{"id":"0a049b3a-de15-4a3c-a6de-46fea039b77b","arxiv_id":"2507.04690","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"MJKAN is a FiLM-modulated RBF layer that beats MLPs on some 1D regression tasks with carefully chosen basis counts, but underperforms MLPs on classification benchmarks.","lead":"This paper introduces MJKAN, a neural network layer that adds feature-wise scaling and shifting to radial basis function activations, combining ideas from Kolmogorov-Arnold Networks and MLPs. The authors report strong fits on toy regression functions but find the model loses to plain MLPs on image and text classification.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Efficiency claim is untested: no KAN baseline appears anywhere, so the promised improvement in computational efficiency over KANs is unsupported by the paper's own experiments.","rationale":"The paper's stated contribution is a hybrid layer that is both more expressive than MLPs and more efficient than KANs. The central claim has two halves. The expressiveness half is at least plausibly grounded: the MJKAN layer computes a sum of univariate RBF-FiLM functions, which is structurally analogous to a KAN layer, and stacking layers can produce non-additive functions. But the efficiency half has no evidential support. In every experiment the only baseline is an MLP; no KAN model is trained. The Introduction explicitly promises comparisons to prior KAN variants, but the experiments never deliver them. A claim that a model is more efficient than X cannot be established by showing it is slower than Y, especially when Y is the model it is supposed to beat in speed. Table 1's timing data show MJKAN slower than the MLP on all three vision datasets, so even the weaker phrase about MLP-level efficiency in the abstract is contradicted by the paper's own numbers. Parameter counts and FLOPs are absent, so the reader cannot assess whether the RBF-FiLM parameterization is actually cheaper than B-spline KANs; from Section 3.2, the cost appears to scale as O(din * K * dout) for gamma and beta, which is not obviously better than a KAN layer's O(din * dout * (G+K+3)) unless K is much smaller than G, and no such comparison is made. This is the load-bearing weakness: if a proper KAN baseline were included and MJKAN were shown to be slower or more memory-hungry, the central claim would collapse. Secondary issues, such as placeholder references and the Section 5 closed-form additive formula being inconsistent with the two-layer stack used in Section 4.2, reinforce the lack of rigor, but the efficiency gap is the decisive one. The reader's weakest_assumption identifies the same concern, so the REJECT verdict should stand.","tokens_in":9064,"tokens_out":7407,"duration_ms":73276,"concrete_test":"Add a standard KAN baseline, for example the original B-spline KAN or an efficient KAN implementation, to the vision, NLP, and regression experiments under matched parameter counts and/or FLOPs, and record wall-clock training and inference time and peak memory. If MJKAN is not faster or cheaper than the KAN baseline on at least the vision benchmarks, the abstract's claim of improved computational efficiency fails. Also report per-layer parameter counts for MJKAN with K in {5, 10, 25, 50} so the comparison is meaningful.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, stated in the abstract and conclusion, is that MJKAN inherits the theoretical advantages of KANs while improving computational efficiency and practical viability. The load-bearing condition for the efficiency half is that the RBF-FiLM layer is cheaper or faster than a standard KAN layer. This condition is never checked. Section 1 promises faster inference and higher predictive accuracy than prior KAN variants, as the experiments will show, but Tables 1 to 4 contain no KAN baseline of any kind (original B-spline KAN, EfficientKAN, or otherwise). The only timing data, Table 1, compares MJKAN to an MLP and shows MJKAN slower on all three vision datasets (MNIST 124.57s vs 120.84s; CIFAR-10 127.7s vs 115.2s; CIFAR-100 125.3s vs 115.5s). No parameter counts, FLOPs, or memory numbers are reported for any model, despite Section 3.1 giving parameter formulas for KAN and MLP. Because the claimed advantage is relative to KANs, comparing only to MLPs cannot support it, and the one timing table actively undermines the broader efficiency rhetoric. The efficiency half of the central claim is therefore unverified, and the paper's own results provide no evidence for it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MJKAN, a neural network layer that combines FiLM-style feature-wise modulation with radial basis function (RBF) activations, intended as a hybrid between KANs and MLPs. The authors claim this layer inherits KAN-like expressiveness while being more computationally efficient and practical. Experiments are reported on MNIST, CIFAR-10/100, AG News, SMS Spam, and five synthetic 1D function-regression tasks, comparing MJKAN with an MLP baseline across varying RBF basis counts. The main conclusions are that MJKAN outperforms MLP on function regression when the basis size is chosen large enough, is competitive with MLP on classification when the basis size is small, and is computationally more efficient than KANs, although no KAN model is included in any experiment.","tokens_in":9326,"tokens_out":4156,"duration_ms":46611,"significance":"If the central claims were established, MJKAN would be a useful contribution to the growing family of KAN-inspired layers, particularly for tasks requiring localized or compositional function approximation. The paper does provide explicit layer equations and a broad set of empirical settings. However, the current manuscript does not substantiate its central efficiency claim (no KAN baseline is ever run), and the regression-superiority claim rests on post hoc selection of basis size with no error bars or repeated-seed analysis. One narrative sentence in Section 4.2 is directly contradicted by the paper's own Table 3. As a result, the claimed significance is not yet supported by the evidence presented.","major_comments":[{"comment":"The paper's central efficiency claim is untested. The abstract and conclusion state that MJKAN improves computational efficiency relative to KANs, but no KAN baseline (B-spline KAN, EfficientKAN, or any other KAN variant) appears in any table or figure. Table 1, the only timing comparison, compares MJKAN with an MLP and shows MJKAN slower on all three vision datasets: MNIST 124.57 s vs 120.84 s, CIFAR-10 127.7 s vs 115.2 s, and CIFAR-100 125.3 s vs 115.5 s. No parameter counts, FLOPs, or memory measurements are reported for any model. Therefore the load-bearing claim that MJKAN is more efficient than KANs is unsupported by the paper's own experiments, and the only timing data actively undermine the broader efficiency rhetoric.","section":"Section 1, Section 4.1, Table 1"},{"comment":"The regression superiority claim is based on post hoc selection of the basis size and is partly contradicted by the reported numbers. For every task in Table 3, the best MJKAN point is at K=50, and the text highlights these best cases, but no error bars, confidence intervals, or multiple-seed results are provided, so it is impossible to tell whether the improvements are statistically meaningful. More seriously, the text says that for Global Pattern, 'MJKAN achieves lower RMSE with moderate basis counts (e.g., K=25),' yet Table 3 lists MJKAN(K=25) RMSE as 0.2329 versus MLP's 0.0840, which is substantially worse. This internal contradiction undermines the narrative of consistent regression improvement and requires correction or clarification.","section":"Section 4.2, Table 3"},{"comment":"The theoretical motivation via the Kolmogorov–Arnold theorem is asserted rather than demonstrated. The paper states that MJKAN 'aligns with' KAT by modeling univariate functions, linear mixing, and outer functions, but no theorem or proof shows that the specific FiLM-modulated RBF architecture can represent or approximate the required univariate functions, nor that it inherits the approximation guarantees of KANs. The claim in Section 1 that fixing the modulation parameters to trivial values reduces the layer to a standard KAN is also not established, since a standard KAN uses B-spline edge functions and a different parameterization. As it stands, the phrase 'inherits the theoretical advantages of KANs' is a rhetorical claim, not a supported result.","section":"Section 3.2"},{"comment":"The empirical comparisons lack standard reproducibility information. The paper does not report learning rates, batch sizes, number of training runs, seeds, hardware, or model parameter counts for the evaluated architectures. Table 3 reports RMSE values without variance, and Tables 1, 2, and 4 report accuracies without error bars or significance tests. Without these details, the claimed accuracy differences (e.g., 96.6 vs 97.9 on MNIST) cannot be distinguished from run-to-run noise, and the experiments cannot be independently reproduced.","section":"Section 4 overall"}],"minor_comments":[{"comment":"Several references appear to be placeholders or non-verifiable entries, notably 'J. Doe and A. Smith' (J. Sci. Comput. 2021) and 'P. Chen et al.' (ICML 2023). These are not standard citation names and should be either replaced with real, verifiable works or removed.","section":"References"},{"comment":"The contribution bullet states 'MJKAN is the first KAN variant to incorporate feature-wise affine modulation,' but no literature search or comparison is provided to support this novelty claim.","section":"Section 1"},{"comment":"The phrase 'Base is an optional nonlinear linear residual update' is self-contradictory; 'nonlinear linear' should be clarified.","section":"Section 3.2"},{"comment":"The regression figure is described in the text as showing predicted versus ground-truth curves, but the subplots are not clearly labeled with the task names and basis sizes in the caption, making it difficult to verify the RMSE values in Table 3.","section":"Figure 1"}],"recommendation":"reject","confidential_remarks":"The manuscript contains what appear to be fabricated or placeholder citations, which is a serious integrity concern that the editor may wish to investigate independently of the technical assessment. The absence of any KAN baseline in a paper whose central claim is efficiency relative to KANs, together with the internal contradiction in the regression discussion, makes the current submission unsuitable for publication in its present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes MJKAN, a layer that feeds each input through an RBF expansion and then applies FiLM-style per-feature modulation before summing. That combination is reasonable and worth a look in itself. Credit where it's due: the paper is unusually honest about classification results, reporting that MJKAN loses to MLP on MNIST, CIFAR-10/100, AG News, and SMS Spam, and it clearly documents the overfitting trend as the basis count grows. The closed-form per-feature expression in the conclusion is a nice touch, even if underdeveloped.\n\nThe soft spots are load-bearing, not cosmetic. The abstract and intro promise faster inference and higher accuracy than prior KAN variants, but no KAN baseline appears in any experiment. Every table compares MJKAN to an MLP, and the only timing table (Table 1) shows MJKAN slower than MLP on all three vision datasets. That directly contradicts the efficiency half of the central claim. The regression results are also weaker than advertised: MJKAN wins only at the largest basis size (K=50) on all five functions, and the authors appear to have selected that setting after seeing the test performance. No error bars, no multiple seeds, no parameter counts. The reference list contains obvious placeholders like \"Doe and Smith [2021]\" and \"Lee and Park [2022]\", which is a serious integrity problem and undermines confidence in the related-work discussion.\n\nI do not think the paper deserves a serious referee in its current form. The missing KAN baseline alone makes the main claim untestable, and the fabricated-looking references are disqualifying. If the authors added a real KAN comparison (e.g., EfficientKAN or a standard B-spline KAN), fixed the references, and reported variance, the architecture would be worth another look. As submitted, it is a modest empirical note with an unsupported headline.\n\nFor a reading group, it could be a useful case study in how KAN papers overclaim without baselines, but I would not cite it in my own work.","headline":"A KAN-inspired layer that never tests its central efficiency claim: no KAN baseline appears anywhere, and the one timing table shows MJKAN slower than MLP.","tokens_in":9840,"tokens_out":2226,"would_cite":false,"duration_ms":25111,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid layer claims KAN expressiveness at MLP-level cost.","keywords":["Kolmogorov-Arnold Networks","FiLM modulation","Radial Basis Functions","hybrid architecture","function approximation","classification","basis size","MLP"],"falsifier":"Run MJKAN and a standard B-spline KAN with matched layer widths on CIFAR-10 for the same number of epochs, measuring wall-clock time, peak memory, and parameter count; if MJKAN is not faster or smaller at equal or better accuracy, the central efficiency claim is refuted.","tokens_in":79,"feed_emoji":"🧠","tokens_out":5038,"duration_ms":70002,"temperature":0.7,"pith_summary":"This paper introduces MJKAN, a neural-network layer that replaces a KAN's spline edges with radial-basis-function expansions modulated by per-feature FiLM scaling and shifting. The authors argue that this design keeps the Kolmogorov-Arnold principle of summing learned univariate functions while reducing the computational burden of standard KANs. Their experiments show MJKAN matches or beats an MLP on function-regression tasks as the number of basis functions grows, and remains competitive with MLPs on image and text classification when the basis count is small. The paper's central conclusion is that MJKAN offers a practical middle ground between KAN expressiveness and MLP efficiency, with basis size acting as a capacity knob that must be tuned to avoid overfitting.","feed_headline":"A hybrid layer claims KAN power at MLP speed","feed_subtitle":"Function regression improves with more basis functions, while classification needs small bases to avoid overfitting.","key_machinery":"The central object is the MJKANLayer: for each input dimension $x_i$, it computes a vector of $K$ Gaussian radial basis functions $\\phi_{ij}(x_i)=\\exp(-(x_i-c_j)^2/2\\sigma^2)$, then forms modulation parameters $\\gamma_i=\\sum_j \\phi_{ij}(x_i)\\gamma_{ij}$ and $\\beta_i=\\sum_j \\phi_{ij}(x_i)\\beta_{ij}$ with learned vectors, producing output $y = \\sum_i (\\gamma_i x_i + \\beta_i)$, optionally plus a base projection. This replaces the B-spline edge functions of a standard KAN with a simpler kernel expansion and an affine feature-wise modulation, which is what the paper claims reduces computation while preserving the additive univariate structure that gives KANs their expressive power.","core_discovery":"The paper's central discovery is that a KAN-style layer can be built from two cheap ingredients—a radial-basis expansion of each input and a FiLM-style affine modulation per basis—so that each output logit is an additive sum of per-feature functions, exactly the Kolmogorov-Arnold decomposition. With this layer, the authors report that MJKAN achieves lower RMSE than an MLP on several symbolic regression tasks (e.g., reducing error on Local Bumps from 0.1955 to 0.1489 at K=50), while on MNIST, CIFAR-10/100, AG News, and SMS Spam it reaches accuracy close to MLP baselines provided the basis count is kept small. The authors interpret this as evidence that MJKAN inherits KAN's approximation strength without KAN's usual cost, at the price of a capacity–generalization trade-off controlled by the number of basis functions.","pith_inferences":["Beyond the paper: the reported basis-size sensitivity suggests that regularizing the modulation parameters (e.g., weight decay on $\\gamma$ and $\\beta$) or using adaptive basis placement could recover some of the lost accuracy at K=50, a testable extension the paper does not run.","Beyond the paper: because the logits decompose additively over features, MJKAN could double as an interpretability tool—feature-attribution maps are exact by construction rather than approximate—though the paper only notes the closed form and does not exploit it.","Beyond the paper: the efficiency claim is only argued structurally; a direct head-to-head timing and parameter benchmark against a B-spline KAN would settle whether 'MLP-like cost' holds, since the paper's only timing table compares MJKAN to an MLP and shows MJKAN slower.","Beyond the paper: the architecture is generic enough that applying FiLM modulation to other kernel families (e.g., Fourier features or wavelets) is a natural next step, potentially preserving the same trade-off profile."],"forward_implications":["If MJKAN works as claimed, it can serve as a drop-in replacement for fully connected layers across vision, language, regression, time-series, and physics tasks, since the layer has no domain-specific assumptions.","Because each output logit is a closed-form additive sum of per-feature functions, model decisions can be symbolically decomposed to show each input feature's exact contribution to each class.","Basis size becomes a key hyperparameter: small bases are needed for classification generalization, while larger bases improve function approximation.","The efficiency claim implies that KAN-style models become practical on standard hardware, assuming the RBF-FiLM layer is cheaper to run than spline-based KAN layers."],"supporting_citations":[{"why":"Supplies the KAN formulation—learnable univariate edge functions—that MJKAN builds on and simplifies.","marker":"Liu et al. [2024a]"},{"why":"Introduces FiLM, the feature-wise linear modulation mechanism MJKAN applies to RBF activations.","marker":"Perez et al. [2018a]"},{"why":"Provides the fair KAN-vs-MLP benchmark framing that motivates the claim that spline activations are the key ingredient.","marker":"Yu et al. [2024]"},{"why":"Documents KAN's empirical cost and accuracy deficits on classification tasks, the problem MJKAN aims to fix.","marker":"Chen et al. [2023]"},{"why":"Introduces kernel activation networks, the broader family that RBF-based KAN variants like MJKAN belong to.","marker":"Kliger et al. [2023]"},{"why":"Quantifies KAN hardware inefficiency (latency, LUTs, DSPs), used to argue that a cheaper KAN layer is needed.","marker":"Kumar and Zhang [2022]"}],"fun_headline_variants":["MJKAN: hybrid layer blends KAN expressiveness with MLP efficiency","KAN power, MLP speed: MJKAN merges both in one layer","MJKAN: RBF bases plus FiLM modulation yields KAN expressivity at MLP cost","MJKAN layer: hybrid of KAN and MLP, with basis size as the dial","MJKAN: fewer basis functions generalizes better in classification"],"cache_read_input_tokens":12032,"weakest_assumption_plain":"The claim that MJKAN is efficient rests on the assumption that an RBF-plus-FiLM layer costs less to train and run than a standard spline-based KAN layer, but the paper never measures that comparison directly.","fun_headline_variants_meta":{"raw":{"variants":["MJKAN: hybrid layer blends KAN expressiveness with MLP efficiency","KAN power, MLP speed: MJKAN merges both in one layer","MJKAN: RBF bases plus FiLM modulation yields KAN expressivity at MLP cost","MJKAN layer: hybrid of KAN and MLP, with basis size as the dial","MJKAN: fewer basis functions generalizes better in classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000682,"raw_usage":{"total_tokens":3126,"prompt_tokens":1005,"completion_tokens":2121,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":2012}},"tokens_in":621,"tokens_out":2121,"duration_ms":14629,"temperature":1.0,"reasoning_tokens":2012,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:41:50.287752+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MJKAN and a standard B-spline KAN with matched layer widths on CIFAR-10 for the same number of epochs, measuring wall-clock time, peak memory, and parameter count; if MJKAN is not faster or smaller at equal or better accuracy, the central efficiency claim is refuted.","supporting_citations":[],"review_version":1}