{"id":"7ab667ca-9936-42d5-a5ad-ad9f67662fea","arxiv_id":"2607.21681","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"CARNet adds cycle-conditioned multihead core aggregation to attention-free multivariate forecasting, reporting best results on many but not all of 12 benchmarks.","lead":"The paper introduces CARNet, a neural network for forecasting many related time series at once that combines learned repeating patterns with a cheap shared 'core' summary. The authors report strong results on 12 benchmarks, but their own tables show the model loses to baselines on several important datasets, so the headline claim is overstated.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed consistent SOTA is not established: Table 2 compares CARNet's 30-epoch training with 10-epoch baseline numbers, and the 38/48 win count is not reproducible from the printed table.","rationale":"The reader's verdict is REJECT, and I agree: the load-bearing assumption is that Table 2 provides a fair, controlled comparison. The paper's own text undermines this: Appendix C specifies CARNet's training budget, while Section 4.3 explicitly contrasts it with 10-epoch baselines and retrains models under a fixed budget to remove the advantage of longer optimization. That makes the main comparison protocol unfair. Moreover, the headline quantitative claim (38/48 best MSE) is internally inconsistent with the table it is derived from, which is a correctness issue independent of fairness. A fixed-budget experiment on a subset of datasets (Table 4) already shows the claimed consistency evaporates: CARNet is not best on several settings. The architecture may still be a useful contribution, and the ablations suggest the components help, but the central claim of consistent state-of-the-art performance is not established as written. No ad hominem is intended; the issue is methodological and evidentiary.","tokens_in":20799,"tokens_out":5397,"duration_ms":45038,"concrete_test":"Re-run every baseline in Table 2 under CARNet's exact protocol — 30 epochs (7 for ETTm1/ETTm2), One-Cycle LR, seed 2024, same look-back length and hyperparameter search budget — and recompute best-setting counts. If CARNet's win rate drops below the claimed 38/48 (or below a majority of the 48 settings), the central SOTA claim is falsified. As a minimal check, recompute the win counts directly from the printed Table 2; the claimed 38/42 should match.","verdict_should_be":"REJECT","load_bearing_attack":"The central 'consistently outperforms' claim rests on Table 2, but its comparisons are not controlled. The caption says baseline results are adopted from TQNet/SOFTS; Appendix C trains CARNet for 30 epochs (7 for ETTm1/ETTm2) with One-Cycle LR and seed 2024. Section 4.3 then states that TimeXer and iTransformer were trained for only 10 epochs in their source evaluations, and the authors retrain TQNet and SOFTS under 10 epochs specifically 'to eliminate any advantage arising from longer optimization.' This admission confirms that the main table did not match training budgets: CARNet had up to 3x more optimization steps than some baselines. The paper's own fixed-budget Table 4 shows the consequence: at 10 epochs CARNet loses to TQNet on ETTm1 avg MSE (0.378 vs 0.377) and to TimeXer on Weather avg MSE (0.243 vs 0.241). Additionally, the claimed '38 out of 48' best-MSE count is not reproducible from the printed table; manual counting from the same numbers yields far fewer CARNet wins (e.g., CARNet has no best MSE on ETTm2 or ETTh2). Thus the headline claim is not supported by the evidence as reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CARNet, an attention-free multivariate time series forecasting model that extends SOFTS-style core aggregation with learnable cycle conditioning and a multi-head core aggregation mechanism. The authors claim that CARNet consistently outperforms strong transformer and non-attention baselines across 12 benchmarks with linear complexity. The paper reports MSE/MAE results, ablations, efficiency analyses, and a fixed-epoch comparison. The central contribution is empirical: the architecture is presented as a new state-of-the-art for periodic multivariate forecasting.","tokens_in":21125,"tokens_out":4643,"duration_ms":42885,"significance":"If the empirical claims were supported, CARNet would be a useful efficient alternative to attention-based channel-dependent models, and the combination of periodic conditioning with core aggregation is a plausible design direction. The complexity analysis and the ablation study provide useful evidence about the contribution of individual components. However, the core empirical claim is not established by the reported experiments, and the paper's own controlled experiments undermine the headline message.","major_comments":[{"comment":"The claimed 'best performance in 38 out of 48 settings' is not reproducible from the printed table. Examples where CARNet is not best: Traffic H=96 (SOFTS 0.376 vs CARNet 0.430), ETTh2 H=96 (TimeXer 0.286 vs CARNet 0.295), ETTm2 H=96 (CycleNet 0.163 vs CARNet 0.171). In fact, CARNet has no best MSE on ETTh2, ETTm2, or Traffic average. The count in §4.2 is therefore contradicted by the paper's own data.","section":"Table 2 / §4.2"},{"comment":"The main comparisons in Table 2 are uncontrolled: CARNet was trained for 30 epochs (7 for ETTm), while baseline numbers were adopted from TQNet/SOFTS, with TimeXer and iTransformer originally trained for only 10 epochs. The authors' own fixed-budget experiment (Table 4) shows that at 10 epochs CARNet loses to TQNet on ETTm1 avg MSE (0.378 vs 0.377) and to TimeXer on Weather avg MSE (0.243 vs 0.241). This directly undercuts the claim that gains come from the architecture rather than from longer optimization.","section":"§4.3, Table 4"},{"comment":"No error bars or significance tests are provided for any baseline. Baseline results are adopted from prior papers rather than rerun under CARNet's training protocol. Appendix F reports seed variance only for CARNet, not for baselines. The claimed 'consistent outperformance' is therefore not statistically established, and the comparisons are not controlled.","section":"§4.1, §4.3, Appendix F"},{"comment":"The ECL average row appears corrupted: SOFTS avg MSE is printed as 0.270 and iTransformer as 0.365, while the individual horizon values for these models are around 0.14–0.22. This is likely a formatting/alignment error, but because the win count is derived from this table, the error further undermines the reliability of the headline result.","section":"Table 2, ECL row"}],"minor_comments":[{"comment":"TimeXer is cited as [4] in the text but appears as [19] in the Table 2 header. 'TQNet' is inconsistently rendered as 'TQnet' in Table 2. The '1st Count' row in Table 2 is a summary rather than a model column and should be clearly separated.","section":"References and notation"},{"comment":"Algorithm 1 line 4 writes V' ← ψ(L_V(ψ(V))), whereas Eq. (2) defines V' = ψ(L_V(F_A(X',Z'))). The extra ψ in the pseudocode appears inconsistent and should be corrected.","section":"Algorithm 1, Eq. (2)"},{"comment":"The parameter counts listed in Figure 5 (124k, 1.04M, etc.) are not clearly mapped to model names in the caption. It is ambiguous which count corresponds to which model.","section":"Figure 5"},{"comment":"'produces predictions that are more accurate with the ground-truth series' should be worded as 'more aligned with the ground-truth series.'","section":"Figure 7 caption"},{"comment":"Appendix C states a 6:2:2 train/validation/test split, but the split sizes in Table 5 (e.g., ETTh1 8545/2881/2881 of 14,400) correspond to approximately 12:4:4. The text and table should be reconciled.","section":"Appendix B"}],"recommendation":"reject","confidential_remarks":"The paper's central claim is contradicted by its own Table 2, and the fixed-epoch comparison in Table 4 shows that the advantage disappears on several datasets when training budgets are matched. The experimental evaluation would need to be redone with matched training protocols, error bars, and a corrected win count before the contribution can be assessed. The architecture itself is plausible, but the current evidence does not support the claimed state-of-the-art performance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: CARNet is a real architectural contribution — combining SOFTS-style core aggregation with learnable cycle conditioning and a multihead partition — and the paper does that part cleanly. The MHCA mechanism is a natural extension, and conditioning both aggregation and redistribution on a phase-aligned cycle is a sensible idea that isn't in the cited prior work. The pseudocode, complexity analysis, and seed-variance results are all useful, and the ablations show each component earns its keep, with Cycle-Conditioned Core Redistribution giving the largest gain. That's solid engineering and honest reporting practice.\n\nThe soft spot is the headline claim. Table 2 compares CARNet trained for 30 epochs (7 for ETTm) against baseline numbers adopted from TQNet/SOFTS, and the paper itself says in Section 4.3 that TimeXer and iTransformer were trained for only 10 epochs in their source evaluations, so the authors retrain TQNet and SOFTS under 10 epochs \"to eliminate any advantage arising from longer optimization.\" That admission effectively concedes that the main table is not budget-controlled. And the fixed-budget comparison in Table 4 tells the real story: CARNet loses to TQNet on ETTm1 average MSE and to TimeXer on Weather average MSE. So \"consistently outperforms\" is not supported by the evidence as presented.\n\nAlso, the claimed 38/48 best-MSE count doesn't hold up against the printed table — CARNet has no best MSE on ETTh2 or ETTm2 at any horizon. That's a reproducibility issue that a referee should flag. Minor formatting hiccup: the ECL average row in Table 2 looks garbled and the reference list appears to label TimeXer as [19] in one spot while [4] is correct.\n\nBottom line: this is a genuine architectural proposal that deserves peer review, but the empirical section needs major revision. A revised version that reruns all baselines under identical training budgets, fixes the win counts, and narrows the claims to the datasets where CARNet actually leads would be a legitimate contribution.","headline":"CARNet is a legitimate architectural combination, but the paper's central SOTA claim is undercut by its own budget-controlled comparison; it deserves a serious referee, not a desk reject.","tokens_in":21640,"tokens_out":1967,"would_cite":false,"duration_ms":22007,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","62M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"CARNet claims that injecting learned cycles into linear-complexity core aggregation outperforms attention-based multivariate forecasters on most benchmark settings.","keywords":["multivariate time series forecasting","cross-variate dependencies","periodic cycles","core aggregation","attention-free","linear complexity","long-term forecasting","channel-dependent modeling"],"falsifier":"Rerun every baseline under CARNet's exact training recipe, including the same number of epochs, one-cycle learning-rate schedule, look-back length, and random seed, then count wins on the same 48 settings; if the 38-of-48 MSE margin shrinks or reverses, the claimed superiority is not established.","tokens_in":20690,"feed_emoji":"📈","tokens_out":3473,"duration_ms":38101,"temperature":0.7,"pith_summary":"CARNet claims that explicit global cycle information can be injected into an attention-free, linear-complexity cross-variate aggregation model without sacrificing efficiency, and that doing so improves forecasting accuracy on real-world multivariate time series. To show this, the authors add a learnable recurrent cycle representation to both the aggregation and redistribution stages of a core-based architecture, and partition the core into multiple heads. On twelve standard benchmarks across four horizons each, they report the best MSE in 38 of 48 settings and best MAE in 42, with ablations attributing the largest gain to cycle-conditioned redistribution. If the comparison holds, periodicity-aware core aggregation could become a competitive alternative to attention-based channel interaction.","feed_headline":"Cycle-aware forecast model tops 38 of 48 benchmark settings","feed_subtitle":"Attention-free and linear-complexity, CARNet injects learned periodic cycles into cross-variate core aggregation.","key_machinery":"Multihead Core Aggregation (MHCA) with cycle-conditioned redistribution. MHCA splits the variate embedding into h heads, transforms each head with a grouped 1x1 convolution, merges head-wise cores, and pools across variates to form a global core; the learned recurrent cycle Z, phase-aligned by the estimated stable cycle length W via autocorrelation analysis, is injected both before aggregation and during redistribution. This carries the argument because it turns the global periodic pattern into a conditioning signal for cross-variate interaction while keeping per-layer complexity linear in both the number of variates and the sequence length.","core_discovery":"The central discovery claimed is that conditioning cross-variate core interactions on a phase-aligned recurrent cycle, rather than simply subtracting the cycle as a residual, lets a linear-complexity, attention-free model capture periodic structure and inter-variable dependencies jointly. CARNet learns a cycle matrix per channel, aligns it with the input window, fuses it into variate embeddings, extracts a shared core through Multihead Core Aggregation by partitioning the feature dimension into h subspaces, pools across variates, and then redistributes the core together with the cycle projection back into each variate. The paper argues that removing any of these stages degrades performance,","pith_inferences":["The cycle-conditioning idea is not inherently tied to this particular core design; a fair test would be porting the same conditioning mechanism into other linear channel-mixing architectures and measuring whether similar gains appear.","Because the paper fixes W at the dataset level, an extension to per-sample or adaptively estimated cycle lengths is natural, and the reported robustness to harmonic values of W suggests that such adaptivity could work without retuning.","The fixed-epoch and multi-seed results suggest the cycle prior accelerates convergence, but the paper does not directly convert this into an accuracy-per-training-epoch comparison; that trade-off is worth quantifying for deployment decisions."],"forward_implications":["On strongly periodic benchmarks such as Electricity and PEMS, CARNet reports the best error in a majority of settings, implying that cycle-conditioned core interaction can match or beat attention-based channel modeling.","Ablations indicate that cycle-conditioned redistribution is the most valuable component, so future core-based designs should inject periodicity at redistribution, not only at input decomposition.","The per-layer complexity is linear in both variates and time, so if the claims hold, the architecture scales to high-dimensional multivariate problems more cheaply than quadratic attention.","The cycle-length ablation shows milder degradation than a comparison model when W is misspecified, suggesting the mechanism tolerates imperfect periodicity estimates."],"fun_headline_variants":["CARNet: cycle-conditioned core aggregation for forecasting","Attention-free model learns periodic cycles to beat baselines","Linear-complexity CARNet wins on 38 of 48 forecasting tasks","Cycle-aware core interactions outperform attention in forecasting"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The comparison assumes that the baseline numbers taken from earlier papers were produced under the same training and evaluation protocol as CARNet's own runs, so the reported win counts reflect a fair race.","fun_headline_variants_meta":{"raw":{"variants":["CARNet: cycle-conditioned core aggregation for forecasting","Attention-free model learns periodic cycles to beat baselines","Linear-complexity CARNet wins on 38 of 48 forecasting tasks","Cycle-aware core interactions outperform attention in forecasting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000184,"raw_usage":{"total_tokens":1111,"prompt_tokens":656,"completion_tokens":455,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":400,"completion_tokens_details":{"reasoning_tokens":389}},"tokens_in":400,"tokens_out":455,"duration_ms":5345,"temperature":1.0,"reasoning_tokens":389,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T07:47:34.706032+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun every baseline under CARNet's exact training recipe, including the same number of epochs, one-cycle learning-rate schedule, look-back length, and random seed, then count wins on the same 48 settings; if the 38-of-48 MSE margin shrinks or reverses, the claimed superiority is not established.","supporting_citations":[],"review_version":1}