acceptodds
Under review as a conference paper at ICLR 2027

Block Disentanglement in CRL: Bridging Identifiability and Visual State Estimation

Abstract

Causal representation learning (CRL) is the process of recovering causally-related latent variables from high-dimensional observations. As a label-free inference method, CRL is particularly attractive for applications where data labels are unavailable or impractical to obtain. While there has been significant progress in understanding the identifiability guarantees of CRL, such guarantees often hold under highly stylized assumptions, which temper the direct application to real-world problems. This paper has a two-fold objective for _interventional_ CRL. First, it establishes identifiability guarantees for substantially weaker interventional assumptions, resulting in _block disentanglement_ of the causal variables, where the block structure depends on the realistically available intervention mechanisms. Secondly, the block disentanglement framework is used for embodied visual state estimation, in which the objective is to recover the latent physical variables of a robotic system directly from visual data (images and videos) _without labeled data._ These two components are critically complementary. The block disentanglement theory delineates identifiability guarantees under weakened assumptions, and the application demonstrates that the resulting objective remains effective in a controlled embodied setting despite further assumption violations, providing a theory-to-practice bridge needed to translate the promise of _label-free_ CRL into practical problems.

Primary area:
causal reasoning
Keywords:
causal representation learning, block disentanglement, identifiability, theory-to-practice bridge

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ozQjOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper aims to relax six of the common strong assumptions of causal representation learning methods. Weaker assumptions imply weaker results and instead of disentangling each generative factors individually, they disentangle a subset of generative factors I\mathcal I in one unique block. They prove that under certain assumptions that the minimization of their loss imply block-disentanglement. Finally they use their method on a moving robotic arm environment.

Critical strengths

The author suggest a new set of assumptions and its corresponding identifiability theorem

Critical weaknesses

The experiments are clearly not enough to assess if the method achieve disentanglement. The classical disentanglement method seems unrelated to the purpose of the paper and don't seem to have a pertinent baseline. The block-disentanglement benchmark do not have any baseline, rely on one seed of one experiments where there is nothing to disentangle and the metric do not measure the correct property.

Questions for authors
  1. Is your "block disentanglement" related to "partial disentangled" of [3] ?
  2. You state in your conclusion that "individual variables cannot be recovered separately", is it theoratically grounded ?
  3. Graphics in Appendix 6 are hard to read
  4. What is the purpose of the scale loss (E[∥h(x)∥]−1)2(\mathbb{E}[\|h(x)\|]-1)^2 ?

Concerning experiments:

  1. The classical disentanglement method seems to be the one introduced by [1] and relies on the assumption that you aim to relax. What is your contribution for this method and how is it related to the purpose of the paper ?
  2. The classical disentanglement experiments have for only baseline a partially supervised method, why not use unsupervised method ? It seems hard to compare RoboPEPP and your method
  3. Why the benchmark of block-disentanglement do not have any baseline ? The R2R^2 metric is highly dependant on the regressor used, therefore it seems impossible to analyze your results without any baseline. You could use other block/partial disentanglement method or you could use classical disentanglement method and aggregate their representation such as done for the baseline in [4] ?
  4. Additionnaly, the metric R2(ZI∣Z^I^)R^2(Z_{\mathcal I}\mid \hat Z_{\hat{\mathcal I}}) only measure one-way regression and not a bijection: if Z^I^\hat Z_{\hat{\mathcal I}} encode all the information ZZ, then R2(ZI∣Z^I^)≈1R^2(Z_{\mathcal I}\mid \hat Z_{\hat{\mathcal I}})\approx 1 despite not being block-disentangled. You should use R2(ZI∣Z^I^)R2(Z^I^∣ZI)\sqrt{R^2(Z_{\mathcal I}\mid \hat Z_{\hat{\mathcal I}}) R^2(\hat Z_{\hat{\mathcal I}}\mid Z_{\mathcal I})} as in [2]
  5. In the experiments described in Appendix C, what is the unactive block Ic\mathcal I^c to disentangle from I\mathcal I ? It looks like Ic=∅\mathcal I^c=\emptyset meaning that there is nothing to disentangle, in such case, why not just learn an auto-encoder in latent dimension d=6d=6 ?
  6. You assess that you empiracally relax assumption C6 (known nn), however you use in your block-disentanglement experiment m=nm=n with mm known, how is it different to know nn ?
  7. Additionally you explain that you relax assumption C1 (atomic intervention) while you use "one nominally targeted joint per environment", how is it really different thant atomic intervention ?

[1] Linear causal representation learning from unknown multi-node interventions (Varici et al, Neurips 2024)
[2] Benchmarks, Algorithms, and Metrics for Hierarchical Disentanglement (Ross et al, ICML 2021)
[3] Nonparametric Partial Disentanglement via Mechanism Sparsity: Sparse Actions, Interventions and Sparse Temporal Dependencies (Lachapelle et al, JMLR 2026)
[4] CITRIS: Causal Identifiability from Temporal Intervened Sequences (Lippe et al, ICML 2022)

3 comments
AuthorsOpenReview ↗

Rebuttal (part 1 of 2)

Q1. Relationship to partial disentanglement

Our definition is a particular partial identification target: only the intervention-affected group ZIZ_{\mathcal I} must be recovered, and its coordinates may be mixed by an arbitrary diffeomorphism within that group.

Q2. Whether individual variables cannot be recovered

What we mean specifically is that our results apply in settings where the assumptions used for existing elementwise-identification guarantees need not hold. We do not prove that individual variables are impossible to recover separately in every such setting, and we have adjusted the language to make the distinction clear.

Q3. Appendix figures

In our camera-ready version, we will enlarge the axes, legends, and labels in Appendix C to improve readability at normal zoom levels.

Q4. Purpose of the scale loss

Without a scale constraint, the regularizer has a simple degeneracy: If the encoder output is rescaled as h′(x)=ah(x)h'(x)=a h(x) and the decoder is adjusted to preserve reconstruction, its transformed score differences scale as a−1a^{-1}. Increasing aa can therefore reduce the score penalty without improving disentanglement. The scale term in Eq. (10) constrains the mean norm of the encoder output and discourages this trivial rescaling.

Q5. Contribution beyond classical disentanglement

We do not claim the classical disentanglement method of Varıcı et al. (2025) as a new theoretical contribution. We use it as a reference setting to investigate the behavior of theoretically motivated CRL objectives in our experimental framework. Our novel contributions are instead (i) the block-identification result under the relaxed intervention conditions of Definitions 2 and 3, and (ii) the associated score-based learning objectives designed to promote the corresponding block structure. The classical experiments provide a point of comparison, while the embodied experiments investigate the objectives under less idealized data conditions.

Q6. Supervised rather than unsupervised baseline

RoboPEPP was selected as a state of the art domain-specific deep learning method for pose estimation. That it uses supervised learning additionally enabled us to evaluate performance across different labeling budgets. Our intent was to present RoboPEPP as a practically relevant task-specific reference rather than a generic unsupervised disentanglement baseline.

Q7. Missing block-disentanglement baseline

While we agree that a block-disentanglement baseline is missing, the reconstruction-only case (λ=0\lambda=0) serves as a controlled ablation of the score penalty under the same architecture and kernel-ridge evaluation protocol, making the relative comparison internally consistent. At db=32d_b=32, active-block R2R^2 increases from 0.230.23 to 0.610.61 at λ=50\lambda=50. This provides evidence that the score regularizer improves target-information concentration within our pipeline, but does not establish superiority over other CRL methods or robustness to the choice of evaluation regressor. We will make this distinction explicit.

Q8. One-way regression versus bijection

Under regularity conditions, exact R2(ZI∣Z^I^)=1R^2(Z_{\mathcal I}\mid\hat Z_{\hat{\mathcal I}})=1 is sufficient to rule out the learned block from depending on nuisance information. However, the reviewer is correct: empirically, under non-ideal R2R^2 values the one-sided metric is not sufficient for showing bijectivity/block identification, and we will include the reverse direction R2(Z^I^∣ZI)R^2(\hat Z_{\hat{\mathcal I}}\mid Z_{\mathcal I}) as well as the metric in [2] in our final revision.

We note that during our initial experiments, we did compute this metric, but did not report them due to the values being generally high (>0.8>0.8) in all settings, and thus not being discriminative. We included these values in the new experiments we conducted (see Appendix C.4.4) and showed that its values are mostly saturated.

Q9. Inactive true block

We do compare against a reconstruction-only autoencoder baseline (albeit with db>nd_b>n) with our method: this corresponds to setting λ=0\lambda = 0 in our experiments. Table 2 shows that enabling the score-based regularizer (λ>0\lambda > 0) does significantly concentrate the information about the active block to the designated block of our estimator.

Furthermore, in Appendix C.4.4 of the revised paper, we provide experiments where the active block size mm is chosen strictly less than latent dimension n=6n=6, specifically for 3 and 5. We observe that when score regularization is set high, both forward and backward R2R^2 metrics increase substantially and come close to the roofline R2R^2 values for that subset, showing improved concentration of target information even when other variations are also present in the data set.

AuthorsOpenReview ↗

Rebuttal (part 2 of 2)

Q10. Unknown latent dimension

Our learning objective requires a selected block size mm and a bottleneck dimension dbd_b, but does not need us to set db=nd_b=n. The new experiments demonstrate that its use is not restricted to m=nm=n (which implies known nn).

Q11. Atomic interventions

While our main results use single-dominant-joint interventions, we had also conducted experiments under multiple-dominant-joint targets in Appendix C.4.3. Furthermore, even our single dominant joint experiments do not correspond formally to the atomic interventional setting of prior work: (i) there are off-target effects on all other joint angles, and (ii) the observational–interventional contrast is formed by contrasting two halves of videos, rather than collecting independent samples from a priori fixed distributions. Therefore, the embodied setting is significantly different from the usual atomic intervention definitions.

Q1: If your method is a specific case of partial identification, you should explain why more precisely and it would be an important baseline in your experiments

Q5: I don't understand to what extend the study of the classical disentanglement of Varici et al. allows a deeper understanding of your method. What are the conclusions of the classical disentanglement benchmark regarding your method as there is no common metric ? If you do not contribute to this method, the following sentence in Section 5 is highly confusing "we compare the performance of our approach" ?

Q6: I can understand that such comparison could be useful, however you should use at least one of the many classical disentanglement method for a cleaner benchmark. But first you should address the concern raised Question 5 on the purpose of this benchmark relatively to your method

Q7: This is an interesting ablation of λ\lambda, however without a baseline those results cannot be interpreted, we cannot know if your method performs well for λ>0\lambda >0 or sub-performs for λ=0\lambda=0

Q9/Q10: My question is the following: when I=[n]\mathcal I=[n], why use your method rather than an auto-encoder with db=nd_b=n ? Concerning results in C.4.4, reverse R2R^2 seems extremely low for the 3 Dataset. But as said earlier, it is hard to draw conclusion without baseline.

Q11: Your main setting is indeed different from the atomic intervention setting, however it feels pretty close to it. Concerning the results in Appendix C.4.3, the Figure 6 uses cryptic notations that I cannot understand

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

This paper proposes a new causal variable identifiability result under known interventions. The key contribution is about switching the goal of element-wise identification to block-wise identification, therefore some required assumptions can be relaxed. A corresponding identification algorithm is also proposed, whose key objective includes a block sparsity loss. Experiments on assumption-matched setting and embodied setting (with some assumption violated) show that, the proposed identification method can achieve similar MSE loss with a baseline method (RoboPEPP) whose uses 5% labeled data.

Critical strengths
  1. The contribution to establish block-wise identifiability result is a natural extension to element-wise identifiability, and has certain practical value.
  2. The idea to estimate log-density ratio using a classifier in loss implementation is elegant and inspiring.
Critical weaknesses
  1. The connection between theoretical results and experimental results seem to be detached. In the theoretical part, the authors propose 2 new conditions (definition 2/3) to replace conventional conditions (C1-C4). In the experiment part, however, the authors test their algorithm under 2 settings (assumption-matched and embodied), none of which matches their theorem.
  2. The evaluation is problematic and the experimental results are unsatisfying. The three metrics in Eqn (11) are insufficient to evaluate the quality of identification. To test whether ZZ and Z^\hat{Z} are equivalent, two-sided R2R^2 scores are necessary. The one-side R2R^2 from Z^\hat{Z} to ZZ only decides whether Z^\hat{Z} contains all information in ZZ, and does not ensure that Z^\hat{Z} entails no redundant information. Moreover, in Table 2, the R2R^2 values are low (<0.7). According to my experience, an ideal R2R^2 should be at least >0.9 to indicate successful identification. In addition, only one scenario (robot pose) is applied in experiments, more scenarios are required for stronger support.
Questions for authors
  1. In Theorem 1, what is the intuitive understanding for “the estimated matrix D is supported on a row set I”?
  2. In the proof (line 784), how is Eqn. (20) derived by “block matrix inversion” from Eqn. (19)? This step is not straightforward to me.
2 comments

W1. Theory and experiments

We recall the rationale for the presented two settings:

  • Assumption-matched: Whether the existing theory is applicable to challenging high-dimensional data. We emphasize that such demonstration was lacking in prior related work.
  • Embodied: We investigate whether the same general approach remains useful under substantial departures from the theoretical assumptions. Our main objective is therefore to investigate the practical relevance of theoretically motivated CRL objectives under progressively more realistic data availability assumptions.

W2. Evaluation metrics and performance

Under regularity conditions, exact R2(ZI∣Z^I^)=1R^2(Z_I \mid \hat Z_{\hat I})=1 is sufficient to rule out the learned block from depending on nuisance information. Specifically, note that Z^=h(X)=h(f(Z))=ϕ(Z)\hat{Z} = h(X) = h(f(Z)) = \phi(Z), therefore Z^I^=H(ZI,ZIc)\hat{Z}_{\hat I} = H(Z_I, Z_{I^c}) for H≜ϕI^H \triangleq \phi_{\hat I}. Exact R2=1R^2=1 implies that there exists a deterministic function FF that maps F(Z^I^)=ZI=F(H(ZI,ZIc))F(\hat Z_{\hat I}) = Z_I = F(H(Z_I, Z_{I^c})). Taking gradient with respect to ZIZ_I and ZIcZ_{I^c} yields JF⋅∂H∂ZI=I,JF⋅∂H∂ZIc=0,J_F\cdot\frac{\partial H}{\partial Z_I} = I, \qquad J_F\cdot\frac{\partial H}{\partial Z_{I^c}} = 0, respectively. When ∣I^∣=∣I∣|\hat I|=|I|, first equation necessitates JFJ_F to be invertible, which implies via the second equation that ∂H∂ZIc=∂Z^I^∂ZIc=0\frac{\partial H}{\partial Z_{I^c}} = \frac{\partial \hat{Z}_{\hat I}}{\partial Z_{I^c}} = 0, i.e., Z^I^\hat{Z}_{\hat I} is only a function of ZIZ_I, such as Z^I^=ϕI^−1(ZI)\hat{Z}_{\hat I} = \phi_{\hat I}{}^{-1}(Z_I) (if all functions are twice differentiable).

However, the reviewer is correct: empirically, under non-ideal R2R^2 values the one-sided metric is not sufficient for showing bijectivity/block identification, and we will include the reverse direction R2(Z^I^∣ZI)R^2(\hat Z_{\hat{\mathcal I}}\mid Z_{\mathcal I}) in our final revision.

We note that during our initial experiments, we did compute this metric, but did not report them due to the values being generally high (>0.8>0.8) in all settings, and thus not being discriminative. We included these values in the new experiments we conducted (see Appendix C.4.4) and showed that its values are mostly saturated.

Q1. Intuition for the row-support assumption

Under Definition 2, the true latent score-difference matrix D(z)D(z) has zero rows outside the active block I\mathcal I (cf. Eq. (5)). The intuition is to impose the same structural sparsity directly on the learned representation: there should exist I^\hat{\mathcal I} with ∣I^∣=∣I∣|\hat{\mathcal I}|=|\mathcal I| such that [DZ^(z^)]I^c,:=0[D_{\hat Z}(\hat z)]_{\hat{\mathcal I}^{c},:}=0. In other words, the environment-dependent score differences should be confined to the designated learned block.

Q2. Block matrix inversion in the proof

Eq. (19) states that Jτ−1J_\tau^{-1} is block lower-triangular, i.e., Jτ−1=(A0\CD).J_\tau^{-1}=\begin{pmatrix}A&0\\\C&D\end{pmatrix}. Since the full matrix is invertible, both square diagonal blocks AA and DD are invertible. Using the block matrix inversion formula yields

=\begin{pmatrix}A^{-1}&0\\\\-D^{-1}CA^{-1}&D^{-1}\end{pmatrix},$$ which can be verified by direct multiplication. Thus the upper-right block of $J_\tau$ is also zero, giving Eq. (20).

Thanks for the rebuttal. My concerns over the two weaknesses have not been addressed well.

For W1, the "assumption-matched" setting is said to have "This setting satisfies (C1)–(C6)", while Theorem 1 is established under "Definitions 2 and 3, together with assumptions (C5)–(C6)". Therefore, there is no setting that exactly verifies Theorem 1.

For W2, though reverse R2R^2 scores are partially reported in the revised paper, they are low in general (<0.8). In my point of view, such low values do not indicate successful identification. I suggest the authors to find other evidences that illustrate the identification quality.

Reviewer_Hm7POpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to translate theoretical causal representation learning (CRL) into practical applications. Existing CRL theories provide elegant identifiability guarantees, but often require strong assumptions or restrictive data access that limit their practical use (the paper catalogs these as (C1)–(C6) in Sec. 2). This work proposes a block-identification theorem (Theorem 1, Sec. 3; proof in App. B) that relaxes several intervention requirements. The theorem claims that the learned representation disentangles the intervened variables as a group from the remaining variables , under assumptions that remaining variables are not ancestors of intervened variables (ancestral closure, Def. 2) and additional identifiability conditions hold (collective rank diversity of the score differences, Def. 3, together with (C5)–(C6)).

On the practical side, the paper develops a theorem-motivated training objective (the block-sparsity loss, Eq. 10) and evaluates it on simulated robotic-arm observations rendered with Panda-Gym. In the experiments, when the standard CRL assumptions hold, an existing score-based objective (Eq. 9, from Varıcı et al., 2025) recovers the individual joint angles from images. When learning from videos with temporally correlated frames and imperfect, multi-joint interventions, the proposed block objective makes the designated six-coordinate block of the representation more predictive of the joint angles than an unregularized autoencoder.

Critical strengths

S1. Bridging CRL identifiability theory and practical visual state estimation is a worthwhile goal, and the explicit catalog of assumptions (C1 to C6) and violations (V1 to V3) is clear and useful for the community.

S2. Theorem 1 has a clean, short proof.

S3. The simulated robot-arm dataset (Panda-Gym renders with controllable, environment-specific changes to the joint angles, in both an i.i.d. and a video version) is itself a useful contribution. With modifications that add more randomness, e.g., varying camera pose, lighting, background, or geometry as nuisance factors, and intervening on only a subset of joints, it could become a very good benchmark for CRL under realistic assumption violations.

Critical weaknesses

Assumption limitations

The paper's goal is to bridge theory and application, and one important move is to weaken the assumptions. However, some of the remaining assumptions are still restrictive or lack justification.

W1. The ancestral-closure assumption (Def. 2) remains restrictive. Although the theorem relaxes several intervention requirements, it assumes that every ancestor of an intervened variable is also directly intervened on in at least one available environment. This may be difficult to satisfy when intervention access is limited. Moreover, because the causal graph and intervention targets are assumed unknown, it is unclear how this condition could be verified in practice. The paper should better justify when ancestral closure is plausible and clarify the tradeoff between relaxing intervention-specific requirements and introducing this structural assumption.

W2. The method assumes the active-block size mm is known and reserves mm learned coordinates for it (Sec. 4.2). In principle mm is identifiable as the rank of the observed score-difference matrix under Def. 3, but the paper does not discuss how mm would be estimated in practice from noisy score estimates. The experiments fix m=6m = 6 while varying dbd_b, so robustness to misspecifying mm is not evaluated.

Soundness of theorem

W3. Theorem 1 fails as stated because of the pseudoinverse transformation in Eq. (8). Perfect reconstruction on the data manifold does not imply that Dz^(z^)=[Jh(x)†]⊤DX(x)D_{\hat z}(\hat z) = [J_h(x)^{\dagger}]^{\top} D_X(x) equals the score difference of the learned representation Z^=h(X)\hat Z = h(X). Consider n=2n = 2 with Z1∼N(0,1)Z_1 \sim \mathcal{N}(0,1) and Z2∣Z1∼N(Z1,1)Z_2 \mid Z_1 \sim \mathcal{N}(Z_1,1), and a single environment (K=1K = 1) that replaces the mechanism of Z1Z_1 by q1(1)=N(μ,1)q_1^{(1)} = \mathcal{N}(\mu,1), where μ≠0\mu \neq 0. By Eq. (4), d(1)(z)=(μ,0)⊤d^{(1)}(z) = (\mu, 0)^{\top}, so the active block I=I1={1}I = I_1 = \lbrace 1 \rbrace is ancestrally closed (Def. 2) and D(z)=d(1)(z)D(z) = d^{(1)}(z) has rank ∣I∣=1|I| = 1 (Def. 3). Samples are i.i.d. and nn is known, so (C5)–(C6) hold. Define f(z1,z2)=(z1,z2,0),h(x)=(x1+x2+x3, x2−x3),g(z^1,z^2)=(z^1−z^2, z^2, 0).f(z_1, z_2) = (z_1, z_2, 0), \quad h(x) = (x_1 + x_2 + x_3, \ x_2 - x_3), \quad g(\hat z_1, \hat z_2) = (\hat z_1 - \hat z_2, \ \hat z_2, \ 0). Then ff is injective, and g(h(X))=Xg(h(X)) = X holds exactly on the data. However, Z^=τ(Z)=(Z1+Z2, Z2)\hat Z = \tau(Z) = (Z_1 + Z_2, \ Z_2) with τ=h∘f\tau = h \circ f, so neither learned coordinate is a function of Z1Z_1 alone.

By Eq. (6), dX(1)(x)=[Jf(z)†]⊤d(1)(z)=(μ,0,0)⊤d^{(1)}_X(x) = [J_f(z)^{\dagger}]^{\top} d^{(1)}(z) = (\mu, 0, 0)^{\top}. The encoder Jacobian JhJ_h has rows (1,1,1)(1, 1, 1) and (0,1,−1)(0, 1, -1), so [Jh†]⊤[J_h^{\dagger}]^{\top} has rows (1/3,1/3,1/3)(1/3, 1/3, 1/3) and (0,1/2,−1/2)(0, 1/2, -1/2). Eq. (8) therefore gives Dz^(z^)=[Jh(x)†]⊤dX(1)(x)=(μ/3, 0)⊤.D_{\hat z}(\hat z) = [J_h(x)^{\dagger}]^{\top} d^{(1)}_X(x) = (\mu/3, \ 0)^{\top}. This is supported on the row set I^={1}\hat I = \lbrace 1 \rbrace with ∣I^∣=∣I∣|\hat I| = |I|, so all hypotheses of Theorem 1 hold, yet Z^1=Z1+Z2\hat Z_1 = Z_1 + Z_2 does not block-disentangle ZI=Z1Z_I = Z_1.

Experiments

W4. The embodied experiments do not directly test disentanglement from varying nuisance factors. All six varying joint angles belong to the active block (App. C: I=[n]I = [n]), while camera pose, lighting, background, and robot geometry remain fixed. Thus, the experiments test whether joint information concentrates in six designated coordinates of an overcomplete representation, rather than whether an intervened subset is disentangled from other varying true variables. Experiments with a proper active subset and varying nuisance factors would more directly evaluate the theorem's central claim.

W5. The invertibility assumption is not explicitly enforced by the model. Theorem 1 requires perfect reconstruction, g(h(X))=Xg(h(X)) = X, which makes the encoder invertible on the data manifold; without it, the identification result does not apply. The paper uses the reconstruction loss to encourage invertibility softly, but there is no analysis of whether a small but nonzero reconstruction loss still makes identification more likely to be achieved.

Presentation

W6. Discussion of some important block-identification work is missing. The paper reviews several block-identification analyses, but other theoretical works with more general results should be discussed. For example, [1] studies general disentanglement maps, which go beyond both point (elementwise) identification and block identification. More importantly, [1] targets disentangling variables in time-series settings, which connects directly to the robotics setting and the non-i.i.d. samples in the experiments. [2] studies multiple interventions within a single environment.

[1] Lachapelle, Sébastien, et al. "Nonparametric partial disentanglement via mechanism sparsity: Sparse actions, interventions and sparse temporal dependencies." arXiv preprint arXiv:2401.04890 (2024).

[2] Li, Adam, Yushu Pan, and Elias Bareinboim. "Disentangled representation learning in non-Markovian causal systems." Advances in Neural Information Processing Systems 37 (2024): 104843–104903.

Questions for authors

Q1. Can you show some experiments result when I⊊[n]I \subsetneq [n]?

I am willing to raise my evaluation if the weaknesses and questions above are adequately addressed.

2 comments

Thank you for the thorough and fair evaluation. Below, we address the weaknesses and your questions.

W1. Ancestral closure

First, we note that our empirical evaluations are intended to investigate whether theoretical requirements can be relaxed in practice. A related ancestral-closure assumptions exist in related block identifiability literature as well (e.g., Ahuja et al. 2024), where the target is usually recovering the invariant block as opposed to our recovery of the “active block”. Nevertheless, it is correct that this can be difficult to verify in real-world settings. We note that our theoretical result highlights that intervention-specific (i.e. data availability) assumptions can be relaxed analytically, albeit replacement assumptions may remain technically restrictive.

W2. Known active-block size

Under Definition 3, active block size mm is characterized by the rank of the score-difference matrix, but estimating this from noisy scores remains a practical challenge. A useful heuristic is motivated by classical CRL results where single-node interventions on a latent variable can identify it (e.g. Varıcı et al. 2025). Accordingly, given KK interventional environments and no other prior information, one might choose m=Km=K as a heuristic starting point. However, our experiments do not yet analyze the effects of a potential misspecification as a result of this heuristic.

W3. Theorem soundness issue

Proof of Theorem 1 (i.e. the identifiability result) is correct but Eq. (8) was inconsistent with the proof. We thank the reviewer for catching the inconsistency. Eq. (8) itself is incorrect for a generic encoder hh because the encoder may have arbitrary derivatives normal to the data manifold, so its unrestricted Jacobian does not align with the data manifold’s tangent directions. The corrected form is DZ^(z^)=Jg(z^)⊤DX(x)=Jτ(z)−⊤DZ(z),τ=h∘f.D_{\hat Z}(\hat z)=J_g(\hat z)^\top D_X(x)=J_\tau(z)^{-\top}D_Z(z),\qquad \tau=h\circ f. where we write the relationship in terms of the Jacobian of the decoder gg. Importantly, the proof in Appendix B already uses the correct equation Jτ(z)−⊤DZ(z)J_\tau(z)^{-\top} D_Z(z), rather than the erroneous pseudoinverse identity. Thus, the correction resolves the inconsistency between Eq. (8) and the proof without requiring a change to the identification argument. The implemented objectives in Eqs. (9)–(10) also use Jg⊤J_g^\top and are unaffected. We have corrected Eq. (8) and the corresponding presentation of Theorem 1.

W4 and Q1. Experiments with varying nuisance factors

The reviewer is correct. Thank you for the suggestion. In Appendix C.4.4 of the revised paper, we provide experiments where the active block size mm is chosen strictly less than latent dimension n=6n=6, specifically 3 and 5. We observe that when score regularization is set high, both forward and backward R2R^2 metrics increase substantially and come close to the roofline R2R^2 values for that subset, showing improved concentration of target information even when other variations are also present in the data set.

W5. Approximate reconstruction and invertibility

Except the controlled toy settings where we have an explicit invertible transform, using a reconstruction loss as a surrogate for ensuring invertibility is rather standard in related literature. We observed that if the reconstruction performance is bad, the disentanglement performance suffers disproportionately — i.e., good reconstruction is a prerequisite for good disentanglement empirically. Theoretically, robustness/stability guarantees for nonlinear CRL is an important but very challenging direction itself with minimal existing literature (e.g., Buchholz & Schölkopf, 2024). Therefore, this type of analysis remains an important direction for future work.

W6. Related work

The temporal CRL literature, which explicitly leverages the underlying dynamics, including [1] and earlier works such as CITRIS/iCITRIS, is closely related to the video domain. We emphasize that our paper investigates a different problem — whether methods from interventional CRL, which instead leverage dataset-level marginal distributional properties, remain applicable to video and temporally dependent data.

On the other hand, [2] is highly relevant in terms of its assumptions, particularly multi-node interventions under nonlinear transformations. However, its objective differs substantially from ours: [2] focuses on identifiability alone, whereas our work additionally investigates how to learn an encoder that promotes the proposed block-identification conditions through a score-based objective. We will expand the literature discussion to more clearly distinguish the assumptions, methods, and results of these approaches.

Thanks for the author's response. Overall, my concerns are partially addressed.

W1, W2. My concern is not resolved. If the empirical study of paper is intended to "investigate whether theoretical requirements can be relaxed in practice." I think these two points should be really tested empirically since the author also claims they are practical challenge. For example, it is suggested to add experiments about the mismatch of the graph by setting the ground truth graph in synthetic data generation and test how violating the graph assumptions will influence the disentanglement results.

W4 and Q1. My concern is partially resolved. I appreciate the new experiments in Appendix C.4.4 with an active block that is a proper subset of the joints. However, the 0.55 R^2 and 0.35 R^2 is not high in absolute value. Thus, to claim disentanglement, some baseline is needed to compare.

Other points are addressed.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank the reviewers for their valuable comments and constructive suggestions. We believe the revised manuscript and responses below address the main technical and soundness concerns raised during review. Critically, we have:

  1. Corrected Eq. (8) in Theorem 1, resolving the inconsistency in the score-difference transformation. The proof in Appendix B already uses the correct identity, and the theoretical identification argument and the stated objective functions remain unchanged.
  2. Added experiments with non-trivial nuisance blocks in Appendix C.4.4, considering active-block sizes m=3m=3 and m=5m=5 for n=6n=6. These experiments evaluate the block regularizer when inactive joint angles also vary, and report both forward and reverse R2R^2 to better characterize the learned representations.

In our responses below, we also clarify questions raised by the reviewers regarding the assumptions, methodology, and evaluation, including the role of ancestral closure, the known active-block size, reconstruction-based invertibility, the choice of baselines, and the relationship to existing CRL literature. We believe these revisions strengthen the paper's theoretical presentation and empirical support while further clarifying the scope and applicability of our results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.