{"id":"91c98419-ea80-4baa-baa3-e5a18a1b1eb4","arxiv_id":"2506.23596","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A2P trains a shared transformer to forecast future time series and detect anomalies in the forecasted signal, using synthetic anomaly prompts, and reports higher F1 than forecasting-plus-detection baselines on four datasets.","lead":"This paper introduces A2P, a deep learning framework that aims to predict not just that something will fail, but exactly when in the future an abnormal event will happen. It combines a forecasting model with an anomaly detector and uses artificially created anomalies during training to make the system better at spotting real ones.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing premise is that synthetic anomalies from §3.3 transfer to real test anomalies; Eq. 6's anomaly-aware weighting is only as valid as that transfer, and the paper never measures it directly.","rationale":"The reader's weakest assumption—that synthetic anomalies from §3.3 must be representative of real anomalies—is exactly the load-bearing point. The method's novel components (AAF, APP, SAP) are trained only on synthetic anomalies, and all of them are frozen or discarded at test time; their only influence on the final predictor is through the learned anomaly-aware weighting in Eq. 6 and through the reconstruction training in Eq. 7. If that synthetic knowledge does not transfer, the method reduces to a forecasting-plus-reconstruction baseline with extra training signal, and the paper's explanation of why it outperforms baselines is unsupported. The paper does provide ablation studies and threshold-free VUS metrics, which are real supporting evidence, but they do not isolate the transfer mechanism. The concrete check—measuring the frozen AAF network's real-window AUC—directly tests the transfer premise and is feasible from the released implementation. Because this is the same concern the reader flagged, my read does not move the verdict; it remains conditional on this premise being verified.","tokens_in":18329,"tokens_out":9298,"duration_ms":103867,"concrete_test":"With the released code, freeze the Anomaly-Aware Forecasting Network trained by Eq. 1 and run it on held-out real test tuples (Xin, Xout, y) before main training; compute the AUC of its predicted anomaly probability against the real anomaly labels. Also compute the same AUC after replacing the APP-selected prompts with random prompts in Eq. 2. If AUC is near 0.5, or if random prompts match learned prompts in downstream AP F1 on a held-out dataset, the synthetic-to-real transfer premise fails and the mechanism in Eq. 6 is not the source of the reported gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"A2P's pre-training (Eqs. 1, 4, 5) injects CARLA-style anomalies and trains the Anomaly Prompt Pool on pseudo-anomalies; at test time (Fig. 5) the model is a forecasting-plus-reconstruction anomaly scorer. The mechanism that should make AP work is the anomaly-aware weight g in Eq. 6, which is trained exclusively on synthetic anomaly labels yz_out. If the distribution of synthetic anomalies differs from real test anomalies, the weighting will emphasize the wrong time steps and the prompt pool will teach the reconstruction module to erase synthetic artifacts rather than real anomalies. The paper's F1 results are indirect evidence at best: ablations show that removing AAF/SAP lowers F1 on the same test sets, but this does not establish that the learned g or prompt selection is attending to real anomalies rather than to some other training-signal artifact. No experiment reports the frozen AAF network's anomaly-probability output on real windows, nor any distributional comparison between synthetic prompt features and real anomaly features. Since the first-effective-method claim depends on this transfer, the central claim is conditional on an unvalidated premise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Anomaly Prediction (AP), a task in which a model must forecast future time series values and then pinpoint which time steps in the forecasted horizon will be anomalous. The authors propose Anomaly to Prompt (A2P), a unified shared-backbone framework with two main components: Anomaly-Aware Forecasting (AAF), which pre-trains a network to predict anomaly probabilities from synthetically injected anomalies and uses this probability to weight the forecasting loss, and Synthetic Anomaly Prompting (SAP), which learns an Anomaly Prompt Pool (APP) that transforms normal embeddings into synthetic anomalous embeddings during training. At test time, only the shared backbone is used: the model forecasts the future window and then reconstructs it, and the reconstruction error serves as the anomaly score. Experiments on MBA, Exathlon, SMD, and WADI compare A2P with combinations of five forecasting models and three anomaly detectors under an F1 metric with a tolerance window, and report that A2P achieves the highest average F1 at all tested prediction horizons. The paper also provides ablations, additional datasets, additional baselines, and parameter-free VUS metrics in the appendix.","tokens_in":18561,"tokens_out":4443,"duration_ms":50193,"significance":"If the technical claims hold, the paper proposes a genuinely new task formulation and a plausible first solution, with a clean unified architecture and public code. The strengths are the breadth of the empirical study (four main datasets, extra datasets in the appendix, multiple baselines), the inclusion of threshold-free VUS metrics in the appendix, and the ablation study that isolates the contribution of each new component. However, the central claim of 'superiority' rests on an evaluation protocol that uses ground-truth anomaly ratios from the test set to set thresholds, and on an unvalidated assumption that synthetic anomalies used in training are representative of the real anomalies that appear in test windows. These issues affect the interpretation of every reported F1 number, so the significance of the result is real but conditional on addressing them.","major_comments":[{"comment":"The threshold for converting anomaly scores to binary predictions is set using the ground-truth anomaly percentage in the test data, per the paragraph 'The threshold for deciding anomalies from anomaly scores is set by following the widely accepted protocol from (Shen et al., 2020a), adjusting for a percentage of anomalies in the test data.' This leaks test-label information into the evaluation and yields F1 scores that are not fully external. Because all methods receive the same threshold treatment, the relative ranking may be less affected, but the reported absolute F1 values are optimistic and the comparison is not a clean forecast evaluation. The paper should either select the threshold on a validation split, or report main-table results with a threshold-free metric such as VUS-PR/VUS-ROC, which are currently delegated to the appendix. In addition, the default tolerance t=50 is generous relative to the horizons Lout=100-400, and the claim of 'pinpointing' anomaly time points should be supported by metrics at t=0 or t=1 in the main text; the appendix figure varying t is a useful start but is not the headline result.","section":"Section 4.1, 'Anomaly Threshold' and 'Evaluation Metrics'"},{"comment":"The core mechanism of the paper is the anomaly-aware weight g(Xin, Xhat_out) in Eq. (6), and the SAP/APP modules that synthesize anomalies for training. Both are trained exclusively on synthetic anomalies injected via the CARLA-style procedure described in Section 3.3. The paper never validates that these synthetic anomalies are representative of the real anomalies that occur in the test windows. This is load-bearing: if the distribution of synthetic anomalies diverges from real anomalies, the learned weight g may emphasize the wrong time steps and the prompt pool may teach the reconstruction module to erase synthetic artifacts rather than real anomalies. The ablations in Tables 2-5 show that the components improve F1 on the same test sets, but this is indirect evidence. The authors should provide direct transfer evidence, for example: (i) the anomaly-probability output of the frozen AAF on real test windows, evaluated as a ranking or AUROC; (ii) a feature-space comparison between real anomaly embeddings and synthetic prompted embeddings; or (iii) an experiment where real anomaly samples (if any) are used to validate or select the prompt pool. Without such evidence, the claim that A2P is an effective AP method remains conditional on an unstated representativeness assumption.","section":"Section 3.3 and Eq. (6)"},{"comment":"The claim of consistent superiority over baselines is not statistically supported for some datasets. For example, at Lout=100 on Exathlon, A2P reports F1=18.64±0.16, while PatchTST+AnomalyTransformer reports 18.10±0.24 and FITS+DCdetector reports 17.38±3.26; these differences are within the reported standard deviations. Similar overlap appears in several other rows. With only three random seeds and no significance tests, the statement in Section 4.2 that 'our model consistently outperforms the baselines' is too strong. The authors should report confidence intervals or significance tests, or at least discuss which differences are reliable, before claiming universal superiority.","section":"Table 1, Exathlon and SMD rows"}],"minor_comments":[{"comment":"The definition of F1 with tolerance t is ambiguous: 'the real ground-truth anomaly time points from [i - t, i + t] are considered to be correctly detected before the calculation of F1-score.' It should state explicitly whether a predicted anomaly at time i is counted as a true positive if any ground-truth anomaly falls in the window, and how multiple matches are handled.","section":"Section 4.1, 'Evaluation Metrics'"},{"comment":"The paper claims to be the first to 'propose a method to deal with the problems of AP,' but You et al. (2024) introduced the AP scenario and is cited as not directly tackling its challenges. To substantiate the 'first effective method' claim, the authors should either compare with a reimplementation of You et al. or explicitly state why it cannot serve as a baseline.","section":"Section 1 and Related Work"},{"comment":"The statement that the additional computational cost 'is only incurred during training, with no extra overhead at inference time' is not fully precise: at test time A2P runs both the forecasting network and the reconstruction network, as shown in Figure 5, but the comparison to the PatchTST+AnomalyTransformer baseline is fair. The sentence should be rephrased to say there is no overhead relative to training a single forecasting-plus-detection pipeline.","section":"Section 4.4 and Figure 7"},{"comment":"The notation table lists symbols such as ~X_in and ~X_p_in but the main text uses eX_in and eX_p_in; please unify the notation across the paper.","section":"Appendix A.2 and Table 7"},{"comment":"The column headers 'F AD' and the spaces in 'W ADI' appear to be formatting artifacts; they should be corrected to 'F1' and 'WADI' for readability.","section":"Tables 1, 10, and 12"},{"comment":"The divergence loss has two terms with different signs; the paper should state explicitly that minimizing LD maximizes the cosine-similarity term, thereby pulling selected keys toward normal features, since this is not immediately obvious from the equation alone.","section":"Section 3.3, Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to be an ICML 2025 proceedings paper; as a journal submission it needs to go substantially beyond the conference version. The main novelty claim is strong, but the evaluation protocol (test-label threshold, tolerance window) and the unvalidated synthetic-to-real transfer are exactly the kinds of issues that a journal reviewer should push on. If the authors can move the VUS/threshold-free results to the main text, add a validation-based threshold or threshold-free metric, and provide direct evidence on the AAF/APP transfer, the paper would be a solid contribution. I would not reject on the basis of disagreement with the task framing; the AP task is well motivated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Jo, here's my read on 2506.23596. The useful thing to know: it's an empirical advance on the Anomaly Prediction task, and the main caveat is that the training signal comes from synthetic anomalies whose realism is never directly checked. The method itself is a clean combination: a shared transformer backbone, an anomaly-aware forecasting loss weighted by a pre-trained anomaly-probability network, and a learnable prompt pool that synthesizes anomalous embeddings during training. At test time only the backbone and heads run, so no extra inference cost. Code is public, ablations are thorough, and they report VUS-PR and VUS-ROC in the appendix, which are threshold-free. That matters because the F1 numbers use a test-set-calibrated threshold and a tolerance window of t=50. The VUS results are weaker than the F1 headline but still consistently ahead of baselines, and they also show results at t=1, so the tolerance choice isn't hiding everything.\n\nWhat's genuinely new: applying the CARLA-style synthetic anomaly injection to training a forecaster for future anomalies, plus the prompt-pool mechanism for diversifying reconstruction training. The shared backbone is a reasonable efficiency story. The paper also does the right thing by comparing against a grid of forecasting x AD pairs and reporting MSE for the forecasting component.\n\nSoft spots, in order. First, they never directly validate that synthetic anomalies resemble real test anomalies. The anomaly-aware weighting in Eq. 6 is trained on pseudo-labels, and while the qualitative plots and forecasting MSE suggest the model captures something real, a distributional comparison of prompt features versus real anomaly features would close the loop. That's a fixable gap, not a fatal one. Second, they don't compare against You et al. (2024), the paper that introduced AP; their claim that You et al. 'does not directly tackle' AP reads as dismissive and should at least be substantiated with a baseline result. Third, the Exathlon gains are tiny (about 1 F1 point at Lout=100), so the method's benefit is concentrated on MBA, SMD, and WADI; that's worth stating honestly.\n\nMy bottom line: this deserves a serious referee. The task is practically relevant, the experiments are extensive, the code is out, and the weaknesses are addressable in revision rather than load-bearing. I'd bring it to a reading group and would probably cite it if I worked on prediction-focused anomaly benchmarking.","headline":"A strong empirical paper on a useful new task; the synthetic-anomaly transfer question is real but not disqualifying, and the threshold-free metrics in the appendix cover part of the gap.","tokens_in":19119,"tokens_out":3111,"would_cite":true,"duration_ms":30080,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A2P claims to be the first effective solution to Anomaly Prediction — naming the exact future time points of anomalies — and reports state-of-the-art F1 on four real-world datasets.","keywords":["anomaly prediction","time series forecasting","anomaly detection","synthetic anomaly prompting","anomaly prompt pool","anomaly-aware forecasting","multivariate time series","prompt tuning"],"falsifier":"Train A2P on windows whose anomalies are injected from the five standard types, then evaluate it on held-out test windows whose anomalies are produced by a different mechanism (for example, only point spikes, or slow drifts, or a sixth shape type never synthesized during training) while keeping the baselines on the same test set; if A2P's F1 advantage over forecast-then-detect baselines shrinks or disappears under that shift, the central claim is an artifact of synthetic-to-real similarity. A stricter variant replaces the default tolerance in the F1 metric with exact time-step matching to confirm the reported ordering is not carried by the tolerance window.","tokens_in":18114,"feed_emoji":"🔮","tokens_out":11185,"duration_ms":105914,"temperature":0.7,"pith_summary":"The paper claims that anomaly prediction — naming the exact future time points at which a system will misbehave, rather than just flagging that it might — is a solvable task, and proposes a framework called A2P that does it. The key move is to stop training forecasting models on normal signals only: A2P pre-trains an Anomaly-Aware Forecasting network that learns how past anomalies shape future signals, and it synthesizes diverse anomalies inside the model by attaching learnable anomaly prompts to signal embeddings, so the detection side has seen abnormal patterns even though no real anomalies were used in training. On four real-world datasets (ECG arrhythmias, distributed-computing workload logs, server metrics, and a water-distribution testbed), A2P reports consistently higher F1 scores than combinations of state-of-the-art forecasting and detection models, and the margin grows as the forecast horizon lengthens from 100 to 400 steps. If the claim holds, anomaly detection becomes a predictive tool instead of a forensic one, which is the step needed for advance warnings in medicine, manufacturing, and infrastructure monitoring.","feed_headline":"Model pinpoints exact future anomaly time points","feed_subtitle":"A2P trains on synthetic anomalies and beats forecast-plus-detect pipelines on four real-world datasets.","key_machinery":"The central object is the Anomaly Prompt Pool (APP): a set of learnable key–prompt pairs in which each prompt is a token sequence that encodes an anomaly pattern. A frozen feature extractor with a [CLS] token turns an input into a query, cosine similarity selects the top-N prompts, and their tokens are concatenated onto the input embedding at the embedding level, transforming a normal reconstruction into a synthetic anomaly; the divergence loss widens the gap between synthetic-anomaly features and pseudo-normal features inside the shared transformer backbone. The Anomaly-Aware Forecasting Network plays the complementary role: cross-attention between the anomaly-injected future embedding and the prior embedding yields per-time-step anomaly probability, which reweights the forecasting loss. The shared backbone is what makes the two tasks one model — forecasting and reconstruction branches share transformer layers and are trained jointly — so that at test time, forecasting and anomaly scoring happen in a single forward path.","core_discovery":"The central claim is that the failure of naive forecast-then-detect pipelines is not a defect of the component models but of training signal: forecasting models trained only on normal data smooth away the very abnormality that must survive into the forecast. A2P therefore couples two mechanisms: Anomaly-Aware Forecasting (AAF), which uses a pre-trained cross-attention network to output per-time-step anomaly probabilities and reweights the forecasting loss so errors at predicted anomaly points are penalized more; and Synthetic Anomaly Prompting (SAP), in which a learnable Anomaly Prompt Pool is queried by a frozen feature extractor, the top-N best-matching prompts are concatenated onto the input embedding, and a divergence loss pushes the resulting synthetic-anomaly features apart from pseudo-normal features. Both components are pre-trained before the main training and frozen at test time, when only the shared transformer backbone is used to forecast, reconstruct, and score anomalies by association discrepancy. The paper reports that A2P outperforms all forecasting-plus-detection baselines on MBA, Exathlon, SMD, and WADI, with its largest advantages at the longest horizon, and that ablations attribute gains to both AAF and SAP, with the divergence loss alone contributing about a 24% F1 improvement on MBA.","pith_inferences":["Because the Anomaly Prompt Pool and the Anomaly-Aware Forecasting network are frozen after pre-training, the framework cannot re-calibrate if deployment reveals anomaly types never synthesized during training; an online variant that updates the prompt pool with confirmed real anomalies is a natural but untested extension.","The top-N prompt retrieval doubles as an interpretation channel: the prompts selected for a given window name the anomaly shape the model is bracing for, which could yield explanations of why a specific time step is flagged without any new machinery.","The learnable injection magnitude suggests the model self-calibrates how abnormal its synthetic anomalies are; a systematic study of how that magnitude tracks each dataset's anomaly ratio could indicate when the approach will need more prompt diversity.","The reported gains are evaluated with a tolerance window in the F1 metric, so the cleanest head-to-head test of the central claim would be exact time-step matching, which the paper's own tolerance sweep suggests A2P would still win."],"forward_implications":["Anomaly-aware forecasting means future abnormal events can be given exact time points rather than just a near-term red flag, making the prediction actionable for proactive maintenance and clinical monitoring.","Training-time synthesis of anomalies removes the need for labeled anomaly examples, so Anomaly Prediction becomes applicable to domains where anomalies are rare or unknown in advance.","The unified shared backbone keeps test-time inference at the cost of a single model, so the prediction capability adds no extra inference overhead over a forecasting model alone.","The improvement over baselines grows with forecast horizon, with the largest gains reported at the longest horizon, suggesting the approach is suited to long-horizon early warning.","Both components ablate positively, and the divergence loss is the single largest contributor, indicating that diversity of synthetic anomalies is what drives detection robustness."],"supporting_citations":[{"why":"Supplies the anomaly-injection procedure (five anomaly types with learnable magnitude placed at high-reconstruction-error regions) used to synthesize anomalies for AAF pre-training.","marker":"(Darban et al., 2025)"},{"why":"Provides the reconstruction-based association-discrepancy anomaly scoring used at test time and is also a strong anomaly-detection baseline.","marker":"(Xu et al., 2022)"},{"why":"Formulates the Anomaly Prediction scenario of pinpointing abnormal future time steps, which this paper claims to solve for the first time.","marker":"(You et al., 2024)"},{"why":"Precursor-of-Anomaly detection that motivates the task but only warns whether an anomaly is coming in the near future, without exact time points.","marker":"(Jhin et al., 2023)"},{"why":"Introduces Point Adjustment, which the paper adapts into a tolerance-based F1 because Anomaly Prediction requires specific time points.","marker":"(Audibert et al., 2020)"},{"why":"Provides the threshold-setting protocol (adjusted for the percentage of anomalies in test data) used to convert anomaly scores into detection outputs.","marker":"(Shen et al., 2020a)"},{"why":"Supplies the MBA electrocardiogram dataset used in the main experiments and qualitative analyses.","marker":"(Moody & Mark, 2001)"},{"why":"Supplies the Exathlon benchmark of distributed-computing workload logs used in the main experiments.","marker":"(Jacob et al., 2020)"},{"why":"Supplies the SMD server-machine dataset used in the main experiments.","marker":"(Su et al., 2019a)"},{"why":"Supplies the WADI water-distribution testbed dataset used in the main experiments.","marker":"(Ahmed et al., 2017)"}],"fun_headline_variants":["A2P forecasts future anomalies with synthetic prompts","Synthetic anomaly prompts improve future forecasting","A2P pinpoints exact anomaly time points via prompting","Anomaly prediction: A2P outperforms forecast-detect baselines","Learnable prompt pool forecasts anomaly timestamps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the synthetic anomalies injected during pre-training (five shape types with learnable magnitude chosen by reconstruction-error regions) resemble the real anomalies that appear in the test windows closely enough that training on them transfers; if the real anomalies are of a different character, A2P learns to recognize its own synthetic artifacts.","fun_headline_variants_meta":{"raw":{"variants":["A2P forecasts future anomalies with synthetic prompts","Synthetic anomaly prompts improve future forecasting","A2P pinpoints exact anomaly time points via prompting","Anomaly prediction: A2P outperforms forecast-detect baselines","Learnable prompt pool forecasts anomaly timestamps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000556,"raw_usage":{"total_tokens":2672,"prompt_tokens":993,"completion_tokens":1679,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":1603}},"tokens_in":609,"tokens_out":1679,"duration_ms":15673,"temperature":1.0,"reasoning_tokens":1603,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:36:57.030409+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train A2P on windows whose anomalies are injected from the five standard types, then evaluate it on held-out test windows whose anomalies are produced by a different mechanism (for example, only point spikes, or slow drifts, or a sixth shape type never synthesized during training) while keeping the baselines on the same test set; if A2P's F1 advantage over forecast-then-detect baselines shrinks or disappears under that shift, the central claim is an artifact of synthetic-to-real similarity. A stricter variant replaces the default tolerance in the F1 metric with exact time-step matching to confirm the reported ordering is not carried by the tolerance window.","supporting_citations":[{"cited_title":"Anomaly prediction: A novel approach with explicit delay and horizon","cited_arxiv_id":null,"evidence_quote":"Formulates the Anomaly Prediction scenario of pinpointing abnormal future time steps, which this paper claims to solve for the first time."}],"review_version":1}