acceptodds
Under review as a conference paper at ICLR 2027

The Batch Size Can Be One: Test-Time Adaptation from Classifier Geometry

Abstract

Test-time adaptation updates a deployed model on an unlabeled test stream. In practice this stream often arrives one sample at a time, with the classes in no particular order. A single sample then provides limited information for estimating distribution-level statistics, and selecting samples by confidence can admit predictions that are wrong yet confident. A corruption, however, shifts the features while leaving the decision regions of the frozen classifier unchanged. We therefore propose Origin-recentered Test-Time Adaptation (OriTTA), which constructs fixed geometric targets from the source model and adapts selected features toward them. Seen from the origin of the feature space, the classes lean toward a common direction. We choose a new origin to bring class scores closer together while keeping it near the mean feature of one synthetic prototype per class. From this origin, prototype directions spread more evenly, and each class obtains an axis and an ideal point before any test data arrives. During adaptation, OriTTA aligns selected features with their predicted class axis and penalizes excessive radius, keeps one input per predicted class in memory to limit feature changes for classes absent from the stream, and resets the model when re-classifying this memory reveals collapse. Without tuning for each setting, OriTTA achieves the highest mean accuracy at batch size 1 on CIFAR-10-C, CIFAR-100-C and ImageNet-C in all sixteen combinations of four backbones, two protocols and two stream orders. On ImageNet-C fog, recent methods collapse in up to two thirds of one hundred runs, while OriTTA never does.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Test-Time Adaptation, Classifier Geometry

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_3u29OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on test-time adaptation when unlabeled examples arrive one at a time, including shuffled and class-blocked streams. It proposes OriTTA, which precomputes a geometric reference frame from the frozen source classifier and one synthetic prototype per class. During adaptation, OriTTA updates only normalization parameters, applies angular and radial losses to samples judged reliable, maintains a class-indexed memory bank, and resets to source weights when some existing method collapse.

Critical strengths
  1. The geometric motivation is clear and partly validated by ablations.

  2. The collapse detector is a potentially reusable contribution.

Critical weaknesses
  1. The experiment results are mixed, and the “best in all sixteen settings” claim is not yet convincing comparing with recent SOTA like Sateen.

  2. The method still depends on predicted labels, so the self-reinforcement risk is reduced but not eliminated. If a sample is confidently misclassified, OriTTA will adapt it toward the wrong ideal point.

  3. The paper emphasizes that OriTTA uses no per-setting tuning. However, the method contains many global parameters, which will make tuning difficult.

  4. The evaluation is too narrow to support the " real-world deployment" motivation. The experiments are conducted on CIFAR-C and ImageNet-C. This is a standard and useful benchmark, but it is synthetic corruption, not natural distribution shift.

  5. The paper says that OriTTA updates “normalization affine parameters,” but it is not clear exactly which parameters are updated in each backbone.

Questions for authors
  1. Exactly which normalization parameters are updated for each backbone?

  2. How were the global constants chosen? a grid search will be very costly.

  3. Why are directly relevant baselines such as NOTE, DELTA, RoTTA omitted? Please either include them or justify their exclusion.

  4. What are the practical costs of the method? Please report anchor-generation time, memory-bank storage, etc.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 3u29 (1/3)

We thank you for the insightful comments and suggestions. We have addressed each of your questions below.

W1. The experiment results are mixed, and the “best in all sixteen settings” claim is not yet convincing comparing with recent SOTA like Sateen.

Response: Thank you for the comment. We believe this concern comes from our wording. The claim in the abstract and the contributions is that OriTTA achieves the highest mean accuracy over the fifteen corruptions in all sixteen settings. However, the Conclusion described it as the most accurate method in every setting, which can be read as a claim about each corruption. We corrected this wording in the Conclusion (Section 5) of the revised version.

On single corruptions, as you pointed out, OriTTA is not always first. However, the gaps are far from symmetric. On ImageNet-C-ResNet50-GN in the episodic shuffled protocol, OriTTA trails SaTeen on eight corruptions by 1.2 points on average and by at most 2.3 points, while it leads on the other seven by 10.7 points on average and by up to 39.7 points. These leads include defocus and glass blur, where SaTeen is stable across seeds and OriTTA is still ahead by 7.7 and 5.7 points, and fog, where SaTeen collapses in 62% of one hundred runs while OriTTA never does, as shown in Table 2 in the revised version. This asymmetry is why OriTTA is ahead in the mean accuracy. On ImageNet-Sketch, a natural shift which we newly added, OriTTA is also ahead of SaTeen on every seed, as presented in Table R3 in our response to W3.

W2. The method still depends on predicted labels, so the self-reinforcement risk is reduced but not eliminated. If a sample is confidently misclassified, OriTTA will adapt it toward the wrong ideal point.

Response: Thank you for the insightful comment. As you pointed out, the risk is reduced rather than eliminated, since a confidently misclassified feature that enters the reliable region can be pulled toward the wrong ideal point. In practice, however, we did not observe such errors reinforcing themselves. If they did, they would appear as abrupt failures on some seeds. Over one hundred seeds on fog, DeYO, ReCAP, and SaTeen collapse in 66%, 33%, and 62% of the runs, while OriTTA never collapses and its standard deviation stays at 0.2 to 0.3 points, as shown in Table 2 in the revised version. The same pattern holds at batch size 64, as shown in Table 33 in the revised version.

Two components keep the remaining risk in check. First, the reliable region admits only the fifth of the cone samples of each class that lie closest to the ideal point. To verify its role, we changed this trust ratio together with the other constants, and the results are presented in the following table.

<Table R1.> Effect of each constant on ImageNet-C-ResNet50-GN, episodic, shuffled. The default row is the ten-seed mean of Table 1 in the revised version, and each variation is a single run with seed 2021 in which one constant is changed. The defaults are $\kappa_0 = 0.11$, trust ratio 0.2, and $w_{\rm rad} = 2$.
fogAvgΔ w.r.t. OriTTA (default)
Source34.0231.44
OriTTA (default)59.0449.59
κ0=0.08\kappa_0 = 0.0856.0247.89−1.70
κ0=0.2\kappa_0 = 0.259.7749.99+0.40
trust ratio = 0.10.159.5749.70+0.11
trust ratio = 0.40.413.2846.34−3.25
wrad=1w_{\rm rad} = 158.9149.42−0.17
wrad=4w_{\rm rad} = 459.2649.74+0.15
learning rate ×0.5\times0.555.8846.66−2.93
learning rate ×2\times261.5751.63+2.04

When the trust ratio is raised to 0.4, the region admits samples farther from the ideal points and fog collapses, while lowering it to 0.1 changes the accuracy little. We added Table R1 with the per-corruption results as Table 15 in Appendix A.4 of the revised version.

Second, the alignment loss pulls one stored input of every class toward its fixed ideal point, which counters the drift of the predicted labels toward a few classes. To verify this, we removed the loss on CIFAR-10-C, and the results are presented in the following table.

<Table R2.> OriTTA with and without $\mathcal{L}_{\rm align}$ on CIFAR-10-C-WRN-28-10, episodic, shuffled, seed 2021. The share is taken over the last 500 predictions, where a balanced stream gives 10%.
OriTTAw/o Lalign\mathcal{L}_{\rm align}
Stored representatives inside their cone, first → last tenth (%)16.9 → 30.813.2 → 16.9
Share of the most predicted class, first → last tenth (%)25.3 → 13.327.7 → 29.2
Pass rate of the reliable region, mean over the last 70% (%)4.62.0
Accuracy over a sliding window, first → last tenth (%)59.9 → 76.957.7 → 62.8
Resets per run03
Accuracy over the stream (%)73.4262.45

Without the loss, the predictions stay concentrated on a few classes until the end of each corruption, and the detector has to reset the model three times. We added Table R2 as Table 25 in Appendix D.3 of the revised version.

AuthorsOpenReview ↗

Response to Reviewer 3u29 (2/3)

W3. The paper emphasizes that OriTTA uses no per-setting tuning. However, the method contains many global parameters, which will make tuning difficult.

Response: Thank you for raising this point. The constants of OriTTA are not meant to be tuned. They were set once on ImageNet-C-ResNet50-GN and are applied unchanged to every dataset and backbone, and a new dataset or checkpoint therefore requires no tuning. As a test on data never used during development, we newly ran OriTTA on ImageNet-Sketch with every constant unchanged, and the results are presented in the following table.

<Table R3.> ImageNet-Sketch-ResNet50-GN, shuffled, batch size 1, ten seeds. The last column counts the seeds on which OriTTA is more accurate than the method of that row on the same stream.
AccuracyGain over sourceSeeds OriTTA wins
Source29.19
TENT22.78±0.43−6.4110/10
EATA30.26±0.62+1.0710/10
SAR31.65±0.18+2.4610/10
DeYO35.44±0.41+6.2510/10
ReCAP37.55±0.38+8.3610/10
SaTeen37.76±0.45+8.5710/10
OriTTA38.56±0.11+9.37

OriTTA is the most accurate method on every seed. Additionally, the defaults lie on the flat side of κ0\kappa_0, the trust ratio, and wradw_{\rm rad} in Table R1. Widening the cone, lowering the trust ratio, or changing wradw_{\rm rad} moves the mean accuracy by at most 0.40 points, and the accuracy drops only when the cone is narrowed or the trust ratio is raised.

In contrast, the published settings of the recent baselines do not transfer from ImageNet-C to CIFAR-100-C. In the episodic shuffled protocol, DeYO, ReCAP, and SaTeen fall far below the source model at 53.25%, while OriTTA reaches 62.16% with the same constants, as shown in Table 3 in the revised version. Since test-time adaptation provides no labeled validation data for a new search, we consider shared constants to be an advantage in practice.

W4. The evaluation is too narrow to support the " real-world deployment" motivation. The experiments are conducted on CIFAR-C and ImageNet-C. This is a standard and useful benchmark, but it is synthetic corruption, not natural distribution shift.

Response: We appreciate your suggestion. Following your suggestion, we newly evaluated OriTTA on ImageNet-Sketch, a natural distribution shift, with every constant unchanged. As presented in Table R3, OriTTA improves the source model from 29.19% to 38.56% and is the most accurate method on every seed. We added this experiment as Table 5 in Section 4.2 of the revised version.

W5 and Q1. The paper says that OriTTA updates “normalization affine parameters,” but it is not clear exactly which parameters are updated in each backbone. Exactly which normalization parameters are updated for each backbone?

Response: Thank you for the question. OriTTA updates only the per-channel scale γ\gamma and shift β\beta of the normalization layers, y=γz^+βy = \gamma \hat{z} + \beta, and keeps every other weight frozen, including the classifier. The normalization statistics are not updated either, since GroupNorm and LayerNorm compute them from each input and BatchNorm keeps those of the source model. The updated parameters of each backbone are listed in the following table.

<Table R4.> Parameters updated by OriTTA. Total counts every parameter of the backbone including the classifier.
BackboneUpdated parametersUpdatedTotalRatio
ImageNet-C-ResNet50-GNscale and shift of all 53 GroupNorm layers53,12025,557,0320.208%
ImageNet-C-ViT-B/16scale and shift of the 24 LayerNorm layers in the 12 blocks, without the final LayerNorm36,86486,567,6560.043%
CIFAR-10-C-WRN-28-10scale and shift of all 25 BatchNorm layers17,95236,479,1940.049%
CIFAR-100-C-WRN-40-2scale and shift of all 37 BatchNorm layers5,4082,255,1560.240%

On ViT-B/16, the final LayerNorm before the classifier is kept frozen. This is stated in Appendix A.1, and we added Table R4 as Table 11 in Appendix A.1 of the revised version.

AuthorsOpenReview ↗

Response to Reviewer 3u29 (3/3)

Q2. How were the global constants chosen? a grid search will be very costly.

Response: Thank you for the good question. We did not run a grid search, and no constant was tuned per setting. During development, we tried a few values for each constant on ImageNet-C-ResNet50-GN, fixed them, and used them unchanged in all sixteen settings and on ImageNet-Sketch. Quantities whose scale differs across backbones, such as ss, δ\delta, τ\tau, κ\kappa, and λ\lambda, are not chosen but computed from the source model by fixed rules (Table 13 in the revised version), and only the remaining dimensionless constants were set on ImageNet-C-ResNet50-GN.

Since the constants were not searched exhaustively, they are not necessarily optimal. In Table R1, several variations, such as a wider cone or a doubled learning rate, are above the defaults. We did not tune them further because test-time adaptation provides no labeled validation data, and a method that needs a search for each dataset could not be deployed in this setting.

Table R1 also shows that a costly search is not needed. κ0\kappa_0 sets the opening of the cone, and the accuracy drops when the cone is narrowed and changes little when it is widened. The trust ratio sets the fraction of cone samples that enter the losses, and the accuracy changes little when it is lowered, while fog collapses when it is raised to 0.4. wradw_{\rm rad} weights the radial loss and has little effect. The learning rate is the only backbone-specific input, with one base rate per normalization type, and it follows the scaling rule for Adam (Malladi et al., 2022) at every batch size, including batch size 1.

Q3. Why are directly relevant baselines such as NOTE, DELTA, RoTTA omitted? Please either include them or justify their exclusion.

Response: We appreciate your suggestion. Following your suggestion, we newly ran RoTTA with its official code on ImageNet-C-ResNet50-GN over the ten seeds of the main tables, and the results are presented in the following table.

<Table R5.> Mean accuracy on ImageNet-C-ResNet50-GN, shuffled, ten seeds. The other baselines are in Table 1 in the revised version.
ImageNet-C-ResNet50-GN, shuffledepisodiccontinual
Source31.4431.44
RoTTA32.2428.01
OriTTA49.5940.81

RoTTA stays near the source model in the episodic protocol and falls below it in the continual one. Together with CoTTA, which falls to the chance level at batch size 1 in Table 4 in the revised version, this indicates that methods designed for continual adaptation with batches do not carry over to a single-sample stream.

NOTE and DELTA cannot be compared under our setting for the following reasons. NOTE replaces BatchNorm with instance-aware batch normalization and re-trains the source model with it, whereas every method in our comparison adapts the same released checkpoints without re-training. DELTA corrects the BatchNorm statistics at test time, which does not apply to ResNet50-GN and ViT-B/16 because they have no BatchNorm. On the CIFAR backbones, every method keeps the source statistics at batch size 1, and the re-weighting of DELTA gives a weight of one to the single sample. Therefore, DELTA reduces to the loss it is combined with, such as the entropy loss of TENT, which is in our tables. We added RoTTA with its per-corruption results as Table 16 in Appendix A.10 of the revised version, together with this discussion, and summarized it in Section 4.2.

Q4. What are the practical costs of the method? Please report anchor-generation time, memory-bank storage, etc.

Response: We appreciate your suggestion. Following your suggestion, we measured the costs of OriTTA outside the stream, which consist of a one-off anchor synthesis and two stored images per class. The results are presented in the following table.

<Table R6.> Costs outside the adaptation time.
ImageNet-C-ResNet50-GNImageNet-C-ViT-B/16CIFAR-10-C-WRN-28-10CIFAR-100-C-WRN-40-2
Anchor synthesis, once per backbone12.2 min28.8 min15 s42 s
Peak GPU memory during synthesis4.3 GiB4.1 GiB0.4 GiB0.3 GiB
Bank and anchors, 32-bit1.20 GB1.20 GB0.25 MB2.5 MB
Bank alone in 8 bits, lossless150 MB150 MB31 kB0.31 MB

The memory bank holds one input image per class, and its size does not grow with the stream, since each slot is overwritten by the latest test input admitted to that class. It stores inputs rather than features because the alignment loss and the detector pass them through the current model. The anchors are kept as well, to re-seed the bank after a reset. For the adaptation time in Table 17 in the revised version in Appendix A.11, OriTTA takes about 26 ms per image on ImageNet-C-ResNet50-GN at batch size 1, compared with 13 ms for TENT and 58 ms for SAR. We added Table R6 as Table 18 in Appendix A.11 of the revised version.

We hope that our response has addressed your concerns. Thank you very much!

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

This paper proposed OriTTA, a framework targeting TTA problem in the situation when batch size was one. The method used the geometry of the frozen classifier to construct the adaptation target. It also built a memory base for samples from different classes to reduce model collapse. The method was experimented on different datasets and conditions, and achieved SOTA average accuracy compared to existing methods. The robustness is further verified through extra experiments spanning multi-seed experiments and ablation studies.

Critical strengths
  • The targeted problem was novel, and the experiments setting had practical values in realworld applications.

  • The method design was well motivated with a definite theoretical explanation. The ablation study further verifies the effect.

  • The experiments covered a wide range of representative TTA methods and different settings, showing the effectiveness and robustness of the proposed method.

Critical weaknesses

The contributions of each individual component are not fully verified. For example, in Table 7 replacing SAR's recovery mechanism with the proposed collapse detector improved its accuracy significantly, close to OriTTA's performance. An ablation of OriTTA without the collapse detector under continual adaptation settings would help clarify the relative contribution of each component.

  • The analysis of the reliability region is not sufficient. Table5 showed that in CIFAR-10-C no anchors fell within the cones but still got higher accuracy. Further analysis is needed to explain such a phenomenon.

  • The extra computational cost was needed to evaluated, which was important in real-world application.

Questions for authors

Same as weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer EiAr (1/2)

We thank you for the insightful comments and suggestions. We have addressed each of your questions below.

W1. The contributions of each individual component are not fully verified. For example, in Table 7 [Table 9 in the revised version] replacing SAR's recovery mechanism with the proposed collapse detector improved its accuracy significantly, close to OriTTA's performance. An ablation of OriTTA without the collapse detector under continual adaptation settings would help clarify the relative contribution of each component.

Response: Thank you for the insightful suggestion! Following your suggestion, we newly conducted the ablation of OriTTA without the collapse detector on CIFAR-10-C-WRN-28-10 in the continual class-blocked protocol, where the detector fires 2.2 times per run. We first removed the detector and then removed each remaining component on top of that model, so that the detector cannot mask a collapse that the component would otherwise prevent. The results are presented in the following table.

<Table R1.> Ablation on CIFAR-10-C-WRN-28-10, continual, class-blocked, ten seeds. The standard frame uses the raw classifier rows, no projection, and the origin at 0.
AverageGain over sourceΔ vs. OriTTA (full)Δ vs. w/o detector
Source56.48
EATA56.94+0.46
OriTTA66.04+9.56
w/o detector63.90+7.42-2.14
+ origin at 063.54+7.06-2.50-0.36
+ no projection59.67+3.19-6.37-4.23
+ standard frame58.17+1.69-7.87-5.73
+ w/o Lang\mathcal{L}_{\rm ang}57.38+0.90-8.66-6.52
+ w/o Lrad\mathcal{L}_{\rm rad}63.88+7.40-2.16-0.02
+ w/o Lalign\mathcal{L}_{\rm align}10.54-45.94-55.50-53.36

Without the detector, OriTTA loses 2.14 points but still keeps about four fifths of its gain over the source model. Among the remaining components, replacing the recentered frame with the standard one and removing Lang\mathcal{L}_{\rm ang} cost the most, while moving the origin to 0 and removing Lrad\mathcal{L}_{\rm rad} change the accuracy only within the seed-to-seed deviation. The origin matters little here because the recentering barely moves it on this backbone, as shown in Table 7 in the revised version. Additionally, Lalign\mathcal{L}_{\rm align} and the detector are complementary safeguards against collapse, since removing Lalign\mathcal{L}_{\rm align} as well drops the accuracy to the chance level.

The accuracy of OriTTA in the episodic protocol does not come from the detector either. As reported in Appendix C, the detector never fires in seven of the eight episodic settings, where OriTTA is the most accurate method without a single reset. We also repeated the ablation in one of these settings, and the results are presented in the following table.

<Table R2.> Ablation on CIFAR-100-C-WRN-40-2, episodic, class-blocked, ten seeds.
AverageGain over sourceΔ vs. OriTTA
Source53.25
OriTTA (full)61.69+8.44
origin at 059.73+6.48-1.96
no projection61.13+7.88-0.56
standard frame56.89+3.64-4.80
w/o Lang\mathcal{L}_{\rm ang}57.72+4.47-3.97
w/o Lrad\mathcal{L}_{\rm rad}61.36+8.11-0.33
w/o Lalign\mathcal{L}_{\rm align}60.27+7.02-1.41

In this setting, the frame contributes the most, since replacing it with the standard one costs more than half of the gain over the source model. Moreover, the two operations of the frame change roles between the two backbones, in line with the geometry in Table 7 in the revised version. On CIFAR-10-C, where the classifier reads only nine of the 640 feature dimensions, the projection matters and the recentering does not. On CIFAR-100-C, where the classifier reads most of the feature dimensions and the recentering moves the origin much further, the recentering matters more than the projection.

Together with the SAR experiment you mentioned, these results indicate that the detector is a safeguard that transfers to other methods, while the frame and the losses provide most of OriTTA’s accuracy. We added Tables R1 and R2 with the per-corruption results as Tables 28 and 27 in Appendices D.7 and D.6 of the revised version.

AuthorsOpenReview ↗

Response to Reviewer EiAr (2/2)

W2. The analysis of the reliability region is not sufficient. Table5 [Table 7 in the revised version] showed that in CIFAR-10-C no anchors fell within the cones but still got higher accuracy. Further analysis is needed to explain such a phenomenon.

Response: Thank you for the good question. OriTTA remains accurate on CIFAR-10-C because test features, unlike the anchors, do enter the cones, and because the alignment loss does not depend on the cone.

The anchors fall outside because the ideal subspace has only nine dimensions on this backbone. The cone rule then gives a half-angle of 18 degrees, while the anchors lie a median of 35 degrees from their axes. As stated in our Limitations, the cone rule becomes too narrow when the number of classes is small.

With the source weights, 5.5% of the frames already fall in the reliable region, and the per-sample losses act from the start. These losses matter on this backbone, as the row without Lang\mathcal{L}_{\rm ang} in Table R1 shows.

Every 30 frames, the alignment loss pulls the stored input of every class toward its ideal point. Since the loss updates normalization parameters shared by every input, the test features move together with these representatives. To verify this, we compared OriTTA with and without Lalign\mathcal{L}_{\rm align} on the same seed and stream, and the results are presented in the following table.

<Table R3.> OriTTA with and without $\mathcal{L}_{\rm align}$ on CIFAR-10-C-WRN-28-10, episodic, shuffled, seed 2021.
OriTTAw/o Lalign\mathcal{L}_{\rm align}
Stored representatives inside their cone, first → last tenth (%)16.9 → 30.813.2 → 16.9
Share of the most predicted class, first → last tenth (%)25.3 → 13.327.7 → 29.2
Pass rate of the reliable region, mean over the last 70% (%)4.62.0
Accuracy over a sliding window, first → last tenth (%)59.9 → 76.957.7 → 62.8
Resets per run03
Accuracy over the stream (%)73.4262.45

With the loss, the share of stored representatives inside their cones nearly doubles within each corruption and the predictions become close to balanced. Without the loss, both change little and the accuracy drops by about 11 points. In other words, the anchors start outside the cones, and the alignment loss draws the stored representatives into them, which keeps the predictions balanced and more of the stream in the reliable region. We added this analysis with Table R3 as Table 25 in Appendix D.3 of the revised version.

W3. The extra computational cost was needed to evaluated, which was important in real-world application.

Response: We appreciate your suggestion. OriTTA takes about 26 ms per image on ImageNet-C-ResNet50-GN at batch size 1, compared with 13 ms for TENT and 58 ms for SAR, according to Table 17 in the revised version (Appendix A.11). The extra time over TENT mainly comes from Lalign\mathcal{L}_{\rm align}. Following your suggestion, we additionally measured the costs outside the stream, which consist of a one-off anchor synthesis and two stored images per class. The results are presented in the following table.

<Table R4.> Costs outside the adaptation time.
ImageNet-C-ResNet50-GNImageNet-C-ViT-B/16CIFAR-10-C-WRN-28-10CIFAR-100-C-WRN-40-2
Anchor synthesis, once per backbone12.2 min28.8 min15 s42 s
Peak GPU memory during synthesis4.3 GiB4.1 GiB0.4 GiB0.3 GiB
Bank and anchors, 32-bit1.20 GB1.20 GB0.25 MB2.5 MB
Bank alone in 8 bits, lossless150 MB150 MB31 kB0.31 MB

We added Table R4 as Table 18 in Appendix A.11 of the revised version.

We hope that our response has addressed your concerns. Thank you very much!

Thank you for your explanation. Most of concerns have addressed and i will keep my score as is,

AuthorsOpenReview ↗

Thank you!

We appreciate your response. We are glad that the additional experiments and discussion have addressed most of your concerns.

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

This paper proposes OriTTA, a geometry-based TTA framework designed for single-sample and non-IID test streams. It constructs fixed class-specific geometric targets from the frozen source classifier and synthetic prototypes, combining selective feature alignment, class-wise memory, and collapse detection to stabilize online adaptation.

Critical strengths

(1) Significant research problem. The paper addresses an important practical challenge in TTA, particularly unreliable adaptation under batch-size-one and temporally correlated test streams.

(2) Well-motivated framework. The limitations of batch-dependent statistics and confidence-based adaptation are clearly identified. Leveraging fixed classifier geometry provides a coherent alternative for guiding feature adaptation without relying on current-batch statistics.

Critical weaknesses

(1) Insufficient justification of novelty. Prior works such as CoMM [1], CPA [2], and the already-cited NCTTA [3] have explored classifier-guided feature alignment for TTA. Although OriTTA introduces origin recentering and subspace projection, Tables 6 and 16 evaluate these components under different datasets or experimental protocols, making their relative contributions difficult to assess. Notably, removing the angular loss leads to a substantial accuracy drop (13.92%), while angular feature alignment is an established concept. It remains unclear whether the improvements primarily stem from existing alignment principles or the proposed recentered geometry. Controlled ablations under identical settings and direct comparisons with prior methods would better substantiate the methodological novelty.

(2) Limited generalization evaluation. The experiments focus primarily on synthetic corruption benchmarks at a fixed severity level. Evaluation under natural distribution shifts and more heterogeneous test streams would better establish the method's practical generalizability. Including a source-only baseline in the main accuracy comparisons would also help clarify the actual benefits of adaptation.

(3) Incomplete discussion of related work. Several relevant studies addressing small-batch and non-IID TTA, including TTN [4], TEMA [5], and TRIBE [6], are overlooked. Discussing these works would provide a more complete research context and strengthen the motivation.

[1] Enhanced Online Test-time Adaptation with Feature-Weight Cosine Alignment

[2] Prototypical Class-wise Test-time Adaptation

[3] Neural Collapse in Test-Time Adaptation

[4] TTN: A Domain-Shift Aware Batch Normalization in Test-Time Adaptation

[5] Unraveling Batch Normalization for Realistic Test-Time Adaptation

[6] Towards Real-World Test-Time Adaptation: Tri-net Self-Training with Balanced Normalization

Questions for authors

see weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer eFvr (1/2)

We thank you for the insightful comments and suggestions. We have addressed each of your questions below.

W1. Prior works such as CoMM [1], CPA [2], and the already-cited NCTTA [3] have explored classifier-guided feature alignment for TTA. Although OriTTA introduces origin recentering and subspace projection, Tables 6 and 16 [Tables 8 and 22 in the revised version] evaluate these components under different datasets or experimental protocols, making their relative contributions difficult to assess. Notably, removing the angular loss leads to a substantial accuracy drop (13.92%), while angular feature alignment is an established concept. It remains unclear whether the improvements primarily stem from existing alignment principles or the proposed recentered geometry. Controlled ablations under identical settings and direct comparisons with prior methods would better substantiate the methodological novelty.

Response: Thank you for the insightful comment. Following your suggestion, we newly compared OriTTA with CoMM and NCTTA, and conducted every ablation in a single setting. First, we ran CoMM and NCTTA with their official code on ImageNet-C-ResNet50-GN in the episodic shuffled protocol at batch size 1, and the results are presented in the following table.

<Table R1.> Comparison with CoMM and NCTTA on ImageNet-C-ResNet50-GN, episodic, shuffled, batch size 1, ten seeds.
AvgGain over sourceΔ w.r.t. OriTTA
Source31.44
CoMM34.30±0.06+2.86-15.29
NCTTA32.54±0.03+1.10-17.05
OriTTA (ours)49.59±0.05+18.15

As you pointed out, aligning features with the classifier is shared by CoMM, CPA, and NCTTA. However, CoMM and NCTTA gain less than 3 points over the source model, while OriTTA gains about 18 points and is the most accurate method on every seed. What OriTTA adds is a set of fixed targets derived from the classifier, which are defined in a frame recovered from the source model before any test data arrive. OriTTA also outperforms both methods on the natural and the mixed shifts, as presented in Tables R3 and R4 in our response to W2. We could not find a public implementation of CPA, and we discussed it in the related work in Section 2 and Appendix B of the revised version. We added CoMM and NCTTA to Table 1 in the revised version, and we will include them in the remaining settings of Tables 1 and 3 in the final version.

Second, we conducted every ablation in a single setting, CIFAR-100-C-WRN-40-2 in the episodic class-blocked protocol, and the results are presented in the following table.

<Table R2.> Ablations on CIFAR-100-C-WRN-40-2, episodic, class-blocked, batch size 1, ten seeds. The standard frame uses the raw classifier rows, no projection, and the origin at 0.
AvgΔ w.r.t. OriTTA (full)
Source53.25
OriTTA (full)61.69
No projection61.13-0.56
Origin at 059.73-1.96
Standard frame56.89-4.80
w/o Lang\mathcal{L}_{\rm ang}57.72-3.97
w/o Lrad\mathcal{L}_{\rm rad}61.36-0.33
w/o Lalign\mathcal{L}_{\rm align}60.27-1.41
Standard frame, w/o Lang\mathcal{L}_{\rm ang}56.41-5.28
Standard frame, w/o Lrad\mathcal{L}_{\rm rad}56.88-4.81
Standard frame, w/o Lalign\mathcal{L}_{\rm align}53.99-7.70

In this setting, replacing the recentered frame with the standard one costs more than removing the angular loss. Moreover, the two are not independent. The angular loss adds only 0.48 points in the standard frame, whereas it adds 3.97 points in the recentered frame. Therefore, in this setting, the angular alignment that OriTTA shares with prior methods is effective mainly in the frame that OriTTA recovers. Within the frame, the recentered origin matters more than the projection, in line with the geometry of this backbone in Table 7 in the revised version, where the recentering raises the fraction of prototypes inside their cones from 47% to 99%. We added Table R2 with the per-corruption results as Table 27 in Appendix D.6 of the revised version.

AuthorsOpenReview ↗

Response to Reviewer eFvr (2/2)

W2. Limited generalization evaluation. The experiments focus primarily on synthetic corruption benchmarks at a fixed severity level. Evaluation under natural distribution shifts and more heterogeneous test streams would better establish the method's practical generalizability. Including a source-only baseline in the main accuracy comparisons would also help clarify the actual benefits of adaptation.

Response: We appreciate your suggestions. Following your suggestions, we newly added a natural distribution shift and a more heterogeneous test stream, presented as Tables 5 and 6 in the revised version (Section 4.2), and added the source model to every accuracy table.

For the natural distribution shift, we evaluated on ImageNet-Sketch with the constants, the anchors, and the learning rate of the main tables unchanged, and the results are presented in the following table.

<Table R3.> ImageNet-Sketch-ResNet50-GN, shuffled, batch size 1, ten seeds. The last column counts the seeds on which OriTTA is more accurate than the method of that row on the same stream.
AccuracyGain over sourceSeeds OriTTA wins
Source29.19
TENT22.78±0.43−6.4110/10
EATA30.26±0.62+1.0710/10
SAR31.65±0.18+2.4610/10
DeYO35.44±0.41+6.2510/10
ReCAP37.55±0.38+8.3610/10
SaTeen37.76±0.45+8.5710/10
CoMM30.08±0.04+0.8910/10
NCTTA29.30±0.05+0.1110/10
OriTTA38.56±0.11+9.37

OriTTA is the most accurate method on every seed. It leads the closest methods, SaTeen and ReCAP, by 0.80 and 1.01 points, with paired 95% intervals that exclude zero.

For a more heterogeneous test stream, we evaluated on the mixed-shift setting of SAR (Niu et al., 2023), a single stream that mixes all fifteen corruptions of ImageNet-C, and the results are presented in the following table.

<Table R4.> ImageNet-C-ResNet50-GN, a single stream mixing the fifteen corruptions, batch size 1, five seeds.
AccuracyGain over sourceΔ w.r.t. OriTTA
Source31.44
TENT7.69±0.76−23.75−29.83
EATA21.11±0.91−10.33−16.41
SAR36.98±0.16+5.54−0.54
DeYO25.56±1.32−5.88−11.96
ReCAP29.03±2.60−2.41−8.49
SaTeen10.78±2.22−20.66−26.74
CoMM36.38±0.11+4.94−1.14
NCTTA16.88±3.49−14.56−20.64
OriTTA (ours)37.52±0.13+6.08

OriTTA is the most accurate method on this stream as well, ahead of SAR, which was designed for such streams. Every method other than SAR and CoMM falls below the source model. Regarding the severity level, the corruption benchmarks of the paper use severity 5, the most severe level.

W3. Incomplete discussion of related work. Several relevant studies addressing small-batch and non-IID TTA, including TTN [4], TEMA [5], and TRIBE [6], are overlooked. Discussing these works would provide a more complete research context and strengthen the motivation.

Response: Thank you for suggesting additional related work. In the revised version, we added TTN, TEMA, and TRIBE to Section 2 and a paragraph on them to Appendix B. All three make batch normalization reliable under small or non-i.i.d. test batches. TTN interpolates the source and test-batch statistics with weights learned in a post-training phase, TEMA extends the test statistics with an exponential moving average over past batches, and TRIBE combines a class-balanced BatchNorm with tri-net self-training. In contrast, OriTTA adapts affine parameters toward targets fixed before the stream and does not use batch statistics. Since our ImageNet backbones use GroupNorm and LayerNorm, and every method keeps the source statistics on CIFAR at batch size 1, we present these methods as complementary to OriTTA on batch-normalized backbones.

We hope that our response has addressed your concerns. Thank you very much!

I appreciate the authors' detailed response. Although some reservations regarding the extent of methodological novelty remain, I believe the major concerns have been addressed. Therefore, I am updating my score.

AuthorsOpenReview ↗

Thank you!

We deeply appreciate your response. We are glad to learn that you have been satisfied with the additional experiments and discussion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely appreciate the reviewers’ time and invaluable feedback. We would like to express our gratitude to the reviewers for their recognition of our research. Below is a summary of the strengths noted in the reviews.

Strengths recognized by reviewers

  • The paper addresses a significant and practical problem, namely test-time adaptation at batch size one under temporally correlated test streams. [eFvr, EiAr]
  • The framework is well motivated, and the fixed classifier geometry provides a coherent alternative for guiding adaptation without relying on batch statistics. [eFvr, EiAr, 3u29]
  • The method design is supported by theoretical explanations and ablation studies. [EiAr]
  • The collapse detector is a potentially reusable contribution. [3u29]
  • The experiments cover a wide range of representative TTA methods and settings, showing the effectiveness and robustness of the proposed method. [EiAr]

We have addressed each reviewer’s comments in detail in our individual responses. In accordance with the reviewers’ suggestions, we revised the paper and uploaded the revised version, which we believe enhanced the paper’s strength. Here, we would like to highlight the major revisions.

Related Work

  • Expanded the discussion of classifier-guided feature alignment as well as small-batch and non-i.i.d. TTA methods, and clarified how OriTTA differs from them in Section 2 and Appendix B, as suggested by Reviewer eFvr.

Experiments

  • Conducted experiments (including CoMM [1], NCTTA [2], and RoTTA [3]) to show the effectiveness of OriTTA compared to prior classifier-guided alignment methods and other baselines suggested by Reviewers eFvr and 3u29.
  • Conducted experiments on a natural distribution shift (ImageNet-Sketch) and a heterogeneous test stream mixing all fifteen corruptions, showing that OriTTA maintains the highest accuracy, as suggested by Reviewers eFvr and 3u29.
  • Performed ablation studies, including OriTTA without the collapse detector, showing that the recentered frame and proposed losses account for most of the performance gains, as suggested by Reviewers EiAr and eFvr.
  • Included hyperparameter sensitivity experiments to show that a costly search is not needed, as suggested by Reviewer 3u29.
  • Measured the one-off anchor synthesis costs and the storage requirements for the anchors and fixed-size memory bank, as suggested by Reviewers EiAr and 3u29.
  • Added all the additional experimental results to the revised version of our paper.

Discussion

  • Additional analysis on why OriTTA remains accurate on CIFAR-10-C although the anchors fall outside their cones, as suggested by Reviewer EiAr.
  • Additional discussion on the global constants, which are shared across settings without per-setting tuning, and on the parameters updated in each backbone, as suggested by Reviewer 3u29.

We hope that the additional experiments and clarifications have convinced the reviewers of the merits of this paper. We appreciate the reviewers’ consideration and remain happy to provide any further clarification.

References

[1] “Enhanced online test-time adaptation with feature-weight cosine alignment.” arXiv 2024.

[2] “Neural collapse in test-time adaptation.” CVPR 2026.

[3] “Robust test-time adaptation in dynamic scenarios.” CVPR 2023.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.