- 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
- 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.
- 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
- 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.
- 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.
- 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.
- 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.
- 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.
- This research is application-oriented, yet neither the codebase nor the constructed pipeline is available.
Questions for authors
See the weaknesses.
7 comments
[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).
| Setting | SMD | SWaT | WaDI |
|---|---|---|---|
| Main model (encoder + update of Eq. 1) | 0.001 / 0.1 | 0.003 / 0.03 | 0.001 / 0.01 |
| Update rule changed: momentum added | 0.001 / 0.1 | 0.003 / 0.03 | 0.001 / 0.01 |
| Update rule changed: normalization removed | 0 / 0.01 | 0 / 0.003 | 0 / 0.003 |
| Predictor changed: linear model without encoder | 0.001 / 0.03 | 0.003 / 0.1 | 0.001 / 0.01 |
| Other TTA method: M2N2 (SGD learning rate) | 0.03 / 0.1 | 0.01 / 0.01 | 0 / 0.001 |
| Other TTA method: CANDI (SGD learning rate) | 0.1 / 0.1 | 0.03 / 0.3 | 0.01 / 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) | SMD | SWaT | WaDI |
|---|---|---|---|
| No gate (main text) | 54.49 | 12.88 | 32.68 |
| 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.
[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 and Shared in Table 3 are single TTT runs with rates 0.001 and 0.01, respectively. By Eq. 13, the SPLIT³ row equals the Shared row in detection AP and the Shared 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.
[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.
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 and Shared μh rows of Table 3. Therefore, the change on SMD is large enough to change the benchmark rank.
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.
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.
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.