{"id":"5c05e6a4-9338-4d95-901c-1e9bfb372c02","arxiv_id":"2505.15228","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CP-KAN reformulates polynomial degree selection in Chebyshev-based Kolmogorov-Arnold networks as a QUBO optimization problem and shows competitive regression performance with fewer parameters on several benchmarks.","lead":"The authors introduce CP-KAN, a neural network that uses Chebyshev polynomials and an optimization routine to automatically pick the right polynomial degree for each neuron. The approach is aimed at regression tasks with little data, such as financial time series, and claims competitive accuracy with far fewer parameters than standard deep models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 7's QUBO objective is separable across neurons and, without a complexity penalty, its per-neuron argmin always selects the maximum degree; the reported adaptive degree distributions imply an undocumented mechanism, so the claimed O(D^N)-to-one-step reduction is not supported.","rationale":"The reader's weakest assumption concerns the staleness of Phase 1 MSE costs after Phase 2 training. My check sharpens the same region: the Phase 1 objective is not merely heuristic, it is separable and monotone. Since Eq. 7 has no cross-neuron terms, the QUBO solves N independent argmin problems; calling this a reduction from exponential joint search is misleading. The nested structure of Chebyshev bases then implies that training MSE cannot increase with d, so without a complexity penalty every neuron gets D_max. The reported degree distributions therefore require an undocumented cost modification. This is internally inconsistent rather than merely overclaimed: the described algorithm cannot produce the presented results. If the concern lands, the paper's primary contribution—degree optimization via QUBO—collapses; the empirical regressions might survive as an architectural study, but the title and abstract promise a mechanism the paper does not actually specify. I therefore recommend REJECT rather than the reader's CONDITIONAL, or at minimum a revision that documents the true cost and compares against per-neuron argmin. This is in partial agreement with the reader: the per-neuron MSE proxy is the right place to look, but the more serious issue is that the proxy is separable and monotone, making the QUBO step superfluous.","tokens_in":17415,"tokens_out":9436,"duration_ms":90567,"concrete_test":"Using the released CP-KAN code, reproduce the Jane Street Phase 1 degree selection while logging the cost matrix. First, compare the QUBO-selected degrees to the per-neuron independent argmin_d MSE_{i,d} on the same training batch; if they coincide, the QUBO adds no coupling. Second, check whether the selected degrees are all equal to D_max; if Fig. 6's varied distribution is reproduced instead, locate the additional complexity term in the code and verify it is absent from Eq. 7. If the argmin baseline matches CP-KAN-QUBO's final validation R² in Table 2, the central complexity-reduction claim is vacuous.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that the QUBO-based degree selection in Eq. 7 reduces degree assignment from O(D^N) to one optimization per layer. As written, Eq. 7 contains no terms coupling different neurons: the fidelity term sums over (i,d) with coefficient MSE_{i,d}, and the one-hot penalty sum_i (sum_d q_{i,d}-1)^2 expands only within neuron i. The objective therefore decomposes into N independent subproblems; for any feasible penalty the optimal assignment is q_{i,d*}=1 with d* = argmin_d MSE_{i,d}. This is O(ND), not a combinatorial search, so the O(D^N) baseline is a strawman and QUBO/simulated annealing is an unnecessary wrapper around per-neuron least-squares model selection. The problem is worse because the Chebyshev bases in Eq. 6 are nested: degree d includes all lower-degree terms, so unconstrained (or coefficient-bounded) least squares on the same training batch gives training MSE nonincreasing in d. Hence the argmin is always the maximum degree D. The paper's Figure 6 and Appendix I.2 show varied and lower-degree selections, which cannot arise from Eq. 7. The hyperparameter tables list a 'Complexity Weight' (Tables 6-8, 11) that is never defined in the objective; the experiments must be using an additional degree penalty or validation-based cost not present in the method description. The degree-selection mechanism credited for parameter efficiency is therefore either trivial (always D_max) or underspecified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CP-KAN, a Kolmogorov-Arnold network whose activation functions are cumulative Chebyshev polynomials, and proposes to select per-neuron polynomial degrees via a QUBO formulation solved by simulated annealing. The claimed contribution is that this reduces degree assignment from O(D^N) to a single optimization step per layer. The authors also report regression experiments on financial and house-price data, classification studies on MNIST/CIFAR-10/Covertype, and a theoretical discussion connecting Chebyshev expansions to Ornstein-Uhlenbeck processes. The empirical sections show competitive regression performance with relatively few parameters, and the appendix contains substantial implementation detail and an available code repository.","tokens_in":17723,"tokens_out":4423,"duration_ms":43251,"significance":"If the degree-selection mechanism worked as advertised, the paper would offer a practical way to adapt polynomial complexity per neuron, with a concrete benefit in low-data regression. The manuscript also provides a broad empirical comparison, transparent hyperparameter tables, runtime measurements, and a public implementation, which are useful strengths. However, the central methodological claim is not supported by the written objective: as stated, the QUBO problem is separable and, without an explicit degree penalty, always selects the maximum degree. This inconsistency affects the core of the paper, not just the presentation. The theoretical link to mean reversion is also asserted rather than derived for the actual network. With those points addressed, the empirical findings could still be of interest, but in its current form the paper overclaims both the optimization contribution and the theoretical foundation.","major_comments":[{"comment":"The QUBO objective in Eq. (7) has no term coupling different neurons. The fidelity term is a sum over (i,d) of MSE_{i,d} q_{i,d}, and the one-hot penalty sum_i (sum_d q_{i,d} - 1)^2 couples degrees only within each neuron i. The problem therefore decomposes into N independent subproblems, each solvable by scanning degrees in O(D). Moreover, the Chebyshev basis in Eq. (6) is nested, so for any fixed projection w,b the least-squares training MSE_{i,d} is nonincreasing in d. Absent any explicit degree penalty, the minimizing assignment is q_{i,D}=1 for every neuron. This cannot produce the varied degree distributions shown in Fig. 6 and Appendix I.2. The hyperparameter tables list a 'Complexity Weight' (Tables 6-8 and 11) that is never defined in Eq. (7) or in Appendix B; this appears to be the undocumented mechanism responsible for the reported adaptive degrees. The authors must define the actual objective, including any degree-complexity term and its weighting, and show how the selected degrees are obtained from it.","section":"Sec. 3.2, Eq. (7); Appendix B; Fig. 6"},{"comment":"The claimed complexity reduction from O(D^N) exhaustive search to 'a single optimization step per layer' is not supported. Since the stated objective is separable, the optimal solution is obtained by N independent O(D) comparisons after the per-neuron MSE costs are computed; no QUBO solver, simulated annealing, or combinatorial search is required. The O(D^N) baseline is therefore a strawman for the objective as written. If the authors intend a genuinely coupled problem, they need to introduce a coupling term (for example, a global degree budget or a layer-wise complexity constraint) and justify why the QUBO formulation is needed. If the problem is instead per-neuron degree selection, the combinatorial claims in the abstract and introduction should be removed and the contribution reframed accordingly.","section":"Sec. 1 and Sec. 3.2; Abstract"},{"comment":"The claim that CP-KAN shows 'robustness to input scales' and performs well on 'raw house prices' is contradicted by the preprocessing described in Appendix D.2, which states that the target is log-transformed with np.log1p() and that features are normalized with StandardScaler(). Table 1 reports 'Raw MSE' and 'Log MSE', but it is unclear what 'Raw' means given this preprocessing. The authors should either report results on genuinely unscaled targets and features, or revise the robustness claim to match the actual protocol. As written, the section overstates a result that the paper's own appendix does not support.","section":"Sec. 4.5 and Appendix D.2"},{"comment":"The theoretical justification is asserted rather than established. Theorem 1 in Appendix K states a standard generator decomposition for the Ornstein-Uhlenbeck SDE, but it is not applied to CP-KAN: no derivation shows that the CP-KAN architecture, its two-phase training, or its degree-selection rule inherits this expansion. Eq. (14) is presented as an error bound for the OU approximation, yet the constants, the norm, and the connection to the trained network are unspecified, and the sentence in the main text describes the bound in ways that do not follow from the appendix. The authors should either provide a formal statement linking the CP-KAN function class to these bounds, or present this material as heuristic motivation rather than 'theoretical analysis establishes connections'.","section":"Sec. 5.3 and Appendix K"}],"minor_comments":[{"comment":"The degree variable d is used both as the maximum degree in Eq. (6) and as the summation bound, but the per-neuron degree d_i is not explicitly defined in the architecture section; distinguish the neuron-specific degree from the global maximum D.","section":"Eq. (6) and surrounding text"},{"comment":"The 'Complexity Weight' hyperparameter is listed in multiple tables but is never defined in the objective; a short definition in Section 3.2 or Appendix B is needed for reproducibility.","section":"Tables 6-8, 11"},{"comment":"The QUBO threshold T and the fixed fallback degree 3 are not specified with concrete values or a procedure for choosing them; state how T is set and how the default degree interacts with the reported degree distributions.","section":"Algorithm 1, Appendix C"},{"comment":"The phrase 'can be found here Fig. 3a' is a broken citation; fix the cross-reference and ensure all figure references in Sections 4.4 and 4.5 point to the correct subfigures.","section":"Fig. 3a and Appendix J"},{"comment":"The main text says experiments focus on a specific trading window, while Appendix D.1 describes a 70/30 split over 200,000 rows; clarify whether the trading window is the same as the data subset and state the temporal ordering of the split.","section":"Appendix D.1 and Sec. 4.4"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the mismatch between Eq. (7) and the reported adaptive degree behavior. This is not a cosmetic problem: the central contribution is the degree-selection mechanism, and the objective as written cannot produce the claimed results. The authors could potentially fix this by defining the missing complexity term, re-running or re-reporting the degree distributions, and substantially revising the complexity claims. If they cannot, the contribution reduces to a per-neuron least-squares degree cap, which would not support the paper's main claims. I would therefore condition acceptance on a full resolution of this point."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the main thing you should know: the central theoretical claim doesn't survive contact with Equation (7). That objective has no terms coupling neurons—the fidelity sum is over (i,d) and the one-hot penalty is per neuron. So the QUBO decouples into N independent argmins. Because the Chebyshev bases are nested, the least-squares MSE is nonincreasing in degree, and the unconstrained argmin is always the maximum degree D. The paper's own Figure 6 and Appendix I.2 show varied, lower-degree selections, which cannot come from Equation (7) as written. The hyperparameter tables list a \"Complexity Weight\" that is never defined in the method. So the actual degree-selection mechanism used in the experiments is either trivial (always D_max) or underspecified.\n\nWhat the paper does well: the writing is clear, the code is available, and the authors test several optimization methods (IP, GA, greedy, QUBO) with a real grid search. The Jane Street results are competitive with far fewer parameters, and the classification limitations on Covertype are honestly reported. The idea of treating polynomial degree selection as a discrete optimization problem is a legitimate and timely direction.\n\nThe soft spots are not minor. Section 4.5 claims robustness \"without explicit preprocessing,\" but Appendix D.2 says features were StandardScaler-normalized and the target log1p-transformed. That is a direct contradiction and undermines the stated advantage. Appendix K quotes a generator bound for Ornstein-Uhlenbeck processes, but it is not applied to CP-KAN; it's an analogy, not a theoretical justification. And the O(D^N) complexity baseline is a strawman: the separable objective means the exact solution is per-neuron model selection, O(ND), and the QUBO/annealing wrapper adds nothing unless there is a coupling or penalty you haven't defined.\n\nIf the authors add the missing complexity term to the objective, describe how it couples neurons, and rerun the ablations against a fixed-degree baseline, the empirical claim might hold. As it stands, the paper deserves a serious referee because the question is relevant and the experiments are extensive, but it needs major revision before it should be accepted. I wouldn't cite it in its current form.","headline":"The paper's central thesis is not supported: Equation (7) is separable and would always pick the maximum degree, while the experiments must be using an unstated complexity penalty; still, the empirical work and the underlying question deserve a rigorous referee.","tokens_in":18274,"tokens_out":4416,"would_cite":false,"duration_ms":38795,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that CP-KAN reduces polynomial degree selection in Chebyshev KANs to a per-layer QUBO optimization, replacing exponential search, and that this yields competitive regression performance with far fewer parameters.","keywords":["Kolmogorov-Arnold networks","Chebyshev polynomials","QUBO","degree selection","simulated annealing","regression","financial time series","parameter efficiency"],"falsifier":"Take a trained CP-KAN, freeze its weights, recompute $\\mathrm{MSE}_{i,d}$ for every degree on the trained representations, and re-solve the QUBO. If the new degree set differs substantially from the Phase-1 set, and retraining from scratch with the new degrees gives no worse validation error, then the pre-training degree choice is not load-bearing. A comparison against random degree assignments of the same parameter count would isolate the QUBO's contribution.","tokens_in":17183,"feed_emoji":"🧮","tokens_out":10266,"duration_ms":83572,"temperature":0.7,"pith_summary":"The paper introduces cumulative polynomial Kolmogorov-Arnold networks (CP-KAN), a neural architecture that uses Chebyshev polynomials as the learnable univariate functions in each neuron. Its central claim is that choosing the right polynomial degree for every neuron—normally an exponential discrete search—can be rewritten as a quadratic unconstrained binary optimization (QUBO) problem and solved once per layer with simulated annealing. If true, this makes polynomial KANs practical, because the architecture can adapt its own complexity instead of requiring a fixed degree chosen by hand. The paper reports that on regression with limited data, CP-KAN reaches competitive accuracy with far fewer parameters than MLPs, recurrent networks, transformers, and other polynomial KAN baselines, and shows more stable training curves.","feed_headline":"One QUBO step replaces exponential degree search in KANs","feed_subtitle":"CP-KAN matches larger MLPs and sequence models on regression with about 10x fewer parameters and stable training.","key_machinery":"The load-bearing device is the QUBO degree-selection objective of Eq. (7), whose fidelity term uses per-neuron least-squares costs $\\mathrm{MSE}_{i,d}$ and whose penalty term enforces exactly one degree per neuron. Each neuron computes $f(x) = \\sum_{i=0}^d c_i T_i(w^\\top x + b)$, where $T_i$ are Chebyshev polynomials of the first kind generated by the recurrence $T_{n+1}(x)=2xT_n(x)-T_{n-1}(x)$, and the sum is cumulative up to the selected degree $d$. The QUBO formulation is what carries the argument: it converts an exponential combinatorial search into a quadratic binary program solvable by simulated annealing, and the resulting degree distribution is fixed before Phase 2 gradient training refines the continuous parameters. The coefficient-decay property of Chebyshev expansions for smooth functions is what justifies truncating at modest degrees.","core_discovery":"At the center of the paper is a constructive reduction. The discrete choice of a Chebyshev polynomial degree $d_i$ for each neuron $i$ is encoded with binary variables $q_{i,d}$, and the optimization objective is the QUBO in Eq. (7): minimize $\\sum_{i,d} \\mathrm{MSE}_{i,d} q_{i,d} + \\alpha \\sum_i (\\sum_d q_{i,d}-1)^2$. The cost $\\mathrm{MSE}_{i,d}$ is the per-neuron least-squares fit error of the cumulative Chebyshev transform $\\Phi_{i,d}$ against the layer's target, so all degree costs are precomputable before any gradient training. Solving this QUBO by simulated annealing assigns a degree to every neuron in a layer in one step, lowering the search complexity from $O(D^N)$ to one optimization pass per layer. The trained network, after gradient fine-tuning, is reported to reach validation $R^2$ of 0.0376 on the financial market prediction benchmark with 1,571-2,547 parameters, compared with 0.0296 for a tuned MLP and negative $R^2$ for other polynomial KAN baselines, and a raw-scale house price MSE of 0.29 versus 1823 for an MLP. The paper also argues that Chebyshev bases are theoretically suited to mean-reverting financial dynamics, using the generator expansion for the Ornstein-Uhlenbeck process and a bound with exponential mean-reversion and $\\log d / d$ approximation terms.","pith_inferences":["Editorial extension: the same QUBO framing applies to any basis with precomputable per-neuron least-squares costs, so the degree-selection machinery is portable to Legendre, Hermite, or Fourier polynomial networks, not just Chebyshev.","Editorial extension: because degrees are fixed before training, alternating degree re-selection with gradient steps is a natural test of whether the optimal degree changes during training; if it does, iterative QUBO re-solving could improve the architecture.","Editorial extension: the paper's ablation of the plain Chebyshev KAN used a different parameter budget and a fixed degree, so the cleanest test of the QUBO's value is a fixed-degree Chebyshev KAN matched to CP-KAN's parameter count; that isolation is left implicit.","Editorial extension: the QUBO formulation is hardware-agnostic, so the same matrices solved by simulated annealing could be submitted to quantum annealers, making CP-KAN a candidate for near-term quantum-assisted architecture search."],"forward_implications":["Degree selection in a polynomial KAN becomes a per-layer combinatorial optimization step, replacing $O(D^N)$ search with one QUBO solve; the paper reports near-linear scaling in network size and feasibility up to roughly 5,000 binary variables on CPU/GPU.","On the regression tasks tested, CP-KAN matches or beats MLPs, LSTMs, GRUs, and transformers with roughly an order of magnitude fewer parameters (1,571-2,547 versus 17,185-506,701), suggesting polynomial KANs can be parameter-efficient on tabular data.","Training stability is a reported advantage: CP-KAN maintains validation performance across extended epochs and learning rates from $10^{-5}$ to $10^{-2}$, while MLPs degrade on the financial task.","The theoretical link to mean-reverting processes gives a domain-specific justification: Chebyshev expansions approximate the Ornstein-Uhlenbeck generator with controlled error, so the architecture is a natural candidate for financial time-series modeling.","Classification remains a known limit: on the Covertype tabular benchmark, smooth polynomial bases are outperformed by tree ensembles, and image tasks only improve with width, so the paper's claim is specifically about regression."],"supporting_citations":[{"why":"Introduces the Kolmogorov-Arnold network architecture that CP-KAN extends with Chebyshev polynomials and degree optimization.","marker":"[Liu et al., 2024]"},{"why":"Supplies the representation theorem that motivates decomposing multivariate functions through univariate functions.","marker":"[Kolmogorov, 1956]"},{"why":"Establishes QUBO as the standard formulation for discrete optimization and annealing, which the paper adapts to degree selection.","marker":"[Yarkoni et al., 2022]"},{"why":"Provides the simulated annealing algorithm used to solve the QUBO in Phase 1.","marker":"[Kirkpatrick et al., 1983]"},{"why":"Supplies the Chebyshev coefficient decay result used to justify truncating the polynomial expansion at modest degrees.","marker":"[Majidian, 2017]"},{"why":"Provides the generator decomposition theorem connecting Ornstein-Uhlenbeck processes to Chebyshev expansions.","marker":"[Applebaum, 2007]"},{"why":"Supplies the error bound for Ornstein-Uhlenbeck approximation used in the paper's theoretical argument.","marker":"[Burtnyak and Malyts'ka, 2018]"},{"why":"Provides the financial market prediction dataset and the weighted $R^2$ metric used in the main experiment.","marker":"[Desai et al., 2024]"},{"why":"Provides the tabular regression and classification benchmarks and the tree-based baselines used for comparison.","marker":"[Grinsztajn et al., 2022]"}],"fun_headline_variants":["QUBO picks KAN polynomial degrees in one pass","Chebyshev KANs with QUBO-optimized degrees","KAN degree search shrinks to one QUBO step","One-shot degree optimization for KANs using QUBO"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that degrees chosen from per-neuron least-squares fits before any joint training remain the right structure after the whole network is trained; if that independence fails, the QUBO step is an expensive heuristic.","fun_headline_variants_meta":{"raw":{"variants":["QUBO picks KAN polynomial degrees in one pass","Chebyshev KANs with QUBO-optimized degrees","KAN degree search shrinks to one QUBO step","One-shot degree optimization for KANs using QUBO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000746,"raw_usage":{"total_tokens":3365,"prompt_tokens":1025,"completion_tokens":2340,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":2268}},"tokens_in":641,"tokens_out":2340,"duration_ms":15532,"temperature":1.0,"reasoning_tokens":2268,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:21:15.993241+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained CP-KAN, freeze its weights, recompute $\\mathrm{MSE}_{i,d}$ for every degree on the trained representations, and re-solve the QUBO. If the new degree set differs substantially from the Phase-1 set, and retraining from scratch with the new degrees gives no worse validation error, then the pre-training degree choice is not load-bearing. A comparison against random degree assignments of the same parameter count would isolate the QUBO's contribution.","supporting_citations":[],"review_version":1}