{"id":"dd07e761-430c-465e-9abd-62302d5b9788","arxiv_id":"2505.20774","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"TimePro forecasts long multivariate series by scanning across variables with a Mamba-like model, then tuning each variable's hidden state at adaptively chosen time points.","lead":"This paper introduces TimePro, a forecast model that scans multivariate time series across variables, then adaptively samples important time points inside the learned state to refine each variable's representation. It reports competitive accuracy on eight benchmarks with linear compute in sequence length, a relevant property for long-horizon forecasting.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4.3's complexity proof is internally inconsistent: the two linear projections in Eq. 8 are stated to cost O(N P^2 D^2), which with P≈L/S is O(N L^2), not O(N L), so the advertised linear-complexity claim is unsubstantiated unless those projections are per-token.","rationale":"The reader's weakest assumption concerns the time-tune mechanism. I think the more load-bearing weakness is the complexity analysis, because the strongest claim explicitly bundles O(NL) complexity with the accuracy results and the paper's own Sec. 4.3 contains a direct contradiction: O(N P^2 D^2) cannot be dropped by treating D as constant. This is an internal-consistency issue, not a contestable modeling choice, and it can be settled by reading the code. The mechanism concern remains relevant: Table 3's adaptive-vs-non-adaptive gain is small and the text/table discrepancy on ETTh1 (0.351 vs 0.451) should be corrected. But even if the time-tune gain were zero, the model might still be competitive; if the complexity claim is false, one of the two advertised properties fails. Since the fix (per-token projection) is plausible and verifiable, conditional acceptance with a mandatory complexity/code check is the right disposition rather than rejection.","tokens_in":18899,"tokens_out":8530,"duration_ms":90950,"concrete_test":"Clone https://github.com/xwmaxwma/TimePro and inspect the HyperMamba forward pass. Print the module and weight shape that produces Et and Ez: if it is nn.Linear(P*D, P*D) applied to a flattened N×(P×D) tensor, its FLOPs scale as N P^2 D^2, confirming quadratic-in-L behavior and invalidating Table 1; if it is nn.Linear(D, D) applied to N×P×D, the O(NL) claim holds and the paper only needs a corrected complexity derivation. Optionally profile with torch.utils.bottleneck on ECL (L=96, H=720, N=321) and compare measured projection FLOPs to N P D.","verdict_should_be":"UNCHANGED","load_bearing_attack":"TimePro's central efficiency claim rests on Section 4.3. The section first states that the two linear projections producing Et and Ez from E ∈ R^{N×(P×D)} cost O(N P^2 D^2). It then says 'these two terms can be ignored' because D and M are constants and concludes O(N L). But with D constant the projection term is O(N P^2), and since P ≈ L/S, this is O(N L^2), not O(N L). The ignored O(N P D) scan term and O(N P M D) time-tune term are indeed linear in P; the projection term is not. Thus Table 1's O(NL) entry for TimePro does not follow from the architecture as written. The claim can be rescued only if the linear projections are applied per patch over the D dimension (cost O(N P D^2)) rather than over the flattened P×D dimension, but the text and Fig. 4 use flattened N×(P×D) tensors and never specify per-token projections. Because the abstract, contributions, and efficiency comparison all advertise linear complexity, this is a load-bearing unsubstantiated step.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TimePro, a Mamba-based architecture for multivariate long-term time series forecasting. The key idea is to scan a patched multivariate series along the variable dimension with a selective state-space model and then refine the resulting hidden state by sampling it at learned per-variable time offsets (the \"time-tune\" step), producing a variable- and time-aware hyper-state. The manuscript reports competitive accuracy on eight standard benchmarks (12 firsts and 2 seconds out of 16 metrics in Table 2) and claims O(NL) complexity (Section 4.3, Table 1), supported by ablations of the time-tune mechanism and hyper-scan variants.","tokens_in":19171,"tokens_out":9937,"duration_ms":103594,"significance":"If the efficiency claim were supported, the time-tune idea—adaptively selecting time points at which to sample the variable-scan state—is a clean and plausible improvement over uniform state transfer in Mamba-based forecasters. The paper specifies the architecture in enough detail to be reimplemented, evaluates on standard benchmarks with a full results table in the appendix, includes multiple ablations, and states code availability; these are genuine strengths. However, the advertised linear-complexity guarantee does not follow from the architecture as written, and the claimed accuracy advantage over strong baselines is often within run-to-run noise, so the headline claims need revision before the paper can be accepted.","major_comments":[{"comment":"Please note that Section 4.3's conclusion is load-bearing for the paper's efficiency claim, and the missing implementation detail for the projections is a separate issue from the missing statistical support. Both should be resolved before publication.","section":"Section 4.3, Eq. (8), Table 1, Appendix B.1"}],"minor_comments":[{"comment":"The text states that the non-adaptive variant obtains MSEs of 0.360 and 0.351 on Exchange and ETTh1, respectively, but Table 3 reports the ETTh1 value as 0.451. Please correct the inconsistency.","section":"Section 5.3, Table 3"},{"comment":"The text says experiments are conducted on 'five widely used real datasets' but then lists eight benchmarks (four ETT subsets plus Exchange, Electricity, Weather, and Solar-Energy). Please make the count consistent.","section":"Section 5.1"},{"comment":"The paragraph about the Weather dataset is duplicated, and 'SOTFS' appears instead of 'SOFTS'. Please clean up the prose.","section":"Section 5.2"},{"comment":"The word 'Concact' should be 'Concat'.","section":"Eq. (8)"},{"comment":"The full-results table is very dense, with MSE and MAE values run together; separating them into distinct columns would greatly improve readability.","section":"Table 6"},{"comment":"The x-axis label 'L' for the number of encoder layers conflicts with the lookback length L used throughout the paper; consider using the symbol gamma as in Eq. (6).","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The complexity issue is the main blocking point. If the projections in Eq. (8) are in fact per-patch, a text correction and a clarified figure suffice; otherwise the O(NL) claim must be withdrawn or the architecture changed. I would also ask the authors to address the mechanism-support gap and to moderate the 'significantly surpassing' language before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I checked the complexity section first, because the abstract and Table 1 both advertise linear complexity. The stress-test note is right: the two linear projections in Eq. 8 are stated as O(N P^2 D^2) on the flattened N×(P×D) input. With D constant and P ≈ L/S, that is O(N P^2) = O(N L^2), not O(N L). The paper's dismissal of those terms because 'D is a relatively small constant' only kills the D^2, not the P^2. The claim can be rescued if those projections are actually per-patch linear layers over the D dimension, but the text and Fig. 4 never say that. Since efficiency is a headline contribution, this is a real weakness and needs to be fixed before the complexity tables can be trusted.\n\nThe actual modeling idea is good. Applying deformable sampling with learned offsets to the state of a Mamba scan along the variable dimension—time-tune—is a new combination, and it is clearly described. The empirical work is thorough: eight benchmarks, multiple horizons, ablations for patch length, dimension, layers, and the time-tune design. TimePro lands best-or-tied on 7 of 8 datasets in Table 2 (12 firsts, 2 seconds out of 16 metrics). That is a solid 'competitive accuracy' claim.\n\nWhere the paper overreaches: the contributions say 'significantly surpassing' baselines, but there are no error bars or multi-seed statistics, and several wins are tiny (0.001–0.005 MSE). The ablation text also has a numerical slip: it says the non-adaptive variant gets 0.351 on ETTh1, while Table 3 shows 0.451. The Pearson-correlation visualization in Fig. 5 is a nice post-hoc check, but it does not directly demonstrate learned delays.\n\nBaseline numbers are taken from prior papers, which is common in this subfield, but it means the comparisons are not re-run under identical conditions. The paper would be stronger with per-dataset hyperparameter settings and a verifiable code release.\n\nOverall, this is a useful method for long-term multivariate forecasting, and the time-tune idea is worth developing. I would send it to review, asking for a corrected complexity analysis and seed-level statistics. The current write-up is not ready as is, but the core contribution is legitimate.","headline":"TimePro's time-tune mechanism is a genuinely new idea and the forecasting results look competitive, but the advertised O(NL) complexity does not follow from the equations as written, and the paper overclaims significance without error bars.","tokens_in":19705,"tokens_out":4365,"would_cite":true,"duration_ms":42007,"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":"TimePro claims that a Mamba variable-scan state, re-sampled at adaptively learned intra-variable time points, captures per-variable delays and reaches state-of-the-art or tied results on seven of eight multivariate long-term forecasting…","keywords":["long-term time series forecasting","state space models","Mamba","multi-delay issue","hyper-state","time-tune strategy","multivariate forecasting","linear complexity"],"falsifier":"Train TimePro and its non-adaptive variant on a synthetic multivariate series where the target is generated from other variables with known, distinct lags, such as variable A with delay 5 and variable B with delay 40 inside the lookback window; if the learned offsets do not concentrate near the true lags, or if the non-adaptive variant matches the adaptive one in MSE, the time-tune mechanism's role in the reported gains is not supported.","tokens_in":18701,"feed_emoji":"📈","tokens_out":6205,"duration_ms":62221,"temperature":0.7,"pith_summary":"TimePro is a Mamba-based architecture for multivariate long-term forecasting that targets the multi-delay problem, where each input variable influences the target over its own time interval. Instead of transferring a plain state uniformly across variables or time points, the model scans along the variable dimension, then uses a learned per-variable offset to sample that state at adaptively selected intra-variable time points and fuses the samples into a hyper-state. The paper reports that this time-tune mechanism achieves 12 firsts and 2 seconds across 16 metrics on eight real-world benchmarks, with state-of-the-art or tied results on seven of them, while keeping computational complexity linear in both the number of variables and the series length. The contribution, as the authors present it, is evidence that learned temporal non-uniformity rather than extra capacity is what lets a linear-complexity forecaster outperform quadratic transformer baselines.","feed_headline":"Forecaster that learns when variables matter tops 12 of 16 benchmarks","feed_subtitle":"Adaptive time-point sampling inside a variable scan captures per-variable delays at linear cost.","key_machinery":"The load-bearing object is the hyper-state produced by the time-tune strategy inside HyperMamba. In place of a plain state propagated uniformly, the variable-scan state $h$ is reshaped, convolved to produce a learnable offset $\\delta_h$, and sampled at reference-plus-offset coordinates by linear interpolation; a linear layer then fuses $M = 9$ sampled time points into the hyper-state $h_o$, which is combined with the Mamba parameter matrix $C$ to yield the output. This construction lets the model keep fine-grained temporal structure per variate while scanning variables, and it is implemented on GPU SRAM so that the added sampling keeps the whole pipeline at $O(NL)$ complexity.","core_discovery":"The central claim is that a Mamba state transferred only along the variable dimension can be made time-aware without losing linear cost. After the Hyper-Scan obtains the plain state $h \\in \\mathbb{R}^{N \\times P \\times D}$ for $N$ variables and $P$ patches, a convolution produces an offset $\\delta_h$, the shifted coordinates $h_{\\text{samp}} = h_{\\text{ref}} + \\delta_h$ are sampled through differentiable linear interpolation $\\psi$ to give $\\hat{h} \\in \\mathbb{R}^{N \\times P \\times D \\times M}$, and a linear mapping fuses the $M$ sampled points into the hyper-state $h_o$. That hyper-state is multiplied by the Mamba output matrix $C$, so each variable's output is conditioned on both the cross-variable scan and locally selected time points. On eight multivariate benchmarks with lookback 96 and horizons 96, 192, 336, and 720, TimePro reports 12 firsts and 2 seconds out of 16 averaged metrics, with the largest consistent margins on the ETT family and Exchange, and the paper credits the gain to the time-tune strategy rather than to wider layers or a deeper stack.","pith_inferences":["Beyond the paper's own claims: if the learned offsets track true physical lags, the same hyper-state could double as an interpretable per-variable delay map, something the paper's post-hoc correlation visualization only hints at.","Beyond the paper's own claims: because the time-tune operator is a differentiable sampler, it could be ported as a generic sample-then-fuse block into other state-space or even non-Mamba forecasting architectures.","Beyond the paper's own claims: a synthetic experiment with known variable-specific lags would separate the two explanations for the reported gains, namely genuine adaptive delay modeling versus added nonlinear capacity, and would show whether the learned offset positions correspond to real delay structure."],"forward_implications":["If the central claim is right, uniform variable-scan or time-scan state transfer leaves accuracy on the table: the adaptive time-tune variant beats its non-adaptive counterpart on Exchange and ETTh1, and full HyperMamba beats variable-only, time-only, and sequential time-and-variable scans.","Linear complexity in both variables and series length would let the architecture scale to much wider channel counts; the channel-scaling experiment shows TimePro's memory and inference time growing roughly linearly where iTransformer grows quadratically.","On the eight benchmarks, the reported results put TimePro at or near the top for horizons 96 to 720, with the largest advantages over transformer-based baselines such as iTransformer and PatchTST.","The design suggests accuracy can come from where the state is sampled rather than from added depth, since most datasets saturate at two to four encoder layers at a moderate feature dimension."],"supporting_citations":[{"why":"Supplies the selective state-space model and hardware-aware scan that HyperMamba modifies.","marker":"Gu & Dao, 2024"},{"why":"Supplies the patching embedding approach that TimePro preserves per variable.","marker":"Nie et al., 2023"},{"why":"Provides the iTransformer baseline and the experimental setup and comparison results TimePro builds on.","marker":"Liu et al., 2024b"},{"why":"Provides the S-Mamba baseline that scans variables and time, the main Mamba competitor in the comparison.","marker":"Wang et al., 2025"},{"why":"Provides the SOFTS state-of-the-art baseline whose average results TimePro is compared against.","marker":"Han et al., 2024"},{"why":"Supplies the differentiable linear interpolation mechanism used for offset-based time-point sampling.","marker":"Xiong et al., 2024"},{"why":"Supplies reversible instance normalization, which TimePro uses to stabilize training and testing inputs.","marker":"Kim et al., 2022"},{"why":"Supplies the Informer benchmark and the ETT datasets used in the evaluation.","marker":"Zhou et al., 2021"},{"why":"Supplies the Autoformer benchmark and dataset sources including Exchange, ECL, and Weather.","marker":"Wu et al., 2021"}],"fun_headline_variants":["Time-aware Mamba state wins 12 of 16 long-term forecasting benchmarks","Per-variable time delays handled by Mamba hyper-state: 12 of 16 top scores","TimePro adaptively samples time points to beat 12 of 16 forecasting sets","Learning when each variable matters: Mamba hyper-state tops 12 of 16"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the premise that learning, per variable, which time points should tune the variable-scan state captures real delayed dependencies rather than merely adding model capacity; if a uniform or non-adaptive state transfer with equal capacity matches TimePro's accuracy, the time-tune mechanism is not what carries the result.","fun_headline_variants_meta":{"raw":{"variants":["Time-aware Mamba state wins 12 of 16 long-term forecasting benchmarks","Per-variable time delays handled by Mamba hyper-state: 12 of 16 top scores","TimePro adaptively samples time points to beat 12 of 16 forecasting sets","Learning when each variable matters: Mamba hyper-state tops 12 of 16"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001058,"raw_usage":{"total_tokens":4445,"prompt_tokens":956,"completion_tokens":3489,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":3400}},"tokens_in":572,"tokens_out":3489,"duration_ms":20709,"temperature":1.0,"reasoning_tokens":3400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:47:54.020356+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train TimePro and its non-adaptive variant on a synthetic multivariate series where the target is generated from other variables with known, distinct lags, such as variable A with delay 5 and variable B with delay 40 inside the lookback window; if the learned offsets do not concentrate near the true lags, or if the non-adaptive variant matches the adaptive one in MSE, the time-tune mechanism's role in the reported gains is not supported.","supporting_citations":[{"cited_title":"SOFTS : Efficient multivariate time series forecasting with series-core fusion","cited_arxiv_id":null,"evidence_quote":"Provides the SOFTS state-of-the-art baseline whose average results TimePro is compared against."},{"cited_title":"Efficient deformable convnets: Rethinking dynamic and sparse operator for vision applications","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable linear interpolation mechanism used for offset-based time-point sampling."}],"review_version":1}