acceptodds
Under review as a conference paper at ICLR 2027

DCD-PFN: A Decoupling-Aware Foundation Model for Causal Discovery

14 views

Abstract

Causal discovery from observational data is critical for understanding complex data-generating mechanisms, yet traditional methods often struggle under strong nonlinearity or noise, and may suffer from computational bottlenecks. Recent tabular foundation models based on Prior-Data Fitted Networks (PFNs) have demonstrated remarkable zero-shot inference capabilities, but their potential for explicit structural causal discovery remains underexplored. To bridge this gap, we propose DCD-PFN, a decoupling-aware foundation model for causal discovery. Rather than directly learning a mapping from data to global graphs, DCD-PFN amortizes decoupling-based local causal discovery. Through pre-training on diverse synthetic Structural Causal Models (SCMs), the model learns to infer sample-wise decoupling weights from observational data, which guide zero-shot Markov boundary (MB) identification in a single forward pass. DCD-PFN recovers global causal structures via a principled reconstruction procedure with parallelized local discovery. Experiments on both synthetic and real-world datasets demonstrate robust zero-shot generalization.

Primary area:
causal reasoning
Keywords:
Causal Discovery, Foundation Models, Prior-Data Fitted Networks, Markov Boundary

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_vB4UOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes DCD-PFN, an approach for causal structure learning from observational data. Building on the decoupled causal discovery (DCD) framework and its Markov boundary (MB)-based reconstruction theory, the authors explore whether local MB discovery can be amortized through pre-training. Specifically, DCD-PFN learns sample-wise decoupling weights and predicts MBs from observational data, which are then used in a parallelized local-to-global procedure to reconstruct a CPDAG. The method is validated on synthetic and real-world datasets.

Critical strengths

The paper proposes a PFN-based Markov boundary (MB) learning method and further integrates it with the existing DCD framework for causal structure learning.

Critical weaknesses
  • The practical scope of the method is limited. Although the paper is motivated by robust causal discovery, the reconstruction framework assumes the absence of unobserved confounders. This strong assumption limits its applicability. Moreover, although computational efficiency is an important motivation, no runtime or scalability results are reported.

  • The main new component is not directly compared with existing MB discovery methods. The major contribution of the paper is the PFN-based MB learner, yet the experiments do not compare it against existing Markov boundary discovery algorithms.

  • The proposed learned MB predictor itself lacks theoretical guarantees. The paper does not establish soundness, completeness, consistency, or error bounds for the PFN-based MB identification procedure.

  • The reconstruction procedure lacks a sufficiently explicit theoretical guarantee in this paper. The paper does not clearly state the soundness and completeness guarantees for Algorithm 1. In addition, the correctness of adjacency recovery and v-structure identification is not explicitly justified.

  • The empirical evaluation is limited and does not sufficiently validate the claimed robustness and efficiency. The experiments are restricted to relatively small and sparse graphs, and no results are provided on larger or denser structures, or in settings with unobserved confounding.

Questions for authors

Please see the weaknesses above for my main concerns.

Some closely related literature on local-to-global causal structure learning is not discussed, e.g., Local-to-Global Bayesian Network Structure Learning (ICML 2017) and A Recursive Method for Structural Learning of Directed Acyclic Graphs (JMLR 2008).

3 comments
AuthorsOpenReview ↗

Response to Reviewer vB4U

We sincerely appreciate your time and effort in reviewing our paper. Below, we provide detailed responses to your comments. If anything remains unclear or if you have further questions, please do not hesitate to let us know.

W1. Causal Sufficiency Assumption

As a new causal discovery method, establishing our methodology under the causal sufficiency assumption serves as a reasonable and standard starting point. Extending this framework to settings with latent variables is a natural next step, which we are actively investigating.

W2. Compare with other MB discovery methods

W1 / W5. Efficiency

Since the ultimate target of our method is the global causal structure, our experiments primarily compare with representative methods across major causal discovery paradigms. Following your suggestion, we have supplemented the evaluation (revised manuscript, Appendix A.7) with variants replacing our MB discovery module with the original DCD [1] and two alternative MB discovery algorithms (GS [2] and IA [3], equipped with KCIT) on the real-world Sachs Flow-Cytometry dataset n=1755n = 1755:

MethodSHD$_c$ ↓Precision ↑Recall ↑F1 ↑time (s)
DCD-PFN141.0000.3000.4621.57
DCD160.6250.2500.357502.08
DCD with GS (KCIT)240.1250.1000.11112179.05
DCD with IA (KCIT)230.1330.1000.11418137.66

(DCD-PFN was run on a single RTX 5090 GPU, while the other methods were executed on CPUs.)

Compared with other MB discovery methods, our method exhibits clear performance advantages. Regarding efficiency, our method achieves a total runtime of 1.57 seconds, representing an improvement of 2–3 orders of magnitude over the original DCD. The speedup is even more pronounced when compared against MB discovery methods employing non-parametric KCIT.

W3. Theory of the Model

MB Identifiability via decoupling: This theoretical foundation has already been established in [1], which directly inspired our work (as noted in lines 105–107).

Direct theoretical guarantees of pre-trained PFN models: We agree that this is a critical and currently under-explored frontier question. While establishing general guarantees for pre-trained PFNs is a substantial undertaking on its own, we consider it an important direction for future theoretical work.

W4. Theory of Algorithm 1

As a parallelized variant, Algorithm 1 preserves the theoretical guarantees (i.e., soundness and completeness) established in [1]. We have briefly noted this in lines 256–258 (right after the assumptions) and will highlight this property more clearly in the revised manuscript.

W5. Scope of Experiments

Our empirical evaluation already includes synthetic data, semi-synthetic data based on real graphs, and the widely used real-world Sachs dataset, which we believe aligns with the established practices of causal discovery literature. If you have any further suggestions regarding specific datasets to test, please feel free to let us know.

Q. Local-to-global Literature

Our related work initially focused primarily on the developing areas of PFNs and amortized causal discovery over recent years. Thank you again for pointing this out, we have incorporated these relevant local-to-global works into Section 2.2 of the revised manuscript.


[1] Zhengkang Guan, Fei Wu, Kun Kuang. Decoupled Causal Discovery. ArXiv, 2026

[2] Dimitris Margaritis, Sebastian Thrun. Bayesian Network Induction via Local Neighborhoods. NIPS, 1999.

[3] Ioannis Tsamardinos, Constantin F. Aliferis, and Alexander Statnikov. Algorithms for Large Scale Markov Blanket Discovery. FLAIRS, 2003.

Thank you for the response. However, most of my major concerns remain unresolved in the current version, and addressing them is essential to the completeness of the work. Therefore, I will maintain my original rating.

Thank you for your reply. We are sorry that our previous response did not fully address your concerns. We would greatly appreciate it if you could point out which specific part remains unresolved, so that we can provide further clarification. Thank you again for your time and feedback.

Reviewer_1E9xOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DCD-PFN proposes an amortized approach to causal discovery that predicts local Markov boundaries from observational data. A transformer trained on synthetic structural causal models jointly learns sample weights and boundary-membership scores, using a pairwise independence objective alongside supervised prediction. These predictions are incorporated into an existing local-to-global reconstruction procedure to estimate a CPDAG. Experiments report improvements over several baselines on small synthetic graphs, benchmark topologies with simulated observations, and one configuration of the Sachs dataset.

Critical strengths
  • Learning a reusable Markov-boundary predictor is a reasonable approach to amortizing local causal discovery and could be useful across datasets.

  • The reported improvements are substantial in several settings, particularly relative to the pretrained AVICI baseline in Table 1 and on Sachs with 1,755 samples.

  • Evaluation consistently uses CPDAGs, and the paper explicitly acknowledges the limitations of approximate boundary predictions, pairwise decoupling, and combinatorial reconstruction.

Critical weaknesses
  • Table 6 shows removing the decoupling loss produces empty graphs on every evaluated dataset, but it only demonstrates that the loss matters under the reported training setup, but leaves open whether it enables meaningful decoupling, stabilizes optimization, or changes score calibration around the fixed 0.5 threshold. Uniform-weight controls, local boundary metrics, and diagnostics across training seeds would make this result substantially more informative.

  • The theoretical motivation remains disconnected from the learned pipeline. Section 3.4 correctly conditions recovery on exact boundary queries, while Appendix B acknowledges that pairwise independence and the embedding readout are only surrogates. However, no local prediction results establish how reliably the model answers the restricted-variable queries used during reconstruction. Those queries are especially important because removing variables changes the distribution seen by a model pretrained on complete synthetic SCMs.

  • Algorithm 1 still enumerates subsets of each initial boundary, giving exponential growth in the number of queries as boundary size increases. Here the efficiency motivation is not supported by measurements. A single forward pass per query does not establish fast graph reconstruction. End-to-end runtime, query counts, memory usage, and scaling with boundary size are missing.

  • The exposition is generally readable, but essential reproducibility details are absent. Section 3.3 and Appendix A.1 do not specify the training distribution over dataset sizes and graph densities, the number of generated training datasets or optimization steps, or the full mechanism/noise sampling distributions. These omissions also make the claimed separation between pretraining and evaluation distributions difficult to assess.

Questions for authors
  • How does DCD-PFN compare with the original DCD method in accuracy and total inference time, and with a direct boundary predictor trained on the same prior?

  • Why does removing the decoupling loss consistently produce empty graphs? Does this persist across training seeds, and what do the local prediction scores and precision–recall curves show?

  • How accurate are boundary predictions on the restricted variable sets actually queried by Algorithm 1, and how do these errors affect skeleton recovery and orientation?

  • Since shared feature attention precedes extraction of the non-target representations, can target information enter the weighting branch? How does this relate to the non-target-only weighting functions in the motivating decoupling theory?

Given the unresoved issues, my initial score is 2.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 1E9x (Part 1/2)

We sincerely thank the reviewer for their insightful questions regarding the methodological details of our model. Below, we provide detailed responses to your comments. If anything remains unclear or if you have further questions, please do not hesitate to let us know.

Q1 / W3. Performance and Efficiency Comparison with Original DCD and Other MB Predictors

Since the original DCD is a concurrent work whose official code had not yet been released, we did not include a direct comparison in our initial submission. Following your suggestion, we have now compared DCD-PFN against our own reproduction of the original DCD (note that implementation differences may exist, so these results serve as a preliminary reference). Furthermore, to the best of our knowledge, there are currently no other pretrained direct boundary predictors. Therefore, to thoroughly address your concern, we also incorporated two MB discovery algorithms (GS [1] and IA [2], equipped with KCIT) as alternative predictors. We conducted these experiments on the real-world Sachs Flow-Cytometry dataset n=1755n=1755 (added in the revised manuscript Appendix A.7):

MethodSHD$_c$ ↓Precision ↑Recall ↑F1 ↑time (s)
DCD-PFN141.0000.3000.4621.57
DCD160.6250.2500.357502.08
DCD with GS (KCIT)240.1250.1000.11112179.05
DCD with IA (KCIT)230.1330.1000.11418137.66

(DCD-PFN was run on a single RTX 5090 GPU, while the other methods were executed on CPUs.)

In terms of efficiency, our method achieves a total runtime of 1.57 seconds, representing a speedup of 2–3 orders of magnitude over the original DCD, and an even more pronounced improvement compared to MB discovery methods employing non-parametric KCIT. In terms of performance, DCD-PFN exhibits clear advantages across all accuracy metrics compared with other MB discovery methods.

Q2 / W1. Effectiveness of the Decoupling Mechanism

Without the decoupling loss, under our current model architecture, the model fails to learn across all random seeds and training hyperparameters we tested. After the same number of training steps, it collapses to outputting uninformative near-zero prediction scores, yielding precision–recall curves equivalent to random guessing. We acknowledge that the explanation in our initial submission (Appendix A.5) was somewhat abrupt and imprecise, and we have refined and clarified this discussion in the revised manuscript.

More importantly, following your insightful suggestion, we added a more direct inference-time ablation experiment in Appendix A.5. Specifically, on the real-world Sachs Flow-Cytometry dataset n=1755n=1755, we fixed a trained checkpoint and evaluated it at inference time using uniform-weight controls as well as randomly shuffling all learned weights across samples (shuffled results are reported as mean±std over 25 runs):

WeightsSHD$_c$ ↓Precision ↑Recall ↑F1 ↑
Learned141.0000.3000.462
Uniform160.8330.2500.385
Shuffled learned16.640 ± 0.8100.667 ± 0.1780.218 ± 0.0410.328 ± 0.068

As shown, losing the learned decoupling weights leads to a noticeable degradation in overall performance, providing more direct evidence for the effectiveness of the decoupled weights.

AuthorsOpenReview ↗

Response to Reviewer 1E9x (Part 2/2)

Q3 / W2. Restricted-variable Queries and Effect on the Final CPDAG

W2 / W4. Separation between pre-training and Evaluation

Indeed, queries on a restricted variable set differ from full-variable queries. However, we would like to clarify that all evaluation distributions are strictly unseen by the model during pre-training. Specifically, our pre-training data is synthesized using mechanisms including random neural networks and decision trees over ER graphs, followed by post-processing including warping and discretization. In contrast, beyond the real-world Sachs dataset, our synthetic evaluation benchmarks are strictly generated from explicit SEMs over BA graphs.

We further conducted an analysis comparing initial MB queries over all variables versus restricted-variable queries on the real-world Sachs Flow-Cytometry dataset n=1755n=1755:

Query typeMB precision ↑MB recall ↑MB F1 ↑
Initial queries over all variables0.8640.2960.427
Restricted-variable queries0.6670.1370.222
Total0.7940.2400.355

Although the overall performance of restricted-variable queries drops, this is primarily because errors from the initial MB queries propagate forward. In particular, variables missed in the initial MB can no longer be recovered in subsequent restricted-variable queries.

Regarding the effect of MB precision/recall on the final CPDAG, the edge precision/recall of the final recovered graph are generally positively associated with those of the predicted MBs (see revised manuscript, Appendix A.6). Specifically, missing a true MB variable directly leads to missing its edge with the target variable (lowering recall), whereas falsely included MB variables will eventually be determined as either adjacent to the target variable or part of a v-structure, introducing spurious edges (lowering precision). Meanwhile, errors in restricted-variable queries affect v-structure identification, thereby impacting edge orientation accuracy.

Q4. Shared Branch

This is a very insightful question. In the initial shared branch of our model, we do not enforce strict attention masking between non-target and target variables. While the scenario you described is architecturally possible, we adopted this design for two main reasons:

  • Cross-feature information interaction in the early layers of PFNs is typically weak [3]. We view the shallow layers (the shared branch) primarily as projecting inputs into a latent space for subsequent layers. The subsequent step of extracting and decoupling the non-target representations is directly motivated by theory.
  • Following existing multi-task PFN designs [4], passing inputs through a shared trunk before branching into separate heads is a validated approach for end-to-end training of multi-output PFNs.

W4. Reproducibility

Thank you for raising this important point. In the main text, we have provided the complete methodological pipeline, and the appendix details the model depth and key training hyperparameters. As for the finer-grained implementation settings you mentioned, many of them are not rigidly fixed. We are currently cleaning up and organizing our codebase, and we will open-source the complete repository upon acceptance to ensure full engineering reproducibility.


[1] Dimitris Margaritis, Sebastian Thrun. Bayesian Network Induction via Local Neighborhoods. NIPS, 1999.

[2] Ioannis Tsamardinos, Constantin F. Aliferis, and Alexander Statnikov. Algorithms for Large Scale Markov Blanket Discovery. FLAIRS, 2003.

[3] Omar Swelam, Lennart Purucker, Jake Robertson, Hanne Raum, Joschka Boedecker, Frank Hutter. Does TabPFN Understand Causal Structures? ArXiv, 2025.

[4] Max Zhu, Martino Mansoldo, Ching-Hao Wang, Stefan Groha. A Causal Foundation Model for Structure and Outcome Prediction. ArXiv, 2026.

Thank you for the detailed responses and additional experiments. These additions have addressed most of my concerns.

I still have following questions:

Regarding Q2

The new ablation provides useful evidence that the learned weights improve performance. However, it remains unclear whether these weights actually achieve the intended decoupling, rather than simply providing better representations for MB prediction. Could the authors add some discussion about how the dependence between non-target variables before and after reweighting?

Q3

Thanks for the additional results. But the substantial decrease in MB F1 from 0.427 to 0.222 remains concerning. Could the authors clarify this? Does this degradation is primarily due to error propagation from the initial MB predictions or the predictor's difficulty with restricted-variable inputs?

Besides, I also encourage the authors to read and discuss several recent relevant works in the paper, on Markov boundary discovery and local-to-global structural reconstruction:

[1] Causal Discovery in the Wild: A Voting-Theoretic Ensemble Approach. ICLR 2026

[2] Global Directional Priors with Local Statistical Validation for Scalable Causal Discovery. ICML 2026

[3] Hierarchical Causal Structure Discovery from Layered Topological Orderings. International Journal of Data Science and Analytics, 2026

[4] FoundCause: Causal Discovery with Latent Confounders from Observational Data. arXiv 2026

[5] Sample, Estimate, Aggregate: A Recipe for Causal Discovery Foundation Models. TMLR 2025

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

This paper proposes DCD-PFN, a decoupling-aware foundation model for zero-shot causal discovery from observational data. Rather than directly predicting a global causal graph, the method amortizes local Markov blanket (MB) identification. It pre-trains on synthetic structural causal models (SCMs), learns sample-wise decoupling weights, and uses these weights to support MB prediction. The predicted local structures are then integrated through a local-to-global reconstruction procedure to obtain a global causal structure.

The method is evaluated on synthetic datasets with different numbers of variables, semi-synthetic datasets including SynTReN and Child, and the real-world Sachs flow-cytometry dataset. The paper compares DCD-PFN with several conventional and amortized causal discovery methods and includes an ablation study of the decoupling loss. The results provide evidence that the method can transfer from synthetic pre-training to several unseen evaluation settings without retraining on each test dataset, although its robustness across different settings remains to be further established.

Critical strengths
  1. A meaningful and technically distinctive approach to zero-shot causal discovery. Rather than treating causal discovery as direct global graph prediction, DCD-PFN combines amortized MB identification with decoupling-aware learning and local-to-global reconstruction. This design has a clear methodological motivation and offers a potentially useful alternative to iterative conditional-independence testing and direct global graph prediction. The contribution is not merely the application of a pretrained tabular model to a new task; it lies in how the model is incorporated into a structured causal discovery pipeline.

  2. Broad empirical evaluation across multiple data settings. The evaluation spans synthetic, semi-synthetic, and real-world data, includes comparisons with multiple causal discovery baselines, and investigates the role of the decoupling loss. The real-world Sachs experiment is particularly useful for assessing whether the method transfers beyond purely synthetic test distributions. The ablation in which removing the decoupling loss leads to degenerate empty-graph predictions also provides evidence that the loss is important under the proposed training setup. Although the evaluation does not establish universal robustness, its breadth provides a meaningful empirical basis for considering the method.

Critical weaknesses
  1. The theoretical guarantees do not directly extend to the learned MB predictor. The reconstruction procedure assumes that Markov blankets (MBs) are known exactly, whereas DCD-PFN predicts MB scores and uses a threshold of 0.5 to select candidate variables. Errors in these predictions may propagate through the reconstruction process and affect the final CPDAG. Although the paper acknowledges this limitation, it does not quantify the impact of MB prediction errors on end-to-end structural recovery.

  2. The decoupling mechanism and structural recovery need further validation. The method uses pairwise weighted HSIC to encourage decoupling, but it remains unclear whether the learned weights achieve the intended properties. The ablation shows that removing the decoupling loss leads to empty-graph predictions, highlighting its importance in the current training setup, but does not directly validate the learned decoupling behavior. Moreover, the relatively low recall on SynTReN and Sachs suggests that the method misses a substantial number of true adjacencies in some settings. Further analysis of the learned weights and residual dependencies, together with an investigation of the causes of missed adjacencies, would help clarify the method's limitations.

Questions for authors
  1. Can the authors provide an analysis of how the precision and recall of predicted MBs affect the final CPDAG reconstruction?

  2. Beyond the ablation of the decoupling loss, can the authors provide direct evidence that the learned sample-wise weights improve the intended decoupling behavior?

  3. How sensitive are the final results to the MB prediction threshold? In particular, can you analyze how threshold changes affect precision and recall, and whether the low recall mainly arises from MB prediction errors or subsequent reconstruction steps?

3 comments
AuthorsOpenReview ↗

Response to Reviewer ZijV

We sincerely thank the reviewer for the constructive and insightful comments regarding the technical details of our method. Below, we provide detailed responses to your comments. If anything remains unclear or if you have further questions, please do not hesitate to let us know.

Q1 / Q3 / W1. Effect of MB Precision/Recall on the Final CPDAG

Q3 / W2. Threshold and Recall

Since the MBs predicted by our model are estimates, estimation errors are indeed unavoidable. This issue is common in causal discovery tasks (e.g. the PC algorithm performs numerous conditional independence tests and similarly accumulates Type I and Type II errors). Following your suggestion, we evaluated our full pipeline across varying MB thresholds on the real-world Sachs Flow-Cytometry dataset n=1755n=1755, recording both the MB and final CPDAG metrics (added in the revised manuscript, Appendix A.6):

Threshold0.10.30.50.70.9
MB precision0.8060.8230.7940.7350.692
MB recall0.3620.2380.2400.2320.199
CPDAG precision0.3641.0001.0001.0000.750
CPDAG recall0.2000.3000.3000.3000.150
CPDAG F10.2580.4620.4620.4620.250
  • Overall performance remains relatively stable across moderate thresholds (0.3–0.7), though overly low or overly high thresholds degrade performance. Notably, precision and recall do not vary strictly monotonically with the threshold. This occurs because the initial MB estimated in Algorithm 1 (line 4) dictates the subsequent restricted-variable queries, causing estimation errors to propagate (whether too many or too few variables are included initially, subsequent sub-MB queries within the initial MB become more error-prone).
  • Effect of MB precision/recall on the final CPDAG: The edge precision/recall of the final recovered graph are generally positively associated with those of the predicted MBs (not strictly proportional due to error propagation). Specifically, omitting a true MB variable directly causes its edge with the target variable to be missed (lowering recall), whereas falsely included MB variables will eventually be determined as either adjacent to the target variable or part of a v-structure, introducing spurious edges (lowering precision).
  • Low Recall: The low recall of the final CPDAG directly stems from the low recall of the predicted MBs, which can be attributed to two factors:
    • Task and data characteristics: As observed in our experiments, many baseline methods also exhibit low recall, suggesting that this behavior is partly driven by the difficulty of causal discovery and the characteristics of the real-world dataset.
    • Decoupling paradigm: While learning decoupling weights directly exposes the MB, reweighting data points under finite samples acts similarly to sample selection, reducing the effective sample size and shifting the high-density regions of the data [1]. This may weaken dependence signals, leading to false negatives.

Q2 / W2 Effectiveness of the Decoupling Mechanism:

Thank you again for this thoughtful question. We added a more direct inference-time ablation experiment in Appendix A.5. On the real-world Sachs Flow-Cytometry dataset n=1755n=1755, we fixed a trained checkpoint and evaluated it at inference time using uniform-weight controls as well as randomly shuffling all learned weights across samples (shuffled results are reported as mean±std over 25 runs):

WeightsSHD$_c$ ↓Precision ↑Recall ↑F1 ↑avgHSIC ↓
Learned141.0000.3000.4620.057
Uniform160.8330.2500.3850.072 (equivalent to the original distribution)
Shuffled learned16.640 ± 0.8100.667 ± 0.1780.218 ± 0.0410.328 ± 0.0680.072 ± 0.002

As shown, losing the learned decoupling weights leads to a noticeable degradation in overall performance, providing more direct evidence for the effectiveness of the decoupled weights. Meanwhile, the learned weights consistently achieve a lower avgHSIC (averaged over the mean normalized weighted HSIC of the weighted distributions across all decoupling processes), confirming the effectiveness of the learned weights in decoupling.


[1] Zhengkang Guan, Fei Wu, Kun Kuang. Decoupled Causal Discovery. ArXiv, 2026

Reviewer_ZijVOpenReview ↗

These results strengthen the empirical support for the proposed method

Thank you for the detailed response and the additional experiments. These results strengthen the empirical support for the proposed method. My overall assessment remains Weak Accept.

Thank you again for your time and valuable feedback, which has significantly enhanced our manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.