{"id":"c1ad8272-0b6d-44f1-9b34-7a454e18a8ee","arxiv_id":"2505.00590","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"AiT forecasts irregular multivariate time series by using time-point-dependent attention weights in place of static linear layers, reporting improved accuracy and runtime on four benchmarks.","lead":"The paper proposes AiT, an irregular time series forecasting model that replaces static linear layers with time-adaptive attention weights, reporting 11% better accuracy and 52% faster runtime than prior state-of-the-art on four public datasets. It is a practical method paper: the core idea is simple, but the claimed gains rely on comparisons to published baseline numbers rather than fully re-run experiments.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline accuracy/runtime margins rest on T-PatchGNN's published numbers, which are not independently reproduced; protocol equivalence is the load-bearing unverified step.","rationale":"I read the paper as making two linked claims: (i) AiT is more accurate and faster than SOTA, quantified as 11%/52%, and (ii) ALinear is a new mechanism enabling this. The model itself is simple and the ablations, hyperparameter sensitivity, and RMTS transfer experiments are internally consistent; the reported Impr. rows also check out arithmetically (Table 1 Impr. = mean of MSE/MAE improvements over T-PatchGNN; Table 2 Impr. = mean of training/inference improvements). I therefore see no fabrication or internal inconsistency. The vulnerable point is comparative: since 18 of 20 baselines and the SOTA denominators are copied from T-PatchGNN, the headline margins stand or fall on protocol equivalence, which the paper does not demonstrate and cannot be checked without code. This is the same load-bearing assumption the reader identified, so my recommendation is unchanged: the paper should be CONDITIONAL on releasing code and reproducing T-PatchGNN under identical conditions. Separately, ALinear is very close to mTAND's multi-time attention; that is a legitimate novelty concern, but it is secondary to the empirical central claim and does not by itself overturn the comparison.","tokens_in":21500,"tokens_out":7176,"duration_ms":69389,"concrete_test":"Run T-PatchGNN's official implementation (https://github.com/usail-hkust/t-PatchGNN) under the exact Appendix B.1/B.2 preprocessing, 6:2:2 splits, query definitions, and the same five seeds plus RTX 4090 environment used for AiT, and recompute Table 1 MSE/MAE and Table 2 training/inference times. If the reproduced T-PatchGNN numbers differ from the copied entries by more than overlapping standard errors (roughly >5% relative on any dataset, or >10% on runtime), the 11%/52% headline margins are not established and the comparison must be re-run end-to-end.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing step in the central claim is the comparison protocol, not the model math. The abstract's '11% accuracy and 52% runtime' margins are computed against T-PatchGNN's published numbers: Table 1's footnote says all results except iTransformer and GraFITi are 'taken from T-PatchGNN,' and Appendix B.2 says the authors 'primarily rely on the results presented in T-PatchGNN.' The 11% figure is the average of Table 1's per-dataset Impr. row (mean of MSE and MAE improvements vs. T-PatchGNN), and 52% is the average of Table 2's Impr. row (mean of training and inference time improvements); both rows therefore inherit any protocol differences in T-PatchGNN's original runs. The paper releases no code (Appendix B.3 says it will be released 'following the release'), and it provides no evidence that T-PatchGNN's numbers were produced under the same random 6:2:2 splits, query definitions, irregular-series preprocessing, metric computation, or, for Table 2, the same hardware. A material difference in any of these would change the denominators of both headline percentages. The arithmetic is internally consistent, so this is a protocol-equivalence gap rather than an internal inconsistency.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AiT, a model for irregular multivariate time series (IMTS) forecasting that replaces the static linear layers of iTransformer with an Adaptive Linear (ALinear) module. ALinear computes a softmax weight matrix from embeddings of input and output time points, enabling variable-length and irregularly sampled series to be handled within a linear-network framework, and a Transformer-based spatial encoder captures variable correlations in a shared latent space. The paper evaluates AiT on PhysioNet, MIMIC, Activity, and USHCN against 20 baselines, reporting an 11% average accuracy improvement and a 52% average runtime reduction relative to prior state of the art, together with ablations, hyperparameter sensitivity, varying-horizon experiments, and an RMTS generality study.","tokens_in":21759,"tokens_out":5892,"duration_ms":56886,"significance":"If the empirical claims are substantiated, AiT is a practically valuable IMTS forecaster that extends the linear-network paradigm to irregular sampling. The paper's strengths are a deliberately simple architecture, a broad baseline suite, ablations that isolate each component, hyperparameter and horizon analyses, and an honest Appendix E limitation statement. However, the headline accuracy and runtime margins are conditional on the comparability of baseline numbers taken from T-PatchGNN, and no statistical significance testing is reported. The potential significance is real but currently unverified; the central claim needs additional empirical support.","major_comments":[{"comment":"The central claim of an 11% average accuracy improvement rests on baseline numbers that were not reproduced by the authors. Table 1's footnote states that results for 18 of 20 baselines are 'taken from T-PatchGNN,' and Appendix B.2 says the authors 'primarily rely on the results presented in T-PatchGNN.' If the data splits, query definitions, irregular-series preprocessing, metric computation, or hardware used in T-PatchGNN differ from those used for AiT, both headline percentages (11% accuracy, 52% runtime) are not established. Please reproduce T-PatchGNN and the strongest baselines under the exact protocol used for AiT, or release code and the exact data-partition and query definitions so that protocol equivalence can be verified.","section":"Section 4.1.2, Table 1, Appendix B.2"},{"comment":"The paper reports mean and standard deviation over five seeds but no statistical significance tests. In particular, Appendix C.3 asserts that the RMTS differences are 'not a statistically significant difference' without describing any test. Because the main comparisons involve overlapping distributions (e.g., T-PatchGNN 4.98±0.08 vs. AiT 4.58±0.06 on PhysioNet MSE), the claimed superiority requires a paired significance test or at least confidence intervals or effect sizes. Please add such tests for the main tables and for the Appendix C.3 claim.","section":"Section 4.2, Table 1; Appendix C.3"},{"comment":"The runtime comparison in Table 2 inherits the same protocol-equivalence problem: the baseline times are copied from T-PatchGNN and were presumably measured on different hardware or with different software configurations. The 52% average runtime reduction in the abstract depends on the comparability of these measurements. Please either reproduce the baselines in the same environment or state explicitly the hardware, software versions, and measurement protocol for every entry.","section":"Table 2"},{"comment":"The headline '11% improvement' is an unweighted average of the four per-dataset Impr. values in Table 1, and each of those values appears to be an average of the MSE and MAE improvements, but this aggregation is not defined anywhere in the paper. Because the per-metric improvements are on different scales, the unweighted average is not a well-defined quantity. Please define the aggregation formula or report per-metric, per-dataset results only.","section":"Abstract and Table 1 'Impr.' row"}],"minor_comments":[{"comment":"The claimed equivalence between ALinear and a static linear layer in Eq. (9) is only schematic: it omits the softmax nonlinearity and the low-rank factorization, so the '≈' should be presented as an intuition rather than a derivation.","section":"Appendix A, Eq. (9)"},{"comment":"The symbol L denotes both the number of Transformer layers in Eq. (6) and the number of observations L_n in Section 3.1; please use different symbols to avoid ambiguity.","section":"Section 3.1 and Section 3.5, Eq. (6)"},{"comment":"The column header 'MA×10−2' in Table 5 is a typo; it should read 'MAE×10−2.'","section":"Table 5"},{"comment":"The label 'TemporalEncoderConcatProjection' in Figure 2 lacks a separator and is difficult to read; please use a clearer label.","section":"Figure 2"},{"comment":"The statement that code will be released 'following the release' is circular; please specify the planned release mechanism and timeline, and provide the code repository or a clear availability statement in the manuscript.","section":"Appendix B.3"},{"comment":"The problem definition refers to 'queries' q_n^j but does not specify how these query time points are generated from the future observation windows; please define their construction explicitly, since this is essential for reproducing the comparison.","section":"Section 3.1 and Section 4.1.3"},{"comment":"Appendix E honestly lists limitations (limited domain coverage and stationarity dependence), but the abstract does not acknowledge these constraints; please temper the abstract's broad claim of superiority or move the limitations into the main text.","section":"Appendix E"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest and well-structured, and the architecture is plausible, but the core comparison depends on borrowed baseline numbers from T-PatchGNN without protocol-equivalence evidence. I recommend major revision because the gap is fixable by reproducing key baselines or releasing code and precise experimental protocols, and because the ablations and auxiliary experiments are informative. Please ask the authors to also clarify the aggregation of the headline '11%' figure and to add significance tests or confidence intervals for the main comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: AiT is iTransformer with its linear layers swapped for mTAND's multi-time attention, and the headline 11%/52% margins rest on T-PatchGNN numbers the authors did not reproduce. That said, this is not a hollow paper: it applies a simple idea across four datasets, runs five seeds, and the ablations tell a coherent story.\n\nNew and good: The combination of time-aware attention as a drop-in replacement for linear layers in an iTransformer-style model is a reasonable engineering choice, and the paper demonstrates it works well on four standard IMTS benchmarks. The ablation study is solid and confirms the spatial encoder and static variable embedding contribute. The authors also show ALinear recovers standard linear forecasts in the RMTS setting, which is a nice sanity check.\n\nSoft spots: The novelty is overstated. Equation 3 is exactly the multi-time attention mechanism from mTAND [25]; the paper cites mTAND as a baseline yet does not discuss the overlap. That is a positioning failure, not a math error. More importantly, the empirical comparison is fragile: 18 of 20 baseline results are taken from T-PatchGNN, and the paper offers no evidence that those numbers were generated under the same splits, preprocessing, query definitions, metric code, or (for runtime) hardware. No code is released, so protocol equivalence cannot be checked. The 11% average gain is an unweighted mean across datasets with no significance tests, and the runtime advantage is even more environment-dependent. These issues do not invalidate the idea, but they do mean the specific headline margins are not established.\n\nBottom line: A useful baseline paper if the authors re-frame the contribution, add a proper comparison to mTAND, and either reproduce competing baselines or mark them clearly as third-party numbers. I'd send it to peer review, but I'd expect substantial revision before I'd trust the quantitative claims.","headline":"AiT is a sensible combination but the core mechanism is not new and the headline margins are built on unverified baseline numbers.","tokens_in":22343,"tokens_out":2698,"would_cite":false,"duration_ms":27108,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adaptive linear networks beat prior state-of-the-art on irregular multivariate time series forecasting, cutting error by 11% and runtime by 52%.","keywords":["irregular multivariate time series","adaptive linear network","time-aware weights","transformer","missing values","variable sampling intervals","efficient forecasting"],"falsifier":"Run AiT and the strongest baselines (especially T-PatchGNN) on the same four datasets under identical preprocessing, splits, and query definitions, and check whether AiT still achieves roughly 11% lower average MSE and 52% lower runtime than the reproduced baselines.","tokens_in":21295,"feed_emoji":"📈","tokens_out":2652,"duration_ms":29239,"temperature":0.7,"pith_summary":"This paper argues that simple linear networks, which are cheap and effective for regular time series, can be made to work on irregular multivariate time series by replacing their static weights with weights that adapt to the actual observation times. The authors introduce AiT, a model that uses an adaptive linear module for temporal encoding and prediction, plus a transformer over variable embeddings for capturing cross-variable correlations. On four benchmark datasets, they report that AiT outperforms 20 baselines, improving average prediction accuracy by about 11% and reducing runtime by 52% relative to the previous state-of-the-art. If these results hold, AiT offers a faster and simpler alternative to ODE-based and graph-based approaches for irregular multivariate forecasting.","feed_headline":"Adaptive linear network cuts IMTS forecast error 11%, runtime 52%","feed_subtitle":"AiT replaces static weights with time-adaptive ones, matching linear-model speed on irregular multivariate data.","key_machinery":"The central mechanism is ALinear, an adaptive linear layer that replaces the fixed weight matrix of a standard linear layer with a dynamically computed matrix. Given input time points and output time points, two small MLP embedders produce keys and queries, and the weight matrix is computed as softmax of the query–key dot product, followed by a weighted sum of the input. This yields a weight matrix whose shape and values adjust to the actual observation and query times, resolving the shape mismatch and numerical misalignment that static weights suffer on irregular data. When time information is absent or constant, learnable default matrices substitute for the embeddings, which lets ALinear reduce to an ordinary linear layer on regular time series.","core_discovery":"The paper's central claim is that the key barrier to using linear networks on irregular time series is the static weight matrix, which cannot align with series of variable length and irregular timing. AiT addresses this with ALinear, an adaptive linear layer that generates its weights dynamically from the input and output time points via dot-product attention and softmax normalization. The temporal encoder uses ALinear to map each variable's irregular observations to a fixed-length embedding, and the predictor uses ALinear again to map that embedding to values at arbitrary future query times. Inter-variable asynchrony is handled by running a transformer over these fixed-size variable embeddings. The paper reports state-of-the-art accuracy and efficiency across four real-world datasets: PhysioNet, MIMIC, Activity, and USHCN.","pith_inferences":["A natural extension the authors do not pursue: applying ALinear as a drop-in replacement for linear layers in other architectures (such as recurrent or state-space models) to make them irregular-data-ready.","The paper's ablation shows the static variable embedding is critical on high-missing-rate datasets, which suggests that a learned per-variable prior complements the adaptive temporal layer; this could be tested further on datasets with extreme missingness.","Because the claimed accuracy gain is 11% averaged over four datasets but varies per dataset, a robustness check across more irregularity patterns (e.g., varying missing rates or sampling jitter) would clarify where the method's advantage is largest."],"forward_implications":["If the reported results hold, AiT provides a computationally cheap linear-architecture alternative to neural-ODE and graph-based methods for irregular multivariate forecasting, with roughly half the runtime of the previous best model.","Because ALinear reduces to a standard linear layer when time points are fixed, the same model can handle both regular and irregular time series without architectural changes.","The approach removes the need for numerical ODE integration and patch-based alignment when modeling irregular samples, potentially simplifying deployment in healthcare and climate monitoring settings.","The success of a linear network with time-adaptive weights suggests that avoiding static weight misalignment matters more than complex nonlinear dynamics for many IMTS forecasting tasks."],"supporting_citations":[{"why":"iTransformer is the base architecture that AiT modifies, providing the transformer-over-variables design that ALinear is plugged into.","marker":"[19]"},{"why":"Linear networks are the foundation ALinear extends; the paper argues static linear weights fail on irregular data and uses this baseline for comparison.","marker":"[36]"},{"why":"T-PatchGNN supplies the benchmark protocol, the data splits, query definitions, and the copied baseline numbers that AiT compares against.","marker":"[39]"},{"why":"The attention mechanism and transformer blocks from this work provide the architectural pattern for ALinear's query-key weighting and the spatial encoder.","marker":"[27]"},{"why":"GraFITi is one of the two baselines reproduced from official implementations, representing a recent graph-based approach for irregular time series forecasting.","marker":"[33]"}],"fun_headline_variants":["Adaptive weights handle irregular time series, 52% faster","AiT: adaptive linear net for irregular series, 11% better","Time-adaptive linear net lifts forecast accuracy 11% on irregular data","AiT cuts error 11% and runtime 52% on irregular time series","Linear net with time-adaptive weights: 11% error cut, 52% speed-up"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes that the baseline numbers not reproduced by the authors, which were taken from a prior paper, were produced under exactly the same data splits, query definitions, preprocessing, and evaluation protocol used for AiT.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive weights handle irregular time series, 52% faster","AiT: adaptive linear net for irregular series, 11% better","Time-adaptive linear net lifts forecast accuracy 11% on irregular data","AiT cuts error 11% and runtime 52% on irregular time series","Linear net with time-adaptive weights: 11% error cut, 52% speed-up"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001363,"raw_usage":{"total_tokens":5501,"prompt_tokens":888,"completion_tokens":4613,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":4512}},"tokens_in":504,"tokens_out":4613,"duration_ms":33515,"temperature":1.0,"reasoning_tokens":4512,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:38:17.278292+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run AiT and the strongest baselines (especially T-PatchGNN) on the same four datasets under identical preprocessing, splits, and query definitions, and check whether AiT still achieves roughly 11% lower average MSE and 52% lower runtime than the reproduced baselines.","supporting_citations":[{"cited_title":"itransformer: Inverted transformers are effective for time series forecasting","cited_arxiv_id":null,"evidence_quote":"iTransformer is the base architecture that AiT modifies, providing the transformer-over-variables design that ALinear is plugged into."},{"cited_title":"Irregular multivari- ate time series forecasting: A transformable patching graph neural networks approach","cited_arxiv_id":null,"evidence_quote":"T-PatchGNN supplies the benchmark protocol, the data splits, query definitions, and the copied baseline numbers that AiT compares against."},{"cited_title":"Gomez, Lukasz Kaiser, and Illia Polosukhin","cited_arxiv_id":null,"evidence_quote":"The attention mechanism and transformer blocks from this work provide the architectural pattern for ALinear's query-key weighting and the spatial encoder."},{"cited_title":"Grafiti: Graphs for forecasting irregularly sampled time series","cited_arxiv_id":null,"evidence_quote":"GraFITi is one of the two baselines reproduced from official implementations, representing a recent graph-based approach for irregular time series forecasting."}],"review_version":1}