{"id":"94248d33-06c6-4e4f-8dd1-5ed4c9465aca","arxiv_id":"2511.11228","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A hybrid KAN-MLP physics-informed network with a trainable convex weight and overlapping domain decomposition improves reported accuracy on high-frequency and multiscale PDE benchmarks.","lead":"This paper combines a Kolmogorov–Arnold network and a multilayer perceptron inside a physics-informed neural network, adding a learnable weight and overlapping subdomains to handle rapidly oscillating solutions. If the benchmark gains hold up, it offers a practical recipe for problems where standard PINNs fail: high-frequency and multiscale PDEs.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Trainable α is never compared against fixed α or manual tuning, so the claimed efficiency/adaptive-balancing benefit is unsupported.","rationale":"The paper's strongest claim has two parts: accuracy and cost. The accuracy comparison could be valid even if α is unnecessary, but the advertised cost reduction is specifically about not manually tuning α. The experiments do not isolate this variable. In Sec. 3.1.1, Figures 4 and 6 show that weighted variants beat the unweighted sum, but not that trainable α beats a fixed good α. Since the authors selected S-function, overlap ratio, subdomain counts, and sampling points per benchmark (e.g., δ=3.3 in Sec. 3.1, δ=2.9 in Sec. 3.2.1), a fixed-α comparison is the missing control. The conclusion's statement that training cost is higher than MLP makes the efficiency claim even more dependent on this missing comparison. This is not a claim of scientific misconduct; it is an empirical gap that can be closed with a straightforward ablation and timing study.","tokens_in":16211,"tokens_out":7793,"duration_ms":64734,"concrete_test":"On Helmholtz ω=32 and the reaction-diffusion problem (3.7), run the modified HPKM with α fixed to 0, 0.25, 0.5, 0.75, 1 and with trainable α, using identical domain decomposition, seeds, epochs, and hardware; also perform a grid search over α to simulate manual tuning and record wall-clock time and final L2. If trainable α does not match or beat the best fixed α and is not cheaper than grid search, the central adaptive-weighting/efficiency claim in the abstract is not supported. Report at least 5 seeds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty is the trainable sigmoid-bounded weight S(α) in Eq. (2.2), advertised as replacing manual α tuning and reducing training costs. However, the experiments in Sec. 3.1.1 compare only different S-shaped functions and an unweighted MLP+KAN sum ('No α'); they do not include fixed α values (e.g., α=0, 0.25, 0.5, 0.75, 1). Thus the reported gains over MLP/KAN baselines could come from the hybrid KAN+MLP architecture or from the domain decomposition (Sec. 2.2) rather than from the adaptive mechanism. This is load-bearing because the abstract specifically claims cost reduction relative to manual hyperparameter tuning, and the conclusion (Sec. 4) concedes the model 'requires higher training costs than traditional MLP architectures.' Without a fixed-α/manual-tuning baseline and runtime measurements, the key advantage of trainable α is not actually demonstrated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a modified Hybrid Parallel Kolmogorov–Arnold Network and Multilayer Perceptron PINN (modified HPKM-PINN) with overlapping domain decomposition. The key modification is replacing the fixed weighting parameter α in the HPKM combination with a trainable parameter transformed by an S-shaped function, so that the convex combination of KAN and MLP outputs is adaptively balanced. The method is tested on Helmholtz (ω=16,32,48), 2D and 5D Poisson, reaction–diffusion, and Allen–Cahn benchmarks, reporting lower L2 errors than MLP-only and KAN-only PINNs. The abstract claims that the method reduces training costs and improves computational efficiency compared with manual hyperparameter tuning.","tokens_in":16482,"tokens_out":6943,"duration_ms":65907,"significance":"If substantiated, the contribution is a simple and inexpensive architectural modification with a plausible intuitive appeal: combining a KAN branch and an MLP branch with a bounded trainable weight within an overlapping domain decomposition. The benchmark suite is reasonably broad, including high-frequency and multiscale problems, and the method is easy to implement. However, the central efficiency claim is unsupported and is in direct tension with the paper's own conclusion that the model requires higher training costs than MLP architectures. The experiments do not isolate the effect of the adaptive weighting mechanism, report no statistical uncertainty, and provide no code. The frequency-separation narrative is also not established. The paper therefore presents a promising empirical idea that requires substantially stronger validation, rather than a demonstrated method.","major_comments":[{"comment":"The central novelty is the trainable weighting parameter S(α). The experiments in Figs. 4–7 compare only different S-shaped functions and the unweighted 'No α' direct sum; they do not include any fixed-α baseline (e.g., α=0, 0.25, 0.5, 0.75, 1) or the original HPKM with manually tuned α. Since the abstract claims reduced cost compared with manual hyperparameter tuning, this missing control is load-bearing. The reported gains could come from the hybrid architecture itself or the domain decomposition rather than from the adaptive mechanism. Add fixed-α baselines and report the cost of tuning α.","section":"§3.1.1, Eq. (2.2)"},{"comment":"The conclusion states that the model 'requires higher training costs than traditional MLP architectures due to its larger parameter size,' while the abstract claims the method 'reduces training costs and improves computational efficiency compared with manual hyperparameter tuning.' No runtime, FLOPs, or parameter counts are reported anywhere. The efficiency claim is therefore unsupported and internally inconsistent. The authors should either substantiate the claim with measurements or remove it from the abstract and introduction.","section":"§4 vs. Abstract"},{"comment":"Overlap ratios, subdomain counts, sampling points, Fourier frequency count K, hard-constraint steepness σ, and the choice of S-shaped function are selected per benchmark with statements such as 'Based on our tests' and 'determined through testing.' The sigmoid function is selected after comparing candidates on the same Helmholtz problems that are later used to claim superiority. This is selection on the test problems and risks circularity. The authors should provide sensitivity analyses and make it clear that all baselines were tuned with the same effort, or conduct validation on separate problems.","section":"§3.1, §3.2.1, §3.3, §3.4"},{"comment":"The hard-constraint operator C is not defined in general. An explicit tanh-based construction is given only for the Helmholtz equation; for the Poisson, reaction–diffusion, and Allen–Cahn examples no hard-constraint construction is provided, even though all experiments rely on it. This omission prevents reproduction. Provide the concrete form of C for each benchmark or switch to soft constraints with reported loss weights.","section":"§2.2.2, Algorithm 1"},{"comment":"The paper claims that a single scalar S(α) automatically balances low-frequency (MLP) and high-frequency (KAN) components. However, both branches receive identical inputs and S(α) is a global scalar per subdomain, not a frequency-dependent filter. The observed improvements may instead reflect an ensemble effect or the higher capacity of having two branches. The S(α) trajectory plots do not demonstrate frequency separation. Please provide a concrete analysis of when the two branches interfere or soften the interpretation.","section":"§2.1.2, Figs. 5 and 7"},{"comment":"All L2 errors in Table 1 are single-run numbers with no seeds, no error bars, and no code. Neural network training is stochastic, so differences such as 2.362×10⁻⁴ vs. 5.115×10⁻⁴ may lie within run-to-run variability. The manuscript should report means and standard deviations over multiple seeds and, ideally, release code to make the comparisons reproducible.","section":"Table 1"}],"minor_comments":[{"comment":"Typo: 'Hemholtz' should be 'Helmholtz'.","section":"§3.1 title"},{"comment":"Typo: 'KAN-baed model' should be 'KAN-based model'.","section":"§3.1.2"},{"comment":"Typo: 'Possion' should be 'Poisson'.","section":"Fig. 9 caption"},{"comment":"The KAN citation is inconsistent: the introduction refers to 'Li et al. [17]' but reference [17] is Kodama (2011); the KAN paper appears as reference [16] with different authors. Reference [6] is cited for a lack of theoretical analysis but actually points to a Fourier features paper.","section":"References [16]–[17]"},{"comment":"The definition of f(x) has unbalanced parentheses, which should be corrected.","section":"Eq. (3.7)"},{"comment":"Line 14 says 'Hard constraints: û^(i)=C u^(i)' but the operator C is not specified there; please clarify that it refers to the construction in §2.2.2 and give the explicit form.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The empirical core of the paper may be salvageable, but the adaptive-α claim needs to be tested against fixed-α baselines and the efficiency claim must be either measured or removed. If the authors cannot provide multi-seed error bars and fixed-α comparisons, the paper should be restricted to an accuracy benchmarking study that does not claim reduced training costs. The selection of hyperparameters on the test problems is a separate but important concern that should be addressed with sensitivity analyses."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it because it is a clean, small extension of the hybrid KAN/MLP idea, and because it illustrates a common gap: the one new knob is never tested against the old knob. The paper replaces the fixed α in HPKM [37] with a trainable α wrapped in a sigmoid, and combines that with FBPINN-style overlap. That is the whole novelty, and the authors are clear about it. On the positive side, the writing is straightforward, the benchmark set is standard, and the reported L2 errors are consistently lower than MLP-only and KAN-only PINNs using half the width. The S(α) evolution plots are a genuinely useful check: the weight drifts toward MLP for low-frequency problems and KAN for high-frequency ones.\n\nThe soft spot is load-bearing. Section 3.1.1 compares different S-shaped functions and an unweighted sum, but never a fixed α. So the claimed advantage over manually tuned α is not demonstrated; the gains could come from the hybrid architecture or the domain decomposition. The abstract also says the method cuts training costs, but Section 4 says it needs higher training costs than traditional MLP. That tension should have been caught before submission. Table 1 is single-run, no seeds, no code, and overlap ratios/subdomain counts are selected on the same problems used for the headline numbers. That is mild circularity, not fatal, but it means the results are not independently checkable. Choosing the sigmoid after seeing the same Helmholtz results adds a bit more selection bias.\n\nThis paper is for researchers working on PINNs for high-frequency or multiscale problems who want a possible cheap upgrade to HPKM. It deserves a serious referee, not a desk reject, because the idea is testable and the baselines are standard. But the referee should ask for fixed-α ablations, error bars, runtime, and code. I would not cite it as it stands.","headline":"Trainable sigmoid-bounded MLP/KAN weight is a plausible idea, but the paper never compares it to fixed-α tuning, and its cost claims contradict its own conclusion.","tokens_in":16957,"tokens_out":3572,"would_cite":false,"duration_ms":34012,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N99","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a trainable, sigmoid-bounded weight between a Fourier-KAN branch and an MLP branch, combined with overlapping domain decomposition, lets a physics-informed neural network solve high-frequency and multiscale PDEs more","keywords":["Physics-informed neural networks","Kolmogorov-Arnold networks","High-frequency multiscale problems","Domain decomposition","Trainable weighting parameter","Spectral bias","Fourier KAN","Hybrid architecture"],"falsifier":"On the high-frequency Helmholtz case (ω=32), record the trained S(α) over many seeds: the paper's mechanism predicts it drifts toward 1, favoring KAN, and that this drift is what produces the accuracy gain. A concrete falsifier is to freeze S(α) at 0.5 and rerun the same benchmarks: if accuracy does not degrade materially, the trainable gate is not the active ingredient. A second check is to replace the Fourier basis in the KAN branch with a static nonlinearity while retraining; if the final L2 errors are statistically unchanged, then the high-frequency claim is not being driven by the Fourier","tokens_in":16081,"feed_emoji":"🧮","tokens_out":6542,"duration_ms":58341,"temperature":0.7,"pith_summary":"The paper proposes a modified hybrid network for physics-informed PDE solving in which a Fourier-based Kolmogorov–Arnold branch and a multilayer perceptron branch are combined by a trainable weight, bounded to [0,1] by an S-shaped function, and the computational domain is split into overlapping subdomains. The central claim is that this automatically balances the KAN's high-frequency resolution with the MLP's low-frequency global structure, removing manual mixing-weight tuning and reducing training cost. The paper reports lower L2 errors than MLP-only or KAN-only PINNs across Helmholtz, two- and five-dimensional Poisson, reaction-diffusion, and Allen–Cahn benchmarks, while using 16 rather than 32 neurons per layer. A sympathetic reader would care because standard PINNs are known to underresolve high-frequency and multiscale solutions, and the paper offers a parameter-light architectural fix.","feed_headline":"Adaptive KAN-MLP blend beats single networks on oscillatory PDEs","feed_subtitle":"A single trainable mixing weight picks KAN for high frequencies, MLP for low ones — and beats both alone.","key_machinery":"The central object is the modified HPKM output u = S(α) u_KAN + (1 − S(α)) u_MLP, where u_KAN is a Fourier-KAN (a Kolmogorov–Arnold network whose univariate activations are truncated Fourier series) and u_MLP is a standard multilayer perceptron. S is an S-shaped map (sigmoid in the final experiments) that confines α to [0,1]; in the domain-decomposed version each subdomain has its own α^(i), and a partition-of-unity window function ω_i blends the local modified-HPKM outputs into a global solution. The mechanism doing the work is the adaptive convex gate: it is trained jointly with the network weights by gradient descent, so the model can tilt toward whichever branch fits the local frequency","core_discovery":"On the paper's own terms, the discovery is that replacing the fixed hyperparameter alpha in the hybrid parallel KAN–MLP architecture with a trainable parameter constrained by S(alpha) in [0,1] yields a self-balancing network: as training proceeds, S(alpha) moves toward the MLP for low-frequency problems and toward the KAN for high-frequency problems. Combined with overlapping domain decomposition and hard-constrained boundary and initial conditions, this modified HPKM-PINN reproduces oscillatory solutions such as sin(32πx)sin(32πy), and in Table 1 it reports the best L2 error among the three architectures on every benchmark considered — for example, Helmholtz with ω=48 at 8.698e-4 versus 1.1","pith_inferences":["Editorial inference: If the learned S(α) is the mechanism, then the same gate idea could be applied at the level of loss terms or frequency bands rather than only subdomains; the paper provides no test of whether per-subdomain weights are essential or merely convenient.","Editorial inference: The paper's benchmark list is separable and low-dimensional; whether the 16-neuron hybrid retains its advantage on non-separable, three-dimensional, or time-critical problems is an untested extrapolation.","Editorial inference: The reaction-diffusion result, where MLP and KAN both fail at about 0.92 while the hybrid reaches 1e-4, suggests the two branches' errors are not just averaged but cancel or complement; a controlled study varying branch widths could show whether the benefit is from the gate or from having two different inductive biases at all."],"forward_implications":["The mixing ratio between KAN and MLP no longer needs manual grid search; training sets it, and the S-shaped bound keeps it a valid convex combination.","On the reported benchmarks, the hybrid delivers smaller L2 error than either single-branch architecture under matched domain decomposition, so it can be read as a more accurate default for oscillatory PDEs.","Because the hybrid matches or beats 32-neuron baselines with 16 neurons per branch, the parameter count and memory footprint are smaller than a pure KAN of comparable width.","Input normalization inside each overlapping subdomain converts global high-frequency features into local lower-frequency ones, which the paper identifies as the reason spectral bias is mitigated.","The evolution of S(α) gives a diagnostic: the learned gate tilts to MLP on low-frequency problems and to KAN on high-frequency problems, supporting the frequency-separation narrative."],"fun_headline_variants":["Trainable KAN-MLP mix tames high-frequency PDEs","Self-tuning KAN-MLP outperforms fixed blends on oscillatory PDEs","Adaptive hybrid network solves high-freq PDEs with domain splitting","Trainable weight in KAN-MLP hybrid bests both alone","KAN-MLP with trainable mixing weight outperforms on oscillatory"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that one scalar weight, squeezed into [0,1], can keep the KAN and MLP branches cooperating so the KAN supplies high frequencies and the MLP supplies low frequencies; if the two branches interfere instead of complementing, the reported advantages could depend on the specific benchmark settings.","fun_headline_variants_meta":{"raw":{"variants":["Trainable KAN-MLP mix tames high-frequency PDEs","Self-tuning KAN-MLP outperforms fixed blends on oscillatory PDEs","Adaptive hybrid network solves high-freq PDEs with domain splitting","Trainable weight in KAN-MLP hybrid bests both alone","KAN-MLP with trainable mixing weight outperforms on oscillatory"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000923,"raw_usage":{"total_tokens":3762,"prompt_tokens":682,"completion_tokens":3080,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":426,"completion_tokens_details":{"reasoning_tokens":2986}},"tokens_in":426,"tokens_out":3080,"duration_ms":19845,"temperature":1.0,"reasoning_tokens":2986,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T22:14:03.020855+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the high-frequency Helmholtz case (ω=32), record the trained S(α) over many seeds: the paper's mechanism predicts it drifts toward 1, favoring KAN, and that this drift is what produces the accuracy gain. A concrete falsifier is to freeze S(α) at 0.5 and rerun the same benchmarks: if accuracy does not degrade materially, the trainable gate is not the active ingredient. A second check is to replace the Fourier basis in the KAN branch with a static nonlinearity while retraining; if the final L2 errors are statistically unchanged, then the high-frequency claim is not being driven by the Fourier","supporting_citations":[],"review_version":1}