acceptodds
Under review as a conference paper at ICLR 2027

CoreFlow: Low-Rank Matrix Generative Models

Abstract

Learning matrix-valued distributions from high-dimensional and possibly incomplete training data is challenging: ambient-space generative modeling is computationally expensive and statistically fragile when the matrix dimension is large but the sample size is limited. We propose CoreFlow, a geometry-preserving low-rank flow model that learns shared row/column subspaces across the matrix distribution, and then trains a continuous normalizing flow only on the induced low-dimensional core. CoreFlow is designed for settings where shared low-rank matrix geometry is present, especially in high-dimensional limited-sample regimes. This separates shared matrix geometry from sample-specific variation, preserves matrix structure, and substantially improves training efficiency. The same framework also handles incomplete training matrices through masked Riemannian updates and iterative completion. Across benchmarks spanning astrophysical imaging, atmospheric reanalysis, ocean climate, and various simulations, CoreFlow substantially improves spectral and moment-level generation quality in few-sample regimes while remaining competitive in data-rich settings, even under compression to 9% of the ambient dimension and with up to 40% missing training entries.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Matrix-valued generative modeling, Low-rank matrix generation, Generative modeling with incomplete data

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jmMeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

CoreFlow is a generative model for matrix-valued data learned from limited and possibly incomplete training samples. It first learns shared orthonormal row and column bases using alternating masked fitting and low-rank imputation, then fits a flow-matching CNF to the vectorized cores. Its main bound decomposes the Wasserstein error into subspace, core-modeling, and residual terms. Experiments on four real and four synthetic datasets report better performance than ambient-space diffusion and flow baselines, particularly with limited data and substantial missingness, as well as faster training.

Critical strengths

S1. Strong robustness to missingness where the held-out evaluation can be verified. ERSSTv5 uses a temporal split and a freeze-then-test pipeline. At 40% missingness, CoreFlow substantially outperforms MissFlow and MissDiff and remains stable across three training seeds.

S2. The ablations include appropriate controls. SMG-Core combines the learned subspaces with a Gaussian core, while PCA-Flow combines the same CNF with a flattened PCA representation. CoreFlow clearly outperforms SMG-Core on Solar and Solar 2.

Critical weaknesses

W1. The stated held-out protocol and several reported values cannot be verified from the released artifacts.**

  • Evaluation reference. Appendix C describes held-out sets for Solar, Solar 2, and LSPF, and §5 says simulations use complete held-out matrices. The released evaluators instead use the training matrices; only ERSSTv5 includes a held-out script and separate evaluation tensors. The artifacts therefore do not establish which reference sets produced the main Solar and LSPF results or permit a memorization check.

  • Configurations and LSPF. Table 7 says that configs/paper contains every setting, but it omits LSPF at 20% missingness and four of twelve simulation settings; configs/baseline and results/paper, both listed in the README, are absent. The README and report script also say that learning rates were selected to match the paper's values. Tables 1 and 23 report 2.70±0.19 for CoreFlow on LSPF at 0% missingness, while the script records 2.88±0.19 and the Table 11 runs average 2.846. At 2.88, PCA-Flow has the lower SVRelL2 mean (2.74), although the intervals overlap.

  • Other discrepancies. The paper describes five independent sets of 500 samples, whereas the code splits one generated batch into 20 subsets. MMD is reported but not implemented in the repository, and the LSPF 40% configuration uses a 72×72 crop rather than the size stated in Appendix C.3.

W2. The source of the gains is only partly isolated.

  • PCA-Flow. Quantitative results cover only LSPF and complete-data simulations; ERSSTv5 appears only qualitatively. PCA-Flow matches CoreFlow on LSPF SVRelL2, while CoreFlow is better on the other metrics. Because PPCA and EM-PCA support missing values, an analogous impute-and-refit baseline could test whether matrix structure matters under missingness.
  • Simulation comparison. The 576-dimensional PCA subspace minimizes training reconstruction error among subspaces of that size, yet CoreFlow is reported to be 2–5.5 times better. The gap therefore requires an explanation beyond retained energy, such as layout, normalization, or tuning.
  • SMG-Core. On Blobs, the true core is Gaussian, the sample size exceeds the core dimension, and the recovered subspaces are exact to the reported precision. Table 3 nevertheless reports 195 for SMG-Core versus 0.43 for CoreFlow, and SMG-Core improves by roughly an order of magnitude when its assumed rank exceeds the true rank. Without the prior settings or code, this comparison is difficult to interpret.
  • The ambient baselines use weak missing-data handling; an impute-then-generate or EM-based baseline would be more informative.
Questions for authors
  1. Please release the non-ERSSTv5 held-out data and evaluator, original configurations, and Solar nearest-training-neighbor distances.
  2. What produced the Solar block entries in Table 12 and the Solar 2 heterogeneous value of 0.59, given the archived value of 0.94?
  3. How does PCA-Flow with EM-PCA perform at 20% and 40% missingness? Please also report its quantitative ERSSTv5 results.
  4. Why does Table 5 show non-negligible 25th singular values for data whose generator implies rank at most 24?

Minor:

  • Please document mean-fill initialization, the use of centering despite the uncentered statement of Proposition 4.3, core standardization, and how the additional Table 11 seeds were set despite a hard-coded seed.
  • Relevant missing work includes MisGAN, Ambient Diffusion, as well as Davis–Kahan perturbation theory for Proposition 4.3.
3 comments
AuthorsOpenReview ↗

Response to Reviewer jmMe [Part 1/2]

We’re sorry for the confusion. We mistakenly submitted the code from our previous OpenReview submission. That version did not include the subsequent revisions and experiments. We have now updated the paper and supplementary code on OpenReview supplementary materials. We take responsibility for this mistake and sincerely appreciate your careful checks.

1. Evaluation references and supplementary materials (W1; Q1).

Section 5 and Appendix C now specify the evaluation references:

DatasetEvaluation reference
Blobs, Bands, Waves, Crosshatch1,000 independent draws per case
LSPF500 test fields, separate from the 8,260 training fields
ERSSTv5The final 84 months, following training and validation
Main Solar / Solar 2 comparisonsThe same 300 images used for training

For Solar and Solar 2, we deliberately use a few-sample setting with 300 training images. All methods use these images as the reference in the main comparisons. This measures fit to the training data. To fully address your concerns, we also added a separate Solar experiment with 240 training images and 60 held-out test images. Table 9 reports three independent training runs on this split at 0% and 40% missingness.

The supplementary code now includes the evaluation reference tensors in data/, the current CoreFlow settings in configs.yaml, and the evaluator. The README links to the larger LSPF and simulation training datasets. Configurations for the independent-training study are in ablations/. Appendix C.5 gives the training settings.

Unless stated otherwise in the paper, our latest results use 500 generated matrices with each of sampling seeds 0, 1, and 2. We calculate each batch’s metrics separately and report their mean and sample SD (Appendix C). Also, our latest code in supplementary materials now includes the MMD implementation in metrics.py.

The revised LSPF evaluation uses matrices before any PCA reconstruction across samples. We split them into 8,260 training and 500 test matrices before fitting. The complete-data spectral error is 10.34 ± 0.02 under this protocol (Table 1, ×100 units). We also corrected the LSPF size description in Appendix C.3. At 20% and 40% missingness, the 9 × 9 patch construction uses a 72 × 72 crop. Every method is evaluated against reference matrices on that same region. Complete-data evaluation uses the full 80 × 80 fields.

Our latest paper’s results use 500 generated matrices with each of sampling seeds 0, 1, and 2. We calculate each batch’s metrics separately and report their mean and sample SD (Appendix C). Also, the supplementary code includes the MMD implementation in metrics.py.

2. Solar nearest-neighbor check (W1; Q1).

Following your suggestion, Appendix D.10 (Table 26) now reports the Solar nearest-training-neighbor distances. On average, generated images are farther from their nearest training image than training images are from their closest other training image. We found no exact copies among the 1,500 generated images per condition. This means CoreFlow is not simply copying and pasting training images under this extremely few-data training setting. Table 9 separately evaluates generation against held-out Solar images.

3. MMD and the LSPF crop (W1).

Appendix C.1 defines the MMD estimator. For the revised evaluations, the RBF bandwidth is computed only from true reference matrices and shared across methods within each dataset and evaluated region. Our latest code in supplementary materials now includes the MMD implementation in metrics.py.

Each LSPF field is originally 80 × 80. With complete training data, we evaluate the whole field. In the missing-data experiments, CoreFlow applied a patchification with 9 × 9 patch size before training. Since 80 is not divisible by 9, preprocessing keeps the upper-left 72 × 72 region, which fits eight patches along each side. We evaluate all methods on this same region at both 20% and 40% missingness. Appendix C.3 now explains this more clearly.

AuthorsOpenReview ↗

Response to Reviewer jmMe [Part 2/2]

[Part 2/2] of our response

4. Completion followed by PCA-Flow, and ERSSTv5 (W2; Q3).

Thank you for this suggestion! We agree that the key question is whether the representation helps beyond the completion step. Following your advice, we implemented two PCA-Flow ablations to examine this:

  • PCA-Flow (Stage I) uses the same completed training matrices as CoreFlow, then replaces the matrix cores with PCA scores. This tests the representation used for generation.
  • PCA-Flow (iterative PCA) alternates PCA fitting with reconstruction of missing entries, keeping observed entries fixed. It then trains the flow on PCA scores. This uses iterative PCA imputation.

Appendix D.8 now describes the implementations. Table 24 gives the full results. The LSPF spectral errors are shown below (SVRelL2 ×100; lower is better). Missingness CoreFlow PCA-Flow (iterative PCA) 20% 4.11 ± 0.07 4.66 ± 0.03 40% 7.12 ± 0.09 10.12 ± 0.11 With the same Stage-I completion, neither representation is best on every metric. At 20% missingness, Stage-I + PCA-Flow has lower spectral error (3.20 versus 4.11). CoreFlow has lower error in the mean Frobenius norm (0.17 versus 0.46).

For ERSSTv5, we now report quantitative results in Table 24. CoreFlow and PCA-Flow have spectral errors of 21.26 ± 0.46 and 23.46 ± 1.16, respectively. The table also reports the moment discrepancies and MMD.

PCA-Flow is a representation ablation we developed within our framework. It tests replacing matrix cores with PCA scores and can also give competitive results.

We keep CoreFlow as our main formulation because it separates shared matrix structure from variation between samples. The bases U,VU,V capture the shared row and column spaces, while the core SS models sample-specific variation. Our completion, compression, and generation use the same bases, without an additional PCA fit. Also, our CoreFlow comes from theoretical interpretations as proved in Section 4.

5. PCA reconstruction and generation (W2).

PCA minimizes training reconstruction error at a fixed dimension. Generation also depends on how well the flow learns the distribution of those coordinates. So lower reconstruction error does not guarantee lower generation error. Appendix D.8 now makes this more clear.

6. SMG-Core on Blobs (W2).

Thank you for catching this! Your concern was correct. The relatively large error came from covariance scaling. The executed sampler fitted an empirical Gaussian to the Stage-I cores. It was divided by the compact-SVD dimension minus one, rather than the training sample count minus one.

Now, we used matched inference seeds 0, 1, and 2 for SMG-Core and CoreFlow, and updated all related tables. Appendix C.7 now describes the corrected implementation. On complete Blobs, SMG-Core has lower spectral error: 0.28 ± 0.04 versus 1.34 ± 0.06 for CoreFlow (Table 3, ×100 units).

The corrected comparisons still show notably lower spectral errors for CoreFlow under missingness (Tables 2–3). We wish to note that SMG-Core already uses the same Stage-I subspaces U,VU,V learned by CoreFlow, including those estimated from incomplete training data. It therefore benefits from CoreFlow’s shared row and column structure before fitting its Gaussian core model.

7. The 25th singular value (Q4).

Thank you for pointing this out. The rank of 24 refers to the simulated data before preprocessing. Table 5 measures the data after preprocessing.

For Waves and Crosshatch, we normalize each matrix to [0, 1] before training. This also shifts all its entries by a constant. The shift differs across matrices, so subtracting the mean training matrix does not fully remove it. That’s why the 25th singular value is non-zero in Waves and Crosshatch.

We checked this by removing the constant component for the spectral calculation only. The 25th-to-first singular-value ratio then falls to 8.64 × 10⁻⁸ for Waves and 7.09 × 10⁻⁸ for Crosshatch, both close to zero. Appendix C.2 (“Rank selection and spectral concentration”), in the paragraph before Table 5, now explains this cause and reports the check.

8. Implementation details and related work (Minor).

Thank you for pointing these out. Section 3 now explains how we initialize missing entries with observed-value means. Appendix C describes centering before the initialization SVD. The supplementary code README now documents core standardization with one mean and one population standard deviation of the training cores. The extra runs in revised Table 9 use training seeds 42, 123, and 2026, as listed in ablations/configs.yaml. Also, we have added MisGAN and Ambient Diffusion to Related Work.

Thank you again for these careful checks! We hope the added experiments and clarifications address your concerns. We are happy to discuss any remaining questions!

AuthorsOpenReview ↗

Final Follow-up on Our Rebuttal

Dear Reviewer jmMe,

As the discussion period ends today, we would really appreciate it if you could take a look at our rebuttal and revised manuscript. We have put considerable efforts into additional experiments and corrections to address your concerns.

We would be very grateful for any final feedback and would appreciate the last opportunity to clarify any remaining concerns before the discussion closes.

Thank you so much for your time and efforts!

Best, Authors

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

Method: CoreFlow is a two-stage generative model for high-dimensional matrices that share a low-rank structure Mi≈USiV⊤M_i \approx U S_i V^\top. Stage I learns shared orthonormal subspaces (U,V)(U, V) on the Stiefel manifold (using alternating masked updates and low-rank completion when entries are missing), and Stage II trains a continuous normalizing flow on the small R×RR \times R core matrices.

Theory: The authors show that the population Stage I loss recovers the true subspaces in the exact low-rank case and bound the Tucker initializer error. They also give a conditional W2W_2 bound (Theorem 4.7 and Corollary B.1) that splits generation error into subspace error, core flow error, rank-RR residual tail energy, and completion shift.

Validation: The model is tested on four rank-24 synthetic datasets and four scientific datasets (Solar, Solar 2, LSPF, and ERSSTv5) at 0%, 20%, and 40% missingness against MissDiff, MissFlow, CSDM, SMG-Core, and a flattened PCA-Flow ablation.

Critical strengths
  1. Efficient, structure-preserving design: Learning the flow in the R2R^2 core space cuts down the generative dimension, and the orthonormal decoder U∗mat(s^)V∗⊤U^* \text{mat}(\hat{s}) V^{*\top} preserves Frobenius distances without amplifying core errors.

  2. On Solar (N=300N=300, 40% missingness), full-space diffusion and flow baselines break down (89.3×10−289.3 \times 10^{-2} and 75.7×10−275.7 \times 10^{-2} spectral error), while CoreFlow reaches 1.01×10−21.01 \times 10^{-2} at 9% dimension and trains 2.9x to 4.5x faster in total wall-clock time.

  3. The paper clearly states which finite-sample rates are not proved in Theorem 4.7 and reports real end-to-end V100 training times that include Stage I optimization (Table 8).

Critical weaknesses
  1. The Blobs synthetic dataset is generated from an i.i.d. Gaussian core S=1.5ZS = 1.5 Z, yet SMG-Core (which uses the exact same Stage I subspaces and fits a conjugate Gaussian NIW core model) gets a massive 195×10−2195 \times 10^{-2} spectral error vs. 0.43×10−20.43 \times 10^{-2} for CoreFlow. Since the true core is Gaussian, this gap points to poorly scaled NIW priors or an implementation issue rather than model misspecification.

  2. Flattened PCA-Flow nearly ties CoreFlow on complete LSPF data (2.74×10−22.74 \times 10^{-2} vs. 2.70×10−22.70 \times 10^{-2}), and it is not tested on Solar or paired with matrix completion for missing data. Also, CSDM only uses naive entrywise mean filling, which mixes up the benefit of the USV⊤U S V^\top representation with the benefit of iterative low-rank completion.

  3. Main table error bars only measure 5 sampling batches from one trained model, the 3-seed runs in Table 11 evaluate Solar and LSPF on the training set instead of the held-out test set, and MMD recomputes the median bandwidth separately for each model's pooled samples.

Questions for authors
  1. Why does SMG-Core get 195×10−2195 \times 10^{-2} error on complete Blobs data when the true cores are i.i.d. Gaussian, and what happens if you fit a simple empirical Gaussian N(sˉ,Σ^)\mathcal{N}(\bar{s}, \hat{\Sigma}) on the Stage I cores instead of the NIW prior?

  2. How does CoreFlow compare to Stage I low-rank completion followed by flattened PCA-Flow on Solar and LSPF across 0%, 20%, and 40% missingness?

  3. What fraction of held-out test Frobenius energy is captured by the shared subspaces at the chosen ranks in Table 4, and how is RR chosen when training data are incomplete?

  4. Can you report 3-seed retraining error bars on the held-out test sets for Solar and LSPF, and re-evaluate MMD (Tables 20 to 22) using a fixed bandwidth σ\sigma computed only from the ground-truth reference set?

3 comments
AuthorsOpenReview ↗

Response to Reviewer ihHx [Part 1/2]

Thank you for checking our experiments carefully!

1. SMG-Core on Gaussian Blobs (W1; Q1).

Thank you for catching this. The large Blobs error came from an incorrect covariance divisor in the empirical Gaussian sampler. It used the compact-SVD dimension minus one, rather than the training sample count minus one.

We corrected the divisor to N−1 and recomputed all SMG-Core results in the revision draft. We used matched inference seeds 0, 1, and 2 for SMG-Core and CoreFlow, and updated all related tables. Appendix C.7 now describes the corrected implementation.

On complete Blobs, SMG-Core has lower spectral error: 0.28 ± 0.04 versus 1.34 ± 0.06 for CoreFlow (Table 3, ×100 units).

The corrected comparisons still show notably lower spectral errors for CoreFlow under missingness (Tables 2–3) on most datasets. We wish to note that SMG-Core already uses the same Stage-I subspaces U and V learned by CoreFlow, including those estimated from incomplete training data. It therefore benefits from CoreFlow’s shared row and column structure before fitting its Gaussian core model.

2. Stage-I completion followed by PCA-Flow (W2; Q2).

Following your suggestion, we newly implemented PCA-Flow with two completion methods for settings with missing data. One uses matrices completed by CoreFlow’s Stage I. The other alternates PCA fitting with reconstruction of missing entries, while keeping observed entries fixed. All details are now given in Appendix D.8.

For LSPF, both representations use 3,136 latent coordinates and the same flow architecture and training settings. The methods share the data split, mask, and evaluation. Table 24 in Appendix D.8 gives the comparisons at 0%, 20%, and 40% missingness.

CoreFlow has lower Frobenius-mean errors in the shared-completion comparison:

LSPF missingnessCoreFlowStage-I + PCA-Flow
20%0.17 ± 0.080.46 ± 0.05
40%0.20 ± 0.020.91 ± 0.02

PCA-Flow has lower spectral error on complete LSPF. At 20% missingness, Stage-I + PCA-Flow also has lower spectral error: 3.20 ± 0.03 versus 4.11 ± 0.07, in ×100 units.

We also evaluated the Solar PCA-Flow ablations on the 240/60 training/test split. The table below uses training seed 42. Its error bars show variation across sampling seeds 0, 1, and 2, with 500 generated images per seed. For comparison, Table 9 reports variation across three independently trained CoreFlow models. Its seed-42 result at 40% missingness is 0.773, matching the 0.77 shown below.

Solar missingnessCoreFlowPCA-Flow / Stage-I + PCA-FlowIterative PCA-Flow
0%1.03 ± 0.030.46 ± 0.04—
20%1.31 ± 0.010.69 ± 0.041.87 ± 0.03
40%0.77 ± 0.010.78 ± 0.012.47 ± 0.02

We developed PCA-Flow as a representation ablation within our framework. CoreFlow uses UU and VV to capture shared row and column structure, while the core SS captures variation between samples. In our CoreFlow, the same bases serve completion, compression, and generation, without an additional PCA fit. Section 4 also provides theoretical guarantees for this factorization under certain assumptions. Therefore, we keep CoreFlow as our main formulation, while PCA-Flow remains a reasonable alternative.

3. Rank choice and held-out energy (Q3).

We fix the rank before fitting each model and keep it unchanged across missingness levels. When training entries are missing, Stage I alternates learning U and V with filling the missing entries. Note that test data are never used to choose the rank. The compression ratio ρ\rho is the number of core coordinates, R2R^2, divided by the number of input entries (Section 5; Appendix C.2).

Following your suggestion, we now report how much energy the learned representation preserves on independent test data, in Appendix D.9 (Table 25). Energy means the sum of squared matrix entries. Across 0%, 20%, and 40% missingness, retention is at least 99.97% for simulations, over 99.8% for ERSSTv5, and about 96% for LSPF within the area used for evaluation.

We hope these explanations help!

AuthorsOpenReview ↗

Response to Reviewer ihHx [Part 2/2]

[Part 2/2] of our response:

4. Independent training seeds and fixed-bandwidth MMD (W3; Q4).

We added three independent training runs with seeds 42, 123, and 2026 at 0% and 40% missingness. Here, Solar uses 240 training and 60 test images. LSPF uses 8,260 training and 500 test fields. Table 9 and Appendix C.6 now report:

DatasetMissingnessSpectral error across training runs, ×100
Solar0%1.049 ± 0.219
Solar40%1.103 ± 0.672
LSPF0%10.328 ± 0.535
LSPF40%7.246 ± 0.367

Each trained model uses inference seeds 0, 1, and 2, generating 500 matrices per seed. We average its three inference seeds first, then calculate the mean and sample SD across trained models.

Following your suggestion, Appendix C.1 now specifies a fixed RBF bandwidth computed only from true reference data for the revised MMD evaluations. Methods share the bandwidth within each dataset and evaluated region. Tables 20–22 now report the MMD comparisons.

CoreFlow has lower MMD than MissDiff and MissFlow at 20% and 40% missingness on Solar, Solar 2, and LSPF (Table 20). This main trend remains unchanged.

Thanks again for these insightful comments! They indeed helped us clarify the evaluation and improve the paper. We are happy to answer any further questions.

Thank you for fixing the SMG-Core implementation and providing the additional comparisons. These results address my primary concerns, though the statistical strength of some comparisons remains limited by the small number of training seeds.

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

This work introduces CoreFlow, a two-stage generative model for matrix-valued data. CoreFlow first learns shared row and column subspaces to project training matrices into low-dimensional core representations. A continuous normalizing flow is then trained on the distribution of these cores; new matrices are synthesized by sampling from the flow and decoding the sampled cores back to the original matrix space. To handle incomplete observations, the model leverages masked subspace updates alongside an iterative matrix completion procedure. The paper establishes population-level theoretical guarantees that disentangle subspace approximation error from the error incurred by core generation. Empirically, CoreFlow is validated on four real-world datasets and four synthetic benchmarks, including low-data regimes and scenarios with up to 40% missing entries. Comparisons are conducted against both ambient-space generative baselines and competing low-dimensional methods, using evaluation metrics such as spectral statistics, entrywise moments, Frobenius norm, and MMD.

Critical strengths

The method offers a coherent way to exploit shared matrix structure: it learns row and column subspaces, then performs generative modeling in a smaller core space. This is a useful design for high-dimensional matrix data when the shared low-rank assumption is appropriate.

Critical weaknesses

The training set and the test set used the same data.

Questions for authors

The code in the supplementary materials shows that the same data samples were used for both training and testing. Could the authors explain this overlap?

5 comments
AuthorsOpenReview ↗

Response to Reviewer PdhG

We’re very sorry for the confusion. When submitting supplementary materials, we mistakenly submitted code from our previous OpenReview submission. That version did not include the subsequent revisions and new experiments.

In the latest main comparisons, we used held-out test set for the simulations (blobs, bands, waves, and crosshatch), LSPF dataset, and ERSSTv5.

For the few-sample Solar Flare (Solar and Solar 2) datasets, we use all 300 images for training and as the reference in the main comparisons. This comparison is fair across all methods, since they used the same set for evaluation. These comparisons measure fit to the training data.

We have updated the paper and supplementary code on OpenReview. We take responsibility for the reported results. The supplement provides the CoreFlow implementation, current configurations, reference data, and evaluator. Its README also links to the larger LSPF and simulation training datasets.

Following the suggestions of Reviewers ihHx and jmMe, we added completion-based PCA-Flow ablations and corrected the SMG-Core covariance scaling. The evaluations rerun for this revision use matched sampling seeds 0, 1, and 2, as described in the latest Appendix C.

To directly reply to your concerns, we have clarified the reference data in Section 5 and Appendix C:

DatasetEvaluation reference
Blobs, Bands, Waves, Crosshatch1,000 independent draws per case
LSPF500 test fields, separate from the 8,260 training fields
ERSSTv5The final 84 months, following training and validation
Main Solar / Solar 2 comparisonsThe same 300 images used for training

We also added a separate Solar experiment with 240 training images and 60 held-out test images. Table 9 reports three independent training runs at 0% and 40% missingness. Here, the 60 test images are not used for training.

We hope these clarifications and the added held-out experiment address your concerns. Thank you for bringing this issue to our attention. We sincerely hope you can take a further look, and we are happy to answer any further questions!

CoreFlow_final/configs.yaml uses the same fields.pt file for both training and evaluation. Doesn’t this data leakage issue still remain?

If I misunderstood, feel free to push back. If I understood correctly, please update the code promptly.

AuthorsOpenReview ↗

Follow-up Response to Reviewer PdhG

Dear Reviewer PdhG,

Thank you for checking our code again! As we explained, only the Solar Flare datasets (Solar and Solar 2) use all 300 images for training and as the reference in the main comparisons. We chose this setting to give every method all available training data under this extremely limited data setting. Every method uses the same reference and the comparisons are fair, and the main tables show clear gains for CoreFlow under this shared evaluations.

For every other dataset: simulations (blobs, bands, waves, crosshatch), LSPF, and ERSSTv5, the evaluation use held-out references.

Below are our train-reference specification in CoreFlow_final/configs.yaml:

Datasettrain_filereference_file
Blobssimulation_A/train.ptsimulation_A/test.pt
Bandssimulation_B/train.ptsimulation_B/test.pt
Wavessimulation_C/train.ptsimulation_C/test.pt
Crosshatchsimulation_E/train.ptsimulation_E/test.pt
LSPFlspf/train.ptlspf/test.pt
ERSSTv5ersstv5/train.ptersstv5/test.pt
Solarsolar/fields.ptsolar/fields.pt

We also evaluated CoreFlow on a separate 240/60 Solar split, using 240 images for training and 60 held-out images as the reference. Those results are in Tables 9 and 27. And this held-out experiment is configured in CoreFlow_final/ablations/configs.yaml:

| Solar | solar_heldout/train.pt | solar_heldout/test.pt |

We hope these clarifications help! Please let us know if you have any further concerns.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General comment to AC and reviewers

Dear Area Chair and reviewers,

Thank you for your careful feedback! During the rebuttal, we updated the supplementary code, clarified the evaluation references, corrected the SMG-Core comparison, and added the requested PCA-Flow controls and quantitative ERSSTv5 results. All changes are reflected in the revised draft in red. Also, the held-out Solar PCA-Flow results discussed with Reviewer ihHx are now in Appendix D.11, Table 27.

Only the Solar Flare datasets (Solar and Solar 2) use all 300 images for training and as the reference in the main comparisons. We chose this setting to give every method all available training data. Every method uses the same reference, and the main tables show clear gains for CoreFlow under this shared evaluations. All other datasets (simulations, LSPF, and ERSSTv5) use held-out references.

We also tested our CoreFlow on a separate 240/60 Solar Flare split (Tables 9 and 27). At 40% missingness, its spectral error is 0.77 against the held-out images, compared with 0.78 against the 300-image reference (both ×10−210^{-2}, training seed 42). Thus, the abstract’s 9% dimension and 40% missingness statement is supported by held-out generation.

With the discussion deadline approaching, we would be sincerely grateful if you could let us know whether any concern remains. We would like to respond while there is still time. Thank you again for helping us improve the paper!

AuthorsOpenReview ↗

Follow-up on Rebuttal and Remaining Concerns

Dear Reviewers PdhG and jmMe,

As the discussion phase will end in around 24 hours, we would sincerely appreciate it if you could let us know whether our response has adequately addressed your concerns. We're happy to clarify anything further!

Thank you so much for your contributions!

Best, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.