected skeletons recovered by these methods into the same TTI procedure used in our method for root identification and causal orientation. The experiments are conducted on the two structures corresponding to Figures 3 and 6(a), with 5,000 samples, exponential noise, and 50 independent datasets for each structure. All distance-based methods use the same empirical information distances.

The results are as follows:

StructureMethodLCER ↑SER ↑SHD ↓
Figure 3Ours (TBP + TTI)1.001.000.00 ± 0.00
Figure 3Neighbor Joining + TTI [3]0.060.069.60 ± 3.51
Figure 3Recursive Grouping + TTI [4]0.060.069.36 ± 3.31
Figure 3SLLT + TTI [5,6]0.560.562.20 ± 2.51
Figure 6(a)Ours (TBP + TTI)1.001.000.00 ± 0.00
Figure 6(a)Neighbor Joining + TTI [3]0.080.087.20 ± 3.22
Figure 6(a)Recursive Grouping + TTI [4]0.080.087.90 ± 3.21
Figure 6(a)SLLT + TTI [5,6]0.640.642.00 ± 2.86

Our method achieves complete recovery on both structures. Among the three classical methods, SLLT performs better than NJ and RG, but still exhibits noticeable structural recovery errors. We will include these experiments, the corresponding literature references, and the relevant discussions in the revised manuscript.

We sincerely thank the reviewer again for these valuable suggestions and guidance on the additional experiments.

References

[1] Structural Estimation of Partially Observed Linear Non-Gaussian Acyclic Model: A Practical Approach with Identifiability. ICLR, 2024.

[2] Efficient and Trustworthy Causal Discovery with Latent Variables and Complex Relations. ICLR, 2025.

[3] The Neighbor-Joining Method: A New Method for Reconstructing Phylogenetic Trees. Molecular Biology and Evolution, 4(4):406–425, 1987.

[4] Learning Latent Tree Graphical Models. Journal of Machine Learning Research, 12:1771–1812, 2011.

[5] Spectral Methods for Learning Discrete Latent Tree Models. Statistics and Its Interface, 10(4):677–698, 2017.

[6] Learning Mixed Latent Tree Models. Journal of Machine Learning Research, 21(249):1–35, 2020.

Thank you for your response. I would appreciate it if you could incorporate a more comprehensive discussion of [1,2] into your paper, addressing the following points:

  1. Explicitly acknowledge that PO-LiNGAM [1], ETCD [2], and LaHiCSL have broader structural applicability to latent-variable causal discovery.

  2. Clarify that your work focuses on developing a differentiable framework for latent-tree causal discovery through continuous optimization, and explain in detail the distinctive advantages of this framework.

  3. Provide experimental comparisons with these methods.

According to the author guideline, authors can revise papers during the rebuttal period. If these points are adequately addressed in the revised paper, I would be willing to increase my score.

[1] Structural estimation of partially observed linear non-gaussian acyclic model: A practical approach with identifiability. ICLR 2024.
[2] Efficient and trustworthy causal discovery with latent variables and complex relations. ICLR 2025.

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

The paper proposes a differentiable objective for completing a latent-tree metric from observed correlations. Given the recovered skeleton, a three-terminal independence criterion then identifies the latent root under a linear, non-Gaussian model; all edges are then oriented away from it. The evaluation covers synthetic trees, a real-world wine dataset, and a VAE integration example.

Critical strengths
  • The root characterization is a useful and interesting theoretical contribution within the stated model: it gives a specific independence test for an otherwise unobserved orientation.
  • The synthetic evaluation covers substantially different tree topologies.
Critical weaknesses
  • It seems that the main guarantee does not yet justify the learning algorithm.

    • Theorem 5.4 assumes the completed distances form a tree metric and their realization has the required latent-tree structure.
    • The authors explicitly note that zero TBP loss does not establish tree-metric membership.
    • However, Algorithm 1 optimizes distances without imposing it and later constructs an MST and collapses nodes.
    • Therefore, I would like to know when this procedure actually reaches the theorem’s conditions.
  • The comparisons also need care: e.g., CDHS and DCD are evaluated on only 106 and 8 of the 190 structures, respectively, so their reported aggregated numbers are not directly comparable.

Questions for authors

Across the synthetic runs, how often does the learned distance matrix satisfy or approximately satisfy the four-point condition before the MST step? How sensitive is root identification to skeleton errors? Also, even for methods with model misspecification (like CDHS and DCD), it would still be interesting to see how they empirically behave.

4 comments
AuthorsOpenReview ↗

Response to Reviewer jJcc

We thank the reviewer for recognizing the theoretical contribution of TTI and for raising important questions. We address the reviewer's questions and suggestions point by point below.

Q1&W1: Theoretical Guarantees and The Four-Point Condition

Theorem 5.1 establishes that a tree metric with zero TBP loss admits an additive-tree realization. In practice, we implement this process in two stages: we first use the differentiable TBP objective to learn information distances involving latent nodes, and then apply MST to extract an explicit tree skeleton from the learned distances. Ideally, when the learned distances satisfy the tree-metric condition and the TBP loss is sufficiently small (ideally zero), the MST-derived skeleton provides the additive-tree realization characterized by Theorem 5.1. We also explored incorporating explicit tree-structure constraints directly into the optimization, but found them less effective in practice. Empirically, we observe that TBP optimization typically produces information-distance matrices that are already close to those induced by the ground-truth tree, allowing the subsequent MST step to recover the skeleton effectively. The strong structural recovery results reported in the main paper further support this two-stage design.

To address Q1, we conduct additional experiments to evaluate how well the learned information-distance matrix satisfies the four-point condition and the three-point structural property captured by TBP, after optimization but before MST construction. We evaluate the ground-truth tree structure in Figure 3 using 50 independent runs, with 5,000 samples per run and exponential noise. The complete learned information-distance matrix DVVD_{VV} contains six observed nodes and four preallocated latent nodes.

First, the four-point condition characterizes whether the pairwise distances among observed nodes admit an additive-tree representation, potentially involving latent internal nodes. Therefore, we evaluate the four-point condition on the observed-node distance submatrix DOOD_{OO}. For any four observed nodes i,j,k,li,j,k,l, we calculate the three pairwise-distance sums and sort them in ascending order as s(1)≤s(2)≤s(3)s_{(1)}\leq s_{(2)}\leq s_{(3)}. We define the normalized four-point violation as g(i,j,k,l)=s(3)−s(2)s(3).g(i,j,k,l)=\frac{s_{(3)}-s_{(2)}}{s_{(3)}}. For each run, we average the normalized violations over all 15 quartets and consider the run to pass if the resulting average does not exceed 0.050.05. The mean normalized four-point violation of DOOD_{OO} is 0.010427, and all 50 runs (100%) satisfy this criterion.

Second, to evaluate how well the learned distances satisfy the three-point structural property of TBP, we report the optimized TBP loss. Across the same 50 runs, the mean and standard deviation of the TBP loss are 0.012811 ± 0.021798 (mean ± sample standard deviation).

Overall, these results provide empirical evidence that the learned distances approximately satisfy both the four-point condition and the three-point structural property captured by TBP.

AuthorsOpenReview ↗

Response to Reviewer jJcc (Continued)

Q2: Sensitivity of TTI to Skeleton Errors

We conduct controlled experiments on the two structures corresponding to Figures 6(a) and 3. Both structures contain six observed nodes, with three and four latent nodes, respectively. We use exponential noise and a sample size of 5,000. First, for each structure, we enumerate all valid candidate trees while keeping the numbers of observed and latent nodes fixed. We then sample two erroneous skeletons at each specified Robinson–Foulds (RF) distance. The RF distance measures the topological difference between an erroneous skeleton and the ground-truth skeleton, with larger distances indicating greater structural differences. Each erroneous skeleton is evaluated on 50 datasets: RF = 1 and 2 for Figure 6(a), and RF = 1, 2, and 3 for Figure 3.

To analyze the effect of skeleton errors on TTI's root-scoring performance, we conduct the following additional experiments. Given an estimated skeleton, TTI computes nHSIC scores for all candidate latent roots (where nHSIC is the average of multiple HSIC evaluations) and selects the candidate with the smallest nHSIC as the root. In each run, we record the nHSIC scores of all candidate latent roots and rank them in ascending order. Rank 1 corresponds to the candidate with the smallest nHSIC, which is the candidate actually selected by TTI. The table reports the mean and standard deviation of nHSIC for each rank.

Furthermore, to account for differences in the scale of the scores, we introduce the Rank-1/Rank-sum metric, defined as the mean nHSIC of Rank 1 divided by the sum of the mean nHSIC values across all candidate ranks. A smaller value indicates that Rank 1 accounts for a smaller proportion of the total candidate scores, reflecting a more pronounced relative scoring advantage over the other candidates.

The results are as follows:

FigureSkeletonRank 1Rank 2Rank 3Rank 4Rank-1 / Rank-sum ↓
Figure 6(a)Oracle2.838 ± 0.23911.649 ± 4.87921.905 ± 8.221—7.798%
Figure 6(a)RF12.779 ± 0.3798.268 ± 2.82017.469 ± 7.947—9.744%
Figure 6(a)RF24.614 ± 2.5097.122 ± 3.26912.447 ± 4.029—19.081%
Figure 3Oracle2.804 ± 0.3389.343 ± 4.07313.671 ± 4.75321.629 ± 8.4505.909%
Figure 3RF15.179 ± 1.8637.666 ± 2.92311.829 ± 4.58419.626 ± 8.12811.692%
Figure 3RF23.457 ± 1.3095.980 ± 2.0337.848 ± 2.69115.151 ± 7.81810.658%
Figure 3RF34.384 ± 2.0866.943 ± 2.7539.670 ± 3.29812.572 ± 5.40213.061%

Note: The nHSIC values are reported in units of 10−310^{-3}.

The results indicate an overall decline in the discriminative power of TTI's candidate-root scores as skeleton errors increase, as reflected by the overall increasing trend in Rank-1/Rank-sum. For the structure in Figure 3, although some fluctuations occur, the overall trend remains consistent: Rank-1/Rank-sum is higher under every erroneous-skeleton setting than under the correct-skeleton setting.

These results suggest that skeleton errors affect TTI's candidate-root scores, reducing the relative scoring advantage of the selected Rank-1 candidate over the other candidates. The extent of this effect also depends on the specific topological errors. We emphasize that these findings are preliminary, and the potential effects of different forms of structural misspecification warrant further investigation.


Q3&W2: Baseline Applicability and Fair Comparison

(1) Comparisons on the Topology Sets Supported by Each Method

Our original benchmark contains 190 topologies, while LaHiCSL, CDHS, and DCD are applicable to 190, 106, and 8 of these topologies, respectively. To ensure a fair comparison, we consider the topology set supported by each baseline and evaluate our method on the corresponding subset.

Comparison ScopeMethodSER ↑SHD ↓
All topologies(190)LaHiCSL0.028.60 ± 4.81
All topologiesOurs0.535.40 ± 4.49
CDHS-supported topologies(106)CDHS0.1415.52 ± 9.30
CDHS-supported topologiesOurs0.483.40 ± 3.67
DCD-supported topologies(8)DCD0.00—
DCD-supported topologiesOurs0.4503.900 ± 4.150

(2) Empirical Performance Beyond the Applicability Conditions

We further evaluate CDHS and DCD on topologies that do not satisfy their respective applicability conditions. Specifically, we test CDHS on the additional 84 topologies and DCD on the additional 182 topologies. Both methods achieve a zero proportion of correctly recovered structures.

MethodNumber of TopologiesLCER ↑SER ↑
CDHS840.00% ± 0.00%0.00% ± 0.00%
DCD1820.00% ± 0.00%0.00% ± 0.00%

Note: SHD is not calculated because the recovered structures are completely incorrect.

We sincerely thank the reviewer for the valuable suggestions. The perspective afforded by these supplementary experiments has helped us reflect more deeply on our work.

I thank the authors for the clear response. I keep my positive score.

Thank you very much for your positive assessment. Your valuable suggestions have helped us further improve the manuscript. We have also included the additional experiments you suggested in the appendix of the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.