acceptodds
Under review as a conference paper at ICLR 2027

Beyond Reverse Denoising: Efficiently Detecting Time Series Anomalies through Prototype-Guided Diffusion Inversion

17 views

Abstract

Most diffusion-based time series anomaly detection methods rely on iterative reverse denoising, coupling detection with generative fidelity even though sample generation is not the goal. Expressive denoisers may reconstruct anomalous patterns, while repeated sampling incurs substantial inference overhead. We argue that the distribution-learning capability of diffusion models can instead define an efficient one-class normality criterion. To this end, we propose PiAD, a prototype-guided framework that detects time series anomalies through deterministic DDIM inversion trajectories without reverse denoising or test-time reconstruction. Our key insight is that a noise predictor learned from normal-dominated data defines normal latent dynamics across noise levels: normal windows tend to follow mode-consistent data-to-noise trajectories, whereas anomalous windows exhibit incompatible transitions. Because normal behavior is often multimodal, a single global trajectory prior may conflate distinct normal modes. PiAD therefore combines prototype-conditioned latent diffusion with prototype-specific gaussian trajectory priors and uses training-calibrated Top- aggregation to retain the most discriminative deviations across inversion states. Experiments on five real-world datasets show that PiAD consistently outperforms existing diffusion-based approaches while requiring substantially less inference time, memory, and model capacity.

Primary area:
learning on time series and dynamical systems
Keywords:
Diffusion Models, Time Series, Anomaly Detection

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_FmPfOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes PiAD, a reconstruction-free time series anomaly detection framework that leverages deterministic DDIM inversion to assess compatibility of test windows with learned normal data-to-noise dynamics. The method augments a lightweight latent diffusion model with prototype conditioning to capture multimodal normal regimes and builds prototype-specific Gaussian priors over multi-step inversion trajectories. Token-level deviations across inversion states are fused via a Top-k strategy to yield anomaly scores. Experiments on five real-world datasets show strong performance and substantial efficiency gains over reconstruction-based diffusion baselines.

Critical strengths
  1. The paper reframes diffusion-based time series anomaly detection from reverse-generation or reconstruction toward detection via deterministic data-to-noise trajectories, thereby removing the need for test-time reconstruction.

  2. The paper addresses two weaknesses of prior diffusion-based TSAD coherently. First, it recognizes that real-world normal behavior is often multimodal and avoids a brittle global trajectory prior by introducing prototype-conditioned latent diffusion with prototype-specific Gaussian trajectory priors. Second, the training-calibrated Top-k aggregation strategy is well motivated: anomaly evidence is not uniformly concentrated at any single inversion state, so selectively fusing the most discriminative states yields a more robust multi-scale criterion. The ablation studies (Tables 2a and 2b) cleanly isolate the contributions of each component and confirm their complementarity.

  3. The method follows a clear, step-by-step algorithm. The mathematical formulation effectively illustrates why multi-step inversion outperforms terminal-state-only scoring and why prototype-specific priors suppress false alarms on normal oscillations.

  4. The evaluation is comprehensive, and PiAD achieves the best average rank across all dataset-metric combinations. The computational analysis is also the best, attributable to eliminating iterative reverse denoising and test-time decoding.

Critical weaknesses
  1. The prototype-specific trajectory priors are modeled as per-token, per-timestep diagonal Gaussians. The assumptions of feature-wise independence and Gaussianity may be restrictive when the latent geometry of dynamics is multimodal or anisotropic. The paper does not justify why a diagonal covariance is sufficient given that the entire motivation of the prototype mechanism is to capture complex, mode-dependent structure.
  2. Deterministic DDIM inversion is assumed to remain well-behaved across multiple steps, but error accumulation in the noise-prediction chain and its stability under covariate shift or distribution drift is not analyzed.
  3. The paper positions itself clearly against reconstruction-based diffusion TSAD but lacks comparison with two baselines that are conceptually close: (i) reconstruction-free, single-step noise-prediction detectors for time series, which are the most direct competitors to the stated contribution. (ii) prototype-/retrieval-conditioned diffusion in time series (e.g., prototype-retrieval), which should be contrasted more explicitly to motivate what is TSAD-specific about the present prototype-conditioned design.
  4. The sensitivity analysis is also partial: the interaction between the number of prototypes K, the number and schedule of inversion steps M, and the Top-k aggregation parameter is not reported comprehensively, which weakens confidence in how tuning-sensitive the reported results are.
Questions for authors
  1. How many inversion steps M are used, and how are the scheduler indices n1,…,nM selected? Please provide a sensitivity analysis of performance and runtime versus M and versus the distribution of steps.
  2. How is the number of prototypes K chosen, how often is the prototype bank recomputed during training, and how sensitive are results to K and update frequency?
  3. Please compare directly against reconstruction-free or low-step diffusion TSAD on accuracy and latency.
  4. How do results change with stronger linear baselines on these datasets? Including them would help contextualize the gains.
  5. What is the impact of the diagonal-Gaussian assumption for the prototype
3 comments
AuthorsOpenReview ↗

Responses to Q1, Q2 and W4

Thank you for your support and the time you have devoted to reviewing our work! Your feedback means a great deal to us. Below, we provide a point-by-point response.

Q1 & W4: “How many inversion steps M are used, and how are the scheduler indices n1,…,nM selected? Please provide a sensitivity analysis of performance and runtime versus M and versus the distribution of steps”

Re: We use M=8M=8 inversion transitions by default, yielding nine scoring states including the clean latent, as specified in Appendix A.4.3. Section 4.3 describes the approximately uniform selection of scheduler indices, including the terminal index. We supplemented the performance sensitivity analysis in Appendix A.5.2 and Figure 4 with runtime measurements and a comparison of index distributions.

Sensitivity to the number of steps. Following your suggestion, we evaluated M∈{2,4,8,10,12}M \in \{2, 4, 8, 10, 12\}. Each entry below reports VUS-PR (%) / test inference time (seconds).

MMSMDSWaTPSMMSLSMAP
255.07 / 2.0062.96 / 2.1331.86 / 7.0192.94 / 1.0699.68 / 1.11
456.10 / 3.3764.56 / 3.7034.63 / 9.5592.11 / 1.3599.64 / 1.99
857.84 / 5.0964.02 / 6.4433.57 / 17.1193.55 / 2.2099.68 / 3.37
1056.21 / 6.1159.26 / 8.6732.20 / 21.2291.57 / 3.0299.69 / 4.09
1253.68 / 8.1561.14 / 10.0032.63 / 27.9390.69 / 3.7399.74 / 5.28

Increasing MM consistently increases inference time without consistently improving detection. Eight steps achieve the highest VUS-PR on SMD and MSL, while four steps perform best on SWaT and PSM; SMAP changes little. These results support sparse inversion, with limited additional benefit from increasing the step count beyond eight.

Sensitivity to the distribution of inversion steps. At a fixed M=8M=8, we compared uniform, low-noise-dense, and high-noise-dense schedules. Let uj=j/Mu_j=j/M for j=1,…,Mj=1,\ldots,M, and let T=N−1T=N-1 denote the terminal scheduler index. We construct target indices from TujT u_j, Tuj2T u_j^2, and T[1−(1−uj)2]T[1-(1-u_j)^2], respectively, followed by rounding and adjustment to ensure exactly MM strictly increasing positive indices with the same terminal index. The underlying diffusion noise schedule remains unchanged.

ScheduleSMDSWaTPSMMSLSMAP
Low-noise-dense49.1560.4031.4085.2199.72
High-noise-dense53.4761.8933.4392.9799.64
Uniform (Ours)56.2064.8234.6292.1799.70

Uniform sampling performs best on SMD, SWaT, and PSM and remains competitive on MSL and SMAP. Concentrating steps at low noise levels reduces performance on four datasets, indicating that coverage across noise levels is useful. We will include these results and specify the scheduler construction in the revised manuscript.

Q2 & W4: “How is the number of prototypes K chosen, how often is the prototype bank recomputed during training, and how sensitive are results to K and update frequency?”

Re: We use K=8K=8 as a shared default across datasets. Appendix A.5.2 and Figure 4 report sensitivity to K∈{2,4,8,16,32}K\in\{2,4,8,16,32\}, supporting a moderate prototype count that balances normal-mode coverage and reliable prototype-specific statistics.

As specified in Appendix A.4.1, the prototype bank is recomputed every five epochs after five warm-up epochs. Following your suggestion, we additionally evaluated update-frequency sensitivity with K=8K=8 and other settings fixed:

Update intervalSMDSWaTPSMMSLSMAP
1 epoch51.9756.7530.1587.3399.48
5 epochs56.2064.8234.6292.1799.70
10 epochs53.7060.8332.4491.9899.69

Updating every five epochs achieves the highest VUS-PR on all five datasets. More frequent updates may destabilize the evolving prototype conditions and consistency targets, while less frequent updates may leave prototypes outdated. These results support our default update interval.

AuthorsOpenReview ↗

Responses to Q3 and W3

Q3: “Please compare directly against reconstruction-free or low-step diffusion TSAD on accuracy and latency.”

Re: Thank you for this suggestion. Although we have not identified a formally published reconstruction-free diffusion TSAD method directly comparable to PiAD, our existing evaluation includes D³R (NeurIPS 2023) [1], which uses direct reconstruction rather than iterative reverse denoising, with accuracy and latency comparisons reported in Tables 1 (Section 5.2) and 3 (Section 5.4).

To better address your questions, we further compared PiAD with DDIM-accelerated IMDiffusion (VLDB 2023) [2] and EnsDiffAD (TKDE 2026) [3], using eight sampling steps for both baselines to match PiAD’s eight inversion steps. Each entry below reports VUS-PR (%) / training time (s) / test time (s).

DatasetIMDiffusion-DDIMEnsDiffAD-DDIMPiAD (Ours)
SMD35.35 / 188.1 / 20.723.74 / 763.0 / 41.856.20 / 15.8 / 5.3
SWaT38.38 / 361.5 / 18.036.53 / 2545.2 / 96.564.82 / 25.9 / 6.8
PSM17.74 / 1047.2 / 76.516.51 / 6127.3 / 204.734.62 / 98.0 / 18.1
MSL37.07 / 114.6 / 8.124.73 / 1396.3 / 17.392.17 / 21.6 / 2.5
SMAP52.83 / 43.9 / 5.936.72 / 336.2 / 12.699.70 / 20.4 / 3.2

PiAD achieves higher VUS-PR, shorter training time, and lower inference latency on all five datasets. DDIM reduces the number of sampling steps at inference, while leaving the baselines’ original training objectives and procedures unchanged, so it does not directly accelerate training. PiAD’s lightweight latent-space diffusion model also reduces training cost, providing an additional efficiency advantage beyond inference-time acceleration.

If you are aware of other specific reconstruction-free or low-step diffusion TSAD methods, we would appreciate the references and would be happy to include direct comparisons.

W3: “The paper positions itself clearly against reconstruction-based diffusion TSAD but lacks comparison with two baselines that are conceptually close: (i) reconstruction-free, single-step noise-prediction detectors for time series, which are the most direct competitors to the stated contribution. (ii) prototype-/retrieval-conditioned diffusion in time series (e.g., prototype-retrieval), which should be contrasted more explicitly to motivate what is TSAD-specific about the present prototype-conditioned design.”

Re: Regarding (i), we have not identified a formally published reconstruction-free, single-step noise-prediction detector for time series directly comparable to PiAD. Nevertheless, as reported in our earlier response, we have additionally compared against eight-step DDIM variants of IMDiffusion (VLDB 2023) [2] and EnsDiffAD (TKDE 2026) [3], matching PiAD’s eight inversion steps. PiAD achieves higher VUS-PR and lower inference latency on all five datasets. If you are aware of other specific methods, we would appreciate the references and would be happy to include direct comparisons.

Regarding (ii), we have added discussions of prototype- and retrieval-conditioned diffusion in Section 2.2 and Appendix A.1.2, covering RATD (NeurIPS 2024) [4], TimeDP (AAAI 2025) [5], and K-ProtoDiff (AAAI 2026) [6]. These methods use retrieved sequences or learned prototypes to guide diffusion-based forecasting or generation. PiAD shares the idea of reference-conditioned diffusion, but its TSAD-specific design couples prototype-conditioned dynamics with prototype-specific normality references for anomaly scoring. Specifically, temporal prototypes condition the noise predictor and regularize predicted clean latents, while each test trajectory is evaluated against its matched prototype’s statistics at multiple inversion states and token positions. These deviations are calibrated and aggregated into point-level anomaly scores without test-time reconstruction. The added discussions clarify that our contribution concerns how prototype conditioning and reference estimation work together for reconstruction-free detection, rather than prototype conditioning alone.

References

[1] Wang et al., "Drift Doesn't Matter: Dynamic Decomposition with Diffusion Reconstruction for Unstable Multivariate Time Series Anomaly Detection," NeurIPS, 2023.

[2] Chen et al., "ImDiffusion: Imputed Diffusion Models for Multivariate Time Series Anomaly Detection," VLDB, 2023.

[3] Ma et al., "EnsDiffAD: Ensemble Diffusion Models for Multivariate Time Series Anomaly Detection," TKDE, 2026.

[4] Liu et al., "Retrieval-Augmented Diffusion Models for Time Series Forecasting," NeurIPS, 2024.

[5] Huang et al., "TimeDP: Learning to Generate Multi-Domain Time Series with Domain Prompts," AAAI, 2025.

[6] Duan et al., "K-ProtoDiff: Key Prototypes-Guided Diffusion for Time Series Generation," AAAI, 2026.

AuthorsOpenReview ↗

Responses to Q4, Q5 & W1 and W2

Q4: “How do results change with stronger linear baselines on these datasets? Including them would help contextualize the gains”

Re: Table 1 (Section 5.2) already includes PCA as a linear baseline. Furthermore, following your suggestion, we have additionally compared PiAD with FITS (ICLR 2024) [1], a lightweight time-series baseline built around a complex-valued linear layer. The results below report VUS-PR / VUS-ROC / Affiliation-F1 (%):

DatasetFITSPiAD
SMD43.14 / 85.47 / 87.2656.20 / 95.24 / 91.40
SWaT43.76 / 75.54 / 78.8264.82 / 90.57 / 81.23
PSM17.48 / 52.65 / 69.4334.62 / 70.96 / 80.45
MSL71.73 / 93.17 / 98.8492.17 / 98.49 / 99.66
SMAP91.34 / 96.84 / 97.4499.70 / 99.94 / 99.84

PiAD outperforms FITS across all 15 dataset–metric combinations, with VUS-PR improvements ranging from 8.36 to 21.06 %. This additional comparison supports the gains over a strong linear-architecture baseline. If you are aware of other strong linear baselines, we would welcome your suggestions and be happy to evaluate them under the same protocol.

Q5 & W1: “What is the impact of the diagonal-Gaussian assumption for the prototype”

Re: The diagonal-Gaussian assumption applies within each prototype, inversion state, and token position, rather than imposing a single Gaussian on all normal trajectories. Intuitively, different prototypes provide separate references for heterogeneous normal modes, while per-dimension variances accommodate axis-aligned anisotropy. Thus, prototype conditioning and diagonal reference statistics serve complementary roles.

In Eq. (16), the practical effect is to standardize each latent dimension by its prototype-specific variability before computing the deviation. This omits explicit cross-dimensional covariance modeling in the scoring reference, but does not impose independence on the encoder or diffusion predictor, which can still represent inter-variable dependencies.

During the development of our method, we actually experimented with using full covariance matrices, but found that this led to a significant increase in computational cost. Maintaining a separate covariance for every prototype, state, and token increases covariance storage from O(D)O(D) to O(D2)O(D^2) per reference and requires regularization when reference samples are limited. We therefore use the diagonal approximation as a trade-off between correlation modeling, sample-efficient estimation, and lightweight scoring.

W2: “Deterministic DDIM inversion is assumed to remain well-behaved across multiple steps, but error accumulation in the noise-prediction chain and its stability under covariate shift or distribution drift is not analyzed.”

Re: Several components of PiAD may help mitigate these effects. The short inversion trajectory limits opportunities for prediction-error accumulation, while prototype conditioning and prototype-specific references accommodate heterogeneous normal regimes rather than relying on a single global reference. Calibration aligns score scales across states, and Top-k aggregation reduces reliance on any single state. These mechanisms may support stability when shifts remain within represented normal regimes.

For severe covariate shift or distribution drift, potential extensions include drift-triggered recalibration of reference statistics, gradual prototype updates using recent, reliably normal windows, and adaptation of the encoder and diffusion predictor when new operating regimes emerge. We will explore these extensions in future work and appreciate your suggestion for highlighting this promising direction.

References

[1] Xu et al., "FITS: Modeling Time Series with 10k Parameters," ICLR, 2024.

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

This paper proposes PiAD, which learns temporal latent representations and prototypes of normal operating modes, then trains a prototype-conditioned diffusion predictor. A short DDIM inversion trajectory produces representations at several noise levels. Prototype-specific deviations are calibrated and aggregated into temporal anomaly scores, without reverse reconstruction at inference.

Critical strengths
  1. Mode-conditioned scoring is well motivated. Multiple normal operating modes are poorly represented by a single global reference distribution. Combining prototypes with local, multilevel deviations provides a coherent approach to this problem and a potentially useful TSAD-specific contribution.

  2. The evaluation supports practical interest in the method. The paper reports VUS and affiliation metrics, multiple runs, and efficiency measurements. It also states that 20 baselines were rerun in a common environment with unlabeled validation loss for checkpoint selection. These are meaningful strengths, subject to clarification of the dataset scope.

Critical weaknesses
  1. As the paper acknowledges, reconstruction-free inversion detection and diffusion-path scoring have already been studied in InvAD and DiffPath. The central missing control is a clean-latent-only prototype score using the same trained encoder, objective, and calibration. The clean state already participates in the proposed score, and neither a positive inversion-step sweep nor Top-1 selection provides this control. Furthermore, Eq. (16) measures marginal state deviations rather than conditional transition inconsistency, leaving the proposed explanation insufficiently validated.

  2. Table 4 lists only 1,525/4,575 training/test points for MSL, 2,052/6,157 for SMAP, and a 66-dimensional SWaT instance with 3,749/11,247 points. These may be legitimate subsets or benchmark versions, but exact identifiers and selection rules are needed. Without them, the breadth of the generalization and efficiency claims cannot be assessed reliably.

Questions for authors
  1. With the trained encoder, prototypes, and calibration fixed, how does clean-latent-only scoring compare with the full score? If this control is unavailable, which existing result isolates the benefit of nonzero inversion states from representation learning and prototype conditioning?

  2. What exact files/entities and preprocessing produce Table 4? Please specify selection rules, split construction, cross-entity aggregation, affiliation threshold selection, and VUS tolerance settings, so that the scope of the reported results can be established.

3 comments
AuthorsOpenReview ↗

Responses to Q1 & W1 and Q2 & W2

We sincerely appreciate your encouraging comments and thoughtful suggestions! We respond to each of your questions below.

Q1 & W1: “With the trained encoder, prototypes, and calibration fixed, how does clean-latent-only scoring compare with the full score? If this control is unavailable, which existing result isolates the benefit of nonzero inversion states from representation learning and prototype conditioning?”

Re: Following your suggestion, we compared clean-latent-only scoring with the full score while keeping the trained encoder, diffusion model, prototypes, reference statistics, and calibration parameters fixed. The clean-latent-only variant uses only the calibrated deviation at z(0)\mathbf{z}^{(0)}, whereas the full method retains the original multistate Top-k scoring. No retraining or separate parameter tuning was performed.

VUS-PR (%)SMDSWaTPSMMSLSMAP
z(0)\mathbf{z}^{(0)}-only47.3960.6928.3781.7097.35
Ours (Full)56.2064.8234.6292.1799.70

The full score improves VUS-PR over clean-latent-only scoring on all five datasets by 2.35–10.47%. With the trained model, prototypes, and calibration held fixed, these gains support the additional value of nonzero inversion states beyond the learned clean representation. This complements Figure 1, which illustrates how anomaly evidence varies across inversion states and motivates selective multistate aggregation.

Q2 & W2: “What exact files/entities and preprocessing produce Table 4? Please specify selection rules, split construction, cross-entity aggregation, affiliation threshold selection, and VUS tolerance settings, so that the scope of the reported results can be established”

Re: We agree that the benchmark scope and evaluation protocol should be stated more explicitly. We have revised the dataset description in Appendix A.4.2 and added a dedicated Evaluation Protocol section in the Appendix A.4.4.

Dataset selection and scope. Given the computational cost of evaluating 20 baselines, particularly diffusion-based methods, we use subset versions of MSL, SMAP, SMD, and SWaT provided by TSB-AD (NeurIPS 2024) [1], together with the processed version of the full PSM dataset. Our selection considers dataset length, dimensionality, application domain, test anomaly ratio, and anomaly type. These datasets include challenging patterns such as pronounced distribution shift in MSL and a sustained anomaly lasting more than 1,000 timestamps in SMAP, providing complementary evaluation settings.

Split construction and preprocessing. As clarified in the new Evaluation Protocol section, we preserve the chronological train/test boundary encoded by tr_N in each filename: the first N timestamps form the training prefix, and the remainder form the test sequence. The training prefix is divided chronologically into 80% for model training and 20% for validation. Normalization statistics are estimated exclusively from the model-training subset, and windows are constructed separately within each split, consistent with Appendix A.2.

Affiliation-F1 threshold selection. As clarified in the new Evaluation Protocol section, we apply the same threshold-selection and evaluation procedure to all compared methods. For each method, we generate candidate thresholds from the unique values at the 0th–99th percentiles of its test anomaly scores. We binarize the scores using st>τs_t>\tau and select the threshold that maximizes standard point-wise F1 against the test labels. Affiliation-F1 is then computed from these binary predictions. The use of test-label-assisted oracle threshold selection for evaluating Affiliation-F1 is also adopted in TSB-AD (NeurIPS 2024) [1] and W1W_1-ACAS (ICLR 2026) [2]. This protocol allows Affiliation-F1 to provide a complementary assessment of temporal localization at the point-wise-F1-optimal operating point, while keeping the evaluation procedure consistent across methods.

VUS tolerance settings. To ensure a consistent evaluation protocol, we use a maximum tolerance radius of 100 for all methods across all datasets. Our implementation evaluates every integer radius from 0 to 100, symmetrically extends anomaly ranges within the sequence boundaries, and evaluates 250 threshold positions sampled uniformly over the descending sorted anomaly scores. The resulting areas are averaged across tolerance radii.

References

[1] Liu et al., "The Elephant in the Room: Towards a Reliable Time-Series Anomaly Detection Benchmark," NeurIPS, 2024.

[2] Gil et al., "Adaptive Conformal Anomaly Detection with Time Series Foundation Models for Signal Monitoring," ICLR, 2026.

Thanks for author's response and the additional experimental results, which address most of my concerns. Therefore, I am maintaining my positive rating.

Thank you very much for your positive assessment and continued support! We sincerely appreciate your thoughtful comments and suggestions, which have helped us clarify and strengthen our work.

Reviewer_W6nWOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a methodology for time series anomaly detection (TSAD) based on diffusion models. The methodology is motivated by the observation that many diffusion-based TSAD methods rely on the generative/reconstruction capabilities of diffusion models, which is unnecessary and might be detrimental. Namely the learned noise predictor of diffusion models might have sufficient information to distinguish typical and atypical behavior: the intuition is that for "in distribution" observations the learned noise predictor would tend to a normal distribution, while anomalous trajectory would push the noise predictor to less typical point. Furthermore, to capture the fact that in real-world scenarios the observed distributions might be multimodal and complex, the method is augmented with a prototype learning procedure (k-means clustering of latent trajectories) with the goal of obtaining a richer representation of these distributions. The main empirical evaluation is based on a series of standard benchmarks dataset for TSAD (SMD, SWaT, PSM, MSL, SMAP) using three metrics VUS-PR (VP) and VUS-ROC (VR) and Affiliation-F1 (AF), comparing with other 20 baseline methods, ranging from standard, deep learning, generative and diffusion-based methods. The proposed method achieve top performance in 12/15 tests (combination of dataset + metric).

Critical strengths
  1. Well develop intuition about the necessary components of diffusion-based models for TSAD.
  2. Principled design for the architecture and components utilized to implement the method.
  3. Strong empirical performance against a diverse baseline.
Critical weaknesses
  • The use of Affiliation-F1 depends on having binary or binarized results. In the case of continuous anomaly scores, the evaluation becomes threshold dependent. A discussion/ablation of such thresholds is necessary for a more robust evaluation and conclusions.
Questions for authors
  • How did you threshold the score to compute the Affiliation-F1? Is the threshold selected on validation data, optimized using test labels, or integrated over thresholds?
  • Is there any type of inductive bias being added by Prototype-Guided Diffusion Inversion that is fundamental to solve the problem? Is it possible to characterize it theoretically?
  • Are trajectory statistics and calibration parameters estimated on the same windows for the encoder and diffusion model? Does estimating them on held-out windows change the results?
4 comments
AuthorsOpenReview ↗

Responses to Q1 & W1 and Q2

Thank you for your positive feedback and support, which are a great encouragement to us! Below, we address your questions one by one.

Q1: “How did you threshold the score to compute the Affiliation-F1? Is the threshold selected on validation data, optimized using test labels, or integrated over thresholds?”

Re: We apply the same threshold-selection and evaluation procedure to all compared methods. For each method, we generate candidate thresholds from the unique values at the 0th–99th percentiles of its test anomaly scores. We binarize the scores using st>τs_t>\tau and select the threshold that maximizes standard point-wise F1 against the test labels. Affiliation-F1 is then computed from these binary predictions. The use of test-label-assisted oracle threshold selection for evaluating Affiliation-F1 is also adopted in TSB-AD (NeurIPS 2024) [1] and W1W_1-ACAS (ICLR 2026) [2]. This protocol allows Affiliation-F1 to provide a complementary assessment of temporal localization at the point-wise-F1-optimal operating point, while keeping the evaluation procedure consistent across methods. We have explicitly documented this procedure in the newly added Evaluation Protocol section of the Appendix A.5.

Q2: “Is there any type of inductive bias being added by Prototype-Guided Diffusion Inversion that is fundamental to solve the problem? Is it possible to characterize it theoretically?”

Re: PiAD introduces a mode-conditioned normality bias: normal time series are assumed to comprise multiple operating regimes, and a window should be evaluated relative to its corresponding regime rather than a single global reference. This bias enters through prototype conditioning of the noise predictor, prototype-consistency regularization of predicted clean latents, and prototype-specific reference statistics for inversion states. Together, these components encourage mode-compatible representations and dynamics while retaining variation within each normal mode.

Following your suggestion, we have attempted to provide some theoretical intuition from the perspective of the law of total covariance. Let Z(m)Z^{(m)} denote a latent token at inversion state mm, and GG its normal-mode assignment. Then

Cov(Z(m))=E[Cov(Z(m)∣G)]+Cov(E[Z(m)∣G]).Cov(Z^{(m)}) = E[Cov(Z^{(m)} | G)] + Cov(E[Z^{(m)} | G]).

Here, the outer expectation and covariance are taken over the normal-mode variable GG.

A single global reference combines within-mode variability with between-mode differences. In particular, a single Gaussian approximation may place its center between normal modes and inflate its variance to accommodate their separation. Prototype-specific references instead estimate within-mode statistics, motivating a more appropriate comparison for heterogeneous normal behavior. This decomposition provides a theoretical motivation for mode-specific references. Empirically, Table 2(b) (Section 5.3) further supports the complementarity of prototype conditioning and prototype-specific priors.

References

[1] Liu et al., "The Elephant in the Room: Towards a Reliable Time-Series Anomaly Detection Benchmark," NeurIPS, 2024.

[2] Gil et al., "Adaptive Conformal Anomaly Detection with Time Series Foundation Models for Signal Monitoring," ICLR, 2026.

AuthorsOpenReview ↗

Responses to Q3

Q3: “Are trajectory statistics and calibration parameters estimated on the same windows for the encoder and diffusion model? Does estimating them on held-out windows change the results?”

Re: In our original procedure, the prototype-specific trajectory statistics and score-calibration parameters are estimated from the same training windows used to optimize the encoder and diffusion model.

To assess the effect of this data reuse, we chronologically divided the original training sequence into 60% for model training (A), 20% for validation (V), and 20% for independent reference estimation (H), preserving the original test split. Splitting was performed before windowing to avoid shared timestamps across subsets. Normalization statistics and the prototype bank were fitted on A only, and V was used for unlabeled checkpoint selection. H was excluded from model training and checkpoint selection.

For each seed, we froze the same encoder, diffusion model, and prototype bank and compared reference statistics estimated from either A (in-sample) or H (held-out). Both settings used equal numbers of reference windows, sampled without replacement. Within each setting, the selected windows were used to estimate both trajectory statistics and calibration parameters. All other scoring settings and test data were identical.

VUS-PR (%)SMDSWaTPSMMSLSMAPAvg.
In-sample49.7662.3930.4088.7299.4166.14
Held-out50.7258.6830.6891.6599.4766.24

These controlled-experiment results may differ from the main table because reserving H reduces model-training data, and matching reference-set sizes reduces the available in-sample reference windows.

Estimating the trajectory statistics and calibration parameters on independent held-out windows yields comparable overall performance. This suggests our method does not rely on reusing model-training windows for reference estimation, although the effect varies across datasets.

Thank you for the detailed response and clarifications. I have no further questions. All the best!

Thank you for your positive feedback! We are glad that our responses have addressed your questions. If the clarifications and additional results have resolved your doubts, we would be deeply thankful if you would continue to support our work and consider raising the score. Thank you again for your time and support!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses

We sincerely thank all the reviewers for their constructive feedback and positive assessments, which are a great encouragement to us! All three reviewers gave positive ratings during the initial review, and the feedback received after the rebuttal period remained positive. In particular, the reviewers recognized:

  • The paper reframes diffusion-based time series anomaly detection and addresses two weaknesses of prior diffusion-based TSAD coherently (Reviewer FmPf).
  • The method has a well developed intuition and a principled design (Reviewer W6nW). Its well-motivated mode-conditioned scoring combines prototypes with local, multilevel deviations, providing a coherent approach to this problem and a potentially useful TSAD-specific contribution (Reviewer KQvz). The training-calibrated Top-k aggregation strategy is well motivated, and the ablation studies confirm their complementarity (Reviewer FmPf).
  • The experiments demonstrate strong empirical performance against a diverse baseline (Reviewer W6nW). The evaluation supports practical interest in the method (Reviewer KQvz) and is comprehensive, with PiAD achieving the best average rank across all dataset-metric combinations. Its computational analysis is also the best (Reviewer FmPf).

The reviewers also raised insightful and constructive concerns. We made every effort to address all the concerns by providing sufficient evidence and requested results.

The main revisions and supplementary results are summarized below.

1. Clarified the inductive bias and modeling assumptions [Reviewers W6nW, FmPf]

We explained PiAD’s mode-conditioned normality bias using the law of total covariance: a global reference combines within-mode variability with between-mode differences, whereas prototype-specific references characterize within-mode variability.

We also clarified diagonal-Gaussian references in Eq. (16), including dimension-wise standardization, complementarity with prototype conditioning, and computational and statistical trade-offs relative to full covariance estimation.

2. Clarified the dataset scope and evaluation settings [Reviewers W6nW, KQvz]

We expanded the dataset description to specify subset versions and selection criteria. We added an Evaluation Protocol section in Appendix A.4.4, documenting Affiliation-F1 threshold selection and VUS tolerance and threshold settings, applied consistently to all compared methods.

3. Added controlled experiments on reference estimation and inversion-state scoring [Reviewers W6nW, KQvz]

As suggested by Reviewer W6nW, we compared in-sample and independent held-out reference estimation using the same frozen model and prototype bank, with matched reference-set sizes. Comparable overall performance suggests that PiAD does not rely on reusing model-training windows for reference estimation.

Following Reviewer KQvz’s suggestion, we compared clean-latent-only and full multistate scoring with the trained model, prototypes, reference statistics, and calibration fixed. The full score consistently performed better across datasets, supporting the additional value of nonzero inversion states.

4. Expanded baseline comparisons, sensitivity analyses, and discussions [Reviewer FmPf]

We compared PiAD with DDIM-accelerated IMDiffusion and EnsDiffAD, matching their sampling steps to PiAD’s inversion steps. PiAD achieved higher detection accuracy, shorter training time, and lower inference latency across all datasets. We additionally evaluated FITS (ICLR 2024), which PiAD consistently outperformed across all datasets and metrics.

We supplemented the inversion-step analysis with inference-time measurements and comparisons of uniform, low-noise-dense, and high-noise-dense index schedules, supporting sparse inversion and broad coverage across noise levels. We also evaluated prototype update intervals of one, five, and ten epochs; updating every five epochs performed best across all datasets, supporting the default alongside the existing prototype-count sensitivity analysis.

We expanded the discussion of RATD (NeurIPS 2024), TimeDP (AAAI 2025), and K-ProtoDiff (AAAI 2026) in Section 2.2 and Appendix A.1.2, clarifying PiAD’s TSAD-specific coupling of prototype-conditioned dynamics and prototype-specific normality references. We also discussed mechanisms that may mitigate inversion-error accumulation and shifts within represented normal regimes, together with potential extensions for severe drift, including reference recalibration, gradual prototype updates, and model adaptation.

We sincerely appreciate all the reviewers’ suggestions, which have helped us improve the clarity, experimental support, and positioning of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.