acceptodds
Under review as a conference paper at ICLR 2027

SPLIT: Separating Test Time Training Plasticity for Anomaly Detection and Localization

Abstract

Multivariate time series (MTS) anomaly diagnosis requires identifying both when anomalies occur and which variables are associated with them. Test Time Training (TTT) continuously adapts a predictive model at test time, but detection and localization can favor different levels of adaptation. We derive an exact recurrence describing how differences in TTT update rates propagate through the evolving fast state and relate the resulting prediction-error differences to the distinct ranking structures of detection and localization. To allow detection and localization to follow task-specific adaptation trajectories within the same trained predictive model, we propose Separated PLasticity In Test Time Training (SPLIT), which shares the encoder and learned initialization while maintaining separate fast states with task-specific update rates. Experiments on the SMD, SWaT, and WaDI datasets consistently show that detection favors lower update rates than localization. Across multiple MTS benchmarks, SPLIT achieves strong localization performance while maintaining competitive detection performance.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time Series, Anomaly Detection, Anomaly Localization, Test Time Training

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_StU6OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies how test-time training (TTT) affects two distinct objectives in multivariate time-series (MTS) anomaly diagnosis: detection (ranking timestamps) and localization (ranking variables within a timestamp). It derives exact recurrences that explain how changing the TTT update rate propagates through a fast-weight state and translates into different prediction-error trajectories, then shows empirically that detection consistently favors lower update rates while localization prefers higher ones. Then, the paper proposes SPLIT$^3$, a simple design that shares a trained encoder and learned initialization but maintains two separate fast states at test time with task-specific update rates.

Critical strengths
  1. SPLIT$^3$ is a minimal, low-overhead design (two fast states with shared encoder/initialization) that operationalizes the insight without requiring separate training objectives or models.
  2. SPLIT$^3$ achieves state-of-the-art or second-best results in multiple cases for detection and top localization on several metrics/datasets.
Critical weaknesses
  1. As a human reviewer, I find the presentation of this manuscript very obscure and poor. To be more specific, first of all, the authors’ motivation warrants further scrutiny. Taking Figure 1 as an example, I find no discernible pattern in the localization performance, and the performance variations across update frequencies are rather marginal, particularly on the SMD dataset. What's more, the manuscript lacks a coherent story. For instance, Section 3 begins its analysis without sufficient motivation or a clear explanation of how the authors’ ideas follow from the preceding discussion. Similar issues persist in Section 4, where the proposed SPLIT$^3$ appears only loosely connected to the extensive theoretical results presented earlier, with almost exclusive reliance on Equation (1). Consequently, the theoretical analysis feels more like an artificially appended component than an integral part of the proposed method.
  2. The derivations rely on a linear fast-weight predictor with a fixed encoder and normalized gradient updates; while this covers a prevalent TTT variant (test-time regression/associative memory), the scope of the claims could be clearer for other TTT forms (e.g., adaptation of non-linear backbones, optimizer momentum, or richer update rules). SPLIT$^3$ essentially instantiates two independent TTT trajectories (one per task). While principled and justified, the algorithmic novelty is modest beyond the analytical framing.
  3. Broader comparisons to localization-centric recent methods could be strengthened, and head-to-head “two independent TTT runs” (separately executed) vs. the authors' proposal numbers would help empirically confirm the stated equivalence beyond the proof. Sensitivity analyses for additional hyperparameters (e.g., memory size m, history length p, stabilization δ) and more granular μ sweeps per dataset are deferred to appendices; including key results in the main text would assist readers. The evaluation treats localization as always-on per-timestamp ranking; many real pipelines gate localization by detection. It would be informative to report gated-localization performance or robustness when detection thresholds vary.
  4. The paper might discuss more extensively how the analysis relates to broader TTT literature outside MTS (e.g., TTT for vision/NLP) and whether the separation principle could carry over. Recent localization-focused methods (e.g., transformer-based localization/diagnosis frameworks) could be discussed in more depth to contextualize SPLIT$^3$’s strengths and limitations relative to approaches that explicitly model cross-variable attributions.
  5. Some implementation details (e.g., encoder architecture specifics, training μtrain selection rationale, and per-dataset tuning protocol) are relegated to the appendix; a concise summary in the main paper would improve readability and reproducibility.
  6. This research is application-oriented, yet neither the codebase nor the constructed pipeline is available.
Questions for authors

See the weaknesses.

7 comments
AuthorsOpenReview ↗

[1/4] Rebuttal

We thank the reviewer for the careful reading and the constructive comments. We respond to each point below. For clarity, we organize our responses as Q1–Q10, with the corresponding weakness indicated in parentheses.

Q1 (W1). The motivation is weak: Figure 1 does not show a localization pattern, and the change on SMD is marginal.

The pattern that Figure 1 is meant to show is the order of the best rates of the two tasks. On all three datasets, the rate with the highest detection AP is lower than the rate with the highest localization HR@100 (SMD 0.001 / 0.1, SWaT 0.003 / 0.03, WaDI 0.001 / 0.01). On SMD in particular, over the entire range of the figure (0.001–0.1), AP keeps decreasing and HR@100 keeps increasing as the rate increases, and all three seeds show the same shape.

The size of the change can be seen in Table 3. When the rate is raised from 0.001 to 0.01, HR@100 increases by 3.51 points and AP decreases by 6.32 points on SMD. For the same change, HR@100 increases by 3.07 points and AP decreases by 25.33 points on SWaT, and HR@100 increases by 5.29 points and AP decreases by 5.74 points on WaDI. That is, the HR@100 change on SMD is larger than that on SWaT, and its AP change is similar to that on WaDI.

This change is much larger than the variation across seeds. The HR@100 gain on SMD is 3.76, 3.42, and 3.35 points for the individual seeds, and the standard deviation of HR@100 over seeds is 0.33–0.53 points. It is also not small compared with the gaps between methods: 3.51 points is about twice the gap between rank 1 and rank 2 in HR@100 on SMD in Table 1 (1.71 points). In terms of rank, for the same rate change, the detection AP on SMD drops from rank 4 to rank 9 among the 10 methods in Table 2 (42.91 → 36.59), and the localization HR@100 rises from rank 2 to rank 1 among the 7 methods in Table 1 (50.98 → 54.49). The values at the two rates are in the Shared μℓ\mu_\ell and Shared μh\mu_h rows of Table 3. Therefore, the change on SMD is large enough to change the benchmark rank.

Q2 (W1). Section 3 starts without motivation, and SPLIT³ in Section 4 is only loosely connected to the theory of Section 3 (it relies almost exclusively on Eq. 1).

We have made the connection between Sections 3, 4, and 5 explicit in the main text. In the revision, we have added sentences in the following four places.

  • Beginning of Section 3: the section now starts from the observation in Figure 1. When only the test-time rate is raised from 0.001 to 0.01 for the same trained model, detection AP decreases and localization HR@100 increases, yet at each rate the two tasks evaluate the same prediction errors. We then state up front that Section 3.1 defines the prediction and the fast-state update used in Figure 1 and in SPLIT³, that Section 3.2 derives the recurrence for the fast-state difference between two trajectories with different rates, and that Section 3.3 shows how the resulting prediction error difference is evaluated in the two tasks.
  • First paragraph of Section 4: we now state what Eq. 3 provides for the design. The prediction errors of two trajectories that share the encoder, the observation sequence, and the initial fast state differ only through the accumulated fast-state difference. Therefore, separating only the fast state allows the two tasks to use different rates.
  • Paragraph containing Eq. 13: we now state that the two fast states of SPLIT³ are the two trajectories analyzed in Section 3.2 (μℓ=μD\mu_\ell = \mu_D, μh=μL\mu_h = \mu_L), and that the difference between the two states and the difference between their prediction errors follow Eq. 3a and Eq. 3b, respectively.
  • Section 5.3: before the results, we now state that the decomposition in Table 4 is motivated by the analysis in Section 3.3. Eq. 5 shows that the detection energy can change at both anomalous and normal timestamps, and Eq. 6 shows that the change in the margin between a labeled variable and a remaining variable depends on the prediction error changes of both variables.

In summary, the analysis in Section 3 plays two roles. It tells us what needs to be separated (Eq. 3), and it provides the framework with which Section 5.3 decomposes and measures the changes in the two tasks (Eq. 5, Eq. 6). Which of the two tasks prefers the lower rate is determined by the experiments in Figure 1 and Sections 5.2 and 5.3.

Q3 (W2). The theory covers only a linear fast state, a fixed encoder, and a normalized gradient update; does it also hold for other forms of TTT, such as a nonlinear backbone or momentum?

The derivation holds in the following scope. Eq. 3b, which expresses the prediction error difference between two rates in terms of the fast-state difference, holds regardless of the update rule as long as the encoder is fixed and the prediction is linear in the fast state.

AuthorsOpenReview ↗

[2/4] Rebuttal

Eq. 3a, the recurrence for the fast-state difference, also holds in the same form for the update without normalization. For the update with momentum, it extends to an exact linear recurrence when the velocity difference is tracked together with the fast-state difference. In contrast, if the encoder is also updated or the prediction is nonlinear in the fast state, the keys or the prediction equations of the two trajectories differ, and these equations do not hold as written.

We therefore checked experimentally whether the rate preference is maintained in other forms as well. We ran rate sweeps in three cases: changing only the update rule on the same trained model, changing the predictor structure, and using other test-time adaptation (TTA) methods, M2N2 and CANDI. Each cell is the rate with the highest detection AP / the highest localization HR@100 (mean over seeds 0–2).

SettingSMDSWaTWaDI
Main model (encoder + update of Eq. 1)0.001 / 0.10.003 / 0.030.001 / 0.01
Update rule changed: momentum added0.001 / 0.10.003 / 0.030.001 / 0.01
Update rule changed: normalization removed0 / 0.010 / 0.0030 / 0.003
Predictor changed: linear model without encoder0.001 / 0.030.003 / 0.10.001 / 0.01
Other TTA method: M2N2 (SGD learning rate)0.03 / 0.10.01 / 0.010 / 0.001
Other TTA method: CANDI (SGD learning rate)0.1 / 0.10.03 / 0.30.01 / 0.01

With momentum added, with normalization removed, and with the linear model without the encoder, the best rate for detection was lower than that for localization on all three datasets. For M2N2 and CANDI, which differ from our model in both architecture and adaptation scheme, there was also no dataset where, based on the mean over seeds, the best rate for detection was higher than that for localization. M2N2 showed the same direction on SMD and WaDI and identical best rates on SWaT; CANDI showed the same direction on SWaT and identical best rates on SMD and WaDI.

However, the trade-off was smaller for these two methods than for the main model: less performance is lost when the best rate of one task is used for the other task. Both methods either exclude some high-error samples from adaptation or adapt mainly on samples that look normal, so the model follows the anomaly less even at a higher rate, which appears to reduce the trade-off.

In summary, the rate preference was clear for a linear fast state that keeps updating inside anomaly events, regardless of the update rule and of whether the encoder is present. For TTA methods that filter anomalous samples during adaptation, the direction was not reversed, but the preference was weaker.

Q4 (W2). The method amounts to two independent TTT trajectories, so the algorithmic novelty is modest.

Please refer to our Common Response.

Q5 (W3). The main text evaluates localization as always on, whereas a real pipeline gates localization by detection; it would be informative to report the gated localization performance or its robustness as the detection threshold varies.

The localization fast state of SPLIT³ is updated at every timestamp regardless of detection alerts. Therefore, gating by detection does not change the localization output itself; it only changes the timestamps at which the output is used. In the main text, we evaluated localization at all anomalous timestamps. As requested, we applied the gate and re-measured while varying the threshold. The threshold was set to a percentile of the detection scores on the normal validation data.

Each cell is HR@100 at alerted timestamps / HR@100 with missed timestamps set to 0 (missed = 0), with the alerted fraction of anomalous timestamps in parentheses (mean over seeds 0–2).

Alert threshold (percentile)SMDSWaTWaDI
No gate (main text)54.4912.8832.68
95%60.56 / 42.68 (67.0%)12.12 / 11.54 (95.2%)60.71 / 27.47 (45.2%)
99%65.79 / 37.32 (56.0%)9.83 / 8.70 (88.5%)65.91 / 23.33 (35.4%)
99.5%67.95 / 33.70 (49.7%)9.81 / 8.66 (88.3%)80.82 / 21.44 (26.8%)
99.9%72.31 / 27.73 (38.6%)9.63 / 8.47 (88.0%)91.83 / 19.46 (21.2%)

Raising the threshold yields fewer alerts, so HR@100 (missed = 0) decreases. In contrast, on SMD and WaDI, HR@100 at alerted timestamps was higher than the value without the gate at every threshold, and it increased further as the threshold increased. This appears to be because timestamps with higher detection scores are more pronounced anomalies whose affected variables are easier to find.

AuthorsOpenReview ↗

[3/4] Rebuttal

On SWaT, HR@100 at alerted timestamps was lower than the value without the gate at every threshold. Of the anomalous timestamps in SWaT, 69% belong to the single longest anomaly event. All timestamps of this event are alerted at every threshold, so the event accounts for 72–78% of the alerted anomalous timestamps. Localization itself is difficult in this event (HR@100 of about 2.6), so when only alerted timestamps are considered, the share of this event grows and HR@100 decreases. Indeed, when we re-measured on the remaining 26 events excluding this one, HR@100 at alerted timestamps was 35.01–37.14, much higher than the SWaT values in the table (9.63–12.12).

Q6 (W3). It would help to strengthen the comparison with recent methods and to confirm the equivalence experimentally by directly comparing the proposed method with two independent TTT runs executed separately.

The six baselines in Table 1 were all published in major conferences and journals between 2023 and 2026: CST-GL (TNNLS 2023), SARAD (NeurIPS 2024), M2N2 (AAAI 2024), AERCA (ICLR 2025), ALoRa (ICLR 2026), and CANDI (AAAI 2026). Among them, SARAD, AERCA, and ALoRa explicitly name diagnosis, root cause analysis, and localization, respectively, in their titles, and CST-GL also provides a procedure for finding root causes from the per-variable scores of a prediction model that learns a variable graph.

In this comparison, SPLIT³ achieved higher HR and NDCG than all six baselines on all three datasets (all 12 columns of HR@100, HR@150, NDCG@100, and NDCG@150). In HR@100, the gap to the best baseline is 1.71 points on SMD (SARAD 52.78), 5.39 points on SWaT (M2N2 7.49), and 1.91 points on WaDI (CANDI 30.77).

Shared μℓ\mu_\ell and Shared μh\mu_h in Table 3 are single TTT runs with rates 0.001 and 0.01, respectively. By Eq. 13, the SPLIT³ row equals the Shared μℓ\mu_\ell row in detection AP and the Shared μh\mu_h row in localization HR@100. As requested, we also confirmed this equivalence experimentally. We ran the SPLIT³ implementation, which updates the two fast states together, and two single TTT runs (one per rate), each in a separate process, and compared them directly. The results all matched Table 3, and the only difference was the computational cost.

Q7 (W4). The paper should further discuss how the analysis relates to the TTT literature outside MTS and whether the separation principle applies to other domains, as well as its strengths and limitations compared with recent localization methods that explicitly model cross-variable attribution (e.g., Transformer-based ones).

First, the two families the reviewer mentions, namely Transformer-based methods and methods that explicitly model the relationships among variables, are already included in the baselines of Table 1.

  • Transformer-based: SARAD learns associations among variables with a Transformer, and ALoRa is a low-rank Transformer. CANDI has an MLP backbone, so it is not a Transformer as a whole, but its adaptation module is an iTransformer-style encoder that treats variables as tokens and applies self-attention across variables.
  • Explicit modeling of relationships among variables: SARAD learns associations among variables, CST-GL learns a variable graph, and AERCA learns a Granger causality structure, and each ranks variables by the per-variable scores obtained from that model. ALoRa ranks variables by the cross-variable contributions to the reconstruction.

That is, four different approaches (attention, graph, causal structure, and contribution) are all in the comparison, and both families include recent methods. The HR@100 of SPLIT³ was higher than the best of these four methods (SARAD, CST-GL, AERCA, and ALoRa) by 1.71 points on SMD, 9.80 points on SWaT, and 10.58 points on WaDI.

Next, we turn to the relation to TTT outside MTS. The form covered by our analysis consists of a fixed encoder, a prediction that is linear in the fast state, and the normalized gradient update of Eq. 1. In that a linear fast state is updated with the gradient of the error, this is the same form as the family of TTT layers and fast weights ([1], [2], [3], [4]) discussed in Related Work, which were proposed for general sequence modeling rather than for MTS.

We separated the two fast states because the two tasks preferred different rates on all three datasets, and our analysis connects this difference to the fact that detection and localization rank the prediction errors of the same model along different axes. Detection ranks timestamps, and localization ranks the variables within a timestamp. We therefore think that the separation principle can also apply to other domains with the same ranking structure.

Reviewer_NTGpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • This paper investigates whether anomaly detection and variable localization in multivariate time series should share the same test-time adaptation state. The authors observe that detection and localization favor different TTT update rates and derive an exact recurrence describing how update-rate differences propagate through evolving fast states. They further relate the resulting prediction-error differences to the distinct ranking structures of the two tasks.

  • The authors propose SPLIT³ (Separated Plasticity in Test Time Training), which shares the trained encoder and learned initialization while maintaining separate fast states with task-specific update rates. Both tasks use the same prediction objective and update rule, allowing different adaptation trajectories within a single trained predictive model.

  • Rate sweeps on SMD, SWaT, and WaDI show that detection favors lower update rates than localization; the derived analysis characterizes the effects of update-rate differences but does not determine which rate each task prefers. SPLIT³ retains the detection performance of the lower-rate trajectory and the localization performance of the higher-rate trajectory, achieving strong localization on these three datasets and competitive detection across five benchmarks, including HAI and MSL.

Critical strengths

1. Robust evidence of task-specific adaptation preferences

  • By varying only the test-time update rate while keeping the trained predictive model fixed, the paper shows that detection AP favors lower rates than localization HR@100 on SMD, SWaT, and WaDI. This pattern remains consistent across training update rates, score normalization choices, and leave-one-dataset-out rate selection, providing empirical evidence of a trade-off when both tasks share a single adaptive state.

2. Controlled characterization of the adaptation trade-off

  • The output-level decomposition shows that higher update rates reduce detection AP primarily through changes at anomalous timestamps, while localization gains are associated with prediction-error changes in the remaining variables. Controlled event replay further shows that higher within-event update rates improve localization across the evaluated datasets. These analyses provide complementary empirical evidence for the different responses of detection and localization to test-time adaptation.
Critical weaknesses

1. Limited Methodological Differentiation and Label-dependent Update-rate Selection

  • Equation (13) establishes that each fast state of SPLIT³ reproduces an independent TTT trajectory, and Table 3 confirms that their respective detection and localization performance is retained. Since different test-time rates can already be applied to the same trained model (Appendix B.4), the methodological contribution beyond task-specific rate selection and parallel TTT execution remains unclear.

  • The main update rates were selected from labeled AP and HR@100 responses during development (Appendices B.1 and B.4), whereas normal validation data do not provide these labels. Although leave-one-dataset-out selection excludes the held-out dataset's labels, it still relies on labeled performance from other datasets (Appendix C.3). A label-independent rate-selection procedure for unseen environments remains unestablished.

2. Limited Empirical Scope and Mechanistic Interpretation

  • The different update-rate preferences are demonstrated on only three localization benchmarks. Detection is also evaluated on HAI and MSL, but corresponding update-rate sweeps are not reported (Figure 1; Tables 2–3).

  • The fast-state analysis characterizes how update-rate differences propagate but does not establish conditions determining which rate either task should prefer (§3.3). Table 4 provides dataset-level output decompositions, while Figure 3 aggregates prediction-error trajectories across events. Whether the proposed ranking explanation holds consistently at the individual-event level remains unverified.

Questions for authors

Q1. Generalizability of Update-rate Preferences (Weakness 2)

  • Figure 1 and Table 3 show different update-rate preferences on SMD, SWaT, and WaDI. Could the authors provide detection AP sweeps on HAI and MSL, which are already evaluated in Table 2, to assess whether the preference for lower rates generalizes?

Q2. Methodological Novelty (Weakness 1)

  • Equation (13) establishes equivalence between SPLIT³'s fast states and independent TTT trajectories, while Appendix B.4 confirms that different rates can be applied to the same trained model. What methodological contribution does SPLIT³ provide beyond task-specific rate selection and shared encoder computation?

Q3. Label Dependence in Rate Selection (Weakness 1)

  • The main update rates were selected using labeled AP and HR@100 responses (Appendices B.1 and B.4). Which data partitions supplied these labels? Since leave-one-dataset-out selection still relies on labeled datasets (Appendix C.3), how would the rates be selected in a new deployment without detection or localization labels?

Q4. Event-level Consistency of the Adaptation Trade-off (Weakness 2)

The analysis does not determine the preferred update rate (§3.3), and Table 4 and Figure 3 report aggregated results. Using the controlled event replay (Appendix C.6), could the authors report the fraction of events where the higher update rate (0.01) improves HR@100 over the lower rate (0.001), alongside the corresponding detection AP changes with outside-event scores held fixed?

3 comments
AuthorsOpenReview ↗

[1/2] Rebuttal

We thank the reviewer for the careful reading and the constructive comments. We respond to each point below.

Reply to Q1 (W2) Generalizability of update-rate preferences

As requested, we report the detection AP sweep on HAI and MSL. Only the test-time rate is varied, and values are means ± standard deviations over seeds 0–2.

Detection AP (×100)μ = 00.0010.0030.010.030.10.3Random AP
HAI69.03 ± 1.1569.40 ± 1.1168.98 ± 1.3468.13 ± 1.6966.57 ± 2.0162.85 ± 2.7159.83 ± 2.482.73
MSL13.51 ± 0.1713.82 ± 0.3314.27 ± 0.3614.81 ± 0.2914.98 ± 0.1515.19 ± 0.2214.74 ± 0.1810.53

On HAI, AP is highest at 0.001, the same rate as in the main text. On MSL, in contrast, AP is highest at 0.1, a higher rate than on the other datasets. Our claim is that, within the same dataset, the rate preferred by detection is lower than the rate preferred by localization, and we restricted its verification to SMD, SWaT, and WaDI, which provide both detection and localization labels. Since MSL does not provide localization labels, we could not test whether the claim holds on this dataset.

Reply to Q2 (W1) Methodological novelty

Please refer to our Common Response.

Reply to Q3 (W1) Label dependence in rate selection

The main rates were chosen during development based on the AP and HR@100 responses observed with labels from the test period. Labels were not used in the training objective, online prediction, score computation, or fast-state updates.

As the reviewer points out, the leave-one-dataset-out (LODO) experiment in Appendix C.3 also relies on the labels of other datasets. We have not established a rate-selection method for settings where no labels are available at all, and we acknowledge this as a limitation of our work.

When existing labeled datasets are available, however, we examined whether the rates for a new dataset can be set without its labels. The table below extends Tables 9 and 10. It compares the performance of the rates transferred by LODO with that of the post-hoc best rates on the target dataset. The best rates are chosen separately for AP and HR@100 within the rate grid of Appendix B.4, {0, 0.001, 0.003, 0.01, 0.03, 0.1, 0.3}.

Target dataset (labels unused)Datasets used for selectionTransferred (μD∗\mu_D^{\ast}, μL∗\mu_L^{\ast})AP (transferred / target best)HR@100 (transferred / target best)
SMDSWaT + WaDI(0.001, 0.01)42.91 / 42.9154.49 / 55.24
SWaTSMD + WaDI(0.001, 0.01)75.35 / 75.3812.88 / 13.25
WaDISMD + SWaT(0.001, 0.03)45.37 / 45.3731.19 / 32.68

Although no target labels were used for rate selection, the gap to the post-hoc best was at most 0.03 points in AP and 1.50 points in HR@100. On the three datasets, rates transferred from other datasets performed close to the best. This suggests that rates determined on existing labeled datasets can serve as defaults for a new dataset.

We would like to clarify that the main goal of our work is not to find the optimal rate. It is to show that, when TTT is used for anomaly diagnosis in MTS, detection can prefer a lower update rate than localization. In Table 9, the two rates were selected without any constraint that they differ, yet in all three cases the localization rate was selected higher than the detection rate. In Table 10, using the localization rate selected on the other datasets instead of sharing the detection rate raised HR@100 by +3.51, +3.07, and +3.80 points.

AuthorsOpenReview ↗

[2/2] Rebuttal

Reply to Q4 (W2) Event-level consistency of the adaptation trade-off

The analysis in Section 3 describes how a rate difference propagates through the fast state into the prediction errors; it does not determine the optimal rate for each task. Likewise, the goal of our work is not to find per-task optimal rates. We use the theoretical analysis to show how a rate difference is reflected differently in the two tasks, use experiments to establish that detection and localization can prefer different update rates, and present a basic structure that performs both tasks in one model. As requested, we checked with the controlled replay of Appendix C.6 whether this direction also holds at the event level. For each event, the preceding fast state is taken from the 0.001 trajectory, the rate is raised from 0.001 to 0.01 only within that event, and scores outside the event are fixed to their 0.001 values.

DatasetEvents (with localization labels)HR@100 up / same / downAP down / up (change > 0.01 points)AP in HR-up events: down / within ±0.01 points / upMean ΔHR / ΔAP
SMD325158 / 113 / 54130 / 2590 / 54 / 14+1.87 / −1.02
SWaT274 / 21 / 212 / 12 / 1 / 1+0.07 / −1.46
WaDI146 / 7 / 110 / 13 / 2 / 1+1.35 / −0.57

On all three datasets, raising the rate increased HR@100 in more events than it decreased it (SMD 158 vs. 54, SWaT 4 vs. 2, WaDI 6 vs. 1). Among events whose AP changed by more than 0.01 points, AP decreased far more often than it increased (130 vs. 25, 12 vs. 1, 10 vs. 1). The same held when we restricted to events where HR@100 increased (90 vs. 14, 2 vs. 1, 3 vs. 1). The mean ΔAP was negative and the mean ΔHR was positive on all three datasets. Thus, between 0.001 and 0.01, the direction in which detection prefers the lower rate largely holds on the three datasets.

Thank you for the detailed responses and additional experiments. I appreciate the authors' efforts to address my concerns.

While the new results strengthen the empirical evidence, my primary concerns regarding methodological novelty and label-dependent update-rate selection remain insufficiently resolved. Therefore, I maintain my original rating.

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

This paper studies the different effects of test-time training on anomaly detection and variable localization. SPLIT³ shares an encoder and learned initialization while maintaining separate fast states with task-specific update rates.

Critical strengths
  1. The paper identifies a clear distinction between the adaptation preferences of detection and localization.
  2. Controlled rate sweeps and event replay analyses provide useful evidence for the observed tradeoff.
  3. The design is simple and computationally efficient, with additional support from leave-one-dataset-out rate selection.
Critical weaknesses
  1. Each fast state is equivalent to an independent TTT run with a different update rate.
  2. The theoretical analysis describes how state differences propagate but does not establish when detection should favor lower rates than localization.
  3. The generality of the central finding remains unclear because localization experiments cover only three datasets and one fast-state update architecture.
  4. Separate detection and localization metrics do not establish performance when localization is applied only to actual detector alerts.
  5. Selecting task-specific rates for new datasets without anomaly localization labels remains an open practical issue.
Questions for authors
  1. What methodological contribution does SPLIT³ provide beyond sharing an encoder between two independent TTT runs with different update rates?
  2. Under what conditions should detection favor lower update rates than localization, and when might this relationship fail?
  3. Does the observed rate preference generalize to other forecasting architectures and adaptation rules?
  4. How does localization perform when evaluated only on actual detector alerts, accounting for missed anomalies and detection delays?
  5. Under matched computational budgets, does state separation outperform a single state with an adaptive update?
4 comments
AuthorsOpenReview ↗

[1/3] Rebuttal

We thank the reviewer for the careful reading and the constructive comments. We respond to each point below.

Reply to Q1 (W1) Methodological contribution beyond two independent TTT runs

Please refer to our Common Response.

Reply to Q2 (W2) When detection favors a lower rate than localization

As the reviewer points out, we have not established theoretically when detection prefers a lower rate than localization. However, the analysis in Section 3 shows that the fast-state difference between two rates accumulates with each update (Eq. 3a) and that this difference appears as a difference in subsequent prediction errors (Eq. 3b). Empirically, with a higher rate the fast state gradually follows the anomaly, so the errors at anomalous timestamps decrease (Figure 3, Appendix C.5), and this effect grows the longer the anomaly lasts. Lower errors mean lower anomaly scores, which can hurt detection, whereas localization only looks at the ranking of the variables within the same timestamp and therefore may not be hurt (Eq. 6).

If so, detection has less reason to prefer the lower rate when anomalies are short. To check this, we divided the anomaly events that have both detection and localization labels into length quartiles (Q1 = shortest, Q4 = longest). We then measured the change when the rate is raised from 0.001 to 0.01, separately for Q1 and Q4. For each quartile, AP uses the anomalous timestamps of that quartile and all normal timestamps, and HR@100 uses the anomalous timestamps of that quartile. Values are changes in points (mean over seeds 0–2).

DatasetQuartile: event length (number of events)Change in detection APChange in localization HR@100
SMDQ1: 3–5 steps (109)+1.64+5.72
SMDQ4: 76–3162 steps (81)−8.45+4.03
SWaTQ1: 2–7 steps (8)+0.33+13.41
SWaTQ4: 16–599 steps (5)−41.72+2.62
WaDIQ1: 10–59 steps (4)−1.70+1.84
WaDIQ4: 86–175 steps (3)−11.70+7.45

On average, localization was better at 0.01 in all four quartiles on the three datasets. In contrast, detection lost substantially in Q4, whereas in Q1 AP increased or the loss was much smaller. In SWaT Q4, the loss comes from a single 599-step event; on the other four events (16–26 steps), AP increased by 6.37 points. A short anomaly ends before the fast state can follow it, so the advantage that the lower rate gives detection appears to diminish or disappear. Indeed, for individual seeds, the order of the best rates was sometimes reversed in Q1 (SMD, one seed: best rate 0.3 for detection and 0.03 for localization; SWaT, two seeds: 0.3 for detection, and 0.1 and 0.03 for localization).

In summary, the preference of detection for a lower rate than localization is clear when anomalies persist for a long time. For short anomalies, the difference between the two tasks weakens or can be reversed depending on the seed. In the revision, we have added this result as Appendix D and state in Section 5.3 when this preference holds and when it may not.

Reply to Q3 (W3) Generalization to other architectures and adaptation rules

SettingSMDSWaTWaDI
Main model (encoder + update of Eq. 1)0.001 / 0.10.003 / 0.030.001 / 0.01
Update rule changed: momentum added0.001 / 0.10.003 / 0.030.001 / 0.01
Update rule changed: normalization removed0 / 0.010 / 0.0030 / 0.003
Predictor changed: linear model without encoder0.001 / 0.030.003 / 0.10.001 / 0.01
Other TTA method: M2N2 (SGD learning rate)0.03 / 0.10.01 / 0.010 / 0.001
Other TTA method: CANDI (SGD learning rate)0.1 / 0.10.03 / 0.30.01 / 0.01

We ran rate sweeps in three cases: changing only the update rule on the same trained model, changing the predictor structure, and using other test-time adaptation (TTA) methods, M2N2 and CANDI. Each cell is the rate with the highest detection AP / the highest localization HR@100 (mean over seeds 0–2).

With momentum added, with normalization removed, and with the linear model without the encoder, the best rate for detection was lower than that for localization on all three datasets and all seeds.

For M2N2 and CANDI, which differ from our model in both architecture and adaptation scheme, there was also no dataset where, based on the mean over seeds, the best rate for detection was higher than that for localization. M2N2 showed the same direction on SMD and WaDI and identical best rates on SWaT; CANDI showed the same direction on SWaT and identical best rates on SMD and WaDI. However, the performance difference between the two best rates was much smaller than for the main model. Both methods either exclude some high-error samples from adaptation or adapt mainly on samples that look normal, so the model follows the anomaly less even at a higher rate, which appears to reduce the trade-off.

AuthorsOpenReview ↗

[2/3] Rebuttal

Therefore, this preference was clear for a linear fast state that keeps updating inside anomaly events, regardless of the update rule and the encoder. For TTA methods that filter anomalous samples during adaptation, the direction was not reversed, but the preference was weaker.

Reply to Q4 (W4) Localization on actual detector alerts

As requested, we re-evaluated localization based on actual detection alerts. The update rates are the same as in the main text: 0.001 for detection and 0.01 for localization. An alert is raised when the detection score exceeds the 99th percentile of the normal validation scores; no test labels are used. "Missed = 0" counts anomalous timestamps without an alert as localization failures. Values are means over seeds 0–2.

DatasetHR@100 (regardless of alerts)HR@100 (alerted timestamps only)HR@100 (missed = 0)Alerted fraction of anomalous timestampsEvent detection rate / mean delay (detected events)False alarm rate on normal timestamps
SMD54.49 ± 0.5365.79 ± 0.4637.32 ± 0.3956.0%98.3% / 3.7 steps6.2%
SWaT12.88 ± 0.109.83 ± 0.448.70 ± 0.4688.5%77.8% / 1.5 steps40.2%
WaDI32.68 ± 0.2165.91 ± 2.2523.33 ± 0.6935.4%57.1% / 18.5 steps1.3%

On SMD and WaDI, HR@100 at alerted timestamps was higher than at missed timestamps. Alerts tend to fire on pronounced anomalies whose affected variables are easier to find. Setting missed timestamps to 0 lowers HR@100 considerably, because the localization result cannot be used at anomalous timestamps where detection raises no alert.

On SWaT, HR@100 was instead lower when only alerted timestamps were considered. About 78% of the alerted anomalous timestamps come from the single longest anomaly event, and localization itself is difficult in this event, which makes the overall value look lower. Indeed, when we re-measured on the remaining 26 events excluding this one, HR@100 at alerted timestamps was 35.21, nearly the same as the value measured regardless of alerts on the same events (35.66).

Reply to Q5 (W5). Adaptive single state and rate selection for new datasets

As asked, we compared SPLIT³ with a single state whose rate changes continuously at every timestamp. The single state uses the same encoder, initial fast state, and update equation (Eq. 1) as SPLIT³, and keeps only one fast state. The rate is set, without labels, between the two rates of SPLIT³ (0.001 and 0.01), using one of the following two signals.

  • Error-driven rate: the rate itself is updated at every timestamp in the direction that reduces the prediction error (a hypergradient update). We evaluated a one-step approximation, which reflects only the effect of the previous update, and a full variant, which maintains the derivative of the fast state with respect to the rate as a matrix of the same size as the fast state and computes the gradient without approximation.
  • Score-driven rate: the higher the moving average of the anomaly score, the lower the rate (the reference point is the 99th percentile of the normal validation scores).

For the two error-driven variants, we varied the step size for updating the rate over 0.001, 0.003, 0.01, 0.03, and 0.1 (10 configurations). For the score-driven rate, we varied the decay factor of the moving average over 0.9 and 0.99 (2 configurations). For each variant, the table reports the best configuration, selected on the test results by the average of AP and HR@100 over the three datasets. The best was step size 0.1 for both error-driven variants and decay factor 0.9 for the score-driven rate. Each cell is detection AP / localization HR@100 (mean over seeds 0–2). FLOPs are relative to SPLIT³ (the range is the minimum–maximum over the three datasets) and were measured in the same way as in Appendix E.

MethodFLOPsSMDSWaTWaDI
SPLIT³ (0.001 / 0.01)1.0042.91 / 54.4975.35 / 12.8845.37 / 32.68
Error-driven rate (one-step, step size 0.1)0.58–0.6936.72 / 54.3849.96 / 12.8839.60 / 32.72
Error-driven rate (full, step size 0.1)1.19–1.2537.03 / 54.0750.88 / 12.8138.80 / 32.64
Score-driven rate (decay factor 0.9)0.58–0.6941.52 / 53.9479.01 / 12.7044.04 / 32.37

For both error-driven variants, HR@100 was similar to that of SPLIT³, but AP was 5.8–25.4 points lower depending on the dataset, nearly the same as that of a single state with the rate fixed at 0.01 (Shared μh\mu_h in Table 3). This is because a higher rate is favorable for reducing the prediction error, so the rate stayed high.

The score-driven rate had higher AP than SPLIT³ on SWaT in all three seeds (+3.66 points on average) with HR@100 lower by 0.18 points, but on SMD and WaDI both metrics were lower than those of SPLIT³ in all three seeds.

AuthorsOpenReview ↗

[3/3] Rebuttal

In summary, the 12 configurations include single-state variants that use fewer FLOPs than SPLIT³ and variants that keep a state of the same size as SPLIT³ and use more FLOPs. On SMD and on WaDI, none of them reached both the AP and the HR@100 of SPLIT³. On SWaT, only the score-driven rate had higher AP, and for the configuration in the table its HR@100 was lower by 0.18 points.

Regarding W5, when other labeled datasets are available, the rates for a new dataset can be set without any of its labels. In the leave-one-dataset-out experiment (Appendix C.3), the rates selected on the other two datasets performed within 0.03 points in AP and 1.50 points in HR@100 of the post-hoc best rates on the target dataset, chosen from the rate grid of Appendix B.4 (per-dataset values are in our reply to Reviewer NTGp, Q3). With these transferred rates, using a separate localization rate instead of sharing the detection rate raised HR@100 by 3.51, 3.07, and 3.80 points on SMD, SWaT, and WaDI, respectively (Table 10). Selecting task-specific rates when no labels are available at all remains a limitation, which we now state in the revised Conclusion.

Thank you for the response. I will maintain my positive rating.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Common Response: Methodological Contribution (ekCT W1/Q1, NTGp W1/Q2, StU6 W2)

Common Response: Methodological Contribution (ekCT W1/Q1, NTGp W1/Q2, StU6 W2)

As the reviewers point out, each fast state of SPLIT³ follows the same trajectory as an independent TTT run with the corresponding update rate. We state this property in Eq. (13) and prove it in Appendix A.4. We do not claim an elaborate algorithm as our contribution. Our contributions are as follows.

First, to the best of our knowledge, prior test-time adaptation (TTA) work on multivariate time series (MTS) has not examined which update rate each task prefers when a single model performs both detection and localization. What we aim to show is not a dataset-specific optimal rate, but the relative preference of the two tasks within the same model, i.e., which task prefers the lower rate. When only the test-time update rate was changed for the same trained model, the best observed rates differed across datasets, but detection preferred a lower rate than localization on all three datasets (Table 13). The same direction also held when the update rule was changed or the encoder was removed.

Second, we analyze how a rate difference propagates differently to the two tasks. A rate difference accumulates in the fast state and changes subsequent prediction errors (Eq. (3)). These error changes are evaluated as rankings across timestamps for detection, and as rankings across variables within a timestamp for localization (Eqs. (5)–(6)). The output decomposition (Table 4) is consistent with this: when the rate increases, the AP drop is driven by score changes at anomalous timestamps, while the HR@100 gain is driven by prediction error changes in the remaining variables, i.e., those other than the labeled variables.

Third, building on these results, we show that a lightweight design that separates only the fast state per task can retain the performance of both tasks simultaneously. Within the same model, the prediction error difference between two rates arises only from the fast state difference (Eq. (3b)), so we keep the encoder and the learned initialization shared. In contrast, sharing a single rate sacrifices performance on one task: sharing 0.001 lowers HR@100 by up to 5.29 percentage points, and sharing 0.01 lowers AP by up to 25.33 percentage points (Table 3). SPLIT³ uses a single fixed pair, μD=0.001\mu_D=0.001 for detection and μL=0.01\mu_L=0.01 for localization, for all datasets and seeds. It retains the performance of both tasks by adding only the prediction and update of one additional fast state. It has the fewest learned parameters and the lowest arithmetic GFLOPs per observation among the compared methods (Table 15), achieves higher HR and NDCG than all six baselines on all three datasets, and shows competitive detection performance (Tables 1 and 2).

In summary, the novelty of this work lies not in an elaborate algorithm, but in identifying the different adaptation preferences of detection and localization, which prior MTS TTA research has not sufficiently addressed, and in filling this gap through analysis and experiments. As TTA becomes more widely used for MTS anomaly diagnosis, we believe that treating the level of adaptation as a task-specific design factor, rather than as a single model-wide setting, will become increasingly important. This perspective can serve as a starting point for future work, such as selecting task-specific rates without labels or adjusting the level of adaptation according to anomaly duration, and we believe our work provides a baseline for these directions.

AuthorsOpenReview ↗

Response to All Reviewers and the AC

We thank the three reviewers for their careful reading and constructive comments. We appreciate the following strengths noted in the reviews.

  • The finding that the two tasks prefer different levels of adaptation. "identifies a clear distinction between the adaptation preferences of detection and localization" (ekCT); "Robust evidence of task-specific adaptation preferences" (NTGp).
  • Controlled analyses. "Controlled rate sweeps and event replay analyses provide useful evidence for the observed tradeoff" (ekCT).
  • A simple and lightweight design. "simple and computationally efficient" (ekCT); "a minimal, low-overhead design" (StU6).

Below we summarize our responses to the concerns raised by more than one reviewer.

A. Methodological contribution (ekCT, NTGp, StU6). Our contribution lies in identifying the different adaptation preferences of detection and localization, which prior MTS TTA research has not sufficiently addressed, and in filling this gap through analysis and experiments. Building on these results, we show that a lightweight design that separates only the fast state can retain the performance of both tasks. Please see our Common Response for details.

B. When the preference holds, and its event-level consistency (ekCT, NTGp). The preference was clear for long anomaly events, whereas for short events it weakened or could be reversed depending on the seed (Appendix D of the revision). In the event-level analysis, the direction in which detection prefers the lower rate also largely held on the three datasets.

C. Generalization to other forms of TTT (ekCT, StU6). With momentum added, with normalization removed, and with a linear model without the encoder, the best rate for detection was lower than that for localization on all three datasets. For two other test-time adaptation (TTA) methods, M2N2 and CANDI, there was no dataset where, based on the mean over seeds, the best rate for detection was higher than that for localization.

Additions to the main text in the revision

  • Section 3: an opening paragraph that starts from the observation in Figure 1 and states the role of each subsection.
  • Section 4: sentences stating what the analysis (Eq. 3) provides for the design, and that the two fast states of SPLIT³ are the two trajectories analyzed in Section 3.2.
  • Section 5.1: an implementation summary, and the result that raising the test-time rate also decreases AP and increases HR@100 for other training update rates (Appendix C.2).
  • Section 5.3: sentences stating that the decomposition in Table 4 is motivated by the analysis in Section 3.3, and that the preference is clearest for long anomaly events (results in the new Appendix D).
  • Appendix D.

Details are given in our Common Response and in the replies to each reviewer. We also thank the AC for the time and effort devoted to our submission.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.