acceptodds
Under review as a conference paper at ICLR 2027

Rewire then Express: Cascaded Flow Matching for Single-Cell Perturbation Response Prediction

Abstract

Predicting cellular responses to genetic and chemical perturbations is a central challenge in functional genomics and drug discovery. Many recent models formulate this task as a single-stage transport from unperturbed to perturbed states. Inspired by the role of upstream regulatory changes in shaping transcriptional responses, we introduce *Rewire-then-Express*, a generative paradigm that structures perturbation prediction as an ordered cascade. The model first predicts an intermediate response representation, supervised by perturbation-induced shifts in pretrained single-cell embeddings, and then generates transcriptomic profiles conditioned on this representation. We instantiate this paradigm as **scCasFlow**, a Cascaded Flow Matching framework. Dual temporal coordinates separately parameterize intermediate-response and transcriptomic generation, while an ordered scheduling strategy coordinates their progression according to the intermediate-to-expression ordering. Flow matching operates in latent space, with a lightweight decoder enabling prediction across the retained gene vocabulary without highly variable gene selection. On the Norman and ComboSciPlex benchmarks, scCasFlow improves predictive performance over the evaluated methods under held-out perturbation splits and achieves more than 100× inference throughput of the evaluated expression-space flow matching baseline. Post hoc analyses further identify associations between the intermediate representations and transcription-factor activity scores inferred from held-out expression profiles, alongside agreement between predicted and observed pathway-aggregated expression changes. Together, these results support the predictive value of an explicitly supervised, ordered generative factorization for cellular perturbation modeling. Code is available at [https://anonymous.4open.science/r/scCasFlow-80E8](https://anonymous.4open.science/r/scCasFlow-80E8).

Primary area:
generative models
Keywords:
single-cell perturbation response prediction

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_hYDtOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper introduces a novel framework, named scCasFlow, that aims to improve upon flow matching-based single-cell perturbation prediction approaches by explicitly modeling intermediate regulatory steps in its generative process. Specifically, scCasFlow aims to first predict population-level perturbation responses and then generates single-cell gene expression profiles conditioned on these predictions, treating the two as separately parameterized flows. The authors demonstrate that scCasFlow outperforms several other perturbation prediction models in generalization to unseen genes and perturbation combinations across a large battery of evaluation metrics. They also perform several ablation analyses to confirm the merits of leveraging intermediate supervision as well as various design choices aimed at formulating the regulatory cascade and corresponding generation procedure.

Critical strengths
  • The paper is clearly written and does a good job of contextualizing its contributions with respect to the broader literature on single-cell perturbation prediction methodologies. The Methodology section is written in a manner that is easy to understand, and the related modeling decisions appear to be generally accompanied by convincing rationales.
  • The authors leverage a comprehensive suite of evaluation metrics to assess the performance of scCasFlow against a number of state-of-the-art models, demonstrating consistently better performance than all benchmarked models. The diversity of the evaluation metrics (covering reconstruction, differential expression, and discrimination tasks) provides strong evidence for scCasFlow’s outperformance in the evaluated settings.
  • The authors introduce a number of creative ablations that help to assess the advantages of the various modeling choices used in scCasFlow. The results of these ablation studies provide support for the utility of the core components of this framework (e.g., use of intermediate flow, intermediate supervision, generation order).
Critical weaknesses
  • It seems as if a cell’s intermediate response representation is defined as the difference between the Stack representations of a control state and a corresponding steady-state, post-perturbation state. It’s not clear that this framing is appropriate, as the post-perturbation Stack representation more closely mirrors the downstream gene expression target than an intermediate state.
  • While the paper includes a number of ablation studies that appear to show the benefits of various model design choices, the evaluation metrics used in these analyses only cover a fraction of those that are used in earlier parts of the paper. In addition, the ablation studies focus only on the Norman-Additive dataset.
  • A large portion of the Appendix is not cited in the main text of the paper. For example, Table S2 includes what appears to be additional benchmarking results that supplement Table 1’s results, but scCasFlow’s performance on the single perturbation setting is not as consistently strong relative to other baseline methods, especially on the MSE and MAE metrics.
Questions for authors
  • Is the intermediate response representation appropriately defined, given that it captures the difference between the final post-perturbation state and a control state?
  • What is the rationale behind selecting the Stack encoder to be used for representing intermediate response, while the Stack tokenizer is used for representing expression-flow states?
  • How do the various ablations affect the broader set of evaluation metrics used in Table 1? Which modeling decisions are most beneficial for each category of metrics? How do the ablation results look for datasets besides the Norman-Additive dataset?
  • Table S2 appears to indicate that scCasFlow’s outperformance of other methods is larger and more robust in the double perturbation setting compared to the single perturbation setting. Can the authors discuss this result?
4 comments

We thank the reviewer for the constructive feedback and helpful suggestions.

Q1. Is the intermediate response representation appropriately defined, given that it captures the difference between the final post-perturbation state and a control state?

We thank the reviewer for this valuable comment. The intermediate is not the difference between paired control and perturbed states of individual cells. Instead, it represents a perturbation-induced population-level shift in pretrained Stack embeddings, estimated from randomly sampled perturbed-cell cohorts relative to the control population (Eq. 5). The cells used to construct this target are sampled independently of those used for individual-cell expression generation.

Importantly, this intermediate is a predictive response representation, not a directly observed biological intermediate state. Although derived from post-perturbation transcriptomic measurements during training, it does not directly identify upstream regulatory mechanisms. Its purpose is to separate the prediction of population-level perturbation effects from the generation of individual-cell expression profiles. At inference, it is predicted solely from control-cell information and perturbation identity, without access to post-perturbation measurements.

Our experiments support this formulation. Shuffling and randomizing intermediate targets reduce PΔ from 0.8488 to 0.7819 and 0.7395, respectively (original manuscript Table 4a). Removing the intermediate flow or using simultaneous generation also reduces performance (original manuscript Table 4b). Furthermore, the associations between learned Rewire representations and inferred TF activities (Fig. 5) provide additional support. These results support the intermediate as a useful perturbation-associated response surrogate rather than a direct measurement of causal regulatory rewiring.

Q2. What is the rationale behind selecting the Stack encoder to be used for representing intermediate response, while the Stack tokenizer is used for representing expression-flow states?

We appreciate the reviewer’s thoughtful feedback. The two components serve distinct purposes. The Stack encoder produces contextualized cell representations suitable for estimating population-level perturbation shifts. We therefore use its embeddings to construct the intermediate supervision. In contrast, the Stack tokenizer projects gene expression into compact gene-module tokens, preserving expression information while reducing the dimensionality of the generative process. These tokens are used by the Express flow, with a jointly trained decoder mapping the generated latents back to gene expression.

This separation follows our modeling formulation: the intermediate representation captures population-level perturbation effects, while the expression flow generates individual-cell responses. The shuffled and random target controls (original manuscript Table 4a) support the encoder-derived intermediate supervision, while the tokenizer and decoder ablations (original manuscript Fig. 4) support the latent-space expression generation design.

Q3. How do the various ablations affect the broader set of evaluation metrics used in Table 1? Which modeling decisions are most beneficial for each category of metrics? How do the ablation results look for datasets besides the Norman-Additive dataset?

We thank the reviewer for pointing this out. As shown in Table 1, we expanded the ablation evaluation on Norman-Additive to include additional metrics, covering expression reconstruction (MSE and MAE), perturbation discrimination (DS), perturbation-effect correlations (P20 and P50), and differential-expression directional agreement (Dir).

We initially focused on PΔ because it directly measures the agreement between predicted and observed perturbation-induced gene-expression changes, which is central to our task. Unlike expression-level errors such as MSE and MAE, it focuses on perturbation effects rather than overall expression reconstruction. Other metrics provide complementary assessments, but their differences are generally less pronounced or more metric-dependent for some ablations.

Table 1. Additional ablation metrics on Norman-Additive.

MethodMSE ↓MAE ↓DS ↑P20 ↑P50 ↑Dir ↑
Monolithic Flow0.0009190.0135780.93550.89540.89770.9400
Deterministic Rewire0.0007310.0120600.94770.91370.91930.9481
Decoupled Decoder0.0008170.0128060.94670.91070.90770.9165
w/o Tokenizer0.0053180.0281590.73990.78600.74770.8750
scCasFlow0.0006830.0115810.96460.92960.93740.9587

We further evaluated the same ablation variants on Norman-Holdout to examine whether these findings transfer to another evaluation setting. The results (Table 2) on Norman-Holdout show that scCasFlow achieves the best results on most metrics. Overall, these results support the contributions of the proposed components beyond Norman-Additive.

Table 2. Ablation on Norman-Holdout.

MethodMSE ↓MAE ↓DS ↑PΔ ↑P20 ↑P50 ↑DE LFC ↑Dir ↑O50 ↑
Monolithic Flow0.0014430.0147320.69110.57730.66300.68430.64840.85340.0487
Deterministic Rewire0.0013360.0144360.71840.59970.67740.69250.66210.85730.0507
Decoupled Decoder0.0015880.0167100.68030.55400.67430.69040.58950.82300.0483
w/o Tokenizer0.0013440.0138080.62110.57640.66610.67440.62050.83490.0703
scCasFlow0.0013260.0142010.74360.60540.71750.70780.66710.84620.0800

Q4. Table S2 appears to indicate that scCasFlow’s outperformance of other methods is larger and more robust in the double-perturbation setting compared to the single-perturbation setting. Can the authors discuss this result?

We thank the reviewer for pointing out this observation. As shown in Table S2, scCasFlow achieves more consistent improvements in the double-perturbation setting, outperforming all evaluated methods across all nine metrics. In contrast, its improvements in the single-perturbation setting are more modest, with CellFlow achieving lower MSE and MAE. This shows that single-gene perturbation responses are relatively easier to capture using existing predictive models, whereas double perturbations involve more complex combinatorial effects. The population-level intermediate supervision and ordered generation in scCasFlow may provide a more useful inductive bias in these challenging settings by explicitly modeling perturbation-associated changes before generating expression responses.

Thank you to the authors for their thorough responses to my questions, the additional ablation experiments, and the clearer interpretation of the results. These responses clearly describe the modeling design choices and have strengthened my confidence in the work. I have no further comments and have updated my score to a 4.

Thank you again for your thoughtful feedback and for taking the time to reassess our work.

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

This paper studies single-cell perturbation response prediction, focusing on generalization to held-out genes and unseen perturbation combinations. scCasFlow first generates an intermediate response representation supervised by population-level embedding shifts, then generates cellular expression conditioned on that representation. The method is evaluated on Norman and ComboSciPlex through benchmark comparisons, component ablations, and transcription-factor and pathway analyses.

Critical strengths
  1. Predicting an intermediate perturbation response before generating cellular expression conditioned on it is an interesting modeling approach.
  2. The paper is clearly written and easy to follow.
Critical weaknesses
  1. Rewire target variability comes from resampling population means, but the experiments do not establish its benefit over using a fixed population-mean target for each perturbation.
  2. Training-target ablations change both supervision and conditioning, leaving Express’s reliance on the generated Rewire representation at inference insufficiently tested.
Questions for authors
  1. Could the authors more clearly distinguish the designs adopted from Latent Forcing from those introduced for perturbation prediction, and explain which experiments support each modification?
  2. How does performance change when the Rewire flow is replaced with deterministic regression, retaining the same intermediate supervision, Express module, and decoder?
  3. Are the differences between strictly sequential generation and the alternative schedules in Table 4b stable across training seeds?
2 comments

We thank the reviewer for the constructive feedback and helpful suggestions.

Q1. Could the authors more clearly distinguish the designs adopted from Latent Forcing from those introduced for perturbation prediction, and explain which experiments support each modification?

We thank the reviewer for this important question. Our work builds on the multi-time scheduling principle of Latent Forcing, but introduces a different generative factorization tailored to single-cell perturbation prediction. Specifically, rather than directly transferring the scheduling mechanism to a new data modality, we introduce a perturbation-specific intermediate representation, defined by population-level shifts in pretrained single-cell embedding space. We further formulate this intermediate as a stochastic latent variable, explicitly supervise its generation using resampled perturbed-cell cohorts, and condition downstream expression generation on its predicted value. This design connects population-level perturbation responses with individual-cell expression generation. The shuffled and random target ablations (original manuscript Table 4a) validate the importance of perturbation-specific intermediate supervision, the schedule comparisons (original manuscript Table 4b) examine the benefit of the proposed generation order, and the deterministic regression experiment demonstrates the advantage of distributional intermediate modeling over point prediction. Thus, our main contribution lies in the perturbation-specific intermediate representation and its integration into a supervised, distributional generative framework, while the multi-time scheduling principle is adopted from Latent Forcing.

Q2. How does performance change when the Rewire flow is replaced with deterministic regression, retaining the same intermediate supervision, Express module, and decoder?

Thanks for this valuable suggestion. We replace the Rewire flow with a deterministic regression predictor that directly predicts the same intermediate target using an MSE objective and a single forward pass. The other components remain unchanged. scCasFlow achieves better results on all reported metrics. Unlike deterministic regression, which predicts a single intermediate representation, flow matching models the conditional distribution of intermediate responses, providing distributional conditioning for the express stage.

Table 1. Comparison with a deterministic regression predictor on Norman-Additive.

MethodMSE ↓MAE ↓DS ↑PΔ ↑P20 ↑P50 ↑DE LFC ↑Dir ↑O50 ↑
Deterministic Regression Predictor0.0008690.0130290.94380.82390.91520.92360.81040.95530.0742
scCasFlow0.0006830.0115810.96460.84880.92960.93740.81920.95870.1284

Q3. Are the differences between scCasFlow and the alternative schedules stable across training seeds?

We thank the reviewer for the comment. We additionally evaluate the generation schedules using seeds 43 and 44, alongside the original seed 42. scCasFlow achieves improvements for each seed, indicating that its advantage is consistent across the three training runs.

Table 2. Multi-seed evaluation of different generation schedules (mean ± SD across three seeds).

Generation scheduleSeedPearson Δ ↑DE LFC ↑
Expression-leading (α=0.33)420.81590.8072
Expression-leading (α=0.33)430.81930.8159
Expression-leading (α=0.33)440.81610.8101
Expression-leading (α=0.33)Mean ± SD0.8171 ± 0.00190.8111 ± 0.0044
Simultaneous (α=1)420.82940.8158
Simultaneous (α=1)430.83030.8110
Simultaneous (α=1)440.82590.8100
Simultaneous (α=1)Mean ± SD0.8285 ± 0.00230.8123 ± 0.0031
Intermediate-leading (α=5)420.83770.8148
Intermediate-leading (α=5)430.83960.8106
Intermediate-leading (α=5)440.83780.8131
Intermediate-leading (α=5)Mean ± SD0.8384 ± 0.00110.8128 ± 0.0021
scCasFlow420.84880.8192
scCasFlow430.85030.8270
scCasFlow440.85150.8254
scCasFlowMean ± SD0.8502 ± 0.00140.8239 ± 0.0041

Weakness 1. Rewire target variability comes from resampling population means, but the experiments do not establish its benefit over using a fixed population-mean target for each perturbation.

Thanks for the thoughtful comment. We compare cohort resampling with fixed population-mean target for each perturbation, keeping other components unchanged. Our approach performs better than the fixed mean variant, supporting the use of resampled intermediate targets rather than a single population-mean target.

Table 3. Fixed vs. resampled intermediate targets on Norman-Additive.

MethodMSE ↓MAE ↓DS ↑PΔ ↑P20 ↑P50 ↑DE LFC ↑Dir ↑O50 ↑
Fixed mean0.0009420.0131920.95210.82610.91770.92420.80950.95560.0774
scCasFlow0.0006830.0115810.96460.84880.92960.93740.81920.95870.1284

Weakness 2. Training-target ablations change both supervision and conditioning, leaving Express’s reliance on the generated Rewire representation at inference insufficiently tested.

We appreciate this helpful suggestion. We conduct three inference time ablations by replacing the generated Rewire representation with zeros, random values, or representations shuffled across perturbations. Each intervention degrades performance relative to scCasFlow. Shuffling causes the largest drop in DS, indicating that perturbation-specific alignment of the rewire representation is important for the express stage.

Table 4. Inference-time Rewire ablations on Norman-Additive.

MethodMSE ↓MAE ↓DS ↑PΔ ↑P20 ↑P50 ↑DE LFC ↑Dir ↑O50 ↑
Random0.0012250.0151210.93960.78310.90530.91510.79840.93660.1155
Zero0.0008320.0128070.95940.80460.91990.92830.80810.94390.1135
Shuffled0.0008410.0125750.92920.82540.91690.92750.81030.94940.1181
scCasFlow0.0006830.0115810.96460.84880.92960.93740.81920.95870.1284
Reviewer_ZARBOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces scCasFlow, a cascaded flow-matching model for single-cell perturbation response prediction. It first predicts an intermediate response representation supervised by population-level shifts in frozen Stack embeddings, then generates expression latents conditioned on that representation. A shared vector-field network implements the two stages, and a jointly optimized decoder maps the generated latents to gene expression. Experiments on Norman and ComboSciPlex report improvements over evaluated competitors, with architectural, intermediate-target, and temporal-schedule ablations supporting the proposed design.

Critical strengths
  1. Substantial improvements across complementary predictive metrics. On Norman-Additive, scCasFlow achieves MSE 0.000683 and Pearson Δ 0.8488, compared with 0.001156 and 0.7480 for Stack (Table 1). ComboSciPlex also shows substantial improvements over the evaluated methods (Table 3). Agreement across expression error, perturbation-effect correlations, and differential-expression metrics supports practical predictive value.

  2. Useful controls investigate the intermediate representation and generation order. The monolithic variant retains the pretrained tokenizer, while shuffled/random intermediate targets and matched training/inference schedules examine additional explanations for the gains (Figure 4; Table 4). These controls support benefits beyond tokenizer reuse alone. The biological analyses are appropriately qualified, including acknowledgment that expression-derived TF associations do not establish causal regulatory recovery.

Critical weaknesses
  1. Missing key baselines. An additive baseline is essential for Norman-Additive, where all constituent single-gene perturbations are observed during training. A deterministic predictor of the same supervised intermediate would also clarify whether stochastic Rewire generation is necessary; the existing “Deterministic Rewire” ablation does not test this.

  2. Limited evidence for distributional fidelity and robustness. Most metrics assess condition-level averages, leaving cellular heterogeneity insufficiently evaluated. Missing repeated-seed uncertainty weakens the schedule comparisons, while mixed held-out single-perturbation results (Table S2) limit broad generalization claims.

Questions for authors
  1. Does scCasFlow outperform additive and linear baselines under exactly the same splits and preprocessing?

  2. How does a supervised deterministic intermediate predictor compare with the Rewire flow?

  3. Do generated cells reproduce within-condition distributions and heterogeneity?

  4. Are the main improvements, especially generation-order gains, stable across seeds? How were validation data used to select cohort size and other hyperparameters?

4 comments

We thank the reviewer for the constructive feedback and helpful suggestions.

Q1. Does scCasFlow outperform additive and linear baselines under exactly the same splits and preprocessing?

We appreciate the reviewer’s concern. We implement additive and linear baselines following Ahlmann-Eltze et al. [1]. The linear baseline uses 10-dimensional gene embeddings with ridge regression, with double perturbations modeled by summing the embeddings of the two perturbed genes. On Norman-Additive, scCasFlow achieves better MSE, MAE, DE LFC, and directional agreement, while the additive baseline performs better on the other five metrics. Both scCasFlow and the additive baseline outperform the linear baseline on the reported metrics.

Table 1. Additive and linear baselines on Norman-Additive.

MethodMSE ↓MAE ↓DS ↑PΔ ↑P20 ↑P50 ↑DE LFC ↑Dir ↑O50 ↑
Additive0.0009940.0137250.98340.86130.96460.96380.61820.91080.1432
Linear0.0016590.0154790.66180.66900.76540.78290.59310.86040.0316
scCasFlow0.0006830.0115810.96460.84880.92960.93740.81920.95870.1284

Table 2. Linear baseline on Norman-Holdout.

MethodMSE ↓MAE ↓DS ↑PΔ ↑P20 ↑P50 ↑DE LFC ↑Dir ↑O50 ↑
Linear0.0021250.0161720.57890.48460.52350.53690.42350.76010.0137
scCasFlow0.0013260.0142010.74360.60540.71750.70780.66710.84620.0800

Q2. How does a supervised deterministic intermediate predictor compare with the Rewire flow?

We appreciate this helpful suggestion. We replace the Rewire flow with a deterministic regression predictor that directly predicts the same intermediate target using an MSE objective and a single forward pass. The other components remain unchanged. scCasFlow outperforms the deterministic regression predictor on all nine reported metrics, supporting distributional intermediate modeling over deterministic point prediction.

Table 3. Comparison with a deterministic regression predictor on Norman-Additive.

MethodMSE ↓MAE ↓DS ↑PΔ ↑P20 ↑P50 ↑DE LFC ↑Dir ↑O50 ↑
Deterministic Regression Predictor0.0008690.0130290.94380.82390.91520.92360.81040.95530.0742
scCasFlow0.0006830.0115810.96460.84880.92960.93740.81920.95870.1284

Q3. Do generated cells reproduce within-condition distributions and heterogeneity?

We thank the reviewer for raising this question. We evaluate within-condition distributional fidelity using E-distance following scPerturb [2] and unbiased MMD² [3]. For each test condition, we sampled 64 generated cells and 64 observed cells, computed both metrics separately for each condition, and averaged across conditions. The procedure was repeated 20 times, with results reported as mean ± SD. scCasFlow achieves the lowest E-distance and MMD², indicating closer agreement between generated and real cell distributions.

Table 4. Within-condition distributional evaluation on Norman-Additive.

MethodE-distance ↓MMD² ↓
CONTROL47.540 ± 1.5000.29870 ± 0.00408
CPA24.863 ± 0.7400.17505 ± 0.00246
GEARS20.570 ± 0.7750.24688 ± 0.00295
scGPT33.040 ± 0.9490.27650 ± 0.00354
CellFlow17.800 ± 0.9150.05898 ± 0.00201
STATE29.833 ± 1.2590.06792 ± 0.00253
scDFM11.491 ± 0.4390.05698 ± 0.00127
Stack7.559 ± 0.4080.12788 ± 0.00139
scCasFlow5.565 ± 0.5630.03156 ± 0.00157

Q4. Are the main improvements, especially generation-order gains, stable across seeds? How were validation data used to select cohort size and other hyperparameters?

We thank the reviewer for this question. Our original experiments used seed 42. We additionally conduct experiments using seeds 43 and 44. scCasFlow achieves the highest performance for each seed, indicating that its advantage over the alternative schedules is consistent across the three training runs. Hyperparameters were selected through ablation experiments on the Norman-Additive validation set, using Pearson Δ as the selection criterion.

Table 5. Multi-seed evaluation of different generation schedules.

VariantSeedPearson Δ ↑DE LFC ↑
Expression-leading (α=0.33)420.81590.8072
Expression-leading (α=0.33)430.81930.8159
Expression-leading (α=0.33)440.81610.8101
Expression-leading (α=0.33)Mean ± SD0.8171 ± 0.00190.8111 ± 0.0044
Simultaneous (α=1)420.82940.8158
Simultaneous (α=1)430.83030.8110
Simultaneous (α=1)440.82590.8100
Simultaneous (α=1)Mean ± SD0.8285 ± 0.00230.8123 ± 0.0031
Intermediate-leading (α=5)420.83770.8148
Intermediate-leading (α=5)430.83960.8106
Intermediate-leading (α=5)440.83780.8131
Intermediate-leading (α=5)Mean ± SD0.8384 ± 0.00110.8128 ± 0.0021
scCasFlow420.84880.8192
scCasFlow430.85030.8270
scCasFlow440.85150.8254
scCasFlowMean ± SD0.8502 ± 0.00140.8239 ± 0.0041

Reference

[1] Ahlmann-Eltze et al. Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Nature Methods, 2025.

[2] Peidli et al. scPerturb: harmonized single-cell perturbation data. Nature Methods, 2024.

[3] Gretton et al. A kernel two-sample test. Journal of Machine Learning Research, 2012.

Thanks for the rebuttal, your response addresses my concerns, and I have raised my score to 3.

Thank you for your careful review. We are glad that the response has addressed your concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.