{"id":"525239f8-0dbb-4fdf-a2ea-4f84c400976f","arxiv_id":"2507.01875","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A univariate VAE with dilated convolutions is proposed as a simple 'foundation' model for time-series anomaly detection, with preliminary zero-shot experiments on two datasets.","lead":"This paper introduces FAE, a small one-dimensional time-series model based on variational autoencoders, and tests whether it can spot unusual patterns in data it has never been trained on. The authors show encouraging but mostly visual results on mobile network and public benchmark time series, with no quantitative anomaly detection scores yet.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Decoder as described cannot produce the time-varying reconstructions shown; a constant repeated latent vector yields constant output through causal dilated convolutions, so the core experiments need a mechanism check.","rationale":"The paper is a preliminary workshop paper and mostly qualitative, as the authors acknowledge in Section 5. The central claim is that FAE, a foundation VAE based on causal dilated convolutions, can track normal behavior on previously unseen time series. For that claim, the model must actually produce time-varying predictions. The strongest internal check is whether the described architecture can do so. The decoder takes a single latent vector z, repeats it T times, and processes it with causal dilated convolutions. As argued, any convolutional stack with a constant input has a constant output: linear convolutions map constants to constants, and pointwise nonlinearities preserve constancy. Boundary padding affects only the first (F−1)·d positions. Thus μ_X and σ_X should be essentially constant in t, yet the figures display rich temporal variation including daily seasonality and weekend effects. One possibility is that the prose omits an important mechanism (e.g., a time embedding, an autoregressive loop, or concatenating the input X with z in the decoder). Another is that the plots were produced by a different architecture than the one described. Either way, the published description is insufficient to establish the central claim. This concern is more load-bearing than the missing α threshold: even a fully specified threshold would not rescue a decoder that cannot produce a time-varying baseline. The reader's weakest assumption identified the α rule and the absence of quantitative anomaly-detection metrics; that is a fair and related concern, but it presupposes the reconstructions are meaningful. My proposed test is cheap: run a trained decoder with fixed z and inspect the outputs. If the output is flat, the paper needs either an architectural correction or a re-evaluation; if it varies, the mechanism must be reported, and the concern resolves. In that latter case, the remaining issues are the quantitative validation gaps highlighted by the reader. I therefore recommend UNVERDICTED rather than a categorical rejection, because the paper cannot be assessed as written until this mechanism question is settled.","tokens_in":12625,"tokens_out":5720,"duration_ms":72464,"concrete_test":"Run the decoder forward pass on a fixed z sampled from a trained FAE with T=256 and J=48, recording μ_X(t) for interior timesteps (e.g., t=64..192). If μ_X is flat up to numerical precision, the architecture as described cannot reproduce Figures 3/9/10, confirming the concern. If μ_X varies, identify which component (e.g., positional encoding, residual/skip connection, non-causal padding, or autoregressive input) introduces the variation and document it. A cheaper analytical check: symbolically apply one dilation-2, F=2 causal convolution to a constant R^{J×T} input and verify the output is constant for all interior t.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"Section 3 states that the decoder input is the latent vector z, repeated T times to form an R^{J×T} sequence. The decoder is a stack of causal dilated convolutional layers (Figure 1). For any such layer with kernel length F and dilation d, the pre-activation at time t is a weighted sum over taps at times t−d·f. If every input column is the same vector z, that weighted sum is independent of t for all positions beyond the boundary, so each layer outputs a constant sequence; subsequent layers, receiving constant inputs, also output constants. Thus μ_X and σ_X should be constant over t except for boundary effects. Yet Figures 3, 9, and 10 show strongly time-varying μ_X that tracks daily seasonality and trends. No per-timestep conditioning, positional encoding, autoregressive connection, or skip connection is described to break this symmetry. If this reading is correct, the reported reconstructions cannot be produced by the architecture as written, and the central modeling claim lacks a mechanism. The missing α threshold identified by the reader is secondary: without a time-varying baseline, no fixed threshold can detect point anomalies that deviate from local temporal patterns. If an unstated mechanism exists, it must be disclosed and analyzed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FAE, a VAE with causal dilated convolutional encoder/decoder for univariate time-series modeling, intended as a pretrained 'foundation' model for zero-shot anomaly detection. The authors present qualitative results on the TELCO dataset and on 12 series from the KDD2021 dataset, including a partial zero-shot experiment on TS12, and argue that FAE tracks seasonal and trend behavior and yields interpretable latent representations.","tokens_in":12841,"tokens_out":6602,"duration_ms":68409,"significance":"If the architecture as described is actually able to produce the time-varying reconstructions shown, FAE would be a small-scale, interpretable foundation model for anomaly detection, and the latent-space analysis is a nice demonstration of VAE transparency. The paper also provides a useful departure point for applying VAEs to univariate time series. However, the current manuscript does not provide quantitative validation, and the architectural description contains a serious internal inconsistency; the significance of the contribution therefore cannot be assessed as it stands.","major_comments":[{"comment":"The decoder as described cannot produce time-varying reconstructions. The input to the decoder is the latent vector Z repeated T times, forming a constant sequence along the time axis. Since all decoder layers are causal dilated convolutions, a constant input produces a constant output for all time steps (except for boundary effects). Yet Figures 3, 9, and 10 show strongly time-varying μ_X and σ_X that track daily seasonality and trends. This is a load-bearing discrepancy: without a mechanism that breaks the temporal symmetry (e.g., per-timestep conditioning, a non-constant latent input, or autoregressive connections), the reported results cannot be reproduced by the architecture. The authors must either provide the actual decoder mechanism or correct the description.","section":"Section 3, Figure 1, Figures 3, 9, 10"},{"comment":"The paper defines the anomaly detection rule using a threshold α—a sample is normal only if its deviation from μ_X stays within α times σ_X—but never reports the α value used in Figures 7 and 8, and provides no quantitative anomaly detection results (e.g., precision, recall, F1, or ROC). The evaluation is entirely qualitative, based on visual inspection, with no error bars and no baselines beyond the earlier DC-VAE, which is also not quantified. Without a specified threshold and quantitative metrics, the claim that FAE detects anomalies on unseen data is unsupported.","section":"Section 3 (anomaly rule) and Section 4"},{"comment":"The zero-shot experiment for TS12 is confounded by the high correlation between TS12 and TS11: training on TS1–TS11 gives the model direct evidence of TS12's seasonal and trend patterns, so this is not a clean test of zero-shot generalization. The more challenging setting (training on TS1–TS10) shows degraded performance, but is only evaluated qualitatively. To substantiate zero-shot generalization, the authors need quantitative comparisons across multiple held-out series and domains, with baselines and statistical significance.","section":"Section 4.5"},{"comment":"The KDD2021-12 experiment trains on the train partition and tests on the test partition of the same 12 series, which is within-series transfer, not zero-shot or foundation-model generalization. The paper's stated goal is to model 'previously unseen datasets,' yet the only partially held-out evaluation is the TS12 experiment in TELCO. A proper foundation-model evaluation should pretrain on a large, diverse corpus and test on completely unseen series from different domains without any retraining, which is not done here.","section":"Section 4.6"}],"minor_comments":[{"comment":"The description of the receptive field expansion ('the dilation of each layer at the encoder increases exponentially as the network deepens') is imprecise; consider giving an explicit formula, e.g., receptive field size after N layers with kernel F and base dilation d.","section":"Section 3"},{"comment":"The abbreviation 'ZPC' is used for the principal components of the latent vectors, but it is not defined in the caption or text; please write out 'principal component of Z' or explain the notation.","section":"Figure 4"},{"comment":"The abstract mentions 'preliminary results in different multi-dimensional time-series datasets,' but FAE is a univariate model; please clarify that the datasets are multivariate but the model analyzes each series separately, or rephrase to avoid confusion.","section":"Abstract and Section 1"},{"comment":"Reference [22] for the KDD2021 dataset points to a competition URL; a proper citation to the dataset itself (e.g., the UCR archive or the competition organizers' publication) would be more appropriate.","section":"References"},{"comment":"The captions mention 'flagged anomalies,' but the anomaly markers and the threshold used to flag them are not described in the captions or the main text; adding a legend and the α value would greatly improve interpretability.","section":"Figures 7 and 8"}],"recommendation":"major_revision","confidential_remarks":"The architectural inconsistency described in Major Comment 1 is severe. If the authors cannot demonstrate that the described decoder produces time-varying outputs, the paper's central modeling claim collapses. I recommend that the revision require a corrected and complete architecture description, possibly with a minimal proof-of-concept experiment that isolates the decoder's ability to produce time-varying outputs from a constant latent input. The lack of any quantitative evaluation is also a major concern for a journal submission; the authors should provide anomaly detection metrics, a reasoned choice of α, and comparisons against at least one standard baseline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi, \n\nBottom line: the paper is a candid, clearly written sketch of a univariate VAE for zero-shot time-series anomaly detection, but there's a load-bearing mechanism problem that makes the central reconstructions impossible under the architecture as described. I'd hold off until the authors either correct the description or provide code. \n\nWhat's genuinely there: the zero-shot ablation on TELCO (train on TS1–11, test on TS12, and the harder TS1–10 case) is a sensible way to probe generalization, and the latent-space visualizations (clock-like ordering by hour, workday/weekend separation) are a nice piece of qualitative evidence. The paper is honest about being preliminary and about VAE limitations. The TELCO dataset release is a real service to the community. \n\nThe problem: Section 3 says the decoder input is the latent vector z, repeated T times. With causal dilated convolutions, a constant input sequence produces a constant output sequence at every layer. So mu_X and sigma_X should be flat in time, yet Figures 3, 9, and 10 show strong daily seasonality and trends. No positional encoding, autoregressive connection, or other mechanism is described to break that symmetry. This is not a nitpick; it's the core evidence. Either the figures come from a different model, or a key detail is missing. \n\nBeyond that, the evaluation is entirely qualitative: no anomaly detection metrics, no error bars, no quantitative baselines, and the alpha threshold from Section 3 is never disclosed. The KDD2021-12 experiment trains and tests on the same series, so it's not zero-shot. Calling a ~480k-parameter model trained on 12 series a 'foundation' model is a stretch, though the authors do hedge. \n\nThe stress-test reading matches the paper text; I don't think it's a misreading. If the authors added code or a corrected architecture and one quantitative anomaly detection result, this could be a useful workshop submission. As it stands, the central claim lacks a working mechanism. \n\nRecommendation: don't send this to peer review yet. Ask the authors to clarify or fix the decoder and add a basic quantitative evaluation.","headline":"Univariate VAE sketch with honest intentions, but the decoder as written cannot produce the time-varying reconstructions shown.","tokens_in":13409,"tokens_out":3302,"would_cite":false,"duration_ms":35318,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that a modest dilated-convolutional variational auto-encoder, pretrained on normal univariate time-series, can track and flag anomalies on previously unseen series in a zero-shot fashion, and that its latent space…","keywords":["time-series anomaly detection","variational auto-encoder","dilated convolutions","zero-shot learning","foundation models","univariate time-series","latent space interpretation"],"falsifier":"Run FAE on the labeled TELCO test set and compute the precision-recall curve obtained by sweeping $\\alpha$ in the normality rule; if no single $\\alpha$ yields detection performance comparable to a per-series calibrated threshold (or to a simple residual z-score baseline), then the fixed-rule zero-shot claim is falsified. Applying the pretrained model to all 250 KDD2021 series and checking whether its normality regions align with the dataset's known anomalies would reveal whether the 'foundation' behavior extends beyond the 12 series shown.","tokens_in":12395,"feed_emoji":"📈","tokens_out":10222,"duration_ms":97853,"temperature":0.7,"pith_summary":"FAE is a variational auto-encoder with dilated causal convolutions that is trained on multiple univariate time-series of normal behavior and then applied to previously unseen series. The paper argues that such a model can track the expected mean and variance of normal operation well enough to flag anomalies on unseen series without retraining. If that holds, anomaly detection for a new monitoring system could be bootstrapped by a single pretrained model instead of bespoke per-series models. The paper supports the claim qualitatively with TELCO mobile-network series, a domain-shifted TELCO2, a zero-shot TS12 experiment, and 12 KDD2021 series.","feed_headline":"Pretrained VAE flags anomalies on unseen time series","feed_subtitle":"No retraining, no per-series calibration: a dilated-convolution VAE learns the shape of normal.","key_machinery":"The central object is the dilated-convolutional variational auto-encoder: each encoder layer dilates by a factor $F^h$ (with $F$ the filter length and $h$ the layer depth), so the final output at position $T-1$ has seen the whole input $\\mathbf{X}$. The number of layers $N$ is the minimum integer satisfying $T \\le 2F^{N-1}$, tying architecture size directly to window length. The encoder's two parallel $J$-filter convolutional layers map this output to $\\boldsymbol{\\mu}_\\mathbf{Z}$ and $\\boldsymbol{\\sigma}_\\mathbf{Z}$; the reparameterization trick samples $\\mathbf{Z}$, and the decoder, symmetric to the encoder, expands $\\mathbf{Z}$ back to $\\boldsymbol{\\mu}_\\mathbf{X}$, $\\boldsymbol{\\sigma}_\\mathbf{X}$. This mechanism is what lets one model observe many series at once and produce an interpretable, reconstruction-based normality region.","core_discovery":"FAE adapts the DC-VAE architecture to univariate series: a stack of causal dilated convolutions expands the receptive field exponentially, the encoder compresses the last temporal position into a low-dimensional latent vector $\\mathbf{Z}$, and a symmetric decoder repeats that vector across time to output per-timestep parameters $\\boldsymbol{\\mu}_\\mathbf{X}$ and $\\boldsymbol{\\sigma}_\\mathbf{X}$. A window $\\mathbf{X}_t$ is normal when it stays within $\\alpha$ times the predicted standard deviation of the predicted mean. The paper reports that when the model is trained on several normal series, the latent space organizes by hour of the day, by weekday versus weekend, and by day of the month, and that the decoder tracks seasonality and trends on held-out series; in the zero-shot TS12 experiment, removing TS12 from training leaves predictions almost unchanged as long as a series with similar seasonal and trend behavior is present.","pith_inferences":["The architecture's direct tie between window length $T$ and layer count suggests a simple scaling rule for building FAE variants at different temporal resolutions.","Because the univariate model discards cross-correlations, it may miss anomalies that appear only as spatial deviations; an ensemble that routes between univariate and multivariate detectors could combine both strengths.","The latent-space 'clock' suggests the VAE encodes the phase of daily seasonality; this phase structure could be exploited for explicit forecasting or for explaining why a point is anomalous.","The foundation-model claim would be tested more convincingly on the full 250-series KDD2021 set, since the reported experiments use only 12 of the 250 series and do not quantify detection performance on them."],"forward_implications":["A single FAE model can monitor all series in a multivariate system one at a time, relying only on univariate temporal structure.","When the model has seen a temporal pattern in one series, it can track the same pattern zero-shot in a related unseen series, as shown for TS12 after training on TS1 to TS11.","When the monitored domain shifts from TELCO to TELCO2, the pretrained univariate model keeps tracking, whereas the multivariate DC-VAE requires recalibration.","Projections of the latent space expose hour-of-day, weekday-versus-weekend, and monthly-trend structure, making the model's normality judgments interpretable."],"supporting_citations":[{"why":"Supplies the DC-VAE architecture (dilated-convolutional VAE) that FAE adapts to univariate series.","marker":"[13]"},{"why":"Provides the variational auto-encoder formulation and the reparameterization trick FAE uses.","marker":"[23]"},{"why":"Provides the TELCO mobile-network dataset with seven months of labeled series used for training and evaluation.","marker":"[15]"},{"why":"Provides the KDD2021 multi-dataset time-series anomaly detection data used to test generalization.","marker":"[22]"},{"why":"Supplies the TPE hyperparameter optimization method used to select FAE's configuration.","marker":"[2]"}],"fun_headline_variants":["Zero-shot VAE learns temporal patterns to flag anomalies","Dilated VAE foundation model spots anomalies on new data","VAE learns seasonal patterns, zero-shot anomaly detection","Pretrained VAE with dilated convs flags unseen anomalies","Dilated VAE: zero-shot anomaly detection on unseen series"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"FAE's anomaly detection depends on the rule that a sample is normal only if its deviation from the predicted mean stays within $\\alpha$ times the predicted standard deviation; the paper never gives the $\\alpha$ used in its figures and provides no quantitative evidence that one fixed $\\alpha$ works across heterogeneous series. If this rule needs per-series calibration, the zero-shot detection claim collapses even if the reconstruction tracking is accurate.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot VAE learns temporal patterns to flag anomalies","Dilated VAE foundation model spots anomalies on new data","VAE learns seasonal patterns, zero-shot anomaly detection","Pretrained VAE with dilated convs flags unseen anomalies","Dilated VAE: zero-shot anomaly detection on unseen series"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001157,"raw_usage":{"total_tokens":4767,"prompt_tokens":890,"completion_tokens":3877,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":3797}},"tokens_in":506,"tokens_out":3877,"duration_ms":29054,"temperature":1.0,"reasoning_tokens":3797,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:41:15.005975+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FAE on the labeled TELCO test set and compute the precision-recall curve obtained by sweeping $\\alpha$ in the normality rule; if no single $\\alpha$ yields detection performance comparable to a per-series calibrated threshold (or to a simple residual z-score baseline), then the fixed-rule zero-shot claim is falsified. Applying the pretrained model to all 250 KDD2021 series and checking whether its normality regions align with the dataset's known anomalies would reveal whether the 'foundation' behavior extends beyond the 12 series shown.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DC-VAE architecture (dilated-convolutional VAE) that FAE adapts to univariate series."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the TELCO mobile-network dataset with seven months of labeled series used for training and evaluation."},{"cited_title":"Keogh, T","cited_arxiv_id":null,"evidence_quote":"Provides the KDD2021 multi-dataset time-series anomaly detection data used to test generalization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the TPE hyperparameter optimization method used to select FAE's configuration."}],"review_version":1}