{"id":"2615a76e-6799-4ca6-beb9-e410adbbafbd","arxiv_id":"2502.05104","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"HyperEnergy, a hypernetwork that learns LSTM weights with a learnable polynomial/RBF kernel, reports lower 24-hour-ahead energy forecast errors than 10 baselines on most of 10 building datasets.","lead":"This paper presents HyperEnergy, a forecaster that uses a hypernetwork to generate the weights of an LSTM, along with learnable polynomial and RBF kernels. Across 10 building datasets it reports lower 24-hour-ahead forecast errors than 10 baselines on most datasets, but not all.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own Table VII contradicts the universal-outperformance claim: on the office building AttentionLSTM beats HyperEnergy on all three metrics, and on the retail store the transformer beats HyperEnergy on MAE and RMSE; the medical-clinic narrative is also inconsistent with the table.","rationale":"I read the paper in good faith and identified the abstract's strongest claim: a single architecture (kernelized hypernetwork plus LSTM) achieves the lowest error across all ten datasets and three metrics. For that claim to hold, every cell in Tables V, VI, and VII must show HyperEnergy as best. The reported numbers do not satisfy this condition for Table VII, and the accompanying text for the medical clinic is inconsistent with that table. This concern is more decisive than the reader's secondary worry about missing confidence intervals: even ignoring seed variance, the printed evidence falsifies the headline. The reader's weakest assumption focused on evaluation fairness and stability; my concern overlaps with the Table VII contradiction they noted, but it centers on the internal accuracy of the reported results rather than on the experimental protocol. Since the paper's contribution could still be valuable after correcting the overclaims and adding reproducibility artifacts, the existing CONDITIONAL verdict remains appropriate; this concern does not by itself force a rejection. The proposed concrete test would settle whether the contradiction is a genuine factual error or a transcription artifact, and therefore whether the central claim must be retracted or merely revised.","tokens_in":18291,"tokens_out":5941,"duration_ms":53340,"concrete_test":"Ask the authors to release the test-set prediction files (or rerun the exact grid-search protocol with the same splits and random seeds) and recompute Table VII. If the office row still shows AttentionLSTM winning MAE, RMSE, and SMAPE, and the retail row still shows the transformer winning MAE and RMSE, then the universal-outperformance claim is false as stated. If the recomputed table instead shows HyperEnergy winning those cells, then Table VII contains transcription errors that must be corrected before the paper can be accepted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that HyperEnergy \"consistently outperformed 10 other techniques\" and that \"for all consumers, HyperEnergy achieves the lowest SMAPE\" (abstract and Figure 12 caption)—is contradicted by the paper's own Table VII. For the office building, AttentionLSTM achieves MAE 11.70, RMSE 14.50, SMAPE 2.50%, while HyperEnergy achieves 13.33, 17.16, and 2.82%, so HyperEnergy loses all three metrics. For the retail store, the transformer achieves MAE 20.48 and RMSE 24.48 versus HyperEnergy's 22.22 and 28.13, so HyperEnergy loses MAE and RMSE (it wins SMAPE 7.38% vs 7.67%). For the manufacturing building, the transformer has better RMSE (51.87 vs 52.16) even though HyperEnergy wins MAE and SMAPE. Additionally, the prose in Section V.C states that for the medical clinic \"the transformer achieved the best results in terms of all three metrics,\" but Table VII shows HyperEnergy winning all three (MAE 5.98 vs 9.76, RMSE 9.36 vs 10.70, SMAPE 2.42% vs 3.65%). These are internal inconsistencies, not questions of statistical significance: as printed, the quantitative evidence refutes the universal-superiority claim. At minimum, the abstract, Figure 12's caption, and the conclusion need to be corrected to say that HyperEnergy is competitive and usually best, and the office/retail results need verification or explanation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"HyperEnergy is a consumer energy forecasting method in which a hypernetwork with a learnable adaptive kernel—a convex combination of polynomial and RBF kernels—generates the LSTM parameters of a primary prediction network. The paper compares HyperEnergy against ten baselines on ten datasets spanning student residences, individual houses, and industrial/commercial buildings, using MAE, RMSE, and SMAPE for 24-hour-ahead forecasts. The central claim is that HyperEnergy consistently outperforms all ten baselines across all consumer types.","tokens_in":18680,"tokens_out":5421,"duration_ms":45458,"significance":"If the claimed universal superiority were established, the contribution would be practically relevant: a single forecasting architecture applicable to diverse consumers with improved accuracy over standard sequence models. The paper also provides ablation studies and a computational-time comparison, and the proposed kernelized hypernetwork idea is clearly described. However, the main claim as stated is contradicted by the paper's own Table VII, and the evaluation protocol (single run, no significance tests) does not support the 'consistently' wording. Thus the significance is conditional on a substantial correction of the claims and verification of the reported results.","major_comments":[{"comment":"The central claim that 'HyperEnergy consistently outperformed 10 other techniques' (Abstract) and that 'for all consumers, HyperEnergy achieves the lowest SMAPE' (Figure 12 caption) is contradicted by the results in Table VII. For the office building, AttentionLSTM has lower MAE (11.70 vs 13.33), RMSE (14.50 vs 17.16), and SMAPE (2.50% vs 2.82%) than HyperEnergy. For the retail store, the transformer has lower MAE (20.48 vs 22.22) and RMSE (24.48 vs 28.13); for the manufacturing building, the transformer has lower RMSE (51.87 vs 52.16). In addition, Section V.C states that for the medical clinic 'the transformer achieved the best results in terms of all three metrics,' but Table VII shows HyperEnergy with the best values on all three (MAE 5.98 vs 9.76, RMSE 9.36 vs 10.70, SMAPE 2.42% vs 3.65%). These are internal inconsistencies that refute the universal-superiority claim as printed. The abstract, Figure 12 caption, Section V.C narrative, and conclusion must be revised to describe HyperEnergy as competitive and usually best, and the office and retail results should be verified or explained.","section":"Abstract, Section V.C, Table VII, Figure 12 caption, Section VI"},{"comment":"The evaluation relies on a single run per model/dataset after per-dataset grid search, with no seed averaging, confidence intervals, or significance tests. Many of the reported advantages are small (e.g., House 4 in Table VI: HyperEnergy MAE 0.27 vs LSTM 0.28; SMAPE 37.62% vs 39.75%), so the claim that HyperEnergy 'consistently' outperforms all baselines is not statistically supported. Please report mean and standard deviation over multiple random seeds and apply appropriate statistical comparisons (e.g., paired tests across test samples or across seeds), or explicitly qualify the results as single-run observations.","section":"Section IV.B, Tables V-VII"},{"comment":"The sentence 'For the medical clinic, the transformer achieved the best results in terms of all three metrics' is the opposite of what Table VII reports: HyperEnergy obtains the lowest MAE (5.98 vs transformer 9.76), RMSE (9.36 vs 10.70), and SMAPE (2.42% vs 3.65%). One of the two sources is wrong; the numbers must be reconciled and re-checked, because the contradiction directly affects the conclusions drawn for the industrial/commercial group.","section":"Section V.C"},{"comment":"The conclusion states 'Across nine datasets, the proposed HyperEnergy outperformed 10 other forecasting approaches,' but the evaluation uses ten datasets (Table I: two residences, four houses, and four industrial/commercial buildings). The abstract also lists only residential consumer types while the paper includes industrial/commercial buildings. Please correct the dataset count and make the abstract's coverage description match the actual evaluation.","section":"Section VI and Abstract"}],"minor_comments":[{"comment":"The text says the models take 'the previous 24 hours of five features,' but the features listed in Section III.A are temperature, day of the year, day of the month, day of the week, hour of the day, and energy consumption, which are six. Please clarify the intended feature count.","section":"Section IV.A"},{"comment":"Table III reports the selected hyperparameters for only the six residential datasets, while Section IV.B states that hyperparameter optimization was conducted 'for each dataset and each model.' The selected hyperparameters for the four industrial/commercial datasets are missing and should be provided.","section":"Table III"},{"comment":"Equations (7)-(11) use the notation W1·Ko + b1, but the shapes of W1, b1, and Ko are not defined; in particular, Ko as written is a sum of kernel evaluations over reference points, so its dimensionality relative to the sample dimension should be stated.","section":"Section III.A.2"},{"comment":"Table X reports an unbalanced number of opening parentheses in the first row (the 'HyperEnergy (Traditional RBF' row), and the closing parenthesis is missing; please fix this typographical error.","section":"Section V.D"},{"comment":"Throughout the paper, there are several typographical errors (e.g., 'archived' instead of 'achieved' in Section V.C, 'captioning' in Section VI, 'therms' in Section V.B, 'adaptation' for 'adoption' in Section II). A careful proofread is needed.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"Note that this manuscript appears to be the author's version of a paper already published in IEEE Transactions on Power Delivery (DOI 10.1109/TPWRD.2024.3486010). The internal inconsistencies discussed in the major comments are therefore particularly concerning: if the published version contains the same errors, a correction may be warranted. For the arXiv version, the authors should at minimum issue a revised version correcting the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"HyperEnergy is a reasonable architecture—a hypernetwork with learnable polynomial/RBF kernels that generates LSTM weights—and the paper shows it usually wins on diverse building datasets. But the universal-superiority claim doesn't survive contact with Table VII: on the office building AttentionLSTM beats it on all three metrics, the transformer beats it on MAE and RMSE for the retail store, and the transformer has better RMSE for manufacturing. Section V.C even says the transformer was best for the medical clinic when Table VII shows HyperEnergy winning all three metrics there. That's an internal inconsistency, not a statistical subtlety.\n\nWhat's genuinely new: the kernelized hypernetwork is a real combination I haven't seen in the cited literature, and the ablation studies give some credit to both pieces—on Residence 2, SMAPE drops from 9.29% (LSTM) to 7.76% (hypernetwork without kernel) to 6.70% with the learnable kernel. House 2 shows a similar pattern. The evaluation spans ten datasets including offices, retail, manufacturing, and a clinic, which is a wider spread than most single-building forecasting papers. The parameter-integration details are concrete enough to reimplement, and they report compute times.\n\nWhere it's soft: the headline claim is overbroad and contradicted by the paper's own results. The conclusion says 'Across nine datasets' though there are ten. There are no repeated runs, confidence intervals, or significance tests; every result is a single grid-selected run, so we don't know if the remaining wins are stable. The ablation covers only two datasets. The novelty is incremental—hypernetworks and kernels are both known, and the gains are typically a few percent SMAPE—but that's fine for an applied journal paper.\n\nWho it's for: anyone working on building-level load forecasting who wants a concrete hypernetwork-based baseline, or researchers comparing meta-learned weight generation. It deserves a serious referee: the method is implementable, the comparison is broad, and the analysis is mostly clear. But the authors need to correct the abstract, Figure 12 caption, and the medical-clinic sentence, and ideally add seed averaging before the superiority claims are taken at face value. If I were editing, I'd send it to review and ask for those fixes.","headline":"A solid, incremental hypernetwork architecture with broad evaluation, but the paper's universal-outperformance claim is flatly contradicted by its own Table VII.","tokens_in":19149,"tokens_out":2482,"would_cite":true,"duration_ms":24525,"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":"A kernelized hypernetwork that writes LSTM weights is claimed to cut energy-forecast error across diverse consumer types.","keywords":["hypernetworks","consumer energy forecasting","load forecasting","LSTM","learnable kernels","polynomial kernel","RBF kernel","diverse consumer types"],"falsifier":"Run the same ten models on the same datasets over multiple random seeds and report the distribution of MAE, RMSE, and SMAPE; the paper's Table VII already shows AttentionLSTM at 2.50% SMAPE versus HyperEnergy's 2.82% on the office building and the transformer beating HyperEnergy on MAE and RMSE for the retail store, so a reader can check the universal claim directly.","tokens_in":18119,"feed_emoji":"⚡","tokens_out":7463,"duration_ms":71151,"temperature":0.7,"pith_summary":"Consumer energy forecasting has to cope with consumption patterns that differ sharply between, say, a student residence and a house with an electric vehicle, and standard deep learners often struggle when those patterns shift. This paper proposes HyperEnergy, a method in which a hypernetwork—a meta-network that outputs weights and biases—generates the parameters of an LSTM forecaster, after transforming the input through a learnable combination of polynomial and radial-basis-function kernels. The authors claim that, across ten datasets covering residences, detached homes, a home with EV charging, a townhouse, and commercial buildings, HyperEnergy achieves lower MAE, RMSE, and SMAPE than ten baselines including LSTM, AttentionLSTM, and transformers. If that claim holds, a single architecture could serve many consumer types without per-type model redesign, which is what the paper aims to establish.","feed_headline":"Meta-network-written LSTM weights outperform 10 forecasters","feed_subtitle":"If true, one architecture could replace per-building model design for residences, homes, and commercial sites.","key_machinery":"The load-bearing object is the kernelized hypernetwork $H_k$: a meta-network that predicts the tensor $\\Theta$ of LSTM weights and biases, rather than learning them by gradient descent on the LSTM itself. Its front end is the learnable adaptive kernel, defined as $K_o = \\lambda(\\alpha x r_j^T + c)^d + (1-\\lambda)\\exp(-\\gamma\\|x-r_j\\|^2)$, a learned convex combination of modified polynomial and RBF kernels evaluated against learnable reference points $r_j$; $\\lambda$ and the reference points are updated during training so the kernel can emphasize gradual trends or sudden spikes as the data demand. The parameter integration module then slices $\\Theta$ into the shapes $[4u, kv+u]$ and $[4u]$ and assigns them to the LSTM gates directly, keeping the LSTM outside the gradient update. That mechanism is what the paper claims adapts one architecture to heterogeneous consumer types.","core_discovery":"The central claim is that the parameters of a forecasting LSTM need not be learned directly by backpropagation; they can be produced by a kernelized hypernetwork that sees the same input window. The hypernetwork maps the 24-hour input through a learnable adaptive kernel and fully connected layers, then emits a parameter tensor that a parameter integration module reshapes into the LSTM's four gates; the LSTM is assigned these parameters in a gradient-free way and only the hypernetwork is updated during training. The paper reports that this arrangement outperforms the ten baselines on most datasets, with particularly large gains on individual homes, and its ablation studies attribute the largest part of the improvement to the hypernetwork itself and a smaller part to the learnable kernel.","pith_inferences":["The paper's own Table VII does not support the abstract's 'consistently outperformed' wording: AttentionLSTM has the lowest SMAPE on the office building and the transformer has lower MAE and RMSE on the retail store, so the evidence as printed supports 'best on most datasets' rather than universal superiority.","Because the reported comparisons use one run per dataset and model, the margin between HyperEnergy and the next-best method on several datasets is small enough that seed-to-seed variance could change the ranking; reporting distributions over seeds would make the claim testable.","The reference points and mixing weight $\\lambda$ could be inspected per consumer to see whether spike-heavy homes push $\\lambda$ toward the RBF term and smoother buildings toward the polynomial term, a diagnostic that would connect the mechanism to the data.","A natural extension, already signalled as future work, is transferring a trained kernelized hypernetwork between similar consumers; the learned kernels may carry reusable structure even if the LSTM weights themselves do not."],"forward_implications":["A single HyperEnergy architecture, with per-dataset hyperparameter selection, can be trained for residences, detached homes, an EV-charging home, a townhouse, and commercial buildings without changing the model structure.","Because the LSTM is not updated by backpropagation, the hypernetwork carries all learned behaviour; if the paper's ablation is right, this re-routing of learning is the main source of the accuracy gain, with the learnable kernel contributing a smaller additional gain.","The learned mixing parameter $\\lambda$ and reference points $r_j$ give a built-in knob for trading off smooth polynomial behaviour against spike-sensitive RBF behaviour on each consumer's data.","Practical deployment is plausible at the reported cost: training times between LSTM and transformer, and inference under roughly 0.2 minutes for an entire test set."],"supporting_citations":[{"why":"Supplies the hypernetwork formulation and the pre-integrative integration strategy that HyperEnergy adapts to load forecasting.","marker":"[26]"},{"why":"Provides the kernel-based transformation background and the polynomial kernel formulation used in the learnable adaptive kernel.","marker":"[14]"},{"why":"Gives the contrast between polynomial and RBF kernels for gradual versus sudden shifts in energy consumption.","marker":"[15]"},{"why":"Defines the transformer architecture used as one of the main comparison baselines.","marker":"[21]"},{"why":"Defines the N-BEATS baseline architecture included in the comparison.","marker":"[17]"},{"why":"Provides the AttentionLSTM baseline against which HyperEnergy is compared.","marker":"[41]"},{"why":"Justifies the choice of LSTM as the primary network and its suitability for capturing sudden residential load changes.","marker":"[12]"},{"why":"Supplies the industrial and commercial building datasets, including the manufacturing, medical clinic, retail store, and office building data.","marker":"[36]"}],"fun_headline_variants":["Hypernetwork-written LSTM beats 10 forecasters across diverse homes","Kernelized hypernetwork generates LSTM to top energy forecasts","One hypernetwork fits all consumer types, outdoes 10 rivals","HyperEnergy: hypernetwork-made LSTM beats 10 methods"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single run per dataset and model, with hyperparameters grid-selected per dataset, is enough to show HyperEnergy consistently outperforms every baseline; the paper's own results for the office building, retail store, and manufacturing building already contest that premise.","fun_headline_variants_meta":{"raw":{"variants":["Hypernetwork-written LSTM beats 10 forecasters across diverse homes","Kernelized hypernetwork generates LSTM to top energy forecasts","One hypernetwork fits all consumer types, outdoes 10 rivals","HyperEnergy: hypernetwork-made LSTM beats 10 methods"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000359,"raw_usage":{"total_tokens":1915,"prompt_tokens":891,"completion_tokens":1024,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":951}},"tokens_in":507,"tokens_out":1024,"duration_ms":10175,"temperature":1.0,"reasoning_tokens":951,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T20:14:03.975530+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same ten models on the same datasets over multiple random seeds and report the distribution of MAE, RMSE, and SMAPE; the paper's Table VII already shows AttentionLSTM at 2.50% SMAPE versus HyperEnergy's 2.82% on the office building and the transformer beating HyperEnergy on MAE and RMSE for the retail store, so a reader can check the universal claim directly.","supporting_citations":[{"cited_title":"Hy- permorph: Amortized hyperparameter learning for image registration,","cited_arxiv_id":null,"evidence_quote":"Supplies the hypernetwork formulation and the pre-integrative integration strategy that HyperEnergy adapts to load forecasting."},{"cited_title":"Forecasting residential energy consumption using support vector regressions,","cited_arxiv_id":null,"evidence_quote":"Provides the kernel-based transformation background and the polynomial kernel formulation used in the learnable adaptive kernel."},{"cited_title":"Regression model-based short-term load forecasting for uni- versity campus load,","cited_arxiv_id":null,"evidence_quote":"Gives the contrast between polynomial and RBF kernels for gradual versus sudden shifts in energy consumption."},{"cited_title":"N-beats neural network for mid-term electricity load forecasting,","cited_arxiv_id":null,"evidence_quote":"Defines the N-BEATS baseline architecture included in the comparison."},{"cited_title":"Multi-task short-term reactive and active load forecasting method based on attention-lstm model,","cited_arxiv_id":null,"evidence_quote":"Provides the AttentionLSTM baseline against which HyperEnergy is compared."},{"cited_title":"Short-term residential load forecasting based on LSTM recurrent neural network,","cited_arxiv_id":null,"evidence_quote":"Justifies the choice of LSTM as the primary network and its suitability for capturing sudden residential load changes."},{"cited_title":"The building data genome project 2, energy meter data from the ASHRAE great energy predictor III competition,","cited_arxiv_id":null,"evidence_quote":"Supplies the industrial and commercial building datasets, including the manufacturing, medical clinic, retail store, and office building data."}],"review_version":1}