acceptodds
Under review as a conference paper at ICLR 2027

Test-Time Co-Adaptation as Data Assimilation

Abstract

Test-time adaptation under corruption mainly follows two directions: directly adapting models to corrupted inputs, or restoring inputs before prediction. These two routes are complementary, yet simply combining them does not necessarily outperform model adaptation alone, since corrupted inputs may lose discriminative evidence while restoration can introduce biased or misleading cues. Inspired by data assimilation, we treat the corrupted input as an observation and its diffusion-restored counterpart as a prior estimate, and propose *test-time co-adaptation* (**CoDA**) to address this issue. CoDA regulates cross-view interaction through cross-view gating for reliable updates, selective Jensen–Shannon coupling for consistent samples, and entropy-aware logit fusion for prediction, while leaving the underlying TTA objective unchanged. On ImageNet-C at severity 5, across five TTA methods and five backbone architectures, CoDA improves independent data–model adaptation by 2.8 percentage points on average and up to 4.4 points, reaching **54.7%** accuracy with ZeroSiam. The gains persist under label shift with batch size one and continual adaptation, and extend to natural rendition shifts on ImageNet-R and geometry-aware corruptions on ImageNet-3DCC, showing that CoDA is effective beyond standard 2D corruptions and generalizes to broader natural and structured distribution shifts.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Test-Time Adaptation, Data Assimilation, Diffusion Restoration

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

This paper proposes CoDA, a test-time adaptation framework that jointly adapts classifiers to corrupted inputs and their restored counterparts. It uses cross-view gating, selective prediction consistency, and entropy-aware fusion to control the interaction between the two branches.

Critical strengths
  1. The idea of jointly using corrupted and restored views while controlling unreliable cross-view interaction is intuitive.
  2. The experiments cover multiple TTA methods, backbone architectures, and evaluation protocols, with additional tests on alternative restoration models and distribution shifts.
Critical weaknesses
  1. The reliability of cross-view agreement and entropy is not sufficiently analyzed. Both branches are initialized from the same classifier, and the restored image is generated from the corrupted input, so they may share incorrect predictions. Lower entropy also does not necessarily indicate better restoration or more accurate classification.

  2. The comparison with closely related diffusion-based adaptation methods is limited. Diffusion-Calibrated Continual Test-Time Adaptation and Collaborative Model and Data Adaptation are discussed in Related Work, but are not directly evaluated.

  3. The corruption experiments only consider severity level 5. It remains unclear whether the same strategy is beneficial under milder corruptions, where restoration may alter useful details in otherwise reliable inputs.

  4. The computational cost remains substantial despite the small additional interaction overhead. Table 3 mainly compares runtime with restoration-based DDA and SDA, rather than model-only TTA.

Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Part 1: Response to Weakness 1

We sincerely appreciate your thoughtful comments and your recognition of CoDA’s intuitive design and our extensive experiments across different methods and settings. We have carefully considered all the points you raised and have addressed each of your concerns in detail below.


W1: We agree that neither cross-view agreement nor lower entropy necessarily guarantees correct predictions. CoDA does not rely on these signals alone. Instead, cross-view gating applies secondary filtering to samples already selected for updates by the underlying TTA method, while consistency regularization is restricted to samples passing both gates. Entropy serves as a relative uncertainty indicator rather than a guarantee of correctness. To further validate this design, we added Table 5, showing that the accuracy of samples selected for model updates increases from 45.89% to 56.53% for ZeroSiam and from 55.43% to 62.61% for SAR. These results demonstrate that our mechanism favors more reliable updates, reducing the risk of reinforcing shared errors. The additions are marked in blue in the revised manuscript (L396–L411). The results are shown below:

BackboneZeroSiamSAR
w/o Gatew/ Gatew/o Gatew/ Gate
ResNet5042.7853.2558.7569.62
ViT-B57.8866.6661.9468.42
ViT-S47.0857.5551.3057.15
Conv-T41.5753.9360.1066.20
Swin-T40.1351.2745.0651.66
Avg.45.8956.5355.4362.61
AuthorsOpenReview ↗

Part 2: Response to Weakness 2

W2: We agree that comparisons with CMDA and diffusion-calibrated continual test-time adaptation (DCTA) are important for a more complete evaluation of CoDA. Following your suggestion, we have added the following comparisons to the revised manuscript.

Comparison with CMDA: We have integrated CoDA with CMDA and added a comparison between CMDA and CMDA+CoDA in Section 5.1, Table 4. The results show that incorporating CoDA improves all three evaluated backbones, demonstrating that the proposed cross-view interaction can further enhance an existing joint data–model adaptation framework. These results support the applicability of CoDA beyond the default DDA-based restoration setting. These additions are marked in blue in the revised manuscript (L336–L345, L357–L362). The results are shown below:

MethodGaussShotImputDefocGlassMotionZoomFrostSnowFogBritContrElasticPixelJPEGAvg
ResNet50
source6.17.56.714.37.611.821.516.216.121.455.13.614.533.342.118.5
CMDA51.546.546.518.119.322.229.728.234.216.258.11.243.852.854.634.9
  +CoDA53.150.449.823.428.037.646.045.944.554.665.123.256.459.056.946.3
SwinT
source28.827.127.222.19.523.627.447.136.046.868.034.021.426.950.233.1
CMDA54.950.150.334.937.140.339.743.238.643.664.535.249.558.056.446.4
  +CoDA53.851.550.933.737.641.643.748.247.451.470.441.050.756.257.649.0
ConvT
source40.139.138.725.611.433.031.249.343.841.970.345.022.541.057.239.3
CMDA59.755.454.944.344.849.546.450.643.342.470.443.853.362.563.253.2
  +CoDA60.057.857.945.346.351.553.157.457.461.472.456.859.964.865.057.8

Comparison with DCTA: We have added a comparison with DCTA in Appendix D.2, Table 19. As DCTA's code is not yet publicly available, we integrate CoDA with REM, a baseline used in DCTA, to provide a fair comparison using the same source checkpoint. Meanwhile, we are actively contacting the authors for further implementation details. The additions are marked in blue in the revised manuscript (L1260–L1269). The results are shown below:

MethodGaussShotImputDefocGlassMotionZoomFrostSnowFogBritContrElasticPixelJPEGAvg
source47.048.247.931.521.241.536.745.850.142.373.68.642.562.063.844.2
Rem56.561.960.846.851.056.557.264.862.564.676.853.258.471.169.860.8
CTTA57.859.361.851.748.859.056.766.065.163.779.855.158.969.871.261.7
CODA(Rem)66.364.564.255.058.161.461.965.164.867.477.055.968.874.271.765.1
AuthorsOpenReview ↗

Part 3: Response to Weakness 3, 4

W3: Thank you for your constructive suggestion. We have extended our experiments to ImageNet-C at severity 3 using DPM-Solver++, which accelerates restoration with fewer denoising steps but may compromise restoration quality and introduce less reliable prior information. Despite this, CoDA still achieves performance gains, demonstrating its effectiveness under milder corruptions. The additional experiments are marked in blue in Appendix C.4, Table 17 (L1198–L1221). The results are shown below:

MethodResNet50ViT-BViT-SConvTSwinTAvg.
EATA57.168.657.658.953.559.1
+DDA/CoDA58.0 / 59.568.4 / 69.659.6 / 61.259.8 / 60.756.0 / 57.160.4 / 61.6
SAR54.362.251.158.951.855.7
+DDA/CoDA56.6 / 57.962.6 / 64.554.7 / 56.658.6 / 60.254.5 / 55.857.4 / 59.0
DeYO57.269.858.659.154.459.8
+DDA/CoDA58.3 / 59.767.4 / 70.560.0 / 62.059.7 / 61.156.0 / 57.560.3 / 62.2
COME53.569.858.560.457.960.0
+DDA/CoDA55.2 / 55.966.5 / 70.659.6 / 61.961.1 / 62.158.8 / 59.960.2 / 62.1
ZeroSiam59.368.559.859.754.860.4
+DDA/CoDA60.2 / 61.967.9 / 69.461.3 / 62.860.4 / 61.656.6 / 57.961.3 / 62.7

W4: We agree that the computational cost deserves further clarification. We have added a direct comparison with model-only TTA, confirming that image restoration dominates the total runtime (9,854 ms out of 9,941 ms). Compared with DDA+ZeroSiam, CoDA introduces only 17 ms additional latency and 12 MB additional GPU memory usage. To reduce restoration costs, we further evaluated AdaIR, a non-diffusion restoration model, which substantially reduces computational cost while maintaining performance improvements. We have also explicitly discussed this limitation in the Limitations section. The additional results are presented in Tables 22 and 23 and marked in blue in the revised manuscript (L1374–L1403). The results are shown below:

MethodTotal (ms)Restore (ms)Adapt (ms)Memory (MB)
ZeroSiam32––966
DDA+ZeroSiam992498547016163
CoDA+ZeroSiam994198548716175
MethodBatchTotal (ms)Restore (ms)Adapt (ms)Memory (MB)
AdaIR16738668706892
AdaIR32138612998713665

We hope that our responses and the additional experimental results in the revised manuscript have provided sufficient clarification and addressed your concerns. We sincerely appreciate your valuable feedback and welcome any further questions or suggestions!

AuthorsOpenReview ↗

Supplementary Experimental Results for W1

To provide further empirical evidence regarding W1, we have conducted an additional stress test on CoDA+SAR to evaluate the robustness of Selective Coupling (SC) to shared incorrect predictions.

Specifically, we randomly discard 50% of the correctly agreed samples while retaining all incorrectly agreed samples when computing the SC loss. Ground-truth labels are used only for this diagnostic experiment. The addition results are marked in blue in the revised manuscript (Appendix E.6, Table 25, L1418–L1440). The results are shown below:

BackboneOriginal CoDAStress Test
ResNet5041.4040.92
ViT-B59.4058.87
ViT-S47.3046.62
ConvNeXt-T44.6044.12
Swin-T40.4040.10

Despite the increased proportion of shared incorrect predictions in SC, accuracy decreases by only 0.30–0.68 percentage points across all five backbones. Together with the improved sample selection accuracy reported in Table 5, these results provide further empirical evidence of the robustness of CoDA's cross-view interaction mechanisms.

We sincerely hope that these additional results help address your concerns, and we greatly appreciate your careful consideration of our work.

Reviewer_h6QoOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CoDA, a test-time co-adaptation framework that combines complementary information from corrupted images and their diffusion-restored counterparts by adapting a separate model branch to each view. The framework comprises three modules: cross-view update gating, selective Jensen–Shannon consistency regularization, and entropy-aware prediction fusion. The authors conduct experiments on ImageNet-C using five TTA methods and five backbone architectures, evaluating regular adaptation, label shift, small-batch settings, and continual adaptation.

Critical strengths
  • The paper studies a useful question: how corrupted observations and diffusion-restored views can complement each other during test-time adaptation.

  • CoDA works with existing TTA methods without additional offline training or changes to the backbone architecture.

  • The experiments cover multiple TTA methods, backbones, and distribution shifts, with consistent average gains over independent dual-branch adaptation.

Critical weaknesses
  • The proposed framework mainly combines several existing techniques, including prediction filtering, consistency regularization, and confidence-based fusion. The authors should provide more direct comparisons with closely related methods, such as Diffusion-calibrated CTTA and CMDA, to better demonstrate the unique contribution of CoDA beyond existing combinations of these components.

  • In the submitted code, the observation branch update strategy, JS loss formulation, and output fusion strategy appear to differ from the formulations described in the paper. The authors should clarify these discrepancies and specify which implementation was used to obtain the reported experimental results.

  • Agreement between the two branches does not necessarily imply correct predictions, and the restoration process may sometimes increase the confidence of incorrect predictions.

  • Report more comprehensive efficiency measurements, including the timing protocol, batch size, end-to-end inference latency, and memory consumption.

Questions for authors

Please refer to the weaknesses above.

  • How does the data-assimilation perspective specifically guide the design of the gating, consistency regularization, and fusion rules? Furthermore, since the restored image is generated from the corrupted input, both branches may agree on an incorrect class.
  • How does the method identify or mitigate such shared errors and prevent consistency regularization from reinforcing incorrect predictions?
6 comments
AuthorsOpenReview ↗

Part 1: Response to Weakness 1, 2

We sincerely appreciate your constructive feedback and your recognition of the value of our research question and CoDA’s compatibility with existing TTA methods. We have carefully considered all the points you raised and have addressed each of your concerns in detail below.


W1: We agree that comparisons with CMDA and diffusion-calibrated continual test-time adaptation (DCTA) are important for a more complete evaluation of CoDA. Following your suggestion, we have added the following comparisons to the revised manuscript.

Comparison with CMDA: We have integrated CoDA with CMDA and added a comparison between CMDA and CMDA+CoDA in Section 5.1, Table 4. The results show that incorporating CoDA improves all three evaluated backbones, demonstrating that the proposed cross-view interaction can further enhance an existing joint data–model adaptation framework. These results support the applicability of CoDA beyond the default DDA-based restoration setting. These additions are marked in blue in the revised manuscript (L336–L345, L357–L362). The results are shown below:

MethodGaussShotImputDefocGlassMotionZoomFrostSnowFogBritContrElasticPixelJPEGAvg
ResNet50
source6.17.56.714.37.611.821.516.216.121.455.13.614.533.342.118.5
CMDA51.546.546.518.119.322.229.728.234.216.258.11.243.852.854.634.9
  +CoDA53.150.449.823.428.037.646.045.944.554.665.123.256.459.056.946.3
SwinT
source28.827.127.222.19.523.627.447.136.046.868.034.021.426.950.233.1
CMDA54.950.150.334.937.140.339.743.238.643.664.535.249.558.056.446.4
  +CoDA53.851.550.933.737.641.643.748.247.451.470.441.050.756.257.649.0
ConvT
source40.139.138.725.611.433.031.249.343.841.970.345.022.541.057.239.3
CMDA59.755.454.944.344.849.546.450.643.342.470.443.853.362.563.253.2
  +CoDA60.057.857.945.346.351.553.157.457.461.472.456.859.964.865.057.8

Comparison with DCTA: We have added a comparison with DCTA in Appendix D.2, Table 19. As DCTA's code is not yet publicly available, we integrate CoDA with REM, a baseline used in DCTA, to provide a fair comparison using the same source checkpoint. Meanwhile, we are actively contacting the authors for further implementation details. The additions are marked in blue in the revised manuscript (L1260–L1269). The results are shown below:

MethodGaussShotImputDefocGlassMotionZoomFrostSnowFogBritContrElasticPixelJPEGAvg
source47.048.247.931.521.241.536.745.850.142.373.68.642.562.063.844.2
Rem56.561.960.846.851.056.557.264.862.564.676.853.258.471.169.860.8
CTTA57.859.361.851.748.859.056.766.065.163.779.855.158.969.871.261.7
CODA(Rem)66.364.564.255.058.161.461.965.164.867.477.055.968.874.271.765.1

W2: Thank you for your careful observation. The previous code included some unused experimental strategies. All reported results follow the methodology described in the paper. We have updated the code accordingly, and all hyperparameters are explicitly documented in the manuscript to ensure reproducibility.

AuthorsOpenReview ↗

Part 2: Response to Weakness 3, 4 and Question 1, 2

W3: CoDA does not introduce additional samples for model updates but applies secondary filtering to those already selected for updates by the underlying TTA method, thereby reducing the risk of reinforcing incorrect predictions. To further validate this, we have added Table 5, showing that the accuracy of samples selected for model updates increases from 45.89% to 56.53% for ZeroSiam and from 55.43% to 62.61% for SAR. These additions are marked in blue in the revised manuscript (L396–L411). The results are shown below:

BackboneZeroSiamSAR
w/o Gatew/ Gatew/o Gatew/ Gate
ResNet5042.7853.2558.7569.62
ViT-B57.8866.6661.9468.42
ViT-S47.0857.5551.3057.15
Conv-T41.5753.9360.1066.20
Swin-T40.1351.2745.0651.66
Avg.45.8956.5355.4362.61

W4: We agree that a more detailed efficiency analysis is important. We have expanded the evaluation to include the timing protocol, batch size, latency, and GPU memory usage, with measurements conducted on an NVIDIA RTX 4090 GPU. The computational cost is primarily dominated by image restoration. Compared with DDA+ZeroSiam, CoDA introduces only 17 ms additional latency and 12 MB additional GPU memory usage at batch size 4.

MethodTotal (ms)Restore (ms)Adapt (ms)Memory (MB)
DDA+ZeroSiam992498547016163
CoDA+ZeroSiam994198548716175

To address this bottleneck, we further evaluated AdaIR, a non-diffusion restoration model, substantially reducing restoration costs while maintaining performance gains. CoDA with AdaIR substantially reduces computational cost while improving accuracy by 2.19 and 2.91 percentage points on ViT-S and ConvNeXt-T (Appendix C.1, Table 15). We have also explicitly discussed this computational limitation in the Limitations section. The additional efficiency results are presented in Tables 22 and 23 and marked in blue in the revised manuscript (L1374–L1403). The results are shown below:

MethodBatchTotal (ms)Restore (ms)Adapt (ms)Memory (MB)
CoDA+AdaIR16738668706892
CoDA+AdaIR32138612998713665

Question

Q1: The data-assimilation perspective guides CoDA by emphasizing the reconciliation of observations with prior information rather than treating either as inherently reliable. The corrupted input preserves sample-specific evidence, while its restored counterpart provides prior-informed information but may introduce semantic distortions. This motivates three design choices.

First, cross-view gating uses complementary predictions to validate updates, while its asymmetric design prevents unreliable restored priors from over-constraining the observation branch. Second, selective consistency regularization restricts alignment to mutually supported predictions, avoiding forced agreement between conflicting views. Third, entropy-aware fusion balances their contributions based on predictive uncertainty and restoration-induced entropy changes. These mechanisms follow a common principle: prior information should complement, rather than override, the original observation.

Q2: Cross-view agreement alone does not guarantee correct predictions. To mitigate shared errors, CoDA applies secondary filtering to samples already selected for model updates by the underlying TTA method, rather than introducing additional update samples. Consistency regularization is further restricted to mutually supported predictions, reducing the risk of reinforcing incorrect predictions. The effectiveness of this design is supported by the improved accuracy of selected update samples, as detailed in our response to W3.


We sincerely hope that our detailed responses and the revisions made to the manuscript have satisfactorily addressed your concerns. We thank you again for your constructive suggestions and welcome any further questions or comments!

Thank you for the detailed response, additional experiments, and code updates. These have partially addressed my concerns. However, I would appreciate further clarification on two points:

W1: Code and Experimental Settings. Most of the implementation inconsistencies I previously raised appear to have been corrected in the updated code. However, the main program and run script still seem to use shuffle=False by default, and the models appear to be reloaded at the beginning of each corruption. This seems inconsistent with the experimental protocols described in the paper. Could the authors clarify the actual settings used for Table 1 and Table 8?

W2: Shared Incorrect Predictions. The additional experiments on sample selection are helpful. However, when both branches agree on an incorrect class, could the JS consistency regularization still reinforce these incorrect predictions? I would appreciate further clarification on this point.

Reviewer_vuaEOpenReview ↗
Rating
4(+1)
Confidence
4(+1)
Informativeness
3(+1)
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The authors introduce CoDA, a TTA method to combine both model adaptation and input restoration. The main foundation for CoDA, is that adjusting the input and the model separately can lead to worsened performance. CoDA uses cross-view gating, selective coupling with Jensen-Shannon divergence, and entropy-aware prediction fusion. CoDA delivers promising results when used with different models, model adaptation methods, and datasets. The papers presents results on different kind of distribution shifts, continual TTA, and batch size of one.

Critical strengths
  1. The paper presents an easy to understand, but very effective method to increase the performance of existing TTA methods. The results show significant improvements over the baseline for most experiments, DDA, as well as improving over just performing model adaptation.

  2. The authors performed a wide variety of experiments and present results on different kind of distribution shifts, continual TTA, and batch size of one. CoDA seems to transfer to a wide variety of experiment settings.

Critical weaknesses
  1. The paper left out some close baselines, such as CMDA (Collaborative Model and Data Adaptation at Test Time) or Diffusion-calibrated Continual Test-time Adaptation. It should explicitly be noted how CoDA performs against these models, instead of using naive DDA as a baseline.

  2. When looking at the breakdown per corruption type, it seems that CoDA often worsen performance compared to using just the model adaptation method. For example ZeroSiam on ImageNet-C in Table 9 ViT-B (the main baseline in the paper) worsen in 8/15 corruptions and only has marginal gains in methods that aren't noise corruptions. There also are no confidence intervals on the results, which make it hard to interpret how consistent the results are.

  3. The paper is evaluating against few TTA methods. Especially some simple TTA methods like TENT, NEO, or PEA, which are minimal and it's not clear whether they would benefit from input restoration are left out.

Questions for authors
  1. Why did you use DDA as the main baseline, and aren't evaluating against CMDA or Diffusion-calibrated Continual Test-time Adaptation? Is it possible to add those as baselines?

  2. Why does CoDA seem to perform worse than just model adaptation in many cases? Why is it so strong on noises? Is it possible to add confidence intervals?

  3. Is it possible to add more model adaptations such as TENT, NEO, or PEA?

4 comments
AuthorsOpenReview ↗

Part 1: Response to Weakness 1

We sincerely appreciate your constructive comments and your recognition of our method’s simplicity and effectiveness, as well as our extensive experimental evaluation. We have carefully considered all the points you raised and have addressed each of your concerns in detail below.


W1: We agree that comparisons with CMDA and diffusion-calibrated continual test-time adaptation (DCTA) are important for a more complete evaluation of CoDA. Following your suggestion, we have added the following comparisons to the revised manuscript.

Comparison with CMDA: We have integrated CoDA with CMDA and added a comparison between CMDA and CMDA+CoDA in Section 5.1, Table 4. The results show that incorporating CoDA improves all three evaluated backbones, demonstrating that the proposed cross-view interaction can further enhance an existing joint data–model adaptation framework. These results support the applicability of CoDA beyond the default DDA-based restoration setting. These additions are marked in blue in the revised manuscript (L336–L345, L357–L362). The results are shown below:

MethodGaussShotImputDefocGlassMotionZoomFrostSnowFogBritContrElasticPixelJPEGAvg
ResNet50
source6.17.56.714.37.611.821.516.216.121.455.13.614.533.342.118.5
CMDA51.546.546.518.119.322.229.728.234.216.258.11.243.852.854.634.9
  +CoDA53.150.449.823.428.037.646.045.944.554.665.123.256.459.056.946.3
SwinT
source28.827.127.222.19.523.627.447.136.046.868.034.021.426.950.233.1
CMDA54.950.150.334.937.140.339.743.238.643.664.535.249.558.056.446.4
  +CoDA53.851.550.933.737.641.643.748.247.451.470.441.050.756.257.649.0
ConvT
source40.139.138.725.611.433.031.249.343.841.970.345.022.541.057.239.3
CMDA59.755.454.944.344.849.546.450.643.342.470.443.853.362.563.253.2
  +CoDA60.057.857.945.346.351.553.157.457.461.472.456.859.964.865.057.8

Comparison with DCTA: We have added a comparison with DCTA in Appendix D.2, Table 19. As DCTA's code is not yet publicly available, we integrate CoDA with REM, a baseline used in DCTA, to provide a fair comparison using the same source checkpoint. Meanwhile, we are actively contacting the authors for further implementation details. The additions are marked in blue in the revised manuscript (L1260–L1269). The results are shown below:

MethodGaussShotImputDefocGlassMotionZoomFrostSnowFogBritContrElasticPixelJPEGAvg
source47.048.247.931.521.241.536.745.850.142.373.68.642.562.063.844.2
Rem56.561.960.846.851.056.557.264.862.564.676.853.258.471.169.860.8
CTTA57.859.361.851.748.859.056.766.065.163.779.855.158.969.871.261.7
CODA(Rem)66.364.564.255.058.161.461.965.164.867.477.055.968.874.271.765.1
AuthorsOpenReview ↗

Part 2: Response to Weakness 2, 3 and Question1, 2, 3

W2: The performance differences across corruption types are largely related to the varying restoration quality of the default DDA pipeline. While diffusion-based restoration offers advantages for noise corruptions, its effectiveness on other corruption types can limit the benefits of co-adaptation. Unlike DDA, SDA, and CMDA, which exhibit performance degradation even below the original baseline on certain corruptions, CoDA consistently improves upon independent data–model adaptation, with only slightly reduced gains on some corruption types due to restoration quality limitations.

Our additional experiments with CMDA(response to W1) further support this explanation: by incorporating model-guided restoration to improve the quality of restored views, CoDA achieves substantial gains on previously challenging corruptions and consistent improvements across all corruption categories.

Regarding statistical reliability, we additionally evaluated CoDA with ZeroSiam using five random seeds across five backbones. The standard deviations range from 0.12 to 0.23 percentage points, indicating stable performance. These additions are marked in blue in the revised manuscript (L336–L362, L1404–L1416). The results are shown below:

MethodResNet50-GNViT-BaseViT-SmallConvNeXt-TinySwin-Tiny
CoDA+ZeroSiam52.7 ± 0.1664.3 ± 0.1256.0 ± 0.2351.3 ± 0.1449.4 ± 0.18

W3: We agree that broader TTA evaluation is important, and we have extended evaluation to include TENT, NEO, and PEA, with the results presented in Appendix D.1, Table 18. For TENT, CoDA improves average accuracy from 38.4% to 43.1%. For the optimization-free methods NEO and PEA, we retain their training-free nature by applying Cross-View Gating and Entropy-Aware Fusion without JS coupling. These results confirm that optimization-free TTA methods benefit from input restoration, with further improvements achieved through controlled cross-view interaction. The additions are marked in blue in the revised manuscript (L1226–L1236, L1242–L1256), and the results are shown below:

MethodResNet50ViT-BViT-SConvTSwinTAvg
Source30.629.923.034.729.129.5
DDA38.139.934.041.635.437.8
Tent24.946.533.736.425.933.5
DDA+Tent33.646.135.442.234.638.4
CoDA+Tent38.155.540.143.638.443.1
NEO33.633.226.637.934.633.2
DDA+NEO42.245.040.246.643.343.5
CoDA+NEO43.246.341.447.644.044.5
PEA46.239.631.146.444.641.6
DDA+PEA50.345.740.352.247.647.2
CoDAa+PEA51.248.142.752.448.948.7

Questions

Q1: We adopt DDA as our primary restoration baseline because its modular pipeline allows us to clearly distinguish the contributions of data adaptation, model adaptation, and their combination. By keeping the restoration process and underlying TTA method fixed, we can better isolate the benefits of controlled cross-view interaction. Moreover, CoDA is not limited to DDA, as demonstrated by its extensions to GDA, Restormer, AdaIR, and CMDA. We have also added comparisons with CMDA and DCTA, as detailed in our response to W1.

Q2: The varying gains across corruption types mainly arise from differences in restoration quality. Diffusion-based restoration is particularly effective against noise, while its benefits are more limited for certain other corruptions. Nevertheless, CoDA mitigates the negative effects of unreliable restoration through controlled cross-view interaction. As discussed in W2, our additional CMDA experiments further demonstrate that improved restoration quality leads to more consistent gains across corruption categories. We have also added repeated experiments with five random seeds to assess statistical stability.

Q3: We agree that evaluating simpler TTA methods is important. We have extended our experiments to TENT, NEO, and PEA, as detailed in our response to W3. And We have cited and discussed these methods in Related Work.


We sincerely hope that our responses and the additional evidence provided in the revised manuscript have adequately addressed your concerns. Thank you again for your valuable suggestions. We welcome any further questions or comments!

Reviewer_vuaEOpenReview ↗

Response to the authors

Thank you for the detailed response and additional experiments. The comparisons with CMDA substantially address my concern about closely related baselines, and the added evaluations with TENT, NEO, and PEA address my concern about applicability to simpler TTA methods. The five-seed experiments also provide useful evidence of the stability of CoDA’s average performance. Together, these additions strengthen the empirical contribution.

My concern about per-corruption regressions relative to model-only adaptation remains partially open. The explanation based on restoration quality is plausible, although improvements over independent data–model adaptation do not directly resolve this comparison. I encourage the authors to clearly acknowledge these remaining regressions and distinguish average improvements from improvements on every corruption.

Overall, the additional evidence sufficiently addresses my main concerns and strengthens my assessment of the method’s effectiveness and generality. I am therefore raising my score to reflect stronger support for acceptance.

We greatly appreciate your thoughtful feedback and positive reassessment of our work, particularly your recognition of the additional experiments and their contributions to strengthening the manuscript.

Following your valuable suggestion, we have further clarified and explicitly acknowledged the remaining limitation in the Limitations section of the revised manuscript. Specifically, we clarify that although CoDA improves overall average accuracy, it may still slightly underperform model-only TTA on certain corruption types due to limitations in restoration quality.

Finally, we wish to reiterate our sincere appreciation for your careful review, constructive feedback, and continued support, which have greatly helped us improve 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.