acceptodds
Under review as a conference paper at ICLR 2027

DRIFT: Disentangled Responsive-Invariant Flow Transport for Single-Cell Perturbation Prediction

11 views

Abstract

Predicting cellular responses to perturbations is a central problem in cellular biology, with broad applications in systems biology and drug discovery. This task is challenging because cellular responses can be complex and cell-state dependent, intrinsic cell-to-cell variability can be confounded with perturbation effects, and destructive single-cell RNA sequencing precludes paired measurements of the same cell before and after treatment. Flow matching transports control cells to perturbed states flexibly, but acting on the full cell state can confound perturbation effects with pre-existing cell-to-cell variability. Disentangled approaches separate responsive from invariant components, but model perturbations through prescribed mechanisms, such as latent shifts or graph edits, limiting their flexibility. We address both limitations in a unified framework. A variational encoder disentangles each cell into an invariant block, capturing state unaffected by the perturbation, and a responsive block, capturing state it changes, through conditional priors and an information-theoretic invariance constraint. Conditional flow matching transports only the responsive block, conditioned on the perturbation and invariant state, yielding a flexible, data-driven model of perturbation effects without confounding pre-existing variability. Across several benchmarks, our method outperforms the strongest published method in settings involving combinatorial and unseen perturbation prediction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Perturbation Prediction, single-cell RNA, Representation Learning, Flow Matching, Disentanglement

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_9yWJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the prediction of single-cell perturbation responses, aiming to forecast the expression distribution of control cells under previously unseen perturbations or combinations thereof. The authors propose a two-stage framework called DRIFT. In the first stage, a Variational Autoencoder (VAE)—incorporating a conditional prior and a mutual information penalty—decomposes each cell into an invariant component (znrz_{nr}) and a response component (zrz_r). In the second stage, a flow-matching model, conditioned on znrz_{nr} and the perturbation encoding, is employed to transform only the zrz_r component. During inference, the znrz_{nr} components of the control cells are kept constant, while the zrz_r components are generated.

Critical strengths
  • The approach effectively combines the disentanglement of a "base state" (akin to the CPA model) with the flexible transformation capabilities of flow-based models; this methodology is readily transferable to other architectures.
  • The motivation is intuitive. By separating perturbation-related information from background (contextual) information, the model can specifically focus on the perturbation-driven dynamics via flow matching.
Critical weaknesses
  • The strategy of directly copying znrz_{nr} during inference implicitly assumes invariance at the single-cell level. However, the CLUB objective used only enforces independence at the population level, and the paper provides no empirical evidence to support invariance at the level of individual-cell counterfactuals.
  • The paper lacks a sensitivity analysis for hyperparameters. The actual impact of αnr\alpha_{nr}, αr\alpha_r, and ϵ\epsilon depends on the scale of the latent space—a detail not reported in the paper. Furthermore, the weighting for the perturbation encoding regularization is omitted, and the protocol for hyperparameter selection is not specified.
  • The baselines are drawn from existing literature, yet the model is not compared against scDFM on the Norman dataset.
  • The paper fails to cite scDRP, a framework that is highly similar in design.
Questions for authors

See my Weaknesses.

4 comments

We thank the reviewer for recognizing the method's intuitive design and transferability. We address the concerns raised below.

The strategy of directly copying znrz_{nr} during inference implicitly assumes invariance at the single-cell level. However, the CLUB objective used only enforces independence at the population level, and the paper provides no empirical evidence to support invariance at the level of individual-cell counterfactuals.

We do not claim per-cell counterfactuals, which are not identifiable from unpaired snapshots (Bunne et al., 2023). DRIFT predicts the distribution of perturbed cells, which is what is measured by every metric we report. For this target, copying the invariant block of a control cell requires znrz_{nr} to have the same distribution in control and perturbed cells. This is the population-level independence that CLUB enforces. We will revise the wording in the paper to clarify this aspect.

Empirically, three results support the copy strategy on the Norman additive dataset:

  • znrz_{nr} does not change with the perturbation. In Table 6 of the paper, a classifier recovers the perturbation a cell received from znrz_{nr} only for 4% of cells, against 94 to 96% from zrz_r. The invariant block thus has nearly the same distribution under every perturbation, which is the condition the copy requires.
  • Copying is better than transporting. With the same frozen autoencoder, a flow that also transports znrz_{nr} is outperformed by DRIFT (see the table on transporting both blocks in our response to Reviewer ePSL).
  • A generated cell keeps an attribute of its source cell. In the benchmarks we use, control and perturbed cells are not paired, so a cell's counterfactual cannot be observed and per-cell invariance is hard to assess. A possible check is to assess attributes fixed before the perturbation. One is the gem group. Cells were loaded into 8 separate 10x channels, and each channel leaves a small technical trace in expression. A cell has the same gem group whatever its perturbation. A classifier predicts it from expression, as shown below:
Cells scoredClassifier trained onScored against the gem group ofAccuracy
Real control cells (held out)real control cellsthe cell itself47%
Real perturbed cellsreal control cellsthe cell itself44%
DRIFT reconstructions of held-out control cells (no transport)reconstructions of control cellsthe cell itself41.4 ± 0.3%
DRIFT generated cellsreconstructions of control cellstheir source control cell37.9 ± 0.5%
DRIFT generated cellsreconstructions of control cellsa random control cell (chance)12.8 ± 0.3%

Generated cells carry the gem group of the control cell they were transported from (37.9%, against 12.8% for a random control cell), close to the 41.4% the autoencoder keeps with no transport (5 seeds). This is consistent with the copied znrz_{nr} keeping information on the source cell through the transport.

The paper lacks a sensitivity analysis for hyperparameters. The actual impact of αnr\alpha_{nr}, αr\alpha_r, and ε\varepsilon depends on the scale of the latent space—a detail not reported in the paper. Furthermore, the weighting for the perturbation encoding regularization is omitted, and the protocol for hyperparameter selection is not specified.

Scale. The OT cost is normalized per block, with each term divided by its minibatch mean, so αnr\alpha_{nr}, αr\alpha_r, and ε\varepsilon do not depend on the latent scale. We will add this to the explanation of Eq. (9).

Sensitivity. Norman additive, one value changed at a time over a 16-fold range (defaults αnr=αr=1\alpha_{nr} = \alpha_r = 1, ε=0.5\varepsilon = 0.5). The flow is retrained on the submitted stage-1 model. Results of these new experiments are below:

SettingρΔ ↑ρΔ^D ↑DES ↑PDS ↑MSE ↓
Default (Table 1)0.8870.8910.9020.9170.0057
αnr\alpha_{nr} = 0.25 / 40.886 / 0.8870.871 / 0.8780.900 / 0.9020.915 / 0.9340.0056 / 0.0051
αr\alpha_r = 0.25 / 40.888 / 0.8860.880 / 0.8810.902 / 0.9010.923 / 0.9020.0055 / 0.0059
ε\varepsilon = 0.125 / 20.888 / 0.8880.887 / 0.8830.900 / 0.9020.914 / 0.9330.0056 / 0.0054
scBIG0.8500.8230.8590.8550.0091

Every metric stays within about 0.02 of the default, and DRIFT stays above scBIG on all five metrics for every setting we explore. This agrees with OT-CFM (Tong et al., 2023), on which the pairing we propose builds, where any coupling yields a flow onto the same target distribution.

All settings were selected on the validation set, using several evaluation metrics jointly, such as ρΔ, PDS, and MSE.

For the perturbation-encoder regularization, we use λaux=λiso=10\lambda_{aux} = \lambda_{iso} = 10. These weights are already set in the code attached to the submission and will be added to the appendix (Table 5) with all other settings.

The baselines are drawn from existing literature, yet the model is not compared against scDFM on the Norman dataset.

We follow scDFM's protocol, including the dataset splits, the number of generated cells per condition, and the evaluation metrics, and we reproduce the scDFM numbers from its official released predictions. scDFM's preprocessing keeps 5,029 genes for training, against 2,051 in the scBIG benchmarks. DRIFT keeps the architecture and objective used on the scBIG benchmarks, with hyperparameters selected on a validation split held out from scDFM's training data. Results for these new experiments are in the tables below:

Additive (4 splits)Pearson Δ ↑DE-Sp. ↑DS ↑L2 ↓MSE ↓MAE ↓Δ̂ ↑Δ̂₂₀ ↑
ControlN.A.N.A.0.5143.9710.018230.0394-0.191-0.152
Additive0.9050.6040.9661.9130.004380.02240.8590.925
scDFM0.8850.7740.9741.7040.003150.02150.8430.924
DRIFT0.9290.9140.9831.5840.002950.01940.9030.941
Holdout, single (4 splits)Pearson Δ ↑DE-Sp. ↑DS ↑L2 ↓MSE ↓MAE ↓Δ̂ ↑Δ̂₂₀ ↑
ControlN.A.N.A.0.5232.6880.009530.02630.1550.188
scDFM0.7130.7350.8911.6190.003010.01900.6660.812
DRIFT0.8010.8860.8911.3810.002330.01450.7220.775
Holdout, double (4 splits)Pearson Δ ↑DE-Sp. ↑DS ↑L2 ↓MSE ↓MAE ↓Δ̂ ↑Δ̂₂₀ ↑
ControlN.A.N.A.0.5324.1930.020730.0424-0.206-0.100
scDFM0.8360.7540.9192.0310.004710.02350.7620.860
DRIFT0.8810.9050.9141.7580.003830.01970.8160.859

DRIFT obtains the best performance. In the additive setting it is first on all eight metrics, ahead of scDFM and the additive baseline. On held-out genes it is first on 12 of the 16 metrics, with a large margin on Pearson Δ (0.801 against 0.713 for held-out single genes). The performance of DRIFT therefore also holds under scDFM's protocol.

The paper fails to cite scDRP, a framework that is highly similar in design.

We thank the reviewer for the pointer. The revision will add it to the related work. scDRP is a disentanglement-based method that separates perturbation-responsive from invariant latent variables. It represents each perturbation as a one-hot index, transports cells with an optimal transport plan computed for each measured perturbation, and estimates the effect of a combination by vector addition of the individual treatment effects of each single perturbation. A one-hot encoding does not generalize to unseen genes, and an OT plan computed for each perturbation is less flexible than DRIFT's single flow, conditioned on gene and drug features and shared across perturbations. This allows DRIFT to predict unseen genes and non-additive combinations, which scDRP's one-hot and additive design does not directly support.

Bunne, C., et al. (2023). Learning single-cell perturbation responses using neural optimal transport. Nature Methods, 20(11), 1759-1768.

Tong, A., Fatras, K., Malkin, N., Huguet, G., Zhang, Y., Rector-Brooks, J., ... & Bengio, Y. (2023). Improving and generalizing flow-based generative models with minibatch optimal transport. arXiv preprint arXiv:2302.00482.

Thank you for the additional experiments. I remain unsure which biological features znrz_{\mathrm{nr}} preserves across perturbations, so I lean toward keeping my original rating.

We thank the reviewer for the reply. In short, znrz_{nr} preserves the technical and compositional state of the cell (sequencing depth, ribosomal and mitochondrial content), while the biological programs are encoded mainly in zrz_r and transported with the perturbation.

The initial concern was invariance at the level of individual-cell counterfactuals. As stated in our response, we do not claim per-cell counterfactuals, and they cannot be verified with unpaired data, since each cell is measured once. What can be analysed is which attributes of a cell znrz_{nr} encodes. We already reported the gem group in our response, and we provide further analyses below on the Norman additive dataset.

We use the attributes recorded for each cell in the dataset (total counts, detected genes, mitochondrial and ribosomal fractions, on which the prior of znrz_{nr} is conditioned), and scores computed from expression for the cell cycle and the 38 MSigDB hallmark programs. On held-out control cells, a regressor predicts each attribute from znrz_{nr} or from zrz_r. Results of these new experiments are below:

  • znrz_{nr} encodes the technical and compositional state of the cell, and a generated cell keeps that of its source control cell. The last two columns compare, within each perturbation, the attribute of a generated cell with that of its source control cell (5 seeds):
AttributeR2R^2 from znrz_{nr}Generated vs source control cellGenerated vs random control cell
Total counts0.990.850.00
Detected genes0.960.830.00
Mitochondrial fraction0.980.790.00
Ribosomal fraction0.990.91-0.03
  • Biological programs are encoded in zrz_r. Every program is predicted better from zrz_r than from znrz_{nr} (5 seeds):
ProgramR2R^2 from znrz_{nr}R2R^2 from zrz_r
Cell cycle (G2M score)0.250.84
E2F targets0.250.75
MYC targets0.490.82
Oxidative phosphorylation0.490.77
Heme metabolism (erythroid)0.160.69
p53 pathway0.110.29
All 38 hallmark programs (mean)0.130.37

DRIFT transports these programs with the perturbation. For example, CDKN1C or TP73 increase the fraction of cells in G1 by 28 to 40 points, and on the 32 test combinations, the change in the G1 fraction predicted by DRIFT correlates with the observed change at 0.895 (5 seeds).

These analyses identify the features that znrz_{nr} preserves, and we will add them to the revised paper.

Reviewer_q1THOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a method to predict cellular responses to perturbations by disentangling the latent space of cells into invariant and variant features. The invariant features capture experimental covariates and the variant features capture the response to perturbations.

Critical strengths

The two stage model of invariant and variant features is a good architecture choice for perturbation predictions

Critical weaknesses

The VAE takes in a different number of highly variable genes depending on the dataset. The invariant and variant blocks are trained separately with the disentanglement pretrained and frozen. The training data consists of snapshot gene expression at a single time point.

Questions for authors

How does freezing stage 1 and using it in stage 2 affect the results compared to training both stages together with an invariance loss term added? This might provide a more robust architecture than pre training and freezing the invariance module

1 comment

We thank the reviewer for recognizing the value of separating invariant and responsive features.

How does freezing stage 1 and using it in stage 2 affect the results compared to training both stages together with an invariance loss term added? This might provide a more robust architecture than pre training and freezing the invariance module

Training the representation first and freezing it is a standard practice in latent generative models (e.g., Rombach et al., 2022). With a frozen encoder the flow learns a fixed target, whereas in joint training the latents it must transport move at every step, which makes optimization unstable. To verify this, we trained both stages jointly, keeping the invariance (CLUB) loss, as suggested by the reviewer. The results show that freezing is better and that joint training underperforms.

TrainingρΔ ↑DES ↑PDS ↑MSE ↓
Frozen stage 1 (DRIFT, Table 1)0.88710.90170.91670.0057
Joint0.77510.75210.57560.0315

The VAE takes in a different number of highly variable genes depending on the dataset. The invariant and variant blocks are trained separately with the disentanglement pretrained and frozen. The training data consists of snapshot gene expression at a single time point.

We address each point raised by the reviewer.

Gene counts. The counts differ because we follow each benchmark's own preprocessing, which keeps our numbers comparable with the published baselines. The architecture is the same on every dataset, and its advantage holds from 2,051 genes (Norman, scBIG's preprocessing) to 5,029 (Norman, scDFM's preprocessing, see the scDFM comparison in our response to Reviewer 9yWJ), which shows that DRIFT performs consistently across different gene counts.

Frozen stage 1. Freezing stage 1 is standard practice, and our results above show that it performs better than joint training.

Single time point. Every method and benchmark in the comparison considers the same snapshots, as a perturbation screen measures each cell once. Extending this to multiple time points is an interesting research direction that we leave for future work.

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. (2022). High-resolution image synthesis with latent diffusion models. In CVPR 2022.

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

DRIFT learns invariant and perturbation-responsive latent blocks with a conditional-prior VAE and an invariance penalty. It then uses a conditional flow to transport the responsive block while retaining the control cell's invariant block. The paper evaluates prediction of genetic and chemical perturbations on Norman, Replogle RPE1 and ComboSciPlex.

Critical strengths
  • S1. A well-motivated method with observable predictive gains. Retaining a control-derived state while learning a flexible, state-conditioned response is a reasonable approach to unpaired single-cell perturbation data. The results show gains over the reported scBIG baseline across several metrics: for example, on Norman additive, Pearson Δ improves from 0.8496 to 0.8871 and PDS from 0.8548 to 0.9167. On Replogle RPE1, DRIFT also improves all nine reported metrics for held-out single-gene perturbations.

  • S2. The experimental evaluation is broad and informative. The paper evaluates genetic and drug perturbations, including held-out combinations and unseen single-gene targets, against a range of published methods. It reports multiple response and discrimination metrics, five-seed variability, component and feature ablations, and latent probes.

Critical weaknesses
  • W1. Performance on unseen perturbations remains incompletely characterized. The RPE1 results are encouraging: DRIFT outperforms the reported scBIG baseline on all nine metrics for 384 unseen single-gene conditions, and DES rises from 0.5145 to 0.7021. Nevertheless, Pearson Δ is 0.5085 versus 0.4875 for scBIG, leaving room to improve absolute accuracy and a relatively small gain on this measure. The RPE1 test covers single-gene knockdowns in one cell line. For the harder case of combinations involving unseen genes, the evidence is less specific.
Questions for authors
  • Non-additive genetic interactions. How well does DRIFT predict interactions between two perturbed genes beyond the sum of their individual effects? In the Norman additive split, where each single-gene perturbation is observed in training, could the authors compare the predicted and measured interaction residuals, ΔAB − ΔA − ΔB, using matched controls and training-derived single-gene effects?
  • Selective transport ablation. In the existing “without disentanglement” ablation, the representation, priors and invariance penalty also change when the full latent is transported. Could the authors freeze the same learned encoder and decoder, keep the same conditioning information and training budget, and compare transporting only the responsive block with transporting both blocks?
1 comment

We thank the reviewer for recognizing the method’s well-motivated design, observable gains, and informative experimental evaluation. We address the concerns below.

Performance on unseen perturbations remains incompletely characterized. The RPE1 results are encouraging: DRIFT outperforms the reported scBIG baseline on all nine metrics for 384 unseen single-gene conditions, and DES rises from 0.5145 to 0.7021. Nevertheless, Pearson Δ is 0.5085 versus 0.4875 for scBIG, leaving room to improve absolute accuracy and a relatively small gain on this measure. The RPE1 test covers single-gene knockdowns in one cell line. For the harder case of combinations involving unseen genes, the evidence is less specific.

The gain on ρΔ. In RPE1, most of the response to a knockdown is shared across genes. A prediction that ignores the perturbed gene (control cells plus the mean training response) already reaches ρΔ 0.482 on the 384 unseen genes, close to both scBIG (0.4875) and DRIFT (0.5085). On this benchmark ρΔ therefore mostly measures the shared response, which both methods capture. The gene-specific part of the response is measured by PDS, which asks whether a prediction is closer to the measured response of its own gene than to those of the other genes. There, the gene-agnostic prediction is at chance (0.501), scBIG reaches 0.552, and DRIFT 0.604, twice as far above chance. RPE1 is our main benchmark for unseen genes, and extending the evaluation to other cell lines and to combinations with unseen genes is interesting future work.

How well does DRIFT predict interactions between two perturbed genes beyond the sum of their individual effects? In the Norman additive split, where each single-gene perturbation is observed in training, could the authors compare the predicted and measured interaction residuals, ΔAB − ΔA − ΔB, using matched controls and training-derived single-gene effects?

We ran the comparison as the reviewer suggests, on the Norman additive test split (32 pairs, both single genes of every pair seen in training). ΔA\Delta_A and ΔB\Delta_B are the training single-gene effects, and ΔAB\Delta_{AB} and Δ^AB\hat\Delta_{AB} are the measured and predicted effects of the pair. All are shifts from the same training control mean. The measured interaction is r=ΔAB−ΔA−ΔBr = \Delta_{AB} - \Delta_A - \Delta_B, and the predicted interaction is r^=Δ^AB−ΔA−ΔB\hat r = \hat\Delta_{AB} - \Delta_A - \Delta_B. We compare r^\hat r with rr using our evaluation metrics applied to the interaction, namely ρ (Pearson over all genes), ρ^D (Pearson over the 20 genes with the largest ∣r∣|r|), ACC and ACC^D (fraction of genes, among all and among those 20, where r^\hat r and rr have the same sign), and MSE. For DRIFT we report mean ± sd over 5 seeds. Results of this new experiment are in the table below:

Predicted interactionρ ↑ρ^D ↑ACC ↑ACC^D ↑MSE ↓
None (additive, r^=0\hat r = 0)––––0.0087
Scaled-additive0.3830.3820.6110.7220.0081
DRIFT0.608 ± 0.0080.753 ± 0.0130.731 ± 0.0050.930 ± 0.0160.0053 ± 0.0002

The scaled-additive baseline is Δ^AB=k(ΔA+ΔB)\hat\Delta_{AB} = k(\Delta_A + \Delta_B), with one k=0.76k = 0.76 fitted on training pairs.

The interaction predicted by DRIFT correlates with the measured one at 0.61 over all genes and at 0.75 on the 20 genes with the strongest interaction, where its sign is correct for 93% of the genes. Its interaction error is 39% lower than that of the additive prediction (0.0053 against 0.0087). This shows that DRIFT performs well on complex perturbation effects beyond the additive ones, as it predicts the interaction term well.

In the existing "without disentanglement" ablation, the representation, priors and invariance penalty also change when the full latent is transported. Could the authors freeze the same learned encoder and decoder, keep the same conditioning information and training budget, and compare transporting only the responsive block with transporting both blocks?

Norman additive, scBIG metrics, mean over seeds. The DRIFT row is from Table 1 of the paper (5 seeds), and the flow that transports both blocks is trained on the same frozen autoencoders (3 seeds).

TransportedρΔ ↑ρΔ^D ↑DES ↑PDS ↑MSE ↓L2 ↓
(znr,zr)(z_{nr}, z_r)0.88680.87380.89940.88610.00683.40
zrz_r only (DRIFT)0.88710.89140.90170.91670.00573.12

When znrz_{nr} is also transported, the flow moves the invariant block too. This extra displacement adds error (MSE +19%) and blurs the differences between conditions (PDS −0.03). With zrz_r only, the whole predicted change is the response. These empirical results support the claim that copying znrz_{nr} is better than transporting it, on all the metrics. In particular, DRIFT discriminates better between conditions (PDS) when it transports only zrz_r and keeps znrz_{nr} fixed.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Comment

We thank the reviewers and area chair for their time and work. The reviewers found the method well motivated with observable predictive gains and a broad and informative experimental evaluation (ePSL), the two-stage model of invariant and variant features a good architecture choice (q1TH), and the motivation intuitive and the methodology readily transferable to other architectures (9yWJ). Here we provide a summary of our rebuttal. Every change announced below will be reflected in the revised paper. The main points of our response and the results of the additional experiments are summarized below for each reviewer.

  1. Reviewer(ePSL):
  • Interaction residuals ΔAB−ΔA−ΔB\Delta_{AB} - \Delta_A - \Delta_B: DRIFT predicts their sign on 93% of the most interacting genes.
  • Selective transport with both zrz_r and znrz_{nr}: transporting both blocks loses PDS (0.917 to 0.886) and worsens MSE by 19%.
  1. Reviewer(q1TH):
  • Joint versus frozen training: joint training empirically shown to be unstable (PDS 0.92 to 0.58).
  1. Reviewer(9yWJ):
  • DRIFT predicts distributions and does not claim per-cell counterfactuals. CLUB enforces population-level invariance, which our empirical checks support.
  • DRIFT outperforms scDFM on Norman under scDFM's protocol.
  • Sensitivity over different hyperparameters: empirical results show the method is robust, with every metric within about 0.02 of the default for αnr\alpha_{nr}, αr\alpha_r, and ε\varepsilon. The perturbation-encoder weights we used (λaux=λiso=10\lambda_{aux} = \lambda_{iso} = 10) will be added to the paper.
  • The missing references will be added to the updated version of the paper.
AuthorsOpenReview ↗

Follow-up on our rebuttal

Dear Reviewers, AC,

Thank you once again for your thorough reviews. We believe we have addressed all the questions, requests for clarification, and additional experiments. We believe this additional work has improved our submission, and we thank you for this.

Do you have any additional feedback you’d like us to take into account? If not, based on our rebuttal, could you consider revising your score to reflect the additional evidence we provide?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.