{"id":"e65d9585-d42b-4eb2-adb5-25fa723a5685","arxiv_id":"2505.06657","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A hybrid of MLP-Mixer, Informer, and KAN with two-stage transfer learning is claimed to reduce EV charging load forecast errors by 4% to 8% on a 26-station dataset.","lead":"This paper proposes MIK-TST, a deep learning model that combines three known neural network designs to forecast electric vehicle charging station load, with a pre-training and fine-tuning step for new stations with little data. It reports a 4% to 8% error reduction over standard baselines on charging data from 26 stations in Boulder, USA, but the comparison lacks error bars and treats the baselines and the new model differently.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No transfer-learning baseline: the claimed 4%/8% gains likely stem from extra source data, since removing transfer learning makes MIK-TST worse than all baselines.","rationale":"The reader's weakest assumption correctly isolates the missing transfer-learning control. This is the single most load-bearing issue because Table 1's entire empirical case depends on it. The paper's own ablation (Table 2, 'w/o-Dcl') shows MIK-TST without transfer learning performs worse than every baseline, so the architecture itself is not demonstrated to be advantageous. The only way to support the headline claim is to show MIK-TST's full pipeline beats baselines with equal access to source data. I concur with the reader's REJECT; the concern is not merely cosmetic, and adding a transfer-enabled baseline is a necessary condition for accepting the paper. I also note secondary issues (test-set hyperparameter selection, single-run results) but do not base the verdict on them.","tokens_in":12100,"tokens_out":3055,"duration_ms":28848,"concrete_test":"Run PatchTST (and optionally Informer) with the exact same two-stage transfer protocol used by MIK-TST: pre-train on stations 0-20 with full historical data, fine-tune on pre-2023 data of stations 21-25, evaluate on the same test period. Use identical optimizer settings and early stopping; report MAE/MSE. If transfer-enabled PatchTST matches or beats MIK-TST (0.4450/0.6322), the claimed improvement is not from the MIK architecture.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section 4.4, Table 1) is that MIK-TST beats PatchTST/Autoformer by 4% MAE / 8% MSE. But the comparison is not controlled: MIK-TST is pre-trained on 21 source stations (Section 4.1) and then fine-tuned, while every baseline is trained only on the five target stations' early data. The only ablation that removes the two-stage transfer learning strategy (Table 2, 'w/o-Dcl') yields MAE 0.4772 / MSE 0.7457, which is worse than every baseline in Table 1. Thus the architecture alone (Mixer+Informer+KAN) is not superior; the reported gains are attributable to the extra source-domain data. Without a transfer-learning-enabled baseline, e.g., a fine-tuned PatchTST or Informer, the 4%/8% claim cannot be attributed to the proposed architecture. This is the weakest load-bearing assumption because it directly concerns the fairness of the experiment that supports the headline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes MIK-TST, a hybrid architecture that combines MLP-Mixer, Informer, and Kolmogorov-Arnold Networks with a two-stage transfer learning strategy for long-sequence load forecasting at newly constructed EV charging stations. The authors pre-train on 21 Boulder stations and fine-tune on early data from five target stations, reporting MAE 0.4450 and MSE 0.6322, interpreted as 4% and 8% improvements over baselines. An ablation study removes the Mixer, KAN, and transfer learning components, and a sensitivity analysis varies hidden dimension, attention heads, and layer depth.","tokens_in":12236,"tokens_out":3662,"duration_ms":32949,"significance":"The problem is practically relevant: newly constructed EV stations have scarce data, and transfer learning is a natural remedy. If the reported gains were genuine, this would be a useful contribution to the forecasting literature. However, the paper does not deliver a controlled empirical comparison: all baselines are trained without access to the source-domain data, and the sole ablation that removes transfer learning makes the model worse than every baseline, showing the reported gains come from additional data rather than from the MIK architecture. The paper also lacks standard deviations, multiple seeds, and statistical tests, and the hyperparameters of the proposed model are selected on the evaluation set. The central claim is therefore not established.","major_comments":[{"comment":"The comparison is not controlled. MIK-TST is pre-trained on 21 source stations and then fine-tuned on the target stations, while every baseline is trained only on the early target data. The paper never includes a transfer-learning-enabled baseline, such as a fine-tuned PatchTST or Informer. Consequently, the reported 4% MAE and 8% MSE gains cannot be attributed to the Mixer-Informer-KAN architecture. The ablation w/o-Dcl in Table 2 removes the two-stage transfer learning strategy and yields MAE 0.4772 and MSE 0.7457, which is worse than every baseline in Table 1; the authors even state this is a 7.2% MAE and 18% MSE increase. This demonstrates that the architecture alone is not superior to the baselines and that the headline improvement is driven by extra source-domain data.","section":"§4.4, Table 1 and §4.1"},{"comment":"Hyperparameter selection is performed on the evaluation set, introducing selection bias. The paper reports the best configuration among hidden dimension d in {64, 128, 256, 512, 1024}, attention heads in {1, 2, 4, 6, 8, 10}, and layers in {1, 2, 4, 6} without a held-out validation split. Baseline models are said to use original parameters without any tuning. This makes the comparison unfair and the reported improvements not statistically meaningful. Please provide a proper validation split, report performance for all configurations, and give standard deviations over multiple seeds.","section":"§4.3 and §4.6"},{"comment":"The claim that MIK-TST outperforms 'the best baseline' by 4% in MAE and 8% in MSE is internally inconsistent. The best MAE baseline is PatchTST (0.4638), but the best MSE baseline is Autoformer (0.6869); the 8% MSE improvement is relative to Autoformer, not PatchTST. The paper should state the per-metric baseline and report confidence intervals for both metrics.","section":"§4.4, Table 1"},{"comment":"The experimental protocol is underspecified. The paper does not provide the prediction horizon (e.g., 24 hours, 7 days), the backtest window, how 'early data prior to 2023' is split into training and validation, or the number of samples in the source and target domains. Without this information, the reader cannot reconstruct the experiment or assess the impact of the fine-tuning sample size. Please include a data description table and a precise train/validation/test split.","section":"§4.1"}],"minor_comments":[{"comment":"The square root in the attention formula is rendered as the literal string 'sqrt(dk)' in the text and should be typeset as sqrt(d_k).","section":"Eq. (7)"},{"comment":"The notation is garbled: x ∈ R^n and y ∈ R^m appear in Eq. (3), but the dimension compatibility with the mixing layers is not defined. The regularization function denoted by the symbol that resembles a script 'D' is introduced only as 'an effective regularization technique' and should be explicitly identified.","section":"§3.1.1, Eqs. (4)–(6)"},{"comment":"There are multiple typographical errors, including 'TBy leveraging this mechanism' on the Mixer page, and inconsistent spacing in equations such as '𝑓𝐿(𝑋) = 𝑊𝑋 +𝑏'. Please proofread the text.","section":"§3.1.1 and throughout"},{"comment":"The parameter sensitivity figures are not described in the caption beyond the parameter name; they should include axis labels, units, and the range of the y-axis to be interpretable.","section":"Figs. 5–7"},{"comment":"No code or data availability statement is provided, which limits the usefulness of the empirical results for the community.","section":"Reproducibility"}],"recommendation":"reject","confidential_remarks":"The paper's main experimental contribution is not supported because the model's improvement disappears when transfer learning is removed. Adding transfer-learning baselines would be essential, but the current manuscript also selects hyperparameters on the test set and provides no uncertainty estimates, so a major revision would in effect require a new study. The topic fits the journal's scope, but the empirical rigor is below the standard expected for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know: the paper's headline claim — 4% MAE and 8% MSE gains from the Mixer-Informer-KAN + transfer learning stack — is not established by the experiments. The architecture itself is an incremental assembly of three published modules with a standard pre-train/fine-tune loop. What's genuinely nice is the problem: newly built EV charging stations with little data are a real forecasting pain point, and the split by chronological station order is a sensible way to simulate a new station.\n\nThe paper does a few things right. It compares against a reasonable set of baselines (PatchTST, Autoformer, DLinear, Crossformer, FreTS). The ablation study explicitly shows how much transfer learning contributes. And the writing is honest about limitations. That matters.\n\nBut the central comparison is uncontrolled. MIK-TST is pre-trained on 21 source stations and then fine-tuned; every baseline is trained only on the five target stations. So the 4/8% could simply be the extra source data — and the ablation confirms it: removing transfer learning (w/o-Dcl) gives MAE 0.4772 / MSE 0.7457, worse than every baseline in Table 1. Without a transfer-learning-enabled baseline (e.g., a fine-tuned PatchTST or Informer), you cannot attribute the gains to the MIK architecture. This is a load-bearing flaw.\n\nOther soft spots: single runs, no standard deviations, no significance tests. Hyperparameters were chosen on the same evaluation set used to report the headline numbers. There are also mechanical issues: Eq. 7 is vanilla softmax attention, not ProbSparse; the ablation labels (Mcl/Kcl/Dcl) are unexplained; and the parameter sensitivity section reports a single run per setting.\n\nI want to be fair: none of this is fraudulent. The gains may even survive a cleaner test. But as it stands, the evidence is too thin for an applied forecasting paper. The authors need to add transfer baselines, report multiple seeds, and use a validation split for hyperparameter selection.\n\nWho is this for? Applied researchers in EV charging / smart grid who want a transfer-learning template. With proper controls, it could be a decent niche study. I'd give them a shot in peer review — but only because the fix is clear and the problem is real. If the revision doesn't add those baselines, reject.","headline":"Uncontrolled comparison: without a transfer-learning baseline, the 4/8% gains in MIK-TST likely come from extra source data, not the architecture.","tokens_in":12820,"tokens_out":2687,"would_cite":false,"duration_ms":26631,"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":"MIK-TST reduces EV charging load forecast error by 4% in MAE and 8% in MSE relative to the best baseline.","keywords":["EV charging load forecasting","transfer learning","Mixer","Informer","KAN","long-sequence time-series forecasting","ProbSparse attention","fine-tuning"],"falsifier":"Run PatchTST or Informer through the same two-stage pre-train/fine-tune protocol on the same 21 source and 5 target Boulder stations; if either reaches or beats MAE 0.4450 and MSE 0.6322, the claimed architectural advantage of MIK-TST is not supported.","tokens_in":11834,"feed_emoji":"🔋","tokens_out":5896,"duration_ms":48626,"temperature":0.7,"pith_summary":"This paper proposes MIK-TST, a forecasting model that combines a Mixer feature-fusion layer, an Informer with ProbSparse attention, and a Kolmogorov-Arnold Network, and trains it in two stages: pre-training on 21 established charging stations and fine-tuning on limited early data from five newer stations. The authors claim this setup reduces mean absolute error by 4% and mean squared error by 8% relative to the best baseline, PatchTST, on data from 26 EV charging stations in Boulder, USA. If true, the value is practical: newly built charging stations could get usable load forecasts from weeks of data instead of months or years, supporting grid planning and charging infrastructure expansion. The paper also reports ablation results attributing the gain chiefly to the transfer-learning stage, with the Mixer and KAN modules contributing smaller improvements.","feed_headline":"Hybrid model cuts EV charging forecast error by up to 8%","feed_subtitle":"Pre-trained on 21 stations then fine-tuned on five new ones, the model beats PatchTST in Boulder tests.","key_machinery":"The load-bearing mechanism is the MIK-TST pipeline itself, which stacks three components: the Mixer module, an all-MLP structure with separate temporal-mixing and feature-mixing layers that fuse multi-source inputs; the Informer encoder-decoder, whose ProbSparse attention selects dominant queries to capture long-range dependencies at sub-quadratic cost; and a KAN module, which replaces fixed node activations with learnable spline functions on the edges. These are tied together by a two-stage transfer learning strategy: pre-train on complete histories of 21 source stations with MSE loss, then fine-tune on limited early data from five target stations, freezing Mixer and Informer weights and updating only KAN, or fine-tuning all parameters at a reduced learning rate. The framework's job is to make forecasts accurate enough for newly constructed stations where only weeks of data exist.","core_discovery":"On the paper's own terms, the central discovery is that a hybrid architecture plus a two-stage transfer learning protocol produces the best reported forecasts on the Boulder 26-station dataset: MIK-TST reaches MAE 0.4450 and MSE 0.6322, compared with PatchTST's 0.4638 and 0.7409, the best baseline on the two metrics. The authors interpret the gain as coming from the interaction of the three modules—Mixer for fusing temporal and channel features, Informer's ProbSparse attention for long-range dependencies, and KAN's learnable activations for nonlinear patterns—together with pre-training on 21 source stations before fine-tuning on five target stations. The ablation study supports the transfer stage as the largest contributor: removing transfer learning raises MAE by 7.2% and MSE by 18%, while removing Mixer raises MAE by 3% and MSE by 3.5%, and removing KAN raises MAE by 1.3% with negligible MSE change.","pith_inferences":["The reported comparison may overstate the architecture's contribution: baselines are trained only on target stations, so part of the 4%/8% edge could be the 21-station pre-training corpus. A transfer-enabled PatchTST or Informer baseline would separate data advantage from model advantage.","The KAN module's contribution looks small in the ablation (1.3% MAE, no MSE change), suggesting a simpler nonlinear head might perform almost as well; testing that substitution would clarify whether KAN earns its complexity.","With a single-city, 26-station dataset and station assignment by chronological opening order, the results may not generalize to stations with very different usage profiles; a multi-city or heterogeneous-station test would be more informative.","The paper's own limitations note the absence of weather, traffic, and socio-economic covariates; adding them could shift which module matters most and could make the transfer gains larger or smaller."],"forward_implications":["A newly opened charging station could receive a usable forecasting model after fine-tuning on a short history (the paper uses pre-2023 data), rather than waiting months to accumulate training data.","The ablation numbers imply the biggest single return comes from the transfer protocol, not from any one network module: without it, MSE rises 18%.","Hyperparameter settings that balance capacity against sparse fine-tuning data (embedding 256, 8 attention heads, 2 encoder/decoder layers) are part of the claimed performance, not incidental choices.","Because MSE improves more than MAE, the model reduces large forecasting errors disproportionately, which matters for grid operators worried about peak misprediction."],"supporting_citations":[{"why":"Supplies the Informer ProbSparse attention backbone used for long-range dependency modeling.","marker":"[7]"},{"why":"Defines PatchTST, the best baseline that MIK-TST is compared against.","marker":"[8]"},{"why":"Defines Autoformer, a decomposition/auto-correlation baseline in the comparison.","marker":"[9]"},{"why":"Defines DLinear, the linear decomposition baseline in the comparison.","marker":"[10]"},{"why":"Defines Crossformer, the cross-dimension attention baseline in the comparison.","marker":"[11]"},{"why":"Defines FreTS, the frequency-domain baseline in the comparison.","marker":"[12]"},{"why":"Defines the KAN module with learnable spline activation functions used in the architecture.","marker":"[13]"},{"why":"Defines the Mixer all-MLP mixing module used for temporal and channel feature fusion.","marker":"[14]"},{"why":"Provides the 26-station Boulder, USA charging dataset used for pre-training, fine-tuning, and evaluation.","marker":"[26]"}],"fun_headline_variants":["Transfer learning lifts EV charging forecast accuracy by 8%","Hybrid model with transfer learning beats baselines in EV load forecast","MIK-TST: two-stage transfer learning for EV charging load prediction","Pre-train on 21 stations, fine-tune on 5: better EV load forecasts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The fairness of comparing MIK-TST, pre-trained on 21 source stations, against baselines trained only on the five target stations is the load-bearing premise; if that comparison is unfair, the claimed 4% and 8% gains may be a data effect, not an architecture effect.","fun_headline_variants_meta":{"raw":{"variants":["Transfer learning lifts EV charging forecast accuracy by 8%","Hybrid model with transfer learning beats baselines in EV load forecast","MIK-TST: two-stage transfer learning for EV charging load prediction","Pre-train on 21 stations, fine-tune on 5: better EV load forecasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000741,"raw_usage":{"total_tokens":3287,"prompt_tokens":906,"completion_tokens":2381,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":2301}},"tokens_in":522,"tokens_out":2381,"duration_ms":16819,"temperature":1.0,"reasoning_tokens":2301,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:36:55.362733+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run PatchTST or Informer through the same two-stage pre-train/fine-tune protocol on the same 21 source and 5 target Boulder stations; if either reaches or beats MAE 0.4450 and MSE 0.6322, the claimed architectural advantage of MIK-TST is not supported.","supporting_citations":[{"cited_title":"Informer:Beyondefficienttransformerforlongsequencetime-series forecasting,","cited_arxiv_id":null,"evidence_quote":"Supplies the Informer ProbSparse attention backbone used for long-range dependency modeling."},{"cited_title":"Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting,","cited_arxiv_id":null,"evidence_quote":"Defines Autoformer, a decomposition/auto-correlation baseline in the comparison."},{"cited_title":"Crossformer: Transformer utilizing cross- dimension dependency for multivariate time series forecasting,","cited_arxiv_id":null,"evidence_quote":"Defines Crossformer, the cross-dimension attention baseline in the comparison."},{"cited_title":"Frequency-domain mlps are more effective learners in time series forecasting,","cited_arxiv_id":null,"evidence_quote":"Defines FreTS, the frequency-domain baseline in the comparison."},{"cited_title":"Mlp-mixer: An all-mlp architecture for vision,","cited_arxiv_id":null,"evidence_quote":"Defines the Mixer all-MLP mixing module used for temporal and channel feature fusion."},{"cited_title":"Electric vehicle charging station data,","cited_arxiv_id":null,"evidence_quote":"Provides the 26-station Boulder, USA charging dataset used for pre-training, fine-tuning, and evaluation."}],"review_version":1}