Pith. sign in

REVIEW 3 major objections 6 minor 44 references

A single complementary matrix gate lets quantum-inspired fast-weight models remember each parameter on its own timescale without losing stable, parallel updates.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-31 22:35 UTC pith:YKPJ7CSX

load-bearing objection Solid incremental methods paper: CMG is a clean, scan-compatible gate that beats scalar gating on QKAN fast programmers, with thorough bake-offs but rank-one gates and selected-arm tables as the real limits. the 3 major comments →

arxiv 2607.27945 v1 pith:YKPJ7CSX submitted 2026-07-30 quant-ph cs.AIcs.LG

Complementary Matrix-Gated QKAN Fast-Weight Programmers for Quantum Dynamics Forecasting

Kuo-Chung Peng , Samuel Yen-Chi Chen , Jiun-Cheng Jiang , Chen-Yu Liu , En-Jui Kuo , Yun-Yuan Wang , Tzung-Chi Huang , Prayag Tiwari
show 9 more authors
Chi-Sheng Chen Chun-Hua Lin Yu-Chao Hsu Tai-Yue Li Saif Al-Kuwari Simon See Kuan-Cheng Chen Nan-Yow Chen Hsi-Sheng Goan
This is my paper
classification quant-ph cs.AIcs.LG
keywords fast weight programmingquantum-inspired machine learningKolmogorov-Arnold networksQKANcomplementary matrix gatingsequence modelingquantum dynamics forecastingself-modulating updates
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Sequence models have to choose what to keep from the past and what to write next. In quantum-inspired fast-weight programmers built on QKANs, a single scalar gate forces every fast parameter to share one memory timescale, which limits how well the model can track complex dynamics. This paper replaces that broadcast gate with Complementary Matrix Gating (CMG): one low-rank sigmoid matrix retains the old fast state coordinate-wise, and its complement writes the new proposal. The update stays a bounded convex combination and can still be computed with parallel prefix scans, at the cost of only one modulation head. Across seven forecasting benchmarks and several sequence lengths, CMG is the most consistent improvement over scalar gating when the fast programmer includes a QKAN module. On multi-step forecasts of Jaynes–Cummings and transmon–resonator quantum dynamics, CMG keeps mean-squared error on the order of 0.001 or lower out to 16 steps and cuts error by at least 91% versus scalar-gated counterparts.

Core claim

Coordinate-wise complementary modulation is a stable and effective fast-state update for QKAN-based fast-weight programmers. CMG uses one sigmoid matrix gate so each fast-state coordinate can retain or overwrite independently, while preserving the bounded convex form and affine prefix-scan structure of scalar gating, and it yields the most consistent gains over scalar gating—especially when the fast programmer is QKAN-based—including multi-step quantum-dynamics forecasts with MSE of order 0.001 or lower and at least 91.2% relative improvement.

What carries the argument

Complementary Matrix Gating (CMG): a rank-one outer-product matrix from the slow programmer is passed through a sigmoid to form Gt; the update is Θt = Gt ⊙ Θt−1 + (1−Gt) ⊙ Δt. One gate matrix supplies both retention and write weights, keeping a per-coordinate convex combination and an associative affine recurrence suitable for parallel prefix scan.

Load-bearing premise

The slow programmer’s rank-one outer-product gates are expressive enough to give each fast-state coordinate a useful retain-versus-write schedule; if real memory needs richer structure than that factorization, the reported gains would not carry over.

What would settle it

On the same Jaynes–Cummings and transmon–resonator multi-step tasks (N=64, horizons 4/8/16), replace CMG’s rank-one gate with a scalar gate or with Only-new modulation and check whether mean-squared error stays near 10−3 or below with ≥91% improvement; if CMG no longer wins or errors jump by orders of magnitude, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • QKAN-based fast-weight models can give each fast parameter its own memory timescale without giving up parallel scan training or bounded updates.
  • Write-only element-wise modulation is a poor substitute: long-horizon quantum-dynamics forecasts need stable retention or complementary retain/write control.
  • Architectures whose fast programmer includes a QKAN module benefit most; classical-only fast programmers gain less consistently.
  • Direct multi-step expectation-value forecasting of open and closed cavity QED systems becomes practical at MSE ~10−3–10−4 for horizons up to 16.
  • Extending the same complementary matrix gate to structured spatio-temporal fast states is the natural path to multivariate physical and quantum-control forecasting.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same complementary-gate algebra could be dropped into non-quantum fast-weight or linear-recurrent stacks wherever a scalar forget gate is the bottleneck.
  • If rank-one gates prove too thin on high-dimensional fast states, a modest fixed-rank factorization would test whether the complementary idea scales without losing scan efficiency.
  • CMG’s convex per-coordinate mix may make it a safer default than unbounded old-state multipliers when deploying quantum-inspired sequence models on noisy hardware or long contexts.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The manuscript proposes Self-Modulating QKAN-based fast-weight programmers and, as the central rule, Complementary Matrix Gating (CMG): a single low-rank (rank-one) sigmoid matrix gate that retains the old fast state while its complement writes the new proposal. CMG is shown to preserve the scalar gate’s coordinate-wise bounded convex update (Eq. 2) and the associative affine prefix-scan form shared by all compared rules. Four update families (Only-new, Only-old, Full, CMG) are compared against scalar gating across four slow/fast backbone combinations (classical vs HQKAN), seven univariate forecasting benchmarks, five input lengths, and direct multi-step forecasting of Jaynes–Cummings and transmon–resonator dynamics from CUDA-Q. Empirically, CMG yields the most consistent single-step gains when the fast programmer includes a QKAN module, and on the two quantum-dynamics multi-step tasks it keeps MSE on the order of 10^{-3}–10^{-4} for H∈{4,8,16} with ≥91.2% relative improvement over scalar gating.

Significance. If the results hold under broader scrutiny, the paper supplies a clean, scan-compatible alternative to broadcast scalar gating for QKAN-FWPs and demonstrates large, reproducible gains on physically motivated quantum-dynamics forecasting. Strengths include: (i) explicit unrolling of CMG to a coordinate-wise convex combination and a shared affine prefix-scan algebra; (ii) a systematic 4×5 architecture–rule grid with five seeds, mean±std tables, win-rate and paired-scatter summaries, learning curves, and parameter/AULC accounting; (iii) direct (non-recursive) multi-step evaluation on CUDA-Q trajectories rather than only synthetic univariate series. The contribution is incremental relative to the authors’ gated and self-modulating QFWP/QKAN-FWP line, but the complementary single-head design and the quantum-dynamics multi-step evidence are concrete and useful for quantum-inspired sequence learning.

major comments (3)
  1. [§III.C–D, Abstract, Conclusion] §III.C–D and the abstract’s “coordinate-wise memory control” claim: CMG implements G_t=σ(M^g_t) with the rank-one factorization M^g_t=m^{g,P}_t(m^{g,Q}_t)^T, so gate entries are constrained to G_{pq}=σ(u_p v_q), not free per coordinate. Boundedness and scan algebra hold for any G∈[0,1]^{P×Q}, but all reported gains only support factorized complementary control. The paper should either (a) state this limitation prominently in the abstract/conclusion and retitle claims as “low-rank complementary modulation,” or (b) add a small ablation (rank-2 / dense gate head, matched parameter budget) on at least the two CUDA-Q tasks to show whether the rank-one restriction is load-bearing for the ≥91.2% multi-step gains.
  2. [§V.A, Tables II–V] §V.A and Table II: top arms are selected at N=16 per dataset, then re-evaluated at N∈{4,8,32,64} in Tables III–V. This post-hoc selection can inflate the apparent cross-length dominance of CMG (24/28 best). Please report, for the full grid without N=16 selection, (i) win rates / median Δ_rel stratified by architecture as in Fig. 2, and (ii) whether CMG remains best at N=64 under a fixed rule chosen a priori (e.g., always CMG QKANFWP / QKAN-QKANFWP). If selection is retained, mark Tables III–V explicitly as “selected-arm transfer” rather than primary evidence.
  3. [§V.B, Figs. 5–6, Conclusion] §V.B, Figs. 5–6: On multi-step Jaynes–Cummings and transmon–resonator, Only-old and Full reach the same 10^{-4}–10^{-5} MSE regime as CMG; Only-new fails. The distinctive benefit attributed to complementary write/retain coupling is therefore only partially isolated from generic bounded memory-side modulation. A short analysis (e.g., gate/modulator statistics, effective retention timescales per fast-state block, or a controlled “non-complementary two-gate” baseline with independent σ gates on old and new) is needed to justify the claim that complementarity—not merely old-branch control at single-head cost—is the decisive ingredient.
minor comments (6)
  1. [Abstract, §V.B] Abstract vs §V.B: abstract says MSE “on the order of 0.001 or lower”; body and Fig. 5 often report ~10^{-4}–10^{-5}. Align the order-of-magnitude wording.
  2. [§III.B, §III.D] Eq. (1) uses a scalar g_t broadcast over Θ; Eq. (2) is the coordinate-wise CMG unroll. A one-line remark that scalar gating is the special case G_t=g_t 1 would help readers connect the two.
  3. [§III.A, §IV.D] Table I “HQKAN” and Fig. 1 “QKAN layer” are clear, but the encoder/decoder widths and DARUAN depth/observable used in experiments are not specified in §IV. Add a short hyperparameter appendix or table.
  4. [§V.A, Fig. 2] Fig. 2(c) median improvement by family shows QKAN-FWP near zero; §V.A notes this only briefly. One sentence on why slow-QKAN/fast-classical benefits less would aid interpretation.
  5. [§II] Related work cites a dense 2026 self-citation cluster ([14],[19],[20], etc.). A short explicit delta paragraph (“vs gated QKAN-FWP [14]: matrix vs scalar gate; vs self-modulating QFWP [19],[20]: complementary single-head CMG on QKAN-FWP”) would help non-insider readers.
  6. [Abstract, §I, Tables II–V] Typographical/consistency: “mean-squared error(MSE)” missing space (abstract); “DatA Re-Uploading ActivatioN” capitalization is idiosyncratic; ensure “QKANFWP” vs “QKAN-FWP” naming is consistent in all table headers.

Circularity Check

1 steps flagged

No load-bearing circularity: CMG’s math is a self-contained construction, and the headline gains are held-out MSE comparisons, not tautologies of fitted inputs.

specific steps
  1. self citation load bearing [§II Related work; §III.B–C; baseline G-QKANFWP [14] in Tables II–V and Figs. 2–6]
    "While in ref. [14], scalar gates are stable and parameter-efficient, broadcasting a single retention/write coefficient to every fast-state coordinate... Following the self-modulating QFWP [19], we let the slow programmer emit low-rank-generated element-wise modulators... We bound the old-state modulation branch with tanh so each coordinate can retain, suppress, or sign-adjust memory without geometric amplification [20]."

    The scalar-gated and self-modulating comparison arms are defined by the authors’ own concurrent 2026 line ([14],[19],[20]). Win rates and Δ_rel are therefore partly relative to in-house baselines rather than independent external methods. This is mild and non-load-bearing: the multi-step absolute MSE figures and the CMG convex/scan derivation do not reduce to those citations, and the forecasting targets are external CUDA-Q / classical series.

full rationale

The paper’s central claims are (i) that Complementary Matrix Gating preserves coordinate-wise convex boundedness and affine prefix-scan form, and (ii) that this update yields lower test MSE than scalar gating on external single-step and CUDA-Q multi-step forecasting tasks. Claim (i) is derived directly from the CMG definition Θ_t = G_t ⊙ Θ_{t−1} + (1−G_t) ⊙ Δ_t (Eq. after §III.D; unrolled form Eq. 2): the nonnegative weights summing to a convex combination follow by algebra from G_t ∈ [0,1], not by importing a uniqueness theorem or renaming a fit. Claim (ii) is standard supervised evaluation—min–max normalized sliding windows, 80/20 chronological splits, five seeds, MSE on held-out horizons—against physical/synthetic series (SHM, Bessel, NARMA, DQC, Jaynes–Cummings, transmon–resonator), not a parameter fitted on a subset and re-reported as a prediction of a linearly related quantity. Self-citations to gated QKAN-FWP [14] and self-modulating QFWP [19],[20] supply architectural lineage and the scalar-gated baseline; that is normal incremental work and is not load-bearing in the sense that the MSE deltas would collapse if those citations were removed—the experiments still stand. No step reduces a claimed first-principles prediction to its own definition or fit. Score 1 only for the mild density of in-house baselines, not for any construction-level circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The paper is an empirical methods contribution. Load-bearing content is the CMG update definition plus experimental protocol choices; it inherits QKAN/DARUAN and FWP structure from prior work rather than re-deriving them. Free parameters are training and architecture knobs; axioms are standard optimization and simulation assumptions; the main invented entity is CMG itself.

free parameters (4)
  • Adam learning rate and training horizon = lr=1e-3, epochs=100, batch=4
    Fixed at 10^{-3}, 100 epochs, batch size 4 for all arms; not cross-validated. Final MSE and AULC comparisons depend on this shared schedule.
  • Low-rank (rank-one) modulation factorization dimensions P,Q and head widths
    Fast state reshaped to R^{P×Q}; gates are outer products of two affine heads. Exact latent widths of HQKAN encoder/QKAN/decoder are not tabulated as a full hyperparameter sheet, yet they set capacity and parameter ratios in Table VI.
  • Best-arm selection at sequence length N=16 = N=16 selection checkpoint
    Table II selects one arm per dataset at N=16, then Tables III–V report that arm across other N. This selection choice affects which models are highlighted as ‘selected’.
  • Benchmark physical constants (γ, g, χ, truncations, initial states) = e.g. γ=0.05, g=π, χ=0.025·2π GHz, cavity 5 / resonator 20 levels
    Open JC and dispersive transmon Hamiltonians and Fock truncations are chosen by hand to define the forecasting targets; results are conditional on these simulation settings.
axioms (5)
  • standard math Affine pairs (A_t,B_t) compose associatively under (A',B')∘(A,B)=(A'⊙A, A'⊙B+B'), so all studied updates admit parallel prefix scan.
    Invoked in §III.D to claim CMG preserves scan efficiency of scalar gating; standard parallel-prefix algebra [18].
  • domain assumption With proposals suitably bounded, coordinate-wise convex combinations keep each fast-state coordinate within the same bound as scalar gating.
    Stated after Eq. (2); inherits the bounded-proposal condition from gated QKAN-FWP [14].
  • domain assumption CUDA-Q Dynamics trajectories are faithful enough targets that lower MSE implies better modeling of the intended quantum dynamics.
    §IV.A uses simulated JC and transmon–resonator expectation values as ground truth without experimental device noise.
  • domain assumption Min–max normalization to [-1,1] and chronological 80/20 splits do not distort comparative ranking of update rules.
    §IV.A preprocessing protocol shared across all arms.
  • ad hoc to paper Old-branch multipliers should be bounded by tanh to avoid geometric amplification under recurrence.
    §III.C imposes fM^old=tanh(M^old) following the group’s stable self-modulating QFWP line [20]; design choice, not a theorem forced by the task.
invented entities (2)
  • Complementary Matrix Gating (CMG) no independent evidence
    purpose: Provide element-wise retain/write control with one low-rank sigmoid gate and its complement, matching single-branch head cost while keeping convex bounded updates.
    Core proposed rule Θ_t = G_t⊙Θ_{t-1} + (1-G_t)⊙Δ_t with G_t=σ(M^g_t). Independent evidence is only the paper’s own benchmarks, not an external measurement.
  • Self-Modulating QKAN-based FWP family (Only-new / Only-old / Full / CMG on FWP, QKANFWP, QKAN-FWP, QKAN-QKANFWP) no independent evidence
    purpose: Systematically attach low-rank modulators to classical and HQKAN slow/fast programmers.
    Architectural product of prior QKAN-FWP and self-modulating QFWP lines; treated as the experimental universe rather than an externally validated object.

pith-pipeline@v1.2.0-daily-grok45 · 21464 in / 4053 out tokens · 83048 ms · 2026-07-31T22:35:08.058098+00:00 · methodology

0 comments
read the original abstract

Sequence models must decide what to write into memory and what to retain. In quantum and quantum-inspired sequence learning, nonlinear recurrent updates often require repeated circuit evaluations and sequential backpropagation through time, making long contexts costly. Gated fast-weight programmers (FWPs) based on quantum-inspired Kolmogorov-Arnold networks (QKANs) alleviate this bottleneck by storing context in time-varying fast parameters. However, their scalar gate applies one retention-write balance to every fast-state coordinate, forcing all parameters to share a memory timescale. We introduce Self-Modulating QKAN-based FWPs, which replace this broadcast gate with low-rank-generated element-wise modulation of the new-proposal branch, a bounded old-state branch, or both. We further propose Complementary Matrix Gating (CMG), which uses one sigmoid matrix gate to retain the old state and its complement to write the new proposal. CMG provides coordinate-wise memory control while preserving the bounded convex update and affine prefix-scan structure of scalar gating, at the modulation-head cost of a single-branch rule. We compare four self-modulating rules with scalar gating across four FWP architectures combining classical and QKAN-based slow and fast programmers. Across seven single-step forecasting benchmarks and five sequence lengths, CMG gives the most consistent improvements for architectures whose fast programmer incorporates a QKAN-based module. In direct multi-step forecasting of Jaynes-Cummings and transmon-resonator dynamics simulated with CUDA-Q Dynamics, CMG models maintain mean-squared errors on the order of 0.001 or lower across forecasting horizons of 4, 8, and 16 steps, while improving on their scalar-gated counterparts by at least 91.2%. These results establish coordinate-wise complementary modulation as a stable and effective update for QKAN-based FWPs.

Figures

Figures reproduced from arXiv: 2607.27945 by Chen-Yu Liu, Chi-Sheng Chen, Chun-Hua Lin, En-Jui Kuo, Hsi-Sheng Goan, Jiun-Cheng Jiang, Kuan-Cheng Chen, Kuo-Chung Peng, Nan-Yow Chen, Prayag Tiwari, Saif Al-Kuwari, Samuel Yen-Chi Chen, Simon See, Tai-Yue Li, Tzung-Chi Huang, Yu-Chao Hsu, Yun-Yuan Wang.

Figure 1
Figure 1. Figure 1: Self-modulating update rules and programmer backbones. (a) Full [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Performance of the self-modulating update rules relative to their paired gated baselines. (a) Win rate by sequence length, aggregated across datasets and [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Final test MSE of each self-modulating rule versus its paired gated baseline across all datasets, sequence lengths, and model families: (a) Only-new, (b) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 6
Figure 6. Figure 6: Relative MSE improvement over the paired gated baseline [ Eq. (3)] [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 4
Figure 4. Figure 4: Test MSE versus training epoch for QKANFWP at [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Mean test MSE for direct multi-step forecasting with QKANFWP [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 11 linked inside Pith

  1. [1]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  2. [2]

    Learning to forget: Continual prediction with LSTM,

    F. A. Gers, J. Schmidhuber, and F. Cummins, “Learning to forget: Continual prediction with LSTM,” Neural computation, vol. 12, no. 10, pp. 2451–2471, 2000

    2000

  3. [3]

    Learning phrase representations using rnn encoder–decoder for statistical machine translation,

    K. Cho, B. Van Merriënboer, Ç. Gulçehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio, “Learning phrase representations using rnn encoder–decoder for statistical machine translation,” in Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), 2014, pp. 1724–1734

    2014

  4. [4]

    Gated feedback recurrent neural networks,

    J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Gated feedback recurrent neural networks,” in International conference on machine learning . PMLR, 2015, pp. 2067–2075

    2015

  5. [5]

    Recurrent quantum neural networks,

    J. Bausch, “Recurrent quantum neural networks,” Advances in neural information processing systems , vol. 33, pp. 1368–1379, 2020

    2020

  6. [6]

    Quantum long short- term memory,

    S. Y .-C. Chen, S. Yoo, and Y .-L. L. Fang, “Quantum long short- term memory,” in Icassp 2022-2022 IEEE international conference on acoustics, speech and signal processing (ICASSP) . IEEE, 2022, pp. 8622–8626

    2022

  7. [7]

    Learning temporal data with a variational quantum recurrent neural network,

    Y . Takaki, K. Mitarai, M. Negoro, K. Fujii, and M. Kitagawa, “Learning temporal data with a variational quantum recurrent neural network,” Physical Review A , vol. 103, no. 5, p. 052414, 2021

    2021

  8. [8]

    On the difficulty of training recurrent neural networks,

    R. Pascanu, T. Mikolov, and Y . Bengio, “On the difficulty of training recurrent neural networks,” in International conference on machine learning. Pmlr, 2013, pp. 1310–1318

    2013

  9. [9]

    Evaluating analytic gradients on quantum hardware,

    M. Schuld, V . Bergholm, C. Gogolin, J. Izaac, and N. Killoran, “Evaluating analytic gradients on quantum hardware,” Physical Review A, vol. 99, no. 3, p. 032331, 2019

    2019

  10. [10]

    Learning to control fast-weight memories: An alterna- tive to dynamic recurrent networks,

    J. Schmidhuber, “Learning to control fast-weight memories: An alterna- tive to dynamic recurrent networks,” Neural Computation, vol. 4, no. 1, pp. 131–139, 1992

    1992

  11. [11]

    Using fast weights to attend to the recent past,

    J. Ba, G. E. Hinton, V . Mnih, J. Z. Leibo, and C. Ionescu, “Using fast weights to attend to the recent past,” Advances in neural information processing systems, vol. 29, 2016

    2016

  12. [12]

    Linear transformers are secretly fast weight programmers,

    I. Schlag, K. Irie, and J. Schmidhuber, “Linear transformers are secretly fast weight programmers,” in International conference on machine learning. PMLR, 2021, pp. 9355–9366

    2021

  13. [13]

    Learning to program variational quantum circuits with fast weights,

    S. Y .-C. Chen, “Learning to program variational quantum circuits with fast weights,” in 2024 International Joint Conference on Neural Networks (IJCNN). IEEE, 2024, pp. 1–9

    2024

  14. [14]

    Gated QKAN-FWP: Scalable quantum-inspired sequence learning,

    K.-C. Peng, S. Y .-C. Chen, J.-C. Jiang, C.-Y . Liu, E.-J. Kuo, Y .-Y . Wang, P. Tiwari, A. Ceschini, C.-S. Chen, Y .-C. Hsu, C.-H. Lin, T.-Y . Li, A. Rosato, M. Panella, S. See, S. Al-Kuwari, K.-C. Chen, N.-Y . Chen, and H.-S. Goan, “Gated QKAN-FWP: Scalable quantum-inspired sequence learning,” 2026. [Online]. Available: https://arxiv.org/abs/2605.06734

    Pith/arXiv arXiv 2026

  15. [15]

    Quantum variational activation functions empower Kolmogorov-Arnold networks,

    J.-C. Jiang, M. Y .-C. Huang, T. Chen, and H.-S. Goan, “Quantum variational activation functions empower Kolmogorov-Arnold networks,” arXiv preprint arXiv:2509.14026 , 2025

    Pith/arXiv arXiv 2025

  16. [16]

    KAN: Kolmogorov–Arnold networks,

    Z. Liu, Y . Wang, S. Vaidya, F. Ruehle, J. Halverson, M. Soljacic, T. Hou, and M. Tegmark, “KAN: Kolmogorov–Arnold networks,” inInternational conference on learning representations , vol. 2025, 2025, pp. 70 367– 70 413

    2025

  17. [17]

    Data re-uploading for a universal quantum classifier,

    A. Pérez-Salinas, A. Cervera-Lierta, E. Gil-Fuster, and J. I. Latorre, “Data re-uploading for a universal quantum classifier,” Quantum, vol. 4, p. 226, 2020

    2020

  18. [18]

    Prefix sums and their applications,

    G. E. Blelloch, “Prefix sums and their applications,” 1990

    1990

  19. [19]

    Self-modulating quantum fast-weight programmers for efficient adaptive sequential learning,

    S. Y .-C. Chen, Y . Peng, K.-C. Peng, J.-C. Jiang, C.-H. Lin, J. J. Park, H.- H. Tseng, H.-Y . Lin, K.-C. Chen, C.-Y . Liu, and S. Yoo, “Self-modulating quantum fast-weight programmers for efficient adaptive sequential learning,” 2026. [Online]. Available: https://arxiv.org/abs/2606.24933

    Pith/arXiv arXiv 2026

  20. [20]

    Stable self-modulating quantum fast-weight programmers with bounded memory gates,

    K.-C. Peng, J.-C. Jiang, C.-H. Lin, Y . Peng, J. J. Park, H.-H. Tseng, H.-Y . Lin, K.-C. Chen, C.-Y . Liu, S. Yoo, and S. Y .-C. Chen, “Stable self-modulating quantum fast-weight programmers with bounded memory gates,” 2026. [Online]. Available: https://arxiv.org/abs/2607.02363

    Pith/arXiv arXiv 2026

  21. [21]

    Cuda quantum: The platform for integrated quantum- classical computing,

    J.-S. Kim et al., “Cuda quantum: The platform for integrated quantum- classical computing,” in 2023 60th ACM/IEEE Design Automation Conference (DAC). IEEE, 2023, pp. 1–4

    2023

  22. [22]

    Harnessing disordered-ensemble quantum dynamics for machine learning,

    K. Fujii and K. Nakajima, “Harnessing disordered-ensemble quantum dynamics for machine learning,” Physical Review Applied , vol. 8, no. 2, p. 024030, 2017

    2017

  23. [23]

    Quantum recurrent neural networks for sequential learning,

    Y . Li, Z. Wang, R. Han, S. Shi, J. Li, R. Shang, H. Zheng, G. Zhong, and Y . Gu, “Quantum recurrent neural networks for sequential learning,” Neural Networks, vol. 166, pp. 148–161, 2023

    2023

  24. [24]

    Quantum gated recurrent neural networks,

    Y . Li, Z. Wang, R. Xing, C. Shao, S. Shi, J. Li, G. Zhong, and Y . Gu, “Quantum gated recurrent neural networks,” IEEE transactions on pattern analysis and machine intelligence , vol. 47, no. 4, pp. 2493–2504, 2024

    2024

  25. [25]

    Reinforcement learning for quantum circuit optimization: A review,

    A. Kundu, A. Sarkar, P. Tiwari, and S. Feld, “Reinforcement learning for quantum circuit optimization: A review,” Openreview, 2026

    2026

  26. [26]

    Kolmogorov-Arnold networks for time series: Bridging predictive power and interpretability,

    K. Xu, L. Chen, and S. Wang, “Kolmogorov-Arnold networks for time series: Bridging predictive power and interpretability,” arXiv preprint arXiv:2406.02496, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Are KANs effective for multivariate time series forecasting?

    X. Han, X. Zhang, Y . Wu, Z. Zhang, and Z. Wu, “Are KANs effective for multivariate time series forecasting?” arXiv preprint arXiv:2408.11306 , 2024

    Pith/arXiv arXiv 2024

  28. [28]

    TimeKAN: KAN-based frequency decomposition learning architecture for long-term time series forecasting,

    S. Huang, Z. Zhao, C. Li, and L. Bai, “TimeKAN: KAN-based frequency decomposition learning architecture for long-term time series forecasting,” arXiv preprint arXiv:2502.06910 , 2025

    Pith/arXiv arXiv 2025

  29. [29]

    A practitioner’s guide to Kolmogorov–Arnold networks,

    A. Noorizadegan, S. Wang, L. Ling, and J. P. Dominguez-Morales, “A practitioner’s guide to Kolmogorov–Arnold networks,” Computer Science Review , vol. 62, p. 100991, 2026. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S1574013726000997

    2026

  30. [30]

    QKAN-LSTM: Quantum-inspired Kolmogorov–Arnold long short-term memory,

    Y .-C. Hsu, J.-C. Jiang, C.-H. Lin, K.-C. Peng, N.-Y . Chen, S. Y .-C. Chen, E.-J. Kuo, and H.-S. Goan, “QKAN-LSTM: Quantum-inspired Kolmogorov–Arnold long short-term memory,” in 2026 International Conference on Quantum Communications, Networking, and Computing (QCNC). IEEE, 2026, pp. 650–659

    2026

  31. [31]

    Parameter-efficient quantum-inspired fast weight programmers for traffic-matrix forecasting,

    K.-C. Peng, J.-C. Jiang, C.-H. Lin, T.-Y . Li, N.-Y . Chen, and S. Y .-C. Chen, “Parameter-efficient quantum-inspired fast weight programmers for traffic-matrix forecasting,” 2026. [Online]. Available: https://arxiv.org/abs/2606.27821

    Pith/arXiv arXiv 2026

  32. [32]

    Generative quantum-inspired Kolmogorov-Arnold eigensolver,

    Y .-C. Lin, Y .-C. Hsu, I.-S. Tsai, C.-H. Lin, K.-C. Peng, J.-C. Jiang, Y .-Y . Wang, T.-C. Huang, T.-Y . Li, K.-C. Chen, S. Y .-C. Chen, and N.-Y . Chen, “Generative quantum-inspired Kolmogorov-Arnold eigensolver,”

  33. [33]

    Gated fast weights for on-the-fly neural program generation,

    I. Schlag and J. Schmidhuber, “Gated fast weights for on-the-fly neural program generation,” in NIPS Metalearning Workshop, 2017

    2017

  34. [34]

    Simple recurrent units for highly parallelizable recurrence,

    T. Lei, Y . Zhang, S. I. Wang, H. Dai, and Y . Artzi, “Simple recurrent units for highly parallelizable recurrence,” in Proceedings of the 2018 conference on empirical methods in natural language processing , 2018, pp. 4470–4481

    2018

  35. [35]

    Recurrent neural networks for multivariate time series with missing values,

    Z. Che, S. Purushotham, K. Cho, D. Sontag, and Y . Liu, “Recurrent neural networks for multivariate time series with missing values,” Scientific reports, vol. 8, no. 1, p. 6085, 2018

    2018

  36. [36]

    Eleatt-rnn: Adding attentiveness to neurons in recurrent neural networks,

    P. Zhang, J. Xue, C. Lan, W. Zeng, Z. Gao, and N. Zheng, “Eleatt-rnn: Adding attentiveness to neurons in recurrent neural networks,” IEEE Transactions on Image Processing , vol. 29, pp. 1061–1073, 2019

    2019

  37. [37]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke et al., “Pytorch: An imperative style, high-performance deep learning library,” 2019

    2019

  38. [38]

    CuTe layout representation and algebra,

    C. Cecka, “CuTe layout representation and algebra,” 2026. [Online]. Available: https://arxiv.org/abs/2603.02298

    Pith/arXiv arXiv 2026

  39. [39]

    QKAN: Quantum-inspired Kolmogorov-Arnold network,

    J.-C. Jiang, “QKAN: Quantum-inspired Kolmogorov-Arnold network,”

  40. [40]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014

    Pith/arXiv arXiv 2014

  41. [41]

    The shape of learning curves: a review,

    T. Viering and M. Loog, “The shape of learning curves: a review,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 6, pp. 7799–7819, 2022

    2022

  42. [42]

    Active learning in the presence of unlabelable examples,

    D. Mazzoni and K. Wagstaff, “Active learning in the presence of unlabelable examples,” in European Conference on Machine Learning , 2004

    2004

  43. [2025]

    Available: https://github.com/Jim137/qkan

    [Online]. Available: https://github.com/Jim137/qkan

  44. [2026]

    Available: https://arxiv.org/abs/2605.04604

    [Online]. Available: https://arxiv.org/abs/2605.04604

    Pith/arXiv arXiv