acceptodds
Under review as a conference paper at ICLR 2027

Recurrent Domain Incremental Adaptation of Time-Series Foundation Models

Abstract

Pretrained Time-Series Foundation Models (TSFMs) provide strong representations for downstream classification tasks. Yet adapting a TSFM under changing operating conditions remains challenging. A model must improve on each newly observed condition without sacrificing conditions it has already learned, often from only a small labeled calibration set and without knowing the operating condition at inference time. We find that this difficulty has an important geometric structure. Under an operating-condition shift, the embedding space of a TSFM does not move uniformly: some local regions change substantially, while others remain nearly unchanged. Consequently, a global adaptation can unnecessarily move embeddings that were already correctly classified. To address this, we propose the novel PACT (Piecewise Adaptation via Chart Transport). PACT partitions the frozen source embedding space into local charts and learns lightweight residual transports for the regions affected by each new condition. An explicit identity path is always retained, and a learned gate softly selects among the identity and accumulated transports for each input. The pretrained TSFM encoder, classifier head, and previously learned transports remain fixed. With only 5% labeled calibration data per arriving condition and no condition identity at test time, PACT maintains above 0.95 average macro-F1 across a four-condition bearing-fault stream. Under homogeneous tasks and label spaces, similar retention-adaptation behavior is observed across public benchmarks.

Primary area:
learning on time series and dynamical systems
Keywords:
domain adaptation, time series, classification, foundation models

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

This paper studies how time-series foundation models continually adapt to new domains in classification tasks. In this setting, new domains have limited labels, and the domain identity is unknown during inference. The model must adapt to new domains without losing accuracy on old ones. To solve this, the authors propose PACT. PACT splits the frozen feature space into several local regions. For each new domain, it learns a region-specific, low-rank residual transformation. It then uses a gating mechanism with an identity branch to combine transformations from different domains. The authors test PACT on five datasets and eight backbone models. Across all five datasets, PACT beats the baselines in final average Macro-F1 and maintains low forgetting.

Critical strengths
  • The research problem has practical value. The paper focuses on how time-series foundation models adapt and avoid forgetting when new domains continually arrive and old domains may reappear. It also considers real-world limits where new domains have few labels and the domain identity is unknown during inference. Compared to traditional one-time domain adaptation, this setup better fits the needs of long-term model deployment.

  • The method is well-motivated and clearly structured. The authors observe that domain shift affects different feature regions to different degrees, and based on this, they introduce local low-rank residual transformations. They also freeze past transformations and learn soft routing to balance adapting to new domains and retaining old ones. Furthermore, an explicit identity branch makes it structurally possible to keep the original predictions, making the overall design reasonable.

Critical weaknesses
  • The core novelty over existing modular continual learning methods (e.g., S-Prompts, SEMA) is limited. PACT mainly differs by using a fixed local partition with region-specific residual updates, but it does not clearly separate the gains of local geometric modeling from increased model capacity.
  • The theoretical claim for low-rank transformations needs more care. In Figure 7, rank-8 matches or beats full-rank updates, but since the frozen linear head has at most 6 classes, rank-8 can already express any linear logit update. Thus, this result shows parameter efficiency rather than an intrinsic low-dimensional structure of local domain shifts.
Questions for authors
  • Compared to existing domain-specific modular methods like S-Prompts and SEMA, what unique capabilities does PACT's fixed local partition provide? Given that local region count and model capacity increase together in Figure 3(b), can the authors separate the gains of local partitioning from those of added model capacity?
  • Since the gating network is trained only with domain labels, how does it decide whether a specific sample needs correction? Can the authors analyze how predictions change after adaptation for originally correct versus incorrect samples, along with their identity branch weights, to verify this selective protection?
  • Since the frozen classification head outputs at most 6 classes, how do the authors interpret the comparison between rank-8 and full-rank updates? To what extent does this result truly support the claim that local feature shifts have an intrinsic low-dimensional structure?
  • What class sampling strategy is used for the calibration samples? Can PACT maintain performance when some classes in the new domain have no labels at all or suffer from severe class imbalance?
3 comments
AuthorsOpenReview ↗

Response to Reviewer mUHi 1/2

Thank you for the time and care you put into evaluating PACT. We especially appreciate your question about whether the gains come from the placement of the local charts or from the additional parameters. It motivated the equal-capacity comparison we report below. Your observation about the frozen classifier also helped us interpret the rank ablation more precisely.

Weakness 1 / Q1. Does local geometry contribute beyond modular capacity?

S-Prompts uses domain-specific prompts, and SEMA adds adapters inside the backbone. PACT operates on the final embedding and indexes each correction by both source region and arriving condition. The encoder, original classifier, and earlier corrections stay frozen as new corrections are learned. Our comparison with SEMA is architectural.

We held the number of transports and gates fixed within each comparison and replaced source k-means with bins from a fixed random projection and source quantiles. The table gives AMF1 averaged over four streams.

Dataset and chart countSource k-meansRandom projectionStreams favoring k-means
FD, K=80.95750.89364/4
HHAR, K=40.83670.78394/4
WISDM, K=40.87740.80784/4
UCI-HAR, K=40.78640.75534/4
UCI-HAR, K=80.77250.70234/4

K-means scored higher in all 20 stream-by-configuration pairs. These are five configurations on four datasets, with the UCI-HAR streams evaluated twice. The differences range from approximately 3.1 to 7.0 points. At K=4 on UCI-HAR, the random partition had no masked cells.

Q2. Does a domain-trained gate identify incorrect predictions?

The gate is supervised by preserved chart information of previous increments combined with the new incoming calibration labels. We measured preservation and correction separately on one FD order, C→B→A→D.

a. Among initially correct source predictions, 890/891 were preserved.

b. Among initially incorrect target predictions, 889/1,042 (85.3%) became correct.

c. Among initially correct target predictions, 10/1,661 (0.60%) became incorrect.

The median identity weight was approximately zero both for target examples that stayed correct and for those that became incorrect. The observed behavior supports successful target correction and source retention on this stream. On source examples, median identity weight was 1.0 among the 890 preserved correct predictions and 0.894 for the one damaged prediction. The target medians pool three conditions. They support a distinction between source and target routing here and do not establish that the gate detects its own errors.

On HHAR, the median source identity weight was 1.00 for preserved correct predictions and 0.06 for correct predictions that changed to wrong. On UCI-HAR, the corresponding medians were 0.996 and 0.03. Thus, lower identity weight was associated with damaged source predictions on both activity datasets. The association alone cannot isolate the contributions of incorrect routing, transport size and initial classification margin.

Weakness 2 / Q3. What does rank 8 establish?

We thank the reviewer for the careful evaluation and agree that the frozen head limits what this ablation can tell us. Our tasks have at most six classes, so the prediction depends on at most five independent logit differences. A shift can affect many embedding dimensions while changing only these few quantities at the classifier. In our experiments, rank-8 transports achieved similar AMF1 to full-rank transports with fewer trainable parameters. We will revise Sec. 4.3 to report this as evidence of parameter efficiency and remove the claim that the underlying feature shifts are intrinsically low-dimensional.

AuthorsOpenReview ↗

Response to Reviewer mUHi 2/2

Q4. What happens when calibration labels are imbalanced or missing?

Our original experiments used class-stratified calibration sets shared by all labeled methods. We kept the label set identical across all baselines for comparison. We kept the total label budget fixed and varied which classes received labels.

Calibration settingMetricDomain ordersFDUCI-HAR
Class-stratifiedAMF140.95750.7864
85% of labels from one classAMF140.86670.7726
Class 0 omittedClass 0 recall40.9950.154
Class 1 omittedClass 1 recall10.1380.016

Imbalance reduced overall AMF1 by 9.08 points on FD and 1.38 points on UCI-HAR. Missing labels had a class-specific effect. On FD, PACT still recognized 99.5% of class-0 examples when that class was omitted, averaged over four orders. On UCI-HAR, recall was low for every omitted class tested, ranging from 1.6% to 15.4%. Only the class-0 omission was repeated across all four orders.

PACT learns corrections by region and condition. An omitted class can therefore receive a correction learned from labeled examples of other classes in the same region. The FD result shows that a class can remain recognizable without new labels, while the UCI-HAR results show why reliable recall cannot be assumed.

In the revised version, we will discuss this behavior more explicitly.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer mUHi,

Thank you again for the time and care you have put into reviewing our work. Your questions helped us strengthen the controls for state matching, examine potential side effects, and clarify how the offline training objective relates to next-step routing.

As the discussion period draws to a close, we would be very grateful if you could share your thoughts on our responses and the revised manuscript. If any concerns remain, we would welcome the opportunity to address them. We sincerely appreciate your time and your help in improving the paper.

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

Under recurring operating-condition shifts, pretrained time-series foundation models typically struggle to accommodate all conditions while adapting. The authors attribute this to global updates, which tend to overwrite conditions the model already handles well. Based on this diagnosis, they propose PACT, which partitions the frozen embedding space into k-means charts, learns a low-rank residual transport per chart for each new condition and then freezes it, and uses a per-chart soft gate to decide how much correction each input receives. In the experiments, the authors evaluate PACT on five datasets against replay, test-time/source-free adaptation, regularization, and domain-incremental baselines, test it across eight backbones, and conduct ablations on its components, transport rank, routing strategy, and label budget. Overall, the paper presents a clear motivation and a well-defined target problem, and the method is supported by extensive experiments.

Critical strengths
  1. The paper is well motivated, and the method follows naturally from it. Fig. 3 shows that the shift is local in the begining, and then the use of local charts, frozen transports, and an identity-preserving gate all make intuitive sense.

  2. The experiments are sufficient. PACT performs better on five datasets and works well across eight different backbones.

Critical weaknesses
  1. The observation that condition shifts degrade embedding regions unevenly, and that global updates tend to overwrite these regions, is not entirely new. In continual learning, many works have studied representation distortion and proposed methods such as prototype augmentation [1], which share the core idea of updating the embedding space locally.

[1] F. Zhu, X.-Y. Zhang, C. Wang, F. Yin, and C.-L. Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5867–5876, 2021.

  1. PACT assumes that each condition shifts only within fixed source charts and can be fixed by a small local correction, and similar assumptions have been questioned in continual learning: clustering in euclidean space may fit poorly for non-stationary data [2], cluster partitions fixed on old data gradually go stale as the distribution moves, and methods that freeze the backbone representation also rely heavily on a strong initial backbone and lose plasticity when it is weak.

[2] Goswami, D., Liu, Y., Twardowski, B., & Van De Weijer, J. (2023). Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning. Advances in Neural Information Processing Systems, 36, 6582-6595.

  1. It is unclear whether, or to what extent, the benchmarks actually contain recurring condition shifts. The central claim that adaptation does not hurt source performance needs to be reported directly.

  2. Some design choices need further justification. The chart count is fixed at K = 4 for all datasets, even though Fig. 3b suggests that performance depends strongly on K, so the authors should explain how K is chosen. Ablations on the gate temperature and the masking threshold are also needed.

Questions for authors
  1. How does the observation of non-uniform degradation differ from prior findings on representation drift and local prototype updating in continual learning? Can the authors clarify whether there is any fundamental difference between PACT and prototype-augmentation methods?
  2. Can the authors report source-domain performance after each adaptation step, to show directly that it is preserved? Maybe run an experiment that simulates the scenario described in the Introduction, where earlier conditions actually reappear in the stream after new ones.
4 comments
AuthorsOpenReview ↗

Response to Reviewer FeuE 1/3

Thank you for your thoughtful review and for pointing us to PASS[1] and FeCAM[2]. We appreciate your questions about how PACT differs from these methods and what happens when an earlier operating condition returns. They prompted us to compare the methods directly and report how well PACT retains performance on the source condition.

Weakness 1 / Q1. What distinguishes PACT from prototype methods?

PACT addresses a post-deployment problem: how can a time-series foundation model continue to recognize the same classes as operating conditions change, when only a few labeled examples are available for each new condition and the condition may be unknown at inference? We agree that earlier work established that distribution shifts can affect feature space unevenly. Our contribution is a way to use the structure already present in a pretrained TSFM’s embeddings to adapt its predictions locally while preserving access to the original model.

Consider a bearing monitored under a new machine load. The fault labels are the same, and the encoder may still capture features that distinguish them. Yet the new load can move some vibration patterns across a decision boundary learned under the source condition, while the classifier continues to recognize other patterns correctly. PACT freezes both the encoder and the source prediction head. From the few labeled examples of each arriving condition, it learns a low-rank correction for each supported region of the source embedding space. At inference, a gate uses the input embedding to blend the available corrections with an identity path before applying the original head. This lets two signals from the same operating condition receive different corrections and keeps earlier corrections available when a condition returns.

The distinction from prototype augmentation is the role of the local region. PASS samples around stored class prototypes to provide training examples for updating a classifier as new classes arrive. FeCAM uses class means and covariances to assign the predicted label, including in its domain-incremental setting. In PACT, a region determines how an incoming signal is corrected before the frozen classifier assigns its label. On bearing-fault diagnosis, with 5% labeled calibration data and four charts, PACT achieved 0.9507 AMF1, compared with 0.8392 for the frozen-feature PASS adaptation and 0.9195 for FeCAM, averaged over three seeds and four domain orders. This result illustrates the practical value of learning local corrections from limited calibration data while retaining the deployed classifier.

Direct comparison with the two cited methods. We evaluated a frozen-feature adaptation of PASS and a FeCAM classifier on the same Sundial embeddings, four streams, calibration budgets, and three seeds. The PASS variant uses prototype augmentation. Freezing the encoder disables feature distillation, and we omitted the original image-rotation objective. The FeCAM variant uses one set of class statistics pooled across observed domains, with covariance shrinkage and normalization. These results characterize the implementations in this time-series protocol.

DatasetPACT at common K=4PASS adaptationFeCAM pooled
FD0.95070.83920.9195
WISDM0.85760.80840.8297
UCI-HAR0.79410.76290.7939

Across four charts, PACT exceeds FeCAM by 3.12 points on FD and 2.79 on WISDM. FeCAM leads by 3.99 on HHAR, and the UCI-HAR difference is 0.02 points. The empirical distinction is clearest on FD. We retain embeddings and domain assignments for routing, whereas the prototype methods retain class statistics.

[1] F. Zhu, X.-Y. Zhang, C. Wang, F. Yin, and C.-L. Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5867–5876, 2021.

[2] Goswami, D., Liu, Y., Twardowski, B., & Van De Weijer, J. (2023). Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning. Advances in Neural Information Processing Systems, 36, 6582-6595.

AuthorsOpenReview ↗

Response to Reviewer FeuE 2/3

Weakness 2. Euclidean assignment, aging charts, and backbone plasticity.

Euclidean distance in PACT assigns the region whose transforms will be evaluated. The frozen head makes the class decision after correction. This addresses a different use of distance from nearest-prototype classification. If target samples far from source support enter the same chart, they share its bank of corrections even when they would benefit from different maps. Likewise, if the encoder maps two differently labeled signals to the same embedding, PACT has no input information with which to distinguish them. On UCI-HAR at four charts, source k-means achieved 0.7864 AMF1 versus 0.7553 for a fixed random-projection partition at equal capacity. The k-means partition scored higher on all four tested streams. These results show where the choice of partition helped. They leave open how performance changes as future embeddings move progressively farther from source coverage. The encoder’s feature quality and the coverage of fixed source charts remain limitations.

Our focus on pretrained TSFMs is important because the encoder has already learned temporal features from a broad time-series corpus before deployment. As operating conditions change, incoming signals may occupy different regions of this fixed embedding space, and the source prediction head may become less reliable in some regions. PACT uses a small labeled calibration set to learn local corrections where the pretrained features still distinguish the classes. Our backbone comparison illustrates why representation quality matters: on UCI-HAR, PACT achieved 0.922 AMF1 with Mantis and 0.517 with PatchTST. The TSFM-specific contribution is therefore a way to reuse pretrained features under changing conditions while preserving the deployed encoder and classifier.

Weakness 3 / Q2. How much source performance is retained?

The table gives the final change from initial source macro-F1 in percentage points, averaged over four streams. A value of −0.02 means a 0.02-point decline.

DatasetPACTEmbedding replayHead fine-tuning
FD−0.02−0.33−31.76
HHAR−0.10−1.13−10.38
WISDM−1.40−10.22−34.70
UCI-HAR−6.16−9.64−38.09
Sleep-EDF−0.14−4.39−15.26

We extracted source macro-F1 after every adaptation from the first column of each run’s evaluation matrix. The table reports the mean over the four Table 1 streams. “Start” is the initial source macro-F1. Each numbered column gives the change from that starting score in percentage points after learning one more condition.

DatasetStart F1Step 1Step 2Step 3Step 4Step 5Step 6Step 7Step 8
FD0.9876−0.06−0.02−0.02—————
HHAR0.9673−0.13−0.20−0.05−0.09−0.090.00−0.24−0.10
Sleep-EDF0.7591+0.11−0.24−0.10−0.24−0.43−0.30−0.53−0.14
WISDM0.9345+0.03−0.39−0.50−0.69−0.52−0.72−1.12−1.40
UCI-HAR0.9357−2.90−3.61−4.71−6.68−5.27−5.92−4.92−6.16

On FD, HHAR, and Sleep-EDF, mean source F1 stayed within 0.53% points of its initial value at every step. WISDM showed a gradual decline, ending 1.40% points below its starting value. UCI-HAR showed the clearest limitation: source F1 had already fallen 6.68% points by the fourth update and ended 6.16% points below its starting value. Across all five datasets, PACT had a smaller final source decline than embedding replay and head fine-tuning.

The benchmark streams evaluate the source after every update. To test the return of an earlier operating condition explicitly, we also constructed an FD revisit stream. After adapting from A to B and C, macro-F1 on held-out A was unchanged at 0.9748. After adapting to D, macro-F1 on held-out B was unchanged at 0.9137. Both revisits used the learned gate without a condition ID.

Figure 12 shows how to read the trajectory for an individual stream. Its first column is the source score at each update. The table above gives endpoints across four streams. Figure 9 shows how the average performance at each incremental learning step varies compared to baselines clearly for some example chains.

Recurrence. Our benchmark re-evaluates previously seen conditions after each update. On a simulated FD revisit with no condition ID at inference, A scored 0.9748 after learning B and C, and B scored 0.9137 after learning D. These probes establish retention on held-out examples from returning conditions. Example: The public subject streams do not record natural chronological recurrence.

AuthorsOpenReview ↗

Response to Reviewer FeuE 3/3

Weakness 4. How sensitive is PACT to chart count, gate temperature, and masking threshold?

Chart count KK. We will add a systematic choice to clarify how we chose KK in the revised version. For each dataset, we select the value with the highest validation-split performance. This makes the selection exploratory. Below we report the full sweep on frozen Sundial. The K=4K=4 column gives a common setting across all five datasets.

DatasetK=1K=1K=4K=4K=8K=8
FD0.93860.94840.9575
HHAR0.83180.83690.8090
WISDM0.87170.87740.8420
UCI-HAR0.77430.78620.7725

The identity path and learned gate were present at every KK. Moving from one global chart to four improved AMF1 on FD and all three activity datasets.

Gate temperature tautau. With the minimum support fixed at m=5m=5, we tested tauin{0.25,0.5,1.0}tau in \{0.25, 0.5, 1.0\}. FD was insensitive over this range (0.9589,0.9591,0.95990.9589, 0.9591, 0.9599 AMF1). UCI-HAR benefited from softer routing (0.7834,0.7911,0.80480.7834, 0.7911, 0.8048), gaining 1.37 AMF1 points when tautau increased from the default 0.50.5 to 1.01.0.

Masking threshold mm. PACT enables a condition-specific correction in a chart once it has at least mm calibration examples. At fixed tau=0.5tau=0.5, we tested min{1,5,10}m in \{1,5,10\}. The default m=5m=5 gave the highest AMF1 on both FD (0.9585,0.9591,0.95170.9585, 0.9591, 0.9517) and UCI-HAR (0.7860,0.7911,0.78340.7860, 0.7911, 0.7834). Requiring ten examples reduced AMF1 by 0.74 and 0.77 points, respectively. Both sensitivity sweeps used K=8K=8 on FD and K=4K=4 on UCI-HAR, varying one parameter at a time.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer FeuE,

Thank you again for the time and care you have put into reviewing our work. Your questions helped us strengthen the controls for state matching, examine potential side effects, and clarify how the offline training objective relates to next-step routing.

As the discussion period draws to a close, we would be very grateful if you could share your thoughts on our responses and the revised manuscript. If any concerns remain, we would welcome the opportunity to address them. We sincerely appreciate your time and your help in improving the paper.

Reviewer_af6YOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies domain-incremental time-series classification on top of a frozen foundation-model encoder and a frozen linear head, and proposes PACT, which partitions the source embedding space into k-means charts, learns a rank-8 residual transport per chart for each arriving condition from a small labeled calibration set, and uses a per-chart gate trained as a domain classifier to softly mix the identity branch and the accumulated transports at inference without condition identity.

Critical strengths

Two critical strengths are listed below.

  1. The routing ablation in Figure 7 (right) compares learned soft routing, hard argmax routing, and oracle routing with the true condition identity, presumably with the same charts, transports, and head (the paper does not state this explicitly). The gap between soft and oracle routing therefore isolates the error that comes from routing rather than from transport capacity or from changes to the frozen host. This decomposition, together with the identity branch that reproduces the frozen prediction exactly when the gate assigns all mass to it, makes the source of the remaining error explicit and is the most informative part of the evaluation.

  2. All Table 1 baselines operate on the same frozen Sundial embeddings and receive the same domain orders, splits, and calibration budgets, and Table 1 reports retention (RF, PRF) alongside AMF1. Figure 6 further separates adaptation from retention across the frozen model, head fine-tuning, ungated transports, and full PACT on all five datasets. This makes it possible to see that ungated transports already raise AMF1 and that the gate is what removes most of the forgetting, which supports the stated complementarity of the two components.

Critical weaknesses

Three critical weaknesses are listed below.

  1. The central claim that correction must be local is not isolated within the full method. Figure 3(b) and 3(c) motivate charts by fitting local affine maps and a target-fitted global map, apparently without the identity branch or the gate (the protocol of Figure 3 is not described), on a dataset and backbone that are not named. The ablation in Figure 6 removes the transports or the gate but never sets K=1K = 1, that is, a single global rank-8 transport per condition combined with the same gate and identity branch. The reported gains could therefore come from residual low-rank transports with gated identity routing alone rather than from chart partitioning. The number of charts is also inconsistent: Section 4.1 states K=4K = 4 for all datasets, whereas the released code and README use K=8K = 8 for the FD row of Table 1 (the README's single-seed, four-order mean of 0.9575 AMF1 equals the Table 1 and Table 2 entries to four decimals), and no sensitivity to KK is reported for the full method.

  2. The comparison with the closest baselines does not control for memory. PACT retains all source training embeddings and all calibration embeddings with their domain indices for gate refitting, which is p (∣Z0∣+∑d∣Zd∣)p\,(|Z_0| + \sum_{d} |Z_d|) floating-point values. The paper calls DER++ "memory-matched" and states that embedding replay and DER++ "have more privileges" because they retain class targets, but it does not report the buffer sizes of either replay method relative to the embeddings PACT retains. Since embedding replay is second on four datasets and within 1.5 to 4.8 points of PACT, the stated margin is not attributable to the method until the retained sets are matched. In addition, CoTTA, EATA, S-Prompts, and DualCD are "adapted to this embedding/head setting, with details in the supplement", but Appendix C devotes a single sentence to this and gives no details; S-Prompts and DualCD modify the backbone or prompt pool in their original form, so the adaptation determines whether Table 1 compares methods or re-implementations. The statement that PACT is "the only method that avoids both failure modes" is also not supported by Table 1, where embedding replay keeps RF at or below 0.065 on every dataset (0.0641 on WISDM and 0.044 or lower elsewhere).

  3. The experimental protocol and the released code do not allow the main table to be reconstructed. WISDM, UCI-HAR, and Sleep-EDF contain 36, 30, and 20 domains, but each stream uses only nine conditions; the paper says only that the streams exhibit "observed relative shifts from the source condition", and neither the rule for choosing the nine nor the full set of four "randomly drawn" orders per dataset is given (one or two orders per dataset can be read off Figures 9 and 12, and the FD orders appear only in the README). Calibration budgets differ by dataset (5% on FD, HHAR, and Sleep-EDF; 450 labels on WISDM; 100 on UCI-HAR) without justification. Figure 2 uses MOMENT-Large, Figure 8 uses Mantis, Figure 11 uses TSLANet, and Table 1 uses Sundial, so the motivating and label-efficiency results cannot be compared directly with Table 1; Figure 2 additionally uses the order A, B, C, D, which is not among the four FD orders listed in the README. The anonymous repository reproduces only the PACT row on FD with a single seed (42) and no baselines, and its regularizer penalizes ∥BA∥F2\|BA\|_F^2 rather than the ∥B∥F2\|B\|_F^2 in Eq. (3). Table 2 reports only the across-seed standard deviation (0.003 on FD), whereas the README shows that the across-order standard deviation under a single seed is 0.011; the paper does not report the latter for any dataset.

The experimental scope of this paper is broad (five datasets, sixteen comparison methods, eight backbones, and component, rank, and routing ablations), and the first two weaknesses can each be addressed with one additional experiment. I would be glad to raise my rating to 3 if the authors provide the K=1K = 1 ablation with the gate and identity branch retained (Question 1) and a comparison with embedding replay under matched retained-embedding budgets (Question 2), with the remaining protocol details clarified.

Questions for authors

Five questions are listed below. These questions may influence my understanding of the method and the experiment results.

  1. With the gate and identity branch unchanged, how does K=1K = 1 (a single global rank-8 transport per condition) compare with K=4K = 4 and K=8K = 8 on the five datasets in Table 1? Which KK was used for each dataset in Table 1, given that the released code uses K=8K = 8 for FD?

  2. How many embeddings do embedding replay and DER++ store on each dataset, and how does this compare with the number of source and calibration embeddings PACT retains for gate refitting? How does PACT perform when the retained source set used for gate training is subsampled to the replay buffer size?

  3. How were S-Prompts, DualCD, CoTTA, and EATA adapted to the frozen-embedding and linear-head setting? Specifically, which parameters are updated, what replaces the prompt pool and the causal disentanglement modules, and were their hyperparameters tuned under the same calibration budget as PACT?

  4. How were the nine domains and the four domain orders selected for WISDM, UCI-HAR, and Sleep-EDF, and why do the calibration budgets differ across datasets? Which dataset and backbone are used in Figure 3, and could Figures 2 and 8 be reported with the Sundial backbone used in Table 1?

  5. For each test domain, what fraction of samples receives identity weight π0>0.5\pi_0 > 0.5, what fraction of (chart, condition) cells is masked for having fewer than five calibration embeddings, and for what fraction of correctly classified source samples does the preservation condition of Proposition 1, stated after Eq. (6), hold? This would show whether the identity path is actually used and whether Propositions 1 and 2 describe the operating regime.

3 comments
AuthorsOpenReview ↗

Response 1/2 to Reviewer af6Y

Thank you for the time you spent examining both the manuscript and the released code. We especially appreciate your suggestions for testing the contribution of local charts and comparing PACT with replay under matched memory budgets. Your close reading also helped us identify and correct important inconsistencies and mistakes in our submitted version.

Weakness 1 / Q1. Does a local correction improve on a gated global correction?

We repeated the comparison with an identity path, learned soft routing, and rank-8 residuals at every chart count. At one chart, each arriving condition has one global transport.

DatasetK=1 AMF1K=4 AMF1K=8 AMF1RF at K=1 / 4 / 8
FD0.93860.94840.95750.0018 / 0.0021 / 0.0022
HHAR0.83180.83690.80900.0193 / 0.0171 / 0.0142
WISDM0.87170.87740.84200.0317 / 0.0292 / 0.0458
UCI-HAR0.77430.78620.77250.0549 / 0.0436 / 0.0462
Sleep-EDF0.61160.59830.59910.0229 / 0.0164 / 0.0140

Four charts improved AMF1 over one on FD and the three activity datasets. Sleep-EDF obtained its highest AMF1 with one chart and lower forgetting with more charts. However, our objective is not only higher accuracy but also lower forgetting; therefore, overall, more charts demonstrate this capability. The separate equal-capacity control in our response to mUHi tests whether the placement of chart boundaries also matters.

Response to mUHi. We compared source k-means with a random-projection partition at the same chart count and with identical transport and gate capacity. The table reports AMF1 averaged over four domain orders.

DatasetKSource k-meansRandom projectionOrders won
FD80.95750.89364/4
HHAR40.83670.78394/4
WISDM40.87740.80784/4
UCI-HAR40.78640.75534/4
UCI-HAR80.77250.70234/4

K-means won all 20 paired comparisons across five configurations on four datasets, with mean gains of 3.1–7.0 AMF1 points. On UCI-HAR at K=4, it gained 3.1 points even though the random partition had no masked cells. These results show that chart placement matters at equal capacity too.

Weakness 2 / Q2. What memory did the methods receive?

PACT and embedding replay retained the same source and calibration embeddings. PACT kept domain assignments to refit its gates and discarded past class labels. Replay kept class labels. DER++ used a reservoir of raw windows, class labels, and logits, sized against the final embedding-array byte budget for each stream. PACT’s gates, centroids, transports, and metadata add persistent storage beyond that budget.

DatasetHistorical embedding count per streamDER++ slotsFull-history PACT / replay AMF1, rerun comparison512-embedding PACT / replay AMF1
FD2,1043150.9575 / 0.90960.9556 / 0.9118
HHAR1,630–1,7083,193–3,3460.8367 / 0.86480.8252 / 0.8622
WISDM978–1,2801,916–2,5070.8774 / 0.86290.8206 / 0.8461
UCI-HARapprox. 1,026–1,066679–7050.7864 / 0.81690.7513 / 0.8078
Sleep-EDF1,927–2,524492–6440.6129 / 0.63570.6003 / 0.6255

We thank the reviewer af6Y and reword the claim that PACT alone avoids both poor adaptation and forgetting. As observed, embedding replay maintained RF lower across the original comparisons. Under forced eviction, PACT retained a 4.38-point AMF1 advantage on FD and had lower RF there by 0.0158. This comparison establishes the scope of the FD finding and PACT's sensitivity to a smaller history on the activity datasets.

Further, we would like to emphasize that we chose the embedding-storage architecture to respect privacy by not storing raw samples or raw labels; we store neither. However, we can maintain storage similar to the baselines (DER++ and replay) by following a similar mechanism and then encoding with the backbone when we use it; this does not change performance because the backbone is always frozen during incremental learning steps.

Weakness 2 / Q3. How were baselines adapted?

The S-Prompts-inspired implementation trains a linear head per domain and selects it by nearest domain mean. The DualCD-inspired implementation uses a learned linear feature mask and classifier with intra- and inter-domain losses. The EATA-inspired implementation updates the head using entropy minimization and a Fisher anchor. The CoTTA-inspired implementation uses noisy embeddings and an exponential-moving-average teacher. These are adaptations to frozen embeddings. In particular, the first implementation has no prompt pool. We will name the variants accordingly and specify their losses, update budgets, and optimization settings. The supervised head fits used the same calibration examples as PACT and ran for 80 epochs. Baseline hyperparameters were tuned separately with each validation split.

AuthorsOpenReview ↗

Response 2/2 to Reviewer af6Y

Weakness 3 / Q4. Reconstructing Table 1.

FD uses all four conditions, and HHAR uses all nine users. WISDM, UCI-HAR, and Sleep-EDF use four fixed nine-subject streams each, with subject composition as well as ordering varying between streams. The original selection was random; we then fixed the selection for each baseline at every seed. The exact stream lists are given below. The calibration budgets were 5% on FD, HHAR, and Sleep-EDF, up to 450 labels on WISDM, and up to 100 on UCI-HAR, sampled by class. This is due to the limited sample availability of some subjects (domains) in the activity datasets. This experimental choice enables a fair opportunity for each incremental subject (domain) to supply the same calibration samples.

Stream lists for Q4. Each parenthesis gives an ordered stream, with the first ID used as source.

DatasetFour streams
FD(0,2,3,1) · (2,0,1,3) · (2,1,0,3) · (2,1,3,0)
HHAR(1,6,4,7,8,5,0,2,3) · (5,1,6,3,2,7,4,0,8) · (5,8,0,3,2,6,4,1,7) · (6,0,8,2,5,1,3,4,7)
WISDM(10,24,20,0,3,1,22,14,2) · (12,23,7,9,5,15,31,28,10) · (24,11,0,23,13,21,16,35,3) · (30,3,0,31,20,24,19,13,17)
UCI-HAR(14,24,1,27,26,10,12,7,9) · (19,14,6,12,4,29,9,18,21) · (24,14,19,17,6,5,7,3,30) · (6,20,18,30,12,24,25,3,17)
Sleep-EDF(0,5,18,9,3,10,6,1,14) · (1,5,9,0,4,3,18,8,13) · (15,6,17,5,8,9,18,19,2) · (18,5,12,7,10,17,0,3,1)

Figure 3 used frozen TSLANet on FD with FD-A as source, 5% calibration, and full-rank affine maps fitted without routing. It motivates the methodology; Q1 tests locality within PACT on Sundial. Figure 2 used an illustrative fifth-order on MOMENT, and Figure 8 used Mantis. Figure 11 visualizes TSLANet representations. These cross-backbone figures will be interpreted within their respective experiments. The budgets are dataset-specific experimental settings. The actual WISDM count per arrival can be lower than 450 when fewer training examples are available.

The across-stream standard deviations of PACT AMF1 were 0.0106 on FD, 0.0214 on HHAR, 0.0197 on WISDM, 0.0144 on UCI-HAR, and 0.0316 on Sleep-EDF. On WISDM, UCI-HAR and Sleep-EDF, this variation reflects both ordering and subject composition. For reproducibility, we will supply the baseline specifications, stream lists, per-stream metrics, and aggregation procedure alongside the PACT configuration in the revised version.

Equation (3) specifies a penalty on B, while the released implementation penalized BA. A rerun with the squared Frobenius norm of B explicitly enabled reproduced the archived AMF1 values to within 0.0013 at the historical chart counts. We will make the training penalty explicit in the code and saved configurations.

Q5. Is the identity path used?

On the FD stream C→B→A→D, identity weight exceeded 0.5 for 99.89% of source examples and for 9.32%, 1.55%, and 10.99% of examples from B, A, and D, respectively. Eight of the 24 non-source chart–condition cells were masked. Proposition 1 certified 889 of the 891 originally correct source predictions. All 889 certified predictions were preserved, and 890 of the 891 originally correct predictions were preserved overall.

These measurements establish certificate coverage for Proposition 1 on this FD stream. Proposition 2 gives a separate sufficient condition for target correctness.

We extended the source certificate analysis to two activity datasets. FD uses the single order above at K=8. HHAR and UCI-HAR use all four orders at K=4. Coverage is the fraction of originally correct source predictions satisfying Proposition 1.

Source diagnosticFDHHARUCI-HAR
Identity weight above 0.599.9%97.5%77.9%
Certificate coverage99.8%97.5%71.1%
Predictions preserved among certified examples100%100%100%
Predictions preserved among all originally correct examples99.9%99.7%92.7%
Non-source cells masked at final step33.3%19.5%10.2%

Coverage answers how often the sufficient condition is useful in practice. UCI-HAR has the lowest coverage and the largest source-F1 decline among these datasets. Many uncertified examples still kept their original prediction.

Furthermore, in the revised version, we will make the appropriate edits to:

  1. updating Table 1 with the matched-memory replay results
  2. dropping or rescoping the claims of state-of-the-art performance and of necessary locality
  3. stating how K was selected for each dataset, basde on compute matched validation split performance.
AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer af6Y,

Thank you again for the time and care you have put into reviewing our work. Your questions helped us strengthen the controls for state matching, examine potential side effects, and clarify how the offline training objective relates to next-step routing.

As the discussion period draws to a close, we would be very grateful if you could share your thoughts on our responses and the revised manuscript. If any concerns remain, we would welcome the opportunity to address them. We sincerely appreciate your time and your help in improving the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Response: PACT’s Core Contribution, New Evidence, and Revisions

We thank the reviewers for their thoughtful feedback. Reviewer mUHi highlighted the practical importance of adapting from few labels when earlier conditions can return, FeuE recognized the motivation for local corrections and the breadth of the evaluation, and af6Y highlighted the shared evaluation protocol and the ablations separating adaptation from retention. Their questions led us to test PACT more directly against the closest alternatives.

Core Contribution. PACT addresses a post-deployment problem for emerging time-series foundation models. As operating conditions change, the same fault or activity can become harder to recognize even when the pretrained features remain informative. To our knowledge, PACT is the first study of few-label, recurrent domain-incremental classification that keeps both the TSFM encoder and its deployed classifier frozen and requires no condition identity at inference. PACT learns where within a condition to apply a correction. Two signals from the same machine load can receive different corrections, and the identity path preserves access to the original prediction when an earlier condition returns.

Additional Evidence and Clarifications

  1. The benefit of local correction survives an equal-capacity control. With the gate and identity path retained, four charts improved over a single global transport on FD and all three activity datasets. We then fixed the number of transports and gates and changed only the partition. Source k-means won on all four streams in each of five configurations across four datasets, gaining 3.11–7.02 average macro-F1 points over random partitions.
  2. PACT leads the closest alternatives on bearing-fault diagnosis. Following the reviewers’ suggestions, we compared PACT directly with prototype methods on the same frozen features and calibration examples. With 5% labels, four-chart PACT achieved 0.9507 average macro-F1, versus 0.9195 for pooled FeCAM[1] and 0.8392 for the frozen-feature PASS[2] adaptation. In a separate matched-memory comparison, PACT led embedding replay by 4.79 points with full history and 4.38 points when both retained at most 512 embeddings.
  3. The retention advantage extends across all five datasets. PACT showed a smaller final decline in source macro-F1 than both embedding replay and head fine-tuning on every dataset, including those where replay achieved higher overall scores. On one FD stream, PACT preserved 890 of 891 initially correct source predictions while correcting 889 of 1,042 initially incorrect target predictions. An explicit revisit also confirmed that held-out condition A retained macro-F1 of 0.9748 after learning B and C, without condition identity at inference.

Overall. The new controls establish that chart placement contributes at fixed capacity, that PACT’s strongest performance gain persists under matched storage, and that its source-retention advantage holds across all five benchmarks. We thank the reviewers for prompting these comparisons and provide the full results in the revised manuscript and individual responses.

[1] Goswami, D., Liu, Y., Twardowski, B., & Van De Weijer, J. (2023). Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning. Advances in Neural Information Processing Systems, 36, 6582-6595.

[2] F. Zhu, X.-Y. Zhang, C. Wang, F. Yin, and C.-L. Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5867–5876, 2021.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.