acceptodds
Under review as a conference paper at ICLR 2027

Where Can Normality Go: Learning Admissible Futures for Time Series Anomaly Detection

Abstract

Time-series anomaly detection commonly uses prediction errors or likelihood-based scores to identify deviations from normal temporal evolution. However, even under the same observed history, normal evolution can follow multiple future trajectories, so deviation from a single expected trajectory need not indicate an anomaly. The set of normal continuations also varies with history: the same future trajectory may be admissible after one historical state but inconsistent with another. This motivates representing temporal normality as a history-conditioned region of admissible future trajectories and detecting anomalies when an observed future falls outside it. Learning such a region is challenging because each training history reveals only one realized normal continuation, providing a sample within the region without directly revealing its boundary. To address this, we propose **FutureBound**, a framework that learns this boundary implicitly through local future-space exploration and selective adversarial optimization. A context-conditioned generator produces multiple locally anchored yet diverse trajectories as probes around observed normal continuations. A discriminator then learns context-future compatibility by contrasting the observed continuation with nearby probes under the same history. To refine this distinction, the generator is adversarially optimized to make selected probes increasingly challenging to the discriminator, with proximity-based and direction-aware selection concentrating these updates around normal evolution. This interaction learns an implicit, potentially asymmetric boundary without prescribing its geometry in advance. At inference time, the learned compatibility directly measures whether an observed future agrees with its preceding history. Extensive experiments on four real-world time-series benchmarks demonstrate consistent improvements over existing methods, supporting history-conditioned future admissibility as a flexible reference for distinguishing anomalous deviations from normal temporal variation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Time Series, Anomaly Detection

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_1JacOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a generator-based TSAD framework called FutureBound. The authors claim that “an observed normal continuation can anchor local exploration of the future space, providing comparisons for learning the surrounding boundary.” The paper proposes, rather than scoring anomalies by prediction error or Gaussian likelihood, treating normality as a history-conditioned region of admissible futures. Any observed future that falls outside the region is flagged as anomalous. The method adopts an alternating adversarial training strategy that trains the generator and the discriminator by distinguishing history-conditioned real futures from more realistic generated futures. The proposed method demonstrates significantly improved performance on the selected datasets.

Critical strengths
  1. The improvement on the seleted dataset is significant.
  2. Generator-based TSAD method is a promisning direction.
Critical weaknesses
  1. The paper’s key argument is that normal futures form a multimodal, non-convex, and possibly asymmetric region, which FutureBound can capture but Gaussian or forecasting approaches cannot. I cannot see how this is supported by any experiments.
  2. What is the connection between the two derived propositions and the proposed framework? Propositions 1 and 2 concern irreducible noise and the definition of convexity. Proposition 1 states that any single deterministic predictor is insufficient, while Proposition 2 concerns Gaussian or convex regions. Both suggest that current approaches are insufficient but do not prove the effectiveness of the proposed approach.
  3. The paper argues that single-trajectory-based forecasting is insufficient but does not include an adapted baseline in the experiments.
  4. I am concerned about the validity and fairness of the experiments. Some popular baselines perform surprisingly poorly, with TimesNet scoring 24.3 and PaAno scoring 27.3 on SWaT, and DCdetector scoring around 50 across all datasets. The top-k aggregation looks 15 steps ahead, but the baselines do not receive the same treatment. Some widely used datasets are missing, such as SMAP, WADI, and GECCO. Also, given that the paper is framed as general TSAD, univariate datasets should also be considered.
  5. In principle, the proposed method also generates near-normal negative samples to refine the decision boundary. Learning to discriminate normal data from synthetic negatives has already been explored by NCAD, SPAGD, and learned anomaly generators such as GenIAS. Moreover, NCAD already conditions detection on a preceding context window. The experiments do not show that the proposed adversarial training strategy and history-conditioned futures are more effective.
  6. How do we ensure that the proposed probes provide informative supervision for anomaly detection?
  7. Hyperparameter selection is not described, and no validation split is mentioned. Sensitivity to the top-k aggregation size k is not reported. This aggregation smooths scores and introduces up to H-1 steps of lookahead, so it is important to show that the performance gains do not primarily come from this processing.
Questions for authors

Please refer to the weaknesses.

3 comments

We thank the reviewer for taking the time to evaluate our work. We provide our responses and corresponding revisions below.

To W1: Our motivation is that normal futures may vary under the same historical context, while their admissible geometry is generally unknown. Possible structures include multimodality, non-convexity, asymmetry, etc. Proposition 1 establishes irreducible future variability, while Proposition 2 highlights the limitations of convex regions. Fig. 2 further reveals Gaussian mismatch in prediction residuals. Together, these findings motivate flexible boundary learning.

FutureBound addresses this through a nonlinear discriminator that learns compatibility between historical contexts and future trajectories, allowing decision regions beyond predefined convex shapes. Locally anchored and diverse probes, together with selective adversarial optimization, provide challenging comparisons for learning this boundary.

Our experiments support this design. Figs. 3, 11, and 12 visualize the generated probes and discriminator responses, while Fig. 8 validates the contributions of future-space exploration and adversarial optimization. These results demonstrate effective discrimination rather than exact recovery of the underlying region. Unlike predictive distribution estimation, FutureBound directly learns compatibility for anomaly detection, while flexible probabilistic forecasters can also represent multiple futures.

To W2: Propositions 1 and 2 motivate FutureBound's design rather than guarantee its effectiveness. Proposition 1 establishes irreducible normal future variability, motivating diverse, locally anchored probes beyond a single prediction. Proposition 2 highlights the limitations of convex decision regions, motivating our geometry-unconstrained discriminator, whose ability to represent non-convex regions is demonstrated in our preceding response.

These principles are implemented through local probe exploration and selective adversarial boundary refinement. While the propositions do not guarantee recovery of the true boundary, Figures 3 and 8 and Table 1 provide empirical evidence of the framework's effectiveness. We have clarified these connections in the revision.

To W3: We have supplemented results with GDN, a representative forecasting-based anomaly detector that predicts future observations from historical data and identifies anomalies through prediction errors. As shown in the table below, FutureBound achieves higher metrics than GDN across all four datasets. These results provide empirical evidence that learning history-conditioned future admissibility offers practical advantages over relying on a single predicted trajectory for anomaly detection.

$\begin{matrix} \hline &&SWaT&&&PSM&&&MSL&&&SMD&\\ \hline &AUC&AUPRC&F1-Score&AUC&AUPRC&F1-Score&AUC&AUPRC&F1-Score&AUC&AUPRC&F1-Score\\\hline GDN &85.09\pm2.11&65.75\pm11.46&69.19\pm6.82 &70.48\pm4.72&45.70\pm1.97&52.93\pm4.19 &53.52\pm0.16&12.12\pm0.16&19.48\pm0.59 &66.95\pm1.52&10.90\pm0.88&14.58\pm4.65\\ FutureBound&\bf85.45\pm1.16&\bf75.12\pm0.74&\bf76.84\pm0.02&\bf76.19\pm3.27&\bf54.39\pm4.39&\bf56.77\pm2.37&\bf72.34\pm0.92&\bf24.19\pm1.42&\bf31.67\pm1.34&\bf77.20\pm0.54&\bf21.16\pm3.35&\bf28.05\pm2.93\\ \hline \end{matrix}$

To W4: All methods are compared under a consistent offline TSAD setting, using the same data splits and point-wise evaluation metrics. Specifically, (1) hyper-parameters follow the original papers or public implementation defaults. TimesNet and DCdetector report point-adjusted results, whereas we evaluate raw scores without point adjustment or label-based score modification (see our response to Reviewer 6Mxy, Q4&W6 for more details). PaAno's published results use the TSB-AD-M benchmark, which differs from our datasets. (2) The 15-step look-ahead does not inherently compromise fairness in our offline setting, where the complete observation sequence is available to all methods. Different uses of temporal context and score aggregation reflect algorithmic design choices rather than unequal access to data, thus identical aggregation is not required. Additional comparisons with look-ahead-enhanced baselines are provided in our response to W7.

Finally, additional experiments on SMAP and TSB-AD-U are presented below and the complete version have been incorporated into the revised manuscript.

$\begin{matrix}\hline Dataset&Method&AUC&AUPRC&F1-Score\\\hline&SPAGD&51.86\pm1.60&14.23\pm1.52&19.06\pm7.37\\&AnomalyTrans&53.05\pm4.84&14.06\pm1.86&24.70\pm2.85\\&TimesNet&46.79\pm0.10&11.98\pm0.04&24.28\pm0.15\\SMAP&TranAD&42.51\pm0.96&10.59\pm0.21&23.00\pm0.05\\&PaAno&48.21\pm1.25&13.46\pm0.14&25.01\pm0.22\\&SDA2D&49.52\pm1.85&12.41\pm0.19&22.99\pm0.16\\&FutureBound&\bf60.02\pm1.84&\bf15.58\pm0.03&\bf29.75\pm1.70\\\hline\end{matrix}$

$\begin{matrix}\hline&OCSVM&LOF&IForest&AutoEncoder&Donut&OmniAnomaly&USAD&TFAD&DCdetector&DualTF&CATCH&MTGFlow&PaAno&SDA2D&FutureBound\\\hline VUS-PR&9.37&22.50&22.57&26.71&15.94&25.75&25.41&13.34&23.55&24.01&30.65&41.39&34.55&38.45&\bf42.08\pm0.47\\VUS-ROC&52.55&76.84&74.75&73.06&74.09&69.39&75.59&48.01&70.25&67.76&71.77&72.40&79.89&80.58&\bf82.69\pm0.31\\\hline\end{matrix}$

To W5: These methods differ in sample generation and optimization. NCAD uses predefined perturbations. SPAGD generates adaptive pseudo-anomalies through reconstruction and uses their residuals for anomaly-aware graph learning. GenIAS augments existing detectors through learnable VAE latent-variance perturbations constrained by predefined distance margins and patching rules. Yet FutureBound generates locally anchored, diverse, history-conditioned future probes without prescribing anomaly patterns or separation margins. Trajectory proximity and residual orientation guide their selective adversarial refinement through discriminator feedback, making them progressively harder to distinguish from normal futures. Our contribution thus extends beyond synthetic anomaly generation to directly learning the history-conditioned admissible boundary. We have add these discussions to Section 3.

Table 1 has already demonstrated that FutureBound outperforms NCAD and SPAGD, while Figs. 4 and 13 provide complementary anomaly score visualizations. For clarity, here we present the results again. $\begin{matrix} \hline &&SWaT&&&PSM&&&MSL&&&SMD&&&SMAP&\\\hline &AUC&AUPRC&F1-Score&AUC&AUPRC&F1-Score&AUC&AUPRC&F1-Score&AUC&AUPRC&F1-Score&AUC&AUPRC&F1-Score\\\hline NCAD&81.16\pm2.29&53.54\pm4.19&70.56\pm1.16 &60.94\pm5.39&37.88\pm2.49&46.18\pm3.63 &60.67\pm1.53&15.06\pm0.35&22.77\pm0.77 &60.71\pm0.44&7.81\pm0.28&13.80\pm0.44 &46.44\pm1.18&12.40\pm0.33&23.11\pm0.04\\ SPAGD&73.76\pm4.24&29.55\pm6.51&40.35\pm9.06 &49.64\pm6.94&30.14\pm3.97&48.77\pm1.20 &54.93\pm2.87&16.73\pm3.41&26.33\pm1.70 &51.17\pm0.60&9.77\pm0.84&16.63\pm0.19 &51.86\pm1.60&14.23\pm1.52&19.06\pm7.37\\ FutureBound&85.45\pm1.16&75.12\pm0.74&76.84\pm0.02 &76.19\pm3.27&54.39\pm4.39&56.77\pm2.37 &72.34\pm0.92&24.19\pm1.42&31.67\pm1.34 &77.20\pm0.54&21.16\pm3.35&28.05\pm2.93 &60.02\pm1.84&15.58\pm0.03&29.75\pm1.70\\\hline \end{matrix}$

GenIAS's official implementation requires 1,000 pretraining epochs without early stopping, followed by downstream detector training, exceeding our computational budget. We therefore provide only a methodological comparison, without claiming empirical superiority.

To W6: Lfit\mathcal{L}_{\mathrm{fit}} anchors probes to normal evolution, while Ldiv\mathcal{L}_{\mathrm{div}} promotes diversity and discourages collapse. The discriminator selects nearby probes to avoid trivial discrimination, whereas the generator refines probes from different residual orientations. Adversarial feedback raises their compatibility scores, creating harder local comparisons for history-future discrimination. Their supervisory value therefore comes from adaptive comparisons, not predefined anomaly patterns or assuming every probe is anomalous.

Fig. 8 supports this mechanism: removing fitting or diversity degrades performance, while Generator-Only and Gaussian Perturbation-GAN assess the benefits of adversarial discrimination and learned probes, respectively. Fig. 3 visualizes the resulting probes and boundary responses.

To W7: Appendix E provides the hyperparameter settings, which are fixed across datasets. Training uses only normal data without a separate validation set.

We additionally evaluate k∈{1,3,5,10,All}k\in\lbrace 1,3,5,10,All\rbrace on SMD and MSL (Appendix H, Fig. 17(b)–(c)), where "All" averages all overlapping trajectory scores. Performance remains relatively stable, and the default k=3k=3 is not the best-performing setting, indicating that our results do not rely on selecting a particularly favorable aggregation size.

$\begin{matrix}\hline &&SMD&&&MSL&\\ k&AUC&AUPRC&F1-Score&AUC&AUPRC&F1-Score\\ \hline 1&77.58&22.14&28.98&72.67&24.46&31.57\\ 3&77.20&21.16&28.05&72.34&24.19&31.67\\ 5&77.87&22.49&28.95&73.03&25.55&32.92\\ 10&77.88&23.19&28.54&72.88&26.05&32.70\\ All&77.90&23.71&28.27&70.56&24.25&34.68\\ \hline \end{matrix}$

The maximum H−1H-1 detection latency arises from evaluating complete future trajectories. To assess the influence of lookahead and aggregation, we additionally apply the same 15-step lookahead to SPAGD, PaAno, and SDA2D (Table 3). FutureBound retains the highest mean performance across PSM, SMD, and MSL, providing evidence that its improvements are not solely attributable to this processing.

$\begin{matrix} \hline Dataset&Method&AUC&AUPRC&F1-Score\\\hline PSM&SPAGD&49.45\pm7.48&29.84\pm3.72&50.90\pm3.82\\ &PaAno&67.31\pm3.28&41.56\pm2.90&50.73\pm3.04\\ &SDA2D&71.61\pm0.54&50.85\pm0.56&51.39\pm0.29\\ &FutureBound&\bf76.19\pm3.27&\bf54.39\pm4.39&\bf56.77\pm2.37\\ \hline SMD&SPAGD&52.17\pm2.05&9.83\pm0.73&16.95\pm0.46\\ &PaAno&77.14\pm1.50&10.82\pm0.90&20.21\pm0.73\\ &SDA2D&74.28\pm0.27&16.78\pm0.44&24.21\pm0.15\\ &FutureBound&\bf77.20\pm0.54&\bf21.16\pm3.35&\bf28.05\pm2.93\\ \hline MSL&SPAGD&52.35\pm3.97&17.07\pm2.42&29.45\pm2.53\\ &PaAno&70.70\pm1.98&20.78\pm0.95&28.71\pm2.15\\ &SDA2D&55.81\pm0.15&14.39\pm0.45&21.02\pm0.25\\ &FutureBound&\bf72.34\pm0.92&\bf24.19\pm1.42&\bf31.67\pm1.34\\\hline \end{matrix}$

We hope our explanations can clarify the points raised. Please feel free to reach out if any further clarification is needed.

Reviewer_sgrMOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper detects time-series anomaly by treating normality as a history-conditioned region of admissible future trajectories, rather than a single expected path. It uses a context-conditioned generator to produce multiple local "probes" around observed normal continuations. A discriminator learns context-future compatibility by contrasting observed futures with these probes. Selective adversarial optimization, guided by proximity and residual orientation, refines the implicit boundary.

Critical strengths
  1. The paper is clear and easy to follow.
  2. GAN-like architectures follow existing Time Series Anomaly Detection works and are reasonable.
Critical weaknesses
  1. Training Instability and Hyperparameter Sensitivity: The framework relies on a delicate adversarial min-max game between the generator and discriminator, combined with multiple loss terms (fitting, diversity, adversarial). Figure 5 and the ablation studies reveal extreme sensitivity to almost all hyperparameters. In practice, tuning this multi-component adversarial system for new, unseen datasets will be prohibitively difficult and highly unstable, limiting its practical adoption compared to simpler reconstruction-based methods.
  2. Inference efficiency and Context Rigidity: The model uses fixed historical ( L ) and future ( H ) horizons for all temporal contexts. As acknowledged in the conclusion, this rigidity prevents the model from adapting to multi-scale temporal dynamics (e.g., sudden spikes vs. long-term drifts), which is a severe limitation for real-world multivariate time series where anomalies manifest at vastly different temporal resolutions.
  3. Are the results sensitive to historical ( L ) and future ( H ) horizons?
  4. Weak Baselines and Architectures: The paper fails to compare against modern, highly optimized models (e.g., CrossAD) with more high-quality datasets (e.g., TSB-AD-U and TSB-AD-M). Demonstrating superiority only over older or less capable baselines and old datasets further weakens the claim of performance.
  5. Limited Methodological Novelty.
Questions for authors

See Weaknesses

8 comments

We appreciate the reviewer's thoughtful feedback and constructive suggestions. We have carefully considered the concerns raised and provide our detailed responses below, along with the corresponding revisions to the manuscript.

To W1: Fig. 7 and Appendix H reveal systematic and interpretable trends rather than erratic performance fluctuations. Increasing MM or HH generally yields diminishing returns as future-space exploration becomes redundant or more challenging. Performance is relatively robust to variations in β\beta, while α\alpha has a stronger influence because it directly governs the balance between local anchoring and exploration. Performance degradation under extreme settings should not be mistaken for training instability.

More importantly, FutureBound adopts the same hyper-parameter configuration across all datasets without specific tuning, demonstrating its practical applicability. The training-loss curves in Fig. 5 show no sustained divergence, while results over ten independent runs provide additional evidence of empirical reliability. In general, these findings do not support the concern that FutureBound requires prohibitively delicate tuning.

To W2: Fixed LL and HH would not necessarily restrict FutureBound to anomalies of a particular temporal scale. Through sliding-window evaluation, both the historical context and future trajectory are continuously updated, allowing the learned decision boundary to adapt to changing temporal dynamics. Consequently, the model can detect both short-lived deviations and sustained anomalies without requiring different horizons. Figs. 11 and 12 further demonstrate its responses to abrupt and gradual anomalous behaviors. So we acknowledge that adaptive horizons could offer additional flexibility, but this is distinct from an inability to detect anomalies at different temporal scales.

Besides, the adversarial framework introduces no additional optimization overhead during inference: each window requires only temporal context encoding and a single discriminator evaluation. As analyzed in Appendix C.1, the resulting computational complexity scales linearly with LL, HH, and DD for a fixed network architecture, supporting efficient sliding-window inference.

To W3: The sensitivity to the future horizon HH has already been evaluated in Fig. 7 and Appendix H (Figs. 14 and 15). Short horizons may miss important temporal evolution, while excessively long horizons increase the complexity of future-space exploration, leading to diminishing returns or performance degradation.

We further investigate the historical context length LL on PSM, as shown below. Performance generally improves as LL increases, suggesting that richer historical information helps characterize future admissibility. However, excessively long histories may introduce redundant information and reduce detection effectiveness. Since LL and HH directly determine the historical context and future space being modeled, their influence on performance is both expected and meaningful.

$\begin{matrix}\hline L&16&32&64&100\\\hline AUC&74.23&77.26&\bf80.95&76.19\\ AUPRC&52.05&55.42&\bf60.95&54.39\\ F1-Score&56.55&56.52&\bf60.34&56.77\\\hline \end{matrix}$

To W4&W5: Our experiments already include 20 baselines across five major TSAD paradigms, including three SOTAs from 2026 (PaAno, GBOC, and SDA2D) and two from 2025 (GCAD and SPAGD). Thus, our evaluation covers both established and recent approaches. We further supplement CrossAD as a relevant additional baseline below.

$\begin{matrix} \hline Method&&SWaT&&&PSM&&&MSL&&&SMD&&&SMAP&\\\hline &AUC&AUPRC&F1-Score&AUC&AUPRC&F1-Score&AUC&AUPRC&F1-Score&AUC&AUPRC&F1-Score&AUC&AUPRC&F1-Score\\\hline CrossAD&81.08\pm0.00&71.50\pm0.00&76.16\pm0.00&64.19\pm0.20&46.71\pm0.14&45.99\pm0.25&70.02\pm1.23&\bf26.28\pm0.13&31.59\pm0.03 &77.04\pm0.11&15.92\pm0.04&24.58\pm0.01&56.74\pm0.03&13.61\pm0.01&27.35\pm0.07\\ FutureBound&\bf85.45\pm1.16&\bf75.12\pm0.74&\bf76.84\pm0.02&\bf76.19\pm3.27&\bf54.39\pm4.39&\bf56.77\pm2.37&\bf72.34\pm0.92&24.19\pm1.42&\bf31.67\pm1.34&\bf77.20\pm0.54&\bf21.16\pm3.35&\bf28.05\pm2.93& \bf60.02\pm1.84& \bf15.58\pm0.03& \bf29.75\pm1.70\\\hline \end{matrix}$

We also provide the experimental results on the TSB-AD-U dataset. The results are also added in the manuscript in Table 5. The results are shown as follows:

$\begin{matrix}\hline&OCSVM&LOF&IForest&AutoEncoder&Donut&OmniAnomaly&USAD&TFAD&DCdetector&DualTF&CATCH&MTGFlow&PaAno&SDA2D&FutureBound\\\hline VUS-PR&9.37&22.50&22.57&26.71&15.94&25.75&25.41&13.34&23.55&24.01&30.65&41.39&34.55&38.45&\bf42.08\pm0.47\\VUS-ROC&52.55&76.84&74.75&73.06&74.09&69.39&75.59&48.01&70.25&67.76&71.77&72.40&79.89&80.58&\bf82.69\pm0.31\\\hline\end{matrix}$

To W6: Our theoretical analysis provides the rationale for learning a flexible normal-future region. Proposition 1 establishes irreducible normal future variability, motivating a reference beyond a single predicted trajectory. Proposition 2 identifies the limitations of convex representations when the admissible region is non-convex, motivating boundary learning without predefined convexity constraints.

Our methodological contribution lies in coupling locally anchored, diverse future probes with selective adversarial boundary refinement. The discriminator learns from nearby probes, while the generator adversarially refines probes from different residual orientations to make them harder to distinguish from normal futures. This interaction learns history-conditioned admissibility without prescribing anomaly patterns or separation margins. Section 3 explains how this mechanism differs from existing approaches.

The experiments evaluate both the intended boundary-learning behavior and the effectiveness of the proposed mechanism. Figs. 3, 11, and 12 provide qualitative evidence that probe exploration and boundary responses vary with historical context. Table 2 and Fig. 9 show that retaining residual orientation and using distinct generator-discriminator selection criteria improve detection over their respective alternatives. Together with the component ablations in Fig. 8 and detection comparisons in Table 1, these results support the value of selective adversarial probing beyond simply generating synthetic samples.

We hope these clarifications and revisions can adequately address your concerns. We would be happy to provide further clarification if needed.

This paper is not just Limited in Methodological Novelty. I find this paper raises ethics concerns including research integrity and plagiarism. I keep the original score as strong rejection.

Evidence of Substantial Methodological Similarities Between EVOLVE and FutureBound

[1] https://arxiv.org/abs/2610.32189. EVOLVE: Uncertainty-Aware Multivariate Time Series Anomaly Prediction with Adaptive Prototype Memories

We identify substantial methodological similarities between EVOLVE and FutureBound across several interconnected aspects, including their underlying research motivations, history-conditioned multi-future modeling, the joint consideration of consistency and diversity, challenging synthetic sample generation, and alternating adversarial optimization.

Our concern does not arise from the isolated use of common techniques. Rather, it concerns the cumulative overlap in a coherent methodological design, where multiple closely related ideas are combined to address uncertainty in future temporal evolution and improve anomaly discrimination.

1.1. Similar Core Motivation: Limitations of Deterministic Future Modeling and History-Dependent Evolution

Relevant locations:

  • EVOLVE: Introduction (Challenges 1 and 2); Section III-B, Design Principles.
  • FutureBound: Introduction; Section 2.1, Problem Formulation and Motivation.

1.2. History-Conditioned Generation of Multiple Plausible Future Trajectories

Relevant locations:

  • EVOLVE: Section III-E, Uncertainty-Aware Multi-Hypothesis Future Modeling, Eqs. (11)–(27); Figure 2.
  • FutureBound: Section 2.2, Contextual Future Space Exploration, Eq. (3); Figure 1.

1.3. Joint Modeling of Normal-Future Consistency and Trajectory Diversity

Relevant locations:

  • EVOLVE: Section III-F, Training Objective and Adversarial Optimization, Eqs. (28)–(35).
  • FutureBound: Section 2.2, Multi-Future Trajectory Generation and Diversity Regularizer, Eqs. (4)–(6).

1.4. Adversarial Generation of Challenging Synthetic Patterns Near Normal Evolution

Relevant locations:

  • EVOLVE: Section III-D, Adaptive Anomaly Prototype Modeling; Section III-F, Perturbation Generator Optimization, Eqs. (38)–(40).
  • FutureBound: Section 2.3, Boundary-Oriented Adversarial Learning, Eqs. (7)–(10).

1.5. Closely Related Alternating Generator–Discriminator Optimization

Relevant locations:

  • EVOLVE: Section III-F, Eqs. (37)–(40).
  • FutureBound: Section 2.3, Eqs. (7)–(10); Appendix C, Algorithm 1.

1.6. Cumulative Similarity in the Overall Methodological Design

Relevant locations:

  • EVOLVE: Figure 2; Sections III-C–III-F.
  • FutureBound: Figure 1; Sections 2.2–2.3; Appendix C, Algorithm 1.
Reviewer_6MxyOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that prediction-error and likelihood-based TSAD scores are fundamentally mis-specified because (i) normal evolution admits multiple legitimate future trajectories from the same history, so deviation from a single expected trajectory need not indicate an anomaly (Prop. 1), and (ii) the set of admissible futures is history-dependent — the same continuation may be normal after one context and anomalous after another. The authors therefore recast temporal normality as a history-conditioned region of admissible future trajectories and detect anomalies when the observed future falls outside it.

Because each training history reveals only one realized continuation, the boundary is never directly observed. FutureBound learns it implicitly: a context-conditioned generator produces multiple locally anchored but diverse future probes around the observed continuation; a discriminator learns context–future compatibility by contrasting the real continuation with nearby probes under the same history; and the generator is adversarially optimized on selectively chosen probes (proximity-based and direction-aware) to sharpen the boundary. At inference, the learned compatibility score directly measures whether an observed future agrees with its preceding history. Experiments on SWaT, PSM, MSL and SMD report AUC/AUPRC/F1 gains over 20+ baselines without point adjustment.

Critical strengths
  1. Well-motivated reformulation. The observation that normal futures are multi-trajectory and context-dependent is correct and practically important; framing the detector as a history-conditioned admissibility region is a genuinely different object from the global-latent-ball view of traditional methods, and the paper is explicit about this distinction.
  2. Sensible design for an under-specified supervision problem. Using generated probes exclusively as local boundary probes (not as samples of a global normal distribution) is the right instinct when each history yields one label.
  3. Selective adversarial refinement. Restricting adversarial updates to proximity/direction-selected probes is a reasonable way to concentrate boundary refinement near normal evolution, and the asymmetric-vs-symmetric probe ablation (Fig. 7) provides direct evidence for this component.
  4. Good analytical hygiene in the appendix. Decision-boundary visualizations, generator ablations, and efficiency analysis (SWaT/PSM/MSL) support the claims beyond the main table.
Critical weaknesses

The paper's own premise — that normal futures are multi-trajectory and contextual — is never measured directly, and the theory only motivates rather than constrains the algorithm; combined with weak baselines and unspecified F1 thresholding, the mechanism's success cannot be attributed to the stated reason.

  1. Theory motivates but does not justify the algorithm. Prop. 1/2 establish that a single-trajectory reference is insufficient and that an admissible region exists. Neither shows that the generator–discriminator–selective-adversary loop converges to such a region or estimates its boundary. There is no convergence, consistency, or estimation-error result. Prop. 1 also assumes an optimal deterministic predictor in a (conditional-mean) sense, which is a strong assumption for non-stationary industrial telemetry that the paper does not discuss.
  2. GAN stability and mode collapse are not addressed. Adversarial optimization plus hard-example selection is a recipe for instability; the paper reports no discriminator-collapse diagnostics or mitigation, and the "Gaussian Perturbation-GAN" ablation is itself inconsistent across datasets — a warning sign.
  3. Many hand-set selection knobs. Proximity radius, direction criterion, and selection ratio govern the core mechanism, yet their sensitivity and individual necessity are relegated to the appendix; the main text does not show that the direction rule (vs. proximity alone) matters.
  4. Per-window inference cost and score comparability. Each window requires generating multiple probes and discriminating them; cost scales with probe count. More importantly, different histories induce regions of different "scale," and no per-window normalization/calibration is discussed — which directly affects whether AUC is a meaningful comparison across windows.
  5. Baseline strength is questionable. TimesNet (24.33 AUC on SWaT), DCdetector (49.90), GCAD (45.39), GBOC (36.83) are far below their commonly reported performance, suggesting under-tuned or protocol-mismatched baselines. This inflates the apparent margin.
  6. F1 threshold selection is unspecified. F1 is threshold-dependent; the paper does not state how the operating point is chosen, so the F1 comparisons may carry optimistic bias.
  7. Narrow empirical domain. All four datasets are industrial/server/spacecraft multivariate sensors; no medical, financial, or strongly periodic series. The central claim about multi-modal futures is never quantified directly (e.g., coverage or multi-modality of the learned region); only downstream AUC is reported.
  8. No variable- or interval-level localization. Only window-level detection is reported. For industrial deployment — the stated motivation — identifying which sensor makes the future inadmissible is often the actual requirement.
Questions for authors
  1. Do you have any result characterizing the learned region itself (volume, asymmetry, multi-modality, coverage of held-out normal futures)? Can the central "multi-trajectory normality" claim be measured rather than inferred from AUC?
  2. Prop. 1's optimal-deterministic-predictor assumption: how does the argument degrade under non-stationarity and distribution shift, which are the norm in SWaT/SMD?
  3. How are the probe-selection thresholds (proximity radius, direction rule, selection ratio) set, and how sensitive is performance to them? Please provide an ablation isolating direction-aware from proximity-only selection.
  4. How is the F1 operating threshold chosen across datasets? Is it tuned on validation normal data or on test labels? Why are several baselines (TimesNet, DCdetector, GCAD) so far below their reported literature numbers? Please clarify their configuration and tuning budget.
  5. Are per-window scores normalized to be comparable across histories? If not, how does this interact with AUC? Can the method produce variable- or interval-level localization, and how would it be evaluated?
3 comments

We sincerely thank the reviewer for the comments and suggestions. To address the concerns raised, we provide the following responses and have made corresponding revisions to the manuscript.

To Q1: FutureBound learns an admissible future region through the discriminator's compatibility score. For a historical context hh and threshold bb, the learned region is defined as A^b(h)={F:Dψ(h,F)≥b}\widehat{\mathcal A}_b(h)=\{F:D_\psi(h,F)\geq b\}. The generated probes explore trajectories around observed normal continuations, while the discriminator learns the compatibility function that defines the region.

We provide empirical evidence beyond AUC results. Figs. 3, 11, and 12 visualize how generated probes cover nearby trajectories and how discriminator responses distinguish normal and anomalous evolution. Fig. 8 further demonstrates that learning the compatibility function improves detection compared with relying on generated trajectories alone. These results provide evidence of the learned scoring behavior and the benefit of boundary learning.

Regarding geometric characterization, each history provides only one observed normal continuation. This makes the true admissible region and its geometry difficult to identify or validate. Moreover, the geometry of the learned region depends on the decision threshold and does not necessarily reflect the true normal support.

To Q2: Proposition 1 does not assume that a trained predictor is optimal. Its theoretical reference is the conditional mean under the current normal dynamics, which may change over time. Even if this mean were known exactly, unpredictable normal variation would still cause nonzero expected prediction error. The argument therefore does not require stationarity.

Under distribution shift, a previously trained predictor may additionally fail to match the new conditional mean. Its error then includes both irreducible normal variability and this additional mismatch. The mismatch can increase prediction error, but it does not invalidate the irreducible-error argument in Proposition 1.

To Q3: Our method does not use a proximity radius or distance cutoff. Probes are ranked by the discrepancy in Eq. (4). The direction rule uses the sign of the mean residual in Eq. (8), with a fixed split at zero. Appendix E specifies M=4M=4, KD=2K_{\mathcal{D}}=2, and KG=1K_{\mathcal{G}}=1 per direction. Thus, both updates use two of the four probes: the discriminator selects the two nearest probes, while the generator selects one nearby probe from each direction. Missing positions are filled with the nearest unused candidates. Please find the parameter sensitivity of selection ratio in Fig. 6, and the direction-ware ablation can be found in our responses to your W4.

To Q4&W6: For DCdetector and TimesNet, they used the PA strategy, which would influence performance significantly. To be more specific:

  • For DCdetector, we use the authors’ public implementation with a 3-layer encoder, 256 hidden dimensions, one attention head, Adam with 3 training epochs, and batch size 128 for all datasets. The window/patch settings are SWaT: 105/([3,5,7]), PSM: 60/([1,3,5]), MSL: 90/([3,5]), and SMD: 105/([3,5,7]); input/output channels match the dataset dimensions (51/25/55/38). These settings largely follow the released code.

  • For TimesNet, we use the implementation provided by the TSB-AD benchmark with its default model and training settings: window length 96, one TimesBlock layer, hidden dimension 8, feed-forward dimension 16, top-k=3k=3, and dropout 0.1. The model is trained with Adam and MSE loss using a learning rate of 10−410^{-4}, batch size 128, and up to 10 epochs.

  • For GCAD, an important protocol difference is the label assignment strategy: the original implementation labels each window by taking the maximum ground-truth label over all points, meaning that a window is considered anomalous if any timestamp within it is anomalous. In contrast, our implementation uses only the label of the final timestamp. For the model configuration, we largely follow the paper and released code. The dataset-specific β\beta values are 0 for SMD, 0.5 for SWaT, 0.5 for PSM, and 0 for MSL. Other settings include batch size 128 and score smoothing with a window of 3.

Regarding F1 evaluation, we follow prior TSAD studies [1,2,3,4] and report the best F1-Score obtained by optimizing the decision threshold. This protocol avoids reliance on a predefined threshold and uses test labels solely for evaluation, without affecting model training.

References:

[1] PaAno: Patch-based representation learning for time-series anomaly detection. ICLR, 2026.

[2] GCAD: Anomaly detection in multivariate time series from the perspective of granger causality. AAAI, 2025.

[3] Calibrated one-class classification for unsupervised time series anomaly detection. TKDE, 2024.

[4] CrossAD: Time series anomaly detection with cross-scale associations and cross-window modeling, NeurIPS. 2026.

To Q5&W5: We do not apply per-window score normalization. All trajectory scores are produced by the same context-conditioned discriminator with shared parameters, rather than by separately calibrated models for different histories. Therefore, different windows are evaluated on a common scoring function. Spectral normalization is additionally applied to all discriminator linear layers to control its output scale.

Second, FutureBound already produces timestamp-level anomaly scores by aggregating the overlapping trajectory scores in Eq. (12), so anomalous intervals can be directly localized from consecutive high-scoring timestamps.

To W1&W2: Propositions 1 and 2 provide the theoretical motivation for FutureBound rather than convergence guarantees. FutureBound translates these insights into locally anchored and diverse probes, whose selective adversarial refinement enables learning an implicit history-conditioned boundary.

Proposition 1 does not assume an optimal trained predictor. It considers all deterministic predictors, using the conditional mean only to characterize the minimum achievable MSE. This argument does not require stationarity: the conditional mean may vary over time, while distribution shifts can introduce additional prediction errors.

Since each history reveals only one realized future and the true admissible region is unknown, we do not claim exact boundary recovery or formal convergence, consistency, or estimation-error guarantees. Instead, empirical evidence are provided to show effectiveness. Figs. 3, 11, and 12 visualize the learned boundary responses compared to other, while the ablations in Section 4.5 validate the contributions of individual components. We additionally provide generator and discriminator loss curves on SWaT and PSM (Fig. 5), showing stable training behavior without sustained divergence. Together, these results provide empirical evidence of the proposed framework's effectiveness and optimization stability.

To W3: Unlike conventional GANs, our generator produces local probes rather than modeling the normal-data distribution. Ldiv\mathcal{L}_{\mathrm{div}} discourages probe collapse, while Lfit\mathcal{L}_{\mathrm{fit}} prevents excessive deviation from normal evolution. Probe selection relies on trajectory proximity and residual orientation rather than discriminator scores, reducing direct feedback from unstable discriminator outputs while still allowing discriminator gradients to refine the selected probes. Spectral normalization, gradient clipping, and a lower discriminator learning rate further stabilize training.

We additionally report generator and discriminator training-loss curves on SWaT and PSM (Fig. 5). Both exhibit stable training behavior without sustained divergence, while the distinct normal and anomalous responses in Fig. 3 further support the discriminator's effectiveness.

The inconsistent performance of Gaussian Perturbation-GAN does not necessarily indicate training instability. Instead, it reflects the limitations of generic Gaussian perturbations in adapting to different temporal dynamics, supporting our use of learned context-conditioned probes.

To W4: FutureBound does not use a predefined proximity radius or a tunable direction threshold. Eq. (8) determines their orientation by the sign of the mean residual, thus no need for hand-setting. We will move the relevant parameter analyses (including selection ratio) from the appendix to the main text and provide a more detailed discussion of them.

We further compare direction-aware selection with proximity-only selection, keeping the discriminator strategy and all other settings unchanged. As shown in the table, FutureBound consistently outperforms Proximity Alone in both metrics on PSM and SMD, demonstrating the benefit of incorporating residual orientation beyond trajectory proximity.

$\begin{matrix} \hline & &PSM &&& SMD & \\\hline &AUC & AUPRC & F1-Score&AUC & AUPRC&F1-Score\\\hline Proximity\;Alone &74.60\pm0.50 &51.11\pm2.21 &56.68\pm0.91&72.95\pm0.13 &14.98\pm1.88 &22.22\pm3.26\\ FutureBound & \bf76.19\pm3.27 &\bf54.39\pm4.39&\bf56.77\pm2.37&\bf77.20\pm0.54 &\bf21.16\pm3.35 &\bf28.05\pm2.93\\\hline \end{matrix}$

To W7: To address this concern, we have extended our evaluation to TSB-AD-U, a benchmark covering diverse application domains, including medical, financial, and environmental data, as well as time series with pronounced periodic patterns. We have incorporated these additional results into the revised manuscript.

$\begin{matrix}\hline&OCSVM&LOF&IForest&AutoEncoder&Donut&OmniAnomaly&USAD&TFAD&DCdetector&DualTF&CATCH&MTGFlow&PaAno&SDA2D&FutureBound\\\hline VUS-PR&9.37&22.50&22.57&26.71&15.94&25.75&25.41&13.34&23.55&24.01&30.65&41.39&34.55&38.45&\bf42.08\pm0.47\\VUS-ROC&52.55&76.84&74.75&73.06&74.09&69.39&75.59&48.01&70.25&67.76&71.77&72.40&79.89&80.58&\bf82.69\pm0.31\\\hline\end{matrix}$

To W8: First, our central claim is not that normal futures are necessarily multimodal, but that normal future evolution is uncertain and history-dependent, with an admissible region whose geometry is unknown. Multimodal and non-convex structures are discussed as possible cases illustrating the limitations of predefined geometric assumptions, rather than properties we claim to have established on real-world datasets. Since each history reveals only one realized continuation, the true admissible region and its geometric properties cannot be directly measured. Empirical evidence is provided through multiple complementary analyses. Figs. 3, 11, and 12 visualize the learned probes and boundary responses, while the ablations in Section 4.5 demonstrate the contributions of future-space exploration and adversarial boundary refinement, rather than AUC alone.

Second, FutureBound already supports timestamp- and interval-level anomaly localization. Eq. (12) assigns anomaly scores to individual timestamps, allowing anomalous intervals to be identified from consecutive high-scoring observations. Figs. 4 and 13 further demonstrate sustained responses within anomalous intervals and clear transitions back to normal behavior. Variable-level attribution is a complementary diagnosis task beyond our current objective of detecting history-incompatible multivariate evolution.

Thank you again for your comments. We hope our responses have addressed your concerns. If you have any further questions, please feel free to let us know.

Comments on the paper

2 comments
AuthorsOpenReview ↗

FORMAL COMPLAINT CONCERNING REVIEWER sgrM’S ALLEGATION OF PLAGIARISM

Dear ACs, SACs, PCs,

Grounds of Complaint

Reviewer sgrM publicly accuses the authors of alleged plagiarism, and the authors have repeatedly stated that the accusation is false. This statement has been communicated within the review process and may cause reputational harm, particularly if the review remains accessible when the authors’ identities become public. The authors therefore request that the allegation be independently assessed and expressly resolved on the review record.

Reviewer sgrM’s provided evidence

(1) Reviewer sgrM’s supportive evidence did not show a direct copy, lightly edited content, or that EVOVLE’s work was passed off as FutureBound’s. Hence, the evidence does not support a plagiarism claim. To further assess the provided evidence, the authors of FutureBound have specifically requested established evidence of appropriation. Our complaint is that Reviewer sgrM has not adequately explained why those comparisons substantiate plagiarism rather than methodological overlap or independent development.

(2) The chronology shows that FutureBound could not have appropriated its work from EVOLVE. FutureBound was submitted to ICLR 2027 by September 25, 2026, and EVOLVE was publicly released on arXiv on October 23, 2026. These dates establish that FutureBound’s submission preceded the public release.

Reviewer sgrM’s cited similarities are unsupported

Reviewer sgrM’s allegations rely on misleading comparisons that overlook fundamental differences between the two papers, and several cited similarities are not supported by the actual formulations.

(1) Unrelated research motivations are presented as similar. EVOLVE's Challenge 1 concerns heterogeneous responses to injected anomaly perturbations, whereas FutureBound's Proposition 2 addresses the geometric limitations of convex normality regions. Neither work investigates the other's problem, yet both are grouped under the reviewer's claim of similar core motivations.

(2) Mathematically different mechanisms are treated as evidence of methodological overlap. The cited equations reveal substantial differences rather than equivalent designs. EVOLVE generates future predictions through Gaussian latent sampling and dual-structured decoding, whereas FutureBound directly generates multiple context-conditioned trajectories without either mechanism. Their training objectives also differ: EVOLVE applies consistency and diversity constraints to separate normal and anomaly-aware views, while FutureBound jointly enforces local fitting and pairwise diversity within the same probe set. Moreover, EVOLVE adversarially optimizes historical perturbations through KL-based evolution-mode alignment, whereas FutureBound selectively refines future probes through discriminator compatibility scores. These are differences in the mathematical formulations and optimization mechanisms themselves, yet the reviewer presents them as evidence supporting an allegation of plagiarism.

(3) The claimed cumulative similarity largely rests on established techniques rather than shared methodological innovations. History-conditioned multi-future modeling predates both papers (e.g., TimeGrad, 2021), synthetic sample generation has long been studied in anomaly detection (e.g., MAD-GAN, 2019; NCAD, 2022), and alternating adversarial optimization originates from the GAN framework (Goodfellow et al., 2014). These general techniques are not unique contributions of EVOLVE or FutureBound. More importantly, their distinctive innovations do not overlap: EVOLVE introduces adaptive anomaly prototype memories, uncertainty-aware latent modeling, and dual-structured decoding, whereas FutureBound develops local future-space exploration and selective adversarial learning of an implicit admissible boundary. Treating shared, established concepts as cumulative evidence of plagiarism, while overlooking the distinct contributions of both works, does not substantiate such a serious allegation.

Requested Action

We request that the relevant chairs:

(1) Independently assess the evidential basis of the plagiarism and research-integrity allegation, separately from the assessment of methodological novelty;

(2) Require the reviewer to substantiate the allegation specifically and address our responses;

(3) If the allegation cannot be substantiated, require Reviewer sgrM’s formal and express withdrawal and correction of the review record;

(4) Ensure that an unsubstantiated misconduct allegation does not prejudice the evaluation of our submission

We welcome rigorous scientific criticism and have made every effort to address the questions raised during review. However, methodological disagreement and allegations of research misconduct require distinct assessments. We request your intervention to ensure that this allegation receives a fair, evidence-based, and documented resolution.

Best regards,

Authors

AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair and Reviewers,

As the discussion period approaches its end, we would like to summarize our responses to the reviewers' main concerns, together with the additional experiments, analyses, and revisions provided during the rebuttal.

Reviewer 6Mxy

  • Q1–Q2, W1–W2, W8 (Motivation and theoretical justification): We clarified the definition of the learned admissible future region, the connection between Propositions 1–2 and FutureBound, and the implications of non-stationarity and distribution shift. We also provided further explanations of the learned discriminator responses and timestamp-level anomaly localization.
  • Q3, W3–W4 (Probe selection and training stability): We clarified the probe selection criteria and supplemented the parameter sensitivity analyses. Our additional ablation shows that direction-aware selection improves AUPRC from 14.98 to 21.16 on SMD compared with proximity-only selection. Training-loss analyses further provide evidence of stable adversarial optimization.
  • Q4–Q5, W5–W7 (Evaluation and generalization): We clarified baseline configurations, F1 threshold selection, score comparability, and inference efficiency. We further extended our evaluation to TSB-AD-U, covering more diverse application domains.

Reviewer sgrM

  • W1–W3 (Stability, efficiency, and temporal horizons): We provided additional analyses of training stability, computational complexity, and parameter sensitivity, including the effects of historical and future horizons. The results further demonstrate the effectiveness of FutureBound under different parameter settings.
  • W4–W5 (Experimental evaluation): We supplemented CrossAD as a recent baseline and extended our experiments to TSB-AD-U. These additional results provide broader empirical evidence of FutureBound's detection performance.
  • W6 (Methodological novelty): We further clarified the contributions of local future-space exploration and selective adversarial optimization, supported by additional ablations and comparisons with existing approaches.

Reviewer 1Jac

  • W1–W2 (Motivation and methodology): We clarified how Propositions 1–2 motivate flexible boundary learning and explained the roles of generated probes and the discriminator. Additional visualizations and ablation analyses further support the effectiveness of this design.
  • W3–W4 (Baselines and evaluation protocols): We added GDN as a forecasting baseline, supplemented experiments on SMAP and TSB-AD-U, and clarified the evaluation protocols. We also provided additional comparisons under the same look-ahead setting to examine its influence on detection performance.
  • W5–W7 (Methodological distinctions and ablations): We clarified the differences between FutureBound and NCAD, SPAGD, and GenIAS, and explained how generated probes provide effective supervision. Additional analyses of top-k aggregation and probe selection further support the proposed learning mechanism.

Taken together, we believe these additional experiments, analyses, and clarifications addressed all reviewers' main technical concerns and further support the effectiveness of FutureBound.

Research Ethics Concern

We addressed Reviewer sgrM’s allegation of plagiarism by providing a detailed response (point by point) to all of their six claims on methodological similarities. On top of that, we provided a clear chronology of the works mentioned in the accusation. Reviewer sgrM did not provide a response to our request for substantial evidence to their plagiarism claim. In view of their lack of response, together with the evidence we provided against the claim, we stand firm in our rejection of the allegation.

We requested an independent evaluation in our formal complaint letter to ACs, SACs, and PCs, and we once again request in this rebuttal summary that appropriate action be taken and that a conclusion on this accusation by Reviewer sgrM be provided. The plagiarism allegation is expected to be evaluated separately from methodological novelty, based on verifiable evidence, and without prejudice to the scientific assessment of our work.

We appreciate the Area Chair's consideration of our complete rebuttal and revised manuscript.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.