acceptodds
Under review as a conference paper at ICLR 2027

DTS-Meta: Dual-Timescale Meta-Learning for Continual Test-Time Adaptation

Abstract

Continual test-time adaptation (CTTA) adapts a source-pretrained model to an unlabeled, non-stationary target stream without source data or domain boundaries. Existing feature-subspace methods identify where to adapt, but coordinating persistent task-related adjustments with rapid environment-dependent corrections remains insufficiently explored. We observe that class-sensitive subspaces remain relatively stable across batches, whereas domain-sensitive subspaces exhibit greater temporal variation and environmental sensitivity. We therefore propose DTS-META, a dual-timescale feature adapter with slow class-sensitive and fast domain-sensitive pathways. The slow pathway learns persistent modulation along directions maintained through smoothed classifier-sensitivity statistics to preserve task-relevant feature structure. The fast pathway responds to the current environment via a more frequently refreshed prototype-residual covariance basis, with a MetaNet mapping projected batch statistics and teacher uncertainty to bounded scale and bias. We train the MetaNet through episodic meta-learning, encouraging updates learned from the current batch to remain effective across input variations. Across three classification benchmarks and continual semantic segmentation on CarlaTTA, DTS-META consistently outperforms existing baselines, demonstrating the benefits of timescale coordination in non-stationary adaptation.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual test time adaptation, Meta learning, Dual-Timescale dynamics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_44sJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DTS-META extends GOLD's classifier-sensitive subspace adaptation with a fast pathway with a prototype-residual covariance based slow-fast paradigm and batch-conditioned parameter generation and a first-order support-query update. The slow pathway controls the overall feature abstraction, while the fast pathway quantifies current-batch-specific feature corrections. It evaluate three classifier benchmarks in standard, recurring, gradual and long-horizon streams as well as in a semantic segmentation setting. The proposed approach is not a new formulation of CTTA but an augmentation of the feature-subspace adaptation paradigm.

Critical strengths
  1. The amount of classification improvement is small but consistent. DTS-META consistently outperforms GOLD for all three standard Evaluation benchmarks, with absolute error reductions of between 0.59 and 0.72. The five-seed results give this bit of progress more credence than one would have with a single best run -which is critical to a positive evaluation since individual improvements were not large.

  2. The paper has fairly useful controls for its main additions. Both Forward and Full have the same Meta Net architecture, so there is a fairly good picture of the episodic update. In addition, the separate refresh-schedule experiment does not compare architectures with different capacities, but keeps the full model the same. These controls suggest that the optimization and schedule learning does something in excess of just adding an additional adapter.

Critical weaknesses
  1. Address the discrepancy in the previously published results of GOLD. The original Table 1 of GOLD reports a 14.1% error for GOLD on CIFAR-10-C, while this paper reports 15.04% for GOLD and 14.45% for DTS-META. Therefore, DTS-META is better than the reproduced baseline but does not beat the results originally published in GOLD in terms of numbers; the reproduction gap between the results and the original is around 0.94, which is larger than the reported 0.59.

  2. Episodic meta-learning gains only 0.27, 0.21, 0.23 over Forward (with the same Meta Net structure) in Table 3, and both default dual-timescale schedules only achieve 0.28 gains over frequently refreshing both pathways in Table 10. Though these are good controls, each table does not report an uncertainty estimate for these differences.

Questions for authors
  1. The reason for the difference between GOLD's initial CIFAR-10-C result and the one shown here could be variations in checkpoint selection, order of the samples, hyperparameters, augmentations, and when predictions are made. It will be interesting to look into whether adjusting any of these factors affects the ranking relative to other methods.

  2. For the comparison of Forward vs. Full, and for the refresh-schedule ablation, the authors should give results for each seed. A matched comparison showing that the prototype-residual covariance is likely to exceed the raw-feature covariance independently of extra capacity would also be helpful.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 44sJ

Thank you for taking the time to review our paper and providing valuable feedback. We appreciate your recognition of the consistent improvements and the controlled comparisons of our design choices. In response to your Weaknesses (W) and Questions (Q), we provide the following Explanations (E).

W1 & Q1: GOLD reproduction and the published-result discrepancy

E1: We used GOLD's official implementation and released CIFAR-10-C configuration without additional hyperparameter tuning or structural changes. The relevant settings are:

  • Model and checkpoint: the pretrained WideResNet-28-10 model from the official loading pipeline.
  • Stream and batching: batch size 200, with all 15 corruption types processed sequentially at severity 5.
  • Augmentation and prediction timing: both follow the released implementation.

Table G1 of our Common Response reports the individual results for seeds 1–5 and their paired uncertainty. Reproduced GOLD has an error of 15.04 ± 0.10% on CIFAR-10-C, compared with 14.1% in the original paper. DTS-META reaches 14.45 ± 0.09% and has lower error in every matched run. The mean reduction is 0.59 pp. We will distinguish the originally reported and reproduced GOLD values in the revised manuscript.

More broadly, all baselines in our benchmark were faithfully reproduced strictly following their officially released codebases and default hyperparameters without modification. Across all evaluated benchmarks, our reproduced results closely align with those reported in the literature, ensuring that all methods are evaluated on an identical, fair, and rigorous playing field.

W2 & Q2: Seed-Wise Uncertainty Estimates and Controlled Subspace Comparisons

E2: Table G2 of our Common Response summarizes the Forward, Raw-cov, and Full results. Forward omits the temporary inner update; Raw-cov replaces only the covariance used to estimate the fast basis. Both retain Full's parameter capacity. Table 1 below provides their individual seed errors. The matched Full errors are already reported in Table G1 of the Common Response.

Table 1: Individual-seed error rates (%) for Forward, Raw-cov, and Full.

BenchmarkVariantSeed 1Seed 2Seed 3Seed 4Seed 5Mean ± std
CIFAR-10-CForward14.5714.7514.8614.5814.7414.70 ± 0.12
Raw-cov14.5214.6014.7014.5114.5414.57 ± 0.08
Full14.3514.5014.5714.3714.4614.45 ± 0.09
CIFAR-100-CForward27.9928.1928.0428.1028.1528.09 ± 0.08
Raw-cov28.1128.1728.0628.0228.0828.09 ± 0.06
Full27.9327.8727.8227.8027.9327.87 ± 0.06
ImageNet-CForward59.0658.9059.0959.0258.7358.96 ± 0.15
Raw-cov59.0258.9059.0958.9858.7258.94 ± 0.14
Full58.8358.8458.7858.7358.5758.75 ± 0.11

Full has lower error than each control in every seed on all three benchmarks. Table 2 gives the paired gains and their pointwise 95% confidence intervals. There are five pairs per dataset; the 15 dataset–seed comparisons are not pooled into one test.

Table 2: Paired error reductions (pp) relative to the two controls.

BenchmarkForward − Full [95% CI] (pp)Raw-cov − Full [95% CI] (pp)
CIFAR-10-C0.250 [0.206, 0.294]0.124 [0.080, 0.168]
CIFAR-100-C0.224 [0.097, 0.351]0.218 [0.146, 0.290]
ImageNet-C0.210 [0.083, 0.337]0.192 [0.073, 0.311]

These comparisons support the temporary support update and residual-covariance basis estimation at fixed parameter capacity in the tested settings. They do not attribute the effects to additional parameters or imply that their contributions are independent.

Table G3 of our Common Response reports the five-seed results for the ImageNet-C CDC schedules, and Table 3 below provides the paired comparisons. The default schedule, Ts=50,Tf=5T_s=50, T_f=5, has lower error than Both Frequent, Both Infrequent, and Inverted in every matched seed.

Table 3: Paired gains (pp) from Default on ImageNet-C under CDC.

Alternative schedule (Ts,TfT_s, T_f)Alternative − Default [95% CI] (pp)
Both Frequent (5, 5)0.280 [0.081, 0.479]
Both Infrequent (50, 50)0.370 [0.185, 0.555]
Inverted (5, 50)0.440 [0.294, 0.586]

This consistent ordering supports the benefit of timescale coordination across the tested random stream initializations.

Thank you for your time and for the additional experiments. My concerns regarding uncertainty and controlled comparisons have been largely addressed. The five-seed comparison between Forward/Full and Raw-cov, as well as the ablation study on the refresh schedule, directly respond to my requests.

However, the discrepancy regarding GOLD has only been partially resolved. While using the official implementation and reporting paired results does strengthen the comparison against your reproduced baseline, it does not explain why GOLD achieved a result of 15.04% instead of the published 14.1%.

AuthorsOpenReview ↗

Response to Reviewer 44sJ

Thank you for your follow-up response and for clarifying the remaining concern. We agree that the paired comparison against our reproduced GOLD baseline does not explain the difference from the published 14.1% result. We have not identified the cause of this discrepancy.

We checked our CIFAR-10-C configuration and source-model loading path against the official release. The source model is the Standard WideResNet-28-10 selected by GOLD’s loading code. We retained the released augmentation, prediction timing, and corruption order. Evaluation used batch size 200, with all 15 corruption types processed sequentially at severity 5. After loading the warmup checkpoint, we verified that Adam used learning rate 0.001, betas (0.9, 0.999), and weight decay 0. For the reported GOLD baseline, we retained the released architecture and hyperparameters.

We also repeated the evaluation at seed 1 after regenerating our locally computed source warmup checkpoint and class prototypes. These additional checks yielded errors close to 15.0% and did not recover the published result.

Across the five matched seeds, GOLD yields 15.04 ± 0.10% error and DTS-Meta yields 14.45 ± 0.09%, a mean reduction of 0.59 percentage points. This supports an improvement over our reproduced baseline under the recorded evaluation conditions. Our CIFAR-10-C error remains above GOLD’s published result.

We will report GOLD’s published result and our reproduced result separately in the revised manuscript. The reported 0.59-percentage-point reduction is based on paired runs against reproduced GOLD under the same evaluation protocol.

Dear Reviewer 44sJ,

Thank you very much for your active engagement in the discussion and for acknowledging our previous responses. We have made every effort to address your remaining concern in our latest reply. As the discussion period is nearing its conclusion, could we kindly inquire if our follow-up explanation has resolved this final question? Thanks for your time and efforts in reviewing our work, and we sincerely look forward to your reply.

Best regards,

Authors

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

This paper proposes DTS-META, a dual-timescale framework for continual test-time adaptation. The slow pathway maintains trainable scaling along class-sensitive directions, while the fast pathway uses prototype residuals and a MetaNet to generate batch-conditioned feature corrections. The MetaNet is optimized through first-order episodic meta-learning, where a temporary update on the current batch is evaluated on an augmented view of the same batch using a support–query objective. Experiments on classification and semantic segmentation benchmarks demonstrate improvements over existing CTTA methods.

Critical strengths
  • The paper addresses an important challenge in continual test-time adaptation (CTTA) where balancing stable task-related adaptation with rapid responses to continuously changing target distributions. This is a relevant and well-motivated problem in non-stationary environments like the previous work, GOLD.

  • The paper claimed empirical motivation for the dual-timescale design. Fig. 1 illustrates that class-sensitive subspaces remain relatively stable over time, whereas domain-sensitive subspaces exhibit greater temporal variation and sensitivity to environmental changes.

  • The proposed dual-timescale feature adapter coordinates persistent task-related adjustments with batch-dependent residual corrections through a slow-class sensitive pathway and a fast domain-sensitive pathway.

  • META-NET evaluates temporary updates from the support view on an augmented query view and updates persistent parameters via first-order meta-gradients.

Critical weaknesses
  • The novelty of the proposed method is unclear. In the proposed dual-timescale feature adapter, the slow class-sensitive pathway, including AGOP-based subspace estimation, EMA updates, periodic eigendecomposition, and persistent feature modulation, comes directly from GOLD. Therefore, the slow pathway itself does not constitute a novel contribution. Although the paper introduces an additional fast domain-sensitive pathway, the two pathways are primarily combined through sequential feature modulation with different update frequencies. It remains unclear whether this constitutes a genuinely new timescale coordination framework or simply an extension of GOLD with an additional adaptation module.

  • The ablation study does not adequately establish the necessity of the proposed dual-timescale design. Although Table 3 provides incremental component ablations, fast-only and full-without-slow variants are missing, making it difficult to assess the individual contributions of the two pathways. More importantly, controlled comparisons with identical update timescales are needed to determine whether the improvements arise from timescale separation rather than additional adaptation capacity.

  • The assumption that the fast subspace captures domain-sensitive directions is not sufficiently justified. The fast subspace is derived from the covariance of residuals between test features and source class prototypes, but its dominant eigenvectors may reflect intra-class variations or pseudo-label noise rather than domain shifts. Furthermore, covariance mean-centering removes common residual shifts. Additional analysis is needed to validate the domain-sensitive interpretation and demonstrate that the slow and fast subspaces capture complementary information.

  • Although Table 3 shows improvements from episodic optimization, it does not fully disentangle the benefit of temporary inner-loop adaptation from augmented query evaluation. A controlled comparison with standard augmentation-based consistency training, using identical augmentations and losses but without the temporary inner update, would better establish the necessity of the proposed meta-learning strategy.

Questions for authors
  • Given GOLD's efficiency, could the authors quantify the proposed method’s additional computational overhead relative to its performance gains?

  • Could the authors disentangle the benefits of distinct adaptation timescales from those of batch-conditioned modulation and episodic optimization?

  • The proposed method relies on class-wise prototypes computed using labeled source data. Could the authors clarify whether such additional source-derived statistics are permitted under the source-free CTTA protocol? Under a strict source-free CTTA setting, adaptation should rely exclusively on the pretrained source model and unlabeled target data. The use of separately precomputed class-wise source feature statistics introduces additional source-domain information and thus violates this strict protocol.

6 comments
AuthorsOpenReview ↗

Response to Reviewer ofWu(Part 1)

Thank you for taking the time to review our paper and providing valuable feedback. We appreciate your recognition of the importance of the CTTA problem and the empirical motivation for our approach. In response to your Weaknesses (W) and Questions (Q), we provide the following Explanations (E).

W1: Timescale Coordination Framework or Sequential Module Extension

E1: Our slow pathway follows GOLD's AGOP formulation. DTS-META places this pathway within a dual-timescale design that coordinates persistent task-related modulation with rapid responses to domain shifts. The contribution lies in how the two pathways interact, rather than simply adding another adapter to GOLD.

The slow pathway maintains task-related adjustments across the stream. The fast pathway computes its batch descriptor from prototype residuals after slow modulation (Eq. 10). The MetaNet therefore conditions its scale and bias on the feature state produced by the slow pathway, targeting the residual changes in dispersion and translation that remain after that adjustment.

Tables G2 and G3 of our Common Response provide the corresponding controls. With architecture and parameter count fixed, the default refresh schedule achieves lower error than Both Frequent, Both Infrequent, and Inverted under CDC shifts. The pathway ablations in Table G2 also show a benefit from retaining both pathways. Together, these results support the coordination of persistent and batch-dependent adaptation, with the refresh-schedule comparison separating the effect of timing from additional parameter capacity.

W2: Completing Component and Timescale Ablations

E2: The definitions and equations below Table G2 of our Common Response distinguish the two added pathway controls:

  • Full w/o Slow removes the slow AGOP pathway and applies fast modulation directly to backbone features. It retains the MetaNet and episodic training.
  • Fast-only directly optimizes persistent fast scale and bias with Adam. It has no MetaNet, batch descriptor, or temporary inner update.

Table G2 reports lower error for Full than for either control. The Full w/o Slow comparison tests the slow pathway while retaining the full fast procedure. Fast-only changes parameter generation, output constraints, and optimization together, so its comparison does not isolate one factor. Table G3 separately tests the refresh arrangement at fixed architecture and parameter count. The default schedule has the lowest error in every matched seed. These controls distinguish pathway removal from refresh timing.

W3: Physical Interpretation of Fast Residual Subspace and Mean-Centering

E3: Raw-feature covariance can reflect differences between class means. Prototype subtraction is intended to reduce that contribution before estimating the fast basis. Raw-cov replaces only the basis-estimation statistic, as defined below Table G2 of our Common Response. The rank, EMA, refresh schedule, MetaNet, and losses are retained. Its descriptor remains residual-based. Full has lower error than Raw-cov on all three benchmarks in Table G2, supporting residual covariance within this procedure.

Mean-centering in Eq. 8 removes the common residual translation from the covariance. The descriptor in Eq. 10 separately retains the normalized projection of the residual mean rˉt\bar{\boldsymbol{r}}_t onto the fast subspace in ut\boldsymbol{u}_t. This information is available to the MetaNet when it generates the fast bias btfast\boldsymbol{b}_t^{\mathrm{fast}} and scale stfast\boldsymbol{s}_t^{\mathrm{fast}}. The bias provides an additive correction within the selected subspace, and the scale modulates feature amplitudes along the same basis, as shown in Eq. 12. Both coefficients depend on the full descriptor; the bias does not guarantee correction outside the selected subspace.

The temporal and paired-image diagnostics in Appendices B.2 and B.3 provide further evidence of sensitivity to environmental changes. Within-class variation and pseudo-label errors can still affect the estimated directions. We therefore use domain-sensitive to describe their measured behavior, rather than a complete separation of semantic and domain factors.

W4: Disentangling Inner-Loop Meta-Update from Augmented Consistency Regularization

E4: Forward provides the requested augmentation-based consistency control. As described in Appendix B.7, it retains Full’s MetaNet, support and augmented query views, teacher targets, and outer loss definitions and weights. Both views use the persistent MetaNet parameters, with ηin=0\eta_{\mathrm{in}}=0.

Full instead evaluates the query with parameters obtained from a temporary support update. Full has lower error in every matched seed on each benchmark; Table G2 of our Common Response summarizes the results. The query is present in both variants, so the difference concerns the inner update rather than the addition of augmentation.

AuthorsOpenReview ↗

Response to Reviewer ofWu(Part 2)

Q1: Computational Overhead Relative to Performance Gains

E5: We measured per-sample FLOPs, batch latency, and peak GPU memory on an NVIDIA RTX 4090 D, using the same precision and batch configurations for both methods. Table 1 reports these measurements alongside the error reductions in Table G1 of our Common Response.

Table 1: Measured compute, peak GPU memory, and batch latency of GOLD and DTS-META.

Benchmark (backbone, B)Error, GOLD→DTS (%)FLOPs/sample, GOLD→DTSPeak VRAM, GOLD→DTS (GB)Latency/batch, GOLD→DTS (s)
CIFAR-10-C (WideResNet-28, 200)15.04→14.45 (−0.59 pp)73.4100→73.4111 G (+1.11 M; +0.0015%)8.89→8.56 (−0.33)0.3447→0.3472 (+0.73%)
CIFAR-100-C (ResNeXt-29, 200)28.59→27.87 (−0.72 pp)15.0002→15.0023 G (+2.15 M; +0.014%)11.64→11.33 (−0.31)0.2099→0.2114 (+0.71%)
ImageNet-C (ResNet-50, 64)59.40→58.75 (−0.65 pp)56.8291→56.8415 G (+12.38 M; +0.022%)14.27→14.35 (+0.08)0.3363→0.3427 (+1.90%)

The reported FLOPs increase is below 0.025%, and measured batch latency increases by less than 2% on all three benchmarks. Peak VRAM changes range from −0.33 to +0.08 GB. The design uses Tf=5T_f=5 for fast-basis refreshes and a two-layer MetaNet. The inner step updates only that network, and the outer update uses a first-order meta-gradient. The table reports the total measured costs; it does not assign the overhead to individual components.

Q2: Disentangling Timescales, Batch Conditioning, and Episodic Meta-Learning

E6: We examined the three components using controls with the same parameter count across five independent seeds. The control definitions are given below Table G2.

The Fixed descriptor control replaces the current batch statistics with a constant input while retaining the other components. Table G2 reports higher mean error on all three benchmarks with this control. The residual moments μ^t\hat{\boldsymbol{\mu}}_t and v^t\hat{\boldsymbol{v}}_t allow the MetaNet to adjust the fast bias btfast\boldsymbol{b}_t^{\mathrm{fast}} and scale stfast\boldsymbol{s}_t^{\mathrm{fast}} in response to the current batch.

Forward uses the persistent MetaNet parameters for both views and omits the temporary support update. This gives higher error on all three classification benchmarks in Table G2. The support–query objective regularizes adaptation by evaluating the support-updated parameters on an augmented query view. It encourages updates that remain useful across transformations of the current batch, rather than fitting transient batch noise.

On ImageNet-C under CDC, we vary the refresh periods while keeping the architecture and parameter count fixed. Table G3 shows that the default schedule, Ts=50,Tf=5T_s=50, T_f=5, achieves lower error than the equal-period schedules, 5, 5 and 50, 50. This arrangement allows the slow pathway to maintain persistent class-discriminative structure while the fast pathway responds more quickly to abrupt environmental changes.

Together, these comparisons support distinct contributions from batch conditioning, episodic optimization, and refresh scheduling in non-stationary adaptation.

Q3: Compliance with the Source-Free CTTA Protocol and Class-Wise Prototypes

E7: Source-free CTTA adapts a pretrained model to an unlabeled target stream without revisiting source training examples during online adaptation. GOLD [1] and RMT [2] retain fixed class prototypes computed from labeled source data before deployment. SANTA [3] also uses source-prototype alignment. We follow this setting: the pretrained model and precomputed class references remain available, while source training examples are not accessed during target adaptation.

Our classification experiments use the same source information as GOLD. We compute the fast-subspace statistics and batch descriptors from the unlabeled target stream and the retained prototypes. The teacher supplies pseudo-labels and soft targets without additional annotations. Under these conditions, DTS-META consistently reduces error on all three classification benchmarks without requiring additional source examples, labels, or source-derived statistics.

References

  • [1] Lai, G., et al. The Golden Subspace: Where Efficiency Meets Generalization in Continual Test-Time Adaptation. CVPR, 2026.
  • [2] Döbler, M., Marsden, R. A., and Yang, B. Robust Mean Teacher for Continual and Gradual Test-Time Adaptation. CVPR, 2023.
  • [3] Chakrabarty, G., Sreenivas, M., and Biswas, S. SANTA: Source Anchoring Network and Target Alignment for Continual Test Time Adaptation. TMLR, 2023.

Thank you for your responses. My main concern remains that the contribution appears to be an incremental extension of GOLD. In table G2, the small gap between fixed descriptor and full also raises questions about the benefit of the contributions. Could a fixed descriptor offer comparable performance without full? Futhermore, why must slow and fast modules be applied sequentially in this order? Could they be reversed, applied in parallel, or combined into a single module?

Reviewer_P9s8OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes DTS-Meta, a dual-timescale feature adapter for continual test-time adaptation. It combines a slow AGOP-based class-sensitive pathway with a fast prototype-residual pathway whose modulation is generated by a MetaNet trained through first-order episodic optimization. Experiments on CIFAR-10-C, CIFAR-100-C, ImageNet-C, and CarlaTTA show consistent improvements over existing baselines, including GOLD.

Critical strengths
  1. The separation of persistent class-sensitive modulation and rapidly updated domain-sensitive modulation is well motivated and supported by subspace stability and environment-sensitivity analyses.

  2. The experiments cover three classification benchmarks, recurring and gradual shifts, ten repeated stream rounds, and semantic segmentation, providing broader validation than a single corruption protocol.

  3. The ablation study progressively evaluates the slow pathway, fast pathway, conditional modulation, and episodic optimization. The reported runtime is also close to that of GOLD despite the additional components.

Critical weaknesses
  1. This article is actually not bad, but its writing style bears all the hallmarks of current AI-generated content. If it is to be accepted, the author really needs to overhaul the structural framework and be mindful of the specific vocabulary used by AI—it’s truly painful to read.

  2. The claim that the two subspaces have distinct temporal roles is supported by diagnostics on CIFAR-100-C. It remains unclear whether the same properties hold across different architectures, datasets, and types of distribution shift.

  3. Performance degrades substantially at small batch sizes. On ImageNet-C, the error reaches 97.2% at batch size four, raising concerns about practical deployment under low-throughput conditions.

Questions for authors
  1. Can the authors clarify the methodological novelty of DTS-Meta relative to GOLD, conditional adapters, and existing episodic test-time adaptation methods?

  2. Are the improvements over GOLD statistically significant across paired stream seeds? Can uncertainty estimates be reported for the differences?

  3. Do the class-sensitive and domain-sensitive subspace properties persist across different datasets and backbone architectures?

  4. How does the support-query meta-learning objective compare with ordinary teacher consistency under an otherwise identical training procedure?

  5. Can the authors improve robustness at small batch sizes and evaluate the method on real-world, non-synthetic distribution shifts?

5 comments
AuthorsOpenReview ↗

Response to Reviewer P9s8(Part 1)

Thank you for taking the time to review our paper and providing valuable feedback. We appreciate your recognition of the motivation behind our design and the breadth of our evaluation. In response to your Weaknesses (W) and Questions (Q), we provide the following Explanations (E).

W1: Writing style and presentation

E1: Our AI Use Statement discloses the use of language polishing and grammar correction. We acknowledge that the manuscript’s presentation needed improvement. We have carefully revised the language and wording throughout the paper to reduce repetition, improve clarity, and express technical contributions more directly and precisely, with the changes marked in blue in the revised manuscript.

W2 & Q3: Generality of Dual-Subspace Dynamics Across Datasets and Architectures

E2: We extended the temporal-dynamics and environment-sensitivity analyses to CIFAR-10-C with WideResNet-28-10 and ImageNet-C with ResNet-50. Together with CIFAR-100-C and ResNeXt-29, the three tested combinations show the same qualitative pattern. The class-sensitive subspace has lower temporal drift. The prototype-residual subspace varies more across batches and captures more feature-change energy associated with environmental shifts.

Given the limited rebuttal space, we provide the additional plots in Appendices B.2 and B.3 of the revised manuscript. Figures 4 and 7 report temporal dynamics and energy capture on CIFAR-10-C. Figures 5 and 8 report the corresponding ImageNet-C results. These observations support the distinction across the tested dataset–backbone combinations.

W3 & Q5: Small-Batch Operational Regimes and Benchmark Selection

E3: Performance at very small batch sizes, such as B=4, is a common challenge for batch-based CTTA, as reported in GOLD [1] and CoTTA [2]. Small batches produce unreliable sample statistics and accelerate drift in teacher pseudo-labels during continuous unsupervised adaptation. DTS-META does not worsen this problem: Table 7 shows that it has the lowest error among the evaluated methods at every tested batch size. Our continual semantic segmentation experiments on CarlaTTA process one frame at a time and improve average mIoU from 54.3% to 55.3%. These results show that the framework can adapt stably from single frames when dense spatial cues are available.

We use CIFAR-10-C, CIFAR-100-C, and ImageNet-C because they are standard corruption benchmarks used by GOLD [1], CoTTA [2], OBAO [3], and SANTA [4]. This shared setup allows direct comparisons with prior work. Table G1 of our Common Response reports statistically significant gains over reproduced GOLD on all three datasets. We also evaluate recurring CDC shifts, gradual shifts, and long-horizon streams to assess robustness across different non-stationary conditions.

References

  • [1] Lai, G., Zhou, D.-W., Li, Z., and Ye, H.-J. The Golden Subspace: Where Efficiency Meets Generalization in Continual Test-Time Adaptation. CVPR, 2026.
  • [2] Wang, Q., Fink, O., Van Gool, L., and Dai, D. Continual Test-Time Domain Adaptation. CVPR, 2022.
  • [3] Zhu, Z., et al. Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation. ECCV, 2024.
  • [4] Chakrabarty, G., Sreenivas, M., and Biswas, S. SANTA: Source Anchoring Network and Target Alignment for Continual Test Time Adaptation. TMLR, 2023.
AuthorsOpenReview ↗

Response to Reviewer P9s8(Part 2)

Q1: Methodological Novelty Relative to GOLD, Conditional Adapters, and Episodic TTA

E4: The design follows our observation that classifier-sensitive and prototype-residual subspaces have different temporal behavior in the tested streams. We retain GOLD's AGOP-based persistent pathway [1], but GOLD relies solely on a single slow pathway and does not analyze domain-shift dynamics in feature space. We resolve this by coupling it with a fast prototype-residual covariance pathway, coordinating long-term stability with rapid batch-level responsiveness.

The fast descriptor is computed from residuals after slow modulation, as defined in Eq. 10. This makes the MetaNet's input depend on the feature state produced by the slow pathway. Conditional parameter generation is an established idea [2], but the generated scale and bias in our work act within the evolving fast subspace.

Each target batch and its augmented view form an online support–query episode. A temporary support step updates the MetaNet, and the query evaluates the resulting modulation. Our method can learn to learn: it evaluates a support-derived update on an augmented query view and uses this feedback to improve the modulation rule for online adaptation.

Tables G2 and G3 of our Common Response provide the controls. Full and Forward use the same MetaNet and two-view losses; their comparison tests the temporary inner update. The CDC experiment keeps the complete architecture fixed while varying the refresh periods, including the default Ts=50,Tf=5T_s=50, T_f=5. These comparisons support the episodic update and the chosen refresh arrangement at fixed parameter capacity.

References

  • [1] Lai, G., et al. The Golden Subspace: Where Efficiency Meets Generalization in Continual Test-Time Adaptation. CVPR, 2026.
  • [2] Perez, E., et al. FiLM: Visual Reasoning with a General Conditioning Layer. AAAI, 2018.

Q2: Statistical Significance and Paired Uncertainty Analysis Across Seeds

E5: Table G1 of our Common Response reports all five matched runs and the uncertainty of the GOLD−DTS-META differences. DTS-META has lower error in every pair. The three two-sided paired t-tests remain significant after Holm correction, with all adjusted p < 0.001. The pointwise intervals assume independent seed pairs and normally distributed differences. With only five pairs, exact two-sided sign-flip checks give p = 0.0625 per dataset before correction, so the significance statement depends on the statistical model.

Q4: Support–Query Meta-Learning versus Ordinary Teacher Consistency

E6: DTS-META-Forward is the teacher-consistency control described in Appendix B.7. It uses the same architecture, augmentation pipeline, and outer loss definitions as Full.

Forward uses the persistent MetaNet parameters ϕt\boldsymbol{\phi}_t for both support and query views:

LtForward=λce2[Ct(ϕt;Btsup)+Ct(ϕt;Btqry)]+At(ϕt,ϕt)\mathcal{L}_t^{\mathrm{Forward}} = \frac{\lambda_{\mathrm{ce}}}{2} \left[ \mathcal{C}_t(\boldsymbol{\phi}_t; \mathcal{B}_t^{\mathrm{sup}}) + \mathcal{C}_t(\boldsymbol{\phi}_t; \mathcal{B}_t^{\mathrm{qry}}) \right] + \mathcal{A}_t(\boldsymbol{\phi}_t, \boldsymbol{\phi}_t)

Full takes the temporary support step ϕt′=ϕt−ηin∇Ltinner\boldsymbol{\phi}'_t = \boldsymbol{\phi}_t - \eta_{\mathrm{in}} \nabla \mathcal{L}_t^{\mathrm{inner}} and evaluates the query with ϕt′\boldsymbol{\phi}'_t:

LtFull=λce2[Ct(ϕt;Btsup)+Ct(ϕt′;Btqry)]+At(ϕt,ϕt′)\mathcal{L}_t^{\mathrm{Full}} = \frac{\lambda_{\mathrm{ce}}}{2} \left[ \mathcal{C}_t(\boldsymbol{\phi}_t; \mathcal{B}_t^{\mathrm{sup}}) + \mathcal{C}_t(\boldsymbol{\phi}'_t; \mathcal{B}_t^{\mathrm{qry}}) \right] + \mathcal{A}_t(\boldsymbol{\phi}_t, \boldsymbol{\phi}'_t)

Here, Ct\mathcal{C}_t denotes the mean teacher–student consistency loss on one view, using fixed EMA-teacher targets. At\mathcal{A}_t contains the shared prototype-contrastive and regularization terms. Other trainable parameters are omitted from the notation. Only the MetaNet takes the temporary inner step; the outer update also updates the other persistent parameters. The query supplies feedback on whether the support step remains useful after augmentation. The five paired runs favor Full on each benchmark, and Table G2 summarizes their errors. This comparison supports the inner update while holding architecture and augmentation fixed.

Dear Reviewer P9s8,

Thank you very much for your constructive comments on our work. We have made every effort to address your concerns by providing detailed explanations, additional experiments, and corresponding revisions in the updated manuscript. As the discussion period is nearing its conclusion, we would like to gently follow up to see if our rebuttal has satisfactorily resolved your concerns, or if there are any further questions we can clarify. Thanks for your time and efforts in reviewing our work, and we sincerely look forward to your feedback.

Best regards,

Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common Response to Area Chair and All Reviewers

Dear Area Chair and Reviewers,

We thank the reviewers for their detailed feedback. Several concerns involve shared comparisons, so we collect the corresponding evidence in this Common Response. Table G1 reports the five-seed comparison with reproduced GOLD and the paired uncertainty estimates. Table G2 summarizes the classification ablations. Table G3 reports the five-seed refresh-schedule experiment under CDC. Our individual replies refer to these tables and provide the definitions and additional paired analyses relevant to each concern.

Table G1: Five-seed comparison of DTS-META with reproduced GOLD under standard CTTA.

ItemCIFAR-10-CCIFAR-100-CImageNet-C
Seed 114.96 / 14.3528.60 / 27.9359.36 / 58.83
Seed 215.08 / 14.5028.63 / 27.8759.58 / 58.84
Seed 315.18 / 14.5728.52 / 27.8259.48 / 58.78
Seed 414.93 / 14.3728.52 / 27.8059.44 / 58.73
Seed 515.05 / 14.4628.68 / 27.9359.14 / 58.57
Mean ± std15.04 ± 0.10 / 14.45 ± 0.0928.59 ± 0.07 / 27.87 ± 0.0659.40 ± 0.17 / 58.75 ± 0.11
Mean reduction (pp)0.5900.7200.650
Paired 95% CI (pp)[0.564, 0.616][0.674, 0.766][0.534, 0.766]
Holm-adjusted p1.20 × 10⁻⁶3.24 × 10⁻⁶1.00 × 10⁻⁴

Entries report GOLD / DTS-META error rates (%, lower is better; mean ± std over 5 seeds). Reductions are paired differences (pp) with pointwise 95% CIs (n=5, df=4). Paired t-test p-values are Holm-adjusted. GOLD is reproduced via its official implementation and configurations.

Table G2: Five-seed classification ablations under standard CTTA.

VariantCIFAR-10-CCIFAR-100-CImageNet-C
Fast-only14.85 ± 0.0928.23 ± 0.0759.27 ± 0.19
Full w/o Slow14.66 ± 0.0828.05 ± 0.1159.19 ± 0.18
Fixed descriptor14.52 ± 0.0928.00 ± 0.0658.87 ± 0.13
Forward14.70 ± 0.1228.09 ± 0.0858.96 ± 0.15
Raw-cov14.57 ± 0.0828.09 ± 0.0658.94 ± 0.14
Full14.45 ± 0.0927.87 ± 0.0658.75 ± 0.11

Entries are error rates in percent, reported as mean ± std over five matched seeds.

Variant Definitions and Mathematical Formulations:

  • Fast-only: Removes the slow pathway and MetaNet. Persistent fast scale sfast\boldsymbol{s}^{\mathrm{fast}} and bias bfast\boldsymbol{b}^{\mathrm{fast}} are updated directly via Adam without episodic inner loops:

(st+1fast,bt+1fast)=(stfast,btfast)−ηf∇(stfast,btfast)(λceLtce+λconLtcon)(\boldsymbol{s}_{t+1}^{\mathrm{fast}}, \boldsymbol{b}_{t+1}^{\mathrm{fast}}) = (\boldsymbol{s}_t^{\mathrm{fast}}, \boldsymbol{b}_t^{\mathrm{fast}}) - \eta_f \nabla_{(\boldsymbol{s}_t^{\mathrm{fast}}, \boldsymbol{b}_t^{\mathrm{fast}})} \left( \lambda_{\mathrm{ce}}\mathcal{L}_t^{\mathrm{ce}} + \lambda_{\mathrm{con}}\mathcal{L}_t^{\mathrm{con}} \right)

  • Full w/o Slow: Removes only the slow pathway (Ftslow,v=Ftv\mathbf{F}_t^{\mathrm{slow},v} = \mathbf{F}_t^v), applying fast modulation directly to raw backbone features while preserving the full MetaNet, bounded coefficients, and episodic support–query training:

Ftout=Ft+[(FtVtfast)⊙1B(stfast)⊤+1B(btfast)⊤](Vtfast)⊤\mathbf{F}_t^{\mathrm{out}} = \mathbf{F}_t + \left[ (\mathbf{F}_t \mathbf{V}_t^{\mathrm{fast}}) \odot \mathbf{1}_B(\boldsymbol{s}_t^{\mathrm{fast}})^\top + \mathbf{1}_B(\boldsymbol{b}_t^{\mathrm{fast}})^\top \right] (\mathbf{V}_t^{\mathrm{fast}})^\top

  • Fixed descriptor: Tests direct descriptor conditioning by replacing the dynamic batch descriptor utv\boldsymbol{u}_t^v with a constant vector c=12rf+2\boldsymbol{c} = \mathbf{1}_{2r_f+2}:

(s~t,v,b~t,v)=Hϕtv(c),v∈sup,qry(\tilde{\boldsymbol{s}}_{t,v}, \tilde{\boldsymbol{b}}_{t,v}) = \mathcal{H}_{\boldsymbol{\phi}_t^v}(\boldsymbol{c}), \quad v \in \\{\mathrm{sup}, \mathrm{qry}\\}

  • Raw-cov: Replaces the prototype-residual covariance with standard raw-feature covariance for fast-basis estimation:

Cov⁡i(ft,i−py^t,i)⟶1B−1∑i=1B(ft,i−fˉt)(ft,i−fˉt)⊤\operatorname{Cov}_i(\boldsymbol{f}_{t,i} - \boldsymbol{p}_{\hat{y}_{t,i}}) \quad\longrightarrow\quad \frac{1}{B-1}\sum_{i=1}^{B} (\boldsymbol{f}_{t,i} - \bar{\boldsymbol{f}}_t)(\boldsymbol{f}_{t,i} - \bar{\boldsymbol{f}}_t)^\top

  • Forward: Uses the persistent MetaNet parameters ϕt\boldsymbol{\phi}_t for both views, without a temporary support update (ηin=0\eta_{\mathrm{in}}=0).

Table G3: Five-seed refresh-schedule ablation on ImageNet-C under CDC.

Schedule (Ts,TfT_s, T_f)Seed 1Seed 2Seed 3Seed 4Seed 5Mean ± std
Both Frequent (5, 5)58.7658.4858.8858.5858.1058.56 ± 0.30
Both Infrequent (50, 50)58.9158.5858.8858.6858.2058.65 ± 0.29
Inverted (5, 50)58.8558.5058.9958.8558.4158.72 ± 0.25
Default (50, 5)58.4058.1958.4358.3058.0858.28 ± 0.15

Ts,TfT_s, T_f denote slow and fast basis-refresh periods in batches, and only these periods vary.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.