Pith. sign in

REVIEW 3 major objections 5 minor 34 references

A single shunted superconducting wire acts as a spiking neuron, and three such wirelets classify handwritten digits at 92.9% accuracy.

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

2026-08-02 23:12 UTC pith:IOI3DGI2

load-bearing objection A credible experimental spiking wirelet neuron, but the 92.9% MNIST number is simulation-only and should be read as a proposal, not a measured hardware capability. the 3 major comments →

arxiv 2602.14330 v2 pith:IOI3DGI2 submitted 2026-02-15 cond-mat.supr-con

Pattern recognition with superconducting wirelet neurons

Khalil Harrabi , Leonardo Cadorim , Milorad Milosevic This is my paper
classification cond-mat.supr-con
keywords superconducting wireletspiking neuronintegrate-and-fireneuromorphic computingtime-dependent Ginzburg-Landaupattern recognitionMNISTresistive shunt
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.

The paper argues that a resistively shunted superconducting wirelet—a micrometer-scale filament with a parallel resistor—behaves as a minimal integrate-and-fire neuron: it fires only above a critical current, spikes faster as bias current rises, has a current-dependent refractory period, and stops firing entirely at high current. This claim matters because it would give superconducting neuromorphic hardware a far simpler building block than Josephson-junction circuits, with tunable spike properties and low energy per event. To show utility, the authors train a linear readout on the full time-dependent voltage traces of just three wirelets and report 92.9% accuracy on MNIST digits, with 100% on a reduced 3x3 task in both simulation and experiment. The work also sketches how gated wirelets could implement trainable synaptic weights on-chip, unifying inference and training in the same superconducting layer.

Core claim

A single superconducting filament shunted by a resistor exhibits all the hallmarks of a biological spiking neuron: a threshold set by the critical current, current-dependent firing frequency, a refractory period tied to the hot-spot delay time, and a 'dead' regime where a stable phase-slip line suppresses spiking. These properties emerge from the periodic formation and relaxation of hot spots that divert current into the shunt. Because the temporal voltage waveform encodes the input current sequence, the authors use the complete time trace—not just spike counts—as the neuron's output, and train a linear cross-entropy readout on it. With only three such neurons they report 92.9% accuracy on a

What carries the argument

The resistively shunted superconducting wirelet: a micrometer-scale superconducting stripe in parallel with a tunable resistor, governed by time-dependent Ginzburg-Landau equations coupled to a circuit equation for kinetic inductance and shunt. The shunt creates relaxation oscillations between the superconducting and resistive (hot-spot) states, producing voltage spikes whose timing depends on applied current, temperature, and shunt resistance. The training procedure treats the discretized voltage waveform of each neuron as a high-dimensional feature vector, and a softmax-cross-entropy linear layer maps it to digit classes.

Load-bearing premise

The 92.9% MNIST result rests on simulated time-dependent Ginzburg-Landau voltage traces, and the experimental demonstration at 100% accuracy used only a simpler 3x3 task with a software-trained readout; if simulated traces misrepresent the real devices' temporal output under MNIST-scale inputs, the headline accuracy would not transfer to hardware.

What would settle it

Run the MNIST task using experimentally recorded voltage traces from three shunted NbTiN wirelets instead of simulated ones; if accuracy drops to near chance or far below 92.9%, the simulated temporal responses are not faithful. A simpler check: compare predicted and measured spike arrival times under identical pulse trains and see whether the discrepancies are smaller than the time bins used in training.

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

If this is right

  • If the claims hold, a single shunted filament is sufficient to reproduce neuron-like dynamics, so superconducting neuromorphic hardware no longer needs multi-element Josephson circuits.
  • Spiking frequency, threshold, and refractory time are tunable by bias current, temperature, and shunt resistance, giving a physical knob for analog encoding.
  • Temporal (rather than rate-only) readout can extract classification power from the full voltage waveform; three neurons suffice for a nontrivial MNIST task, implying the wirelet's nonlinear dynamics do substantial feature mapping.
  • The proposed gated-wirelet synapse would allow on-chip training, removing the software-readout bottleneck and enabling fully cryogenic learning hardware.
  • Sub-picojoule per-synaptic-event energies, if realized in scaled devices, would place superconducting wirelets among the most energy-efficient neuromorphic platforms.

Where Pith is reading between the lines

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

  • The MNIST accuracy is achieved with a linear readout on simulated traces; a natural testable extension is whether the same accuracy survives hardware noise and drift, since experimentally recorded traces were only validated on the 3x3 task.
  • The temporal-processing idea may transfer to other relaxation oscillators (e.g., nanowire or photonic) whose voltage traces are rich enough to carry class information, suggesting the key insight is not the superconductor but the time-domain readout.
  • The scaling argument suggests adding more wirelets could push accuracy well beyond 92.9%, but the linear readout also becomes higher-dimensional; whether hardware analog noise saturates the gain is an open question.
  • Because the readout weight matrix is time-resolved, the same network can, in principle, distinguish inputs that share the same integrated pulse train but differ in timing—a property worth probing directly.

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 / 5 minor

Summary. The paper reports that a single resistively shunted superconducting filament ('wirelet') can act as an integrate-and-fire spiking neuron. Experimental NbTiN whiskers show threshold behavior, current-controlled firing frequency, refractory delay, and a death regime, and TDGL simulations reproduce these regimes. For pattern recognition, three wirelets encode pixel intensities as sequential current pulses; a software-trained linear readout on the concatenated voltage traces achieves 100% accuracy on experimental 3x3 images (1000 test images per digit) and, in simulation, 92.9% on 22x20 MNIST. The paper also sketches an on-chip training scheme using gated output wirelets.

Significance. The experimental demonstration of a single-element superconducting spiking neuron with tunable threshold, frequency, and death is a solid and useful contribution. The qualitative agreement between the TDGL simulations and the measured spiking-frequency and death behavior adds credibility to the device model. The 3x3 pattern-recognition experiment with recorded traces is a valid proof-of-concept for the temporal-readout idea, even though the readout is software-based. However, the headline MNIST result is entirely simulation-based: the voltage traces used for training and testing are generated by the authors' own TDGL dynamics, and the linear readout has 1.87M parameters trained on 30k images. Hardware transferability of that result is not established. The gated-wirelet on-chip training scheme is an untested proposal. These issues are addressable by recalibrating the claims and adding robustness evidence, so the paper has value but needs revision.

major comments (3)
  1. [Fig. 5 and Methods 'Synaptic training', Eq. (4)] The 92.9% MNIST accuracy is obtained entirely from simulated voltage traces generated by Eqs. (1)-(3), not from measured wirelet outputs. The concatenated waveform V has 3N = 187,440 elements, so W is a 187,440 x 10 matrix with ~1.87M trainable parameters, trained on 30k images. Such a high-dimensional linear readout can exploit simulator-specific spike timing, ringing, or discretization artifacts that would not survive in noisy hardware. The experimental 3x3 result does not validate this transfer because it uses only 9 distinct, well-separated pixel patterns and an off-chip readout. Please state explicitly that Fig. 5 is a simulation-only demonstration and provide evidence of robustness, e.g., noise-injected traces or an experimental waveform-based test.
  2. [Fig. 4c and 'Pattern recognition with few superconducting neurons'] The experimental neural-network operation is a software readout on recorded voltage traces; the synaptic weights are not implemented in hardware. The figure caption acknowledges 'synaptic multiplexing was software-based,' but the abstract and conclusion state 'demonstrated pattern recognition' without this qualification. This is not a flaw in the raw measurements, but it means the hardware-level neuromorphic claim is currently a proof-of-concept with off-chip training, not an embedded implementation.
  3. [On-chip training with superconducting wirelet neurons] The gated output wirelet is presented as the synaptic element for on-chip training, but no experiment or simulation is shown to support the assumption that an electrostatically gated wirelet can implement a time-dependent multiplicative weight on the aggregate waveform. The claim that 'This gated superconducting synaptic stage enables compact, energy-efficient, and reconfigurable functionality' is therefore speculative. If this is kept as a contribution, it should be explicitly labeled as a proposal, or supported by a feasibility simulation of the gated-wirelet response.
minor comments (5)
  1. [Methods, Eq. (1)] The term printed as '\gamma^2/2 \partial|\psi|^2/\partial t^2' appears to be a typo for '\gamma^2/2 \partial|\psi|^2/\partial t'. As printed, the equation is dimensionally inconsistent. Please correct and confirm the simulated equation.
  2. [Methods, Eq. (3)] The shunt resistance is denoted Rsh in the main text but Rs in Eq. (3); please use consistent notation.
  3. [Fig. 5b caption] The caption refers to '420 image pixels' for a 22 x 20 MNIST image, which has 440 pixels. Please reconcile the numbers.
  4. [Discussion, energy efficiency] The energy estimates (100 fJ spike generation, ~1 pJ voltage-to-current conversion, <0.5 pJ per synaptic event) are stated without a supporting calculation or measurement. Please specify the assumptions (e.g., voltage amplitude, pulse duration, impedance) or clearly label these as order-of-magnitude estimates.
  5. [General] The manuscript does not include a data/code availability statement. Given the central role of TDGL simulations, making the simulation code and processed experimental traces available would strengthen reproducibility.

Circularity Check

0 steps flagged

No circularity: the classification results use held-out test data with a forward simulation model, not a fitted-input prediction.

full rationale

The derivation chain is not circular. The wirelet neuron properties (threshold, refractory period, firing frequency, death) are produced by the TDGL model of Eqs. (1)-(3) and cross-validated against the experimental measurements in Fig. 2 and Extended Data. The pattern-recognition pipeline is a standard supervised linear readout y = VW (Eq. 4), with weights trained on labeled images and evaluated on held-out test images. The weights are not fit to the test labels, and the simulated voltage traces used for the MNIST task are generated by the forward TDGL model with stated parameter values, not constructed from the classifier. Thus the 92.9% MNIST figure is a simulated forward-model demonstration, and its lack of hardware validation is a generalizability limitation rather than a circular reduction. The self-citations (Refs. 13, 33, 34) concern methods and delay-time characterization, but the load-bearing spiking phenomena are also directly measured in this paper, so they are not load-bearing self-citations. No step reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 1 invented entities

The central free parameter is the trained linear readout, which is legitimate but means the 'neural network operation' is a learned projection of the temporal response, not an emergent hardware computation. The simulations use standard TDGL parameters with gamma set by hand. The gated wirelet synapse is an in-principle proposal without experimental support.

free parameters (4)
  • TDGL parameter gamma = 10
    Inelastic electron-phonon scattering time parameter in the TDGL equation; stated as 'without loss of generality, we used γ=10'. It affects spiking dynamics and is chosen by hand.
  • TDGL parameter u = 5.79
    Standard microscopic TDGL constant (related to inelastic scattering) used in the simulations.
  • Software readout weights W = 3N x 10 matrix
    Trained via SGD on simulated/experimental traces; this is the fitting that produces the classification accuracy. This is a fitting parameter, not a prediction, although it is a legitimate trainable readout.
  • Shunt resistances for the 3-neuron network = 0.5, 0.75, 1.0 RGL (sim); 0.3, 1.0, 1.7 Ohm (exp)
    Chosen to provide different responses; they are not fitted to the classification but are design parameters that influence the results.
axioms (3)
  • domain assumption Time-dependent Ginzburg-Landau equations (Eqs. 1-3) accurately model the spatio-temporal dynamics of the NbTiN wirelet and the shunted circuit.
    The entire MNIST classification simulation rests on this. TDGL is a well-established mesoscopic model, but microsecond-scale hot-spot dynamics with strong heating may go beyond its validity; the authors do not quantitatively validate the simulated MNIST traces against experiment.
  • domain assumption The linear softmax readout on the discretized temporal voltage vector (Eq. 4-8) is an adequate model of what a hardware 'gated wirelet' synaptic layer would compute.
    The paper discusses on-chip training with gated wirelets, but the demonstrated training is software linear regression. The equivalence is only schematic (Fig. 4a).
  • ad hoc to paper The 'electrostatically gated output wirelet' can implement a time-dependent multiplicative weight over the aggregate input waveform as assumed.
    Section 'On-chip training' proposes this without experimental demonstration or simulation of the gated wirelet itself; it is a proposal, not a validated element.
invented entities (1)
  • Gated output wirelet synapse no independent evidence
    purpose: A superconducting filament with an electrostatic gate that modulates its local critical current to serve as a trainable temporal synaptic weight in a fully on-chip architecture.
    The authors cite works on field-effect control of superconducting channels, but do not demonstrate that such gating works for their NbTiN wirelets at MHz spiking rates nor that it can realize the time-dependent weights. It is a proposed extension.

pith-pipeline@v1.3.0-alltime-deepseek · 11553 in / 7964 out tokens · 62518 ms · 2026-08-02T23:12:32.517854+00:00 · methodology

0 comments
read the original abstract

Neuromorphic computing aims to reproduce the energy efficiency and adaptability of biological intelligence in hardware. Superconducting devices are an attractive platform due to their ultra-low dissipation and fast switching dynamics. Here we employ a resistively shunted superconducting wirelet as a minimal artificial neuron for temporal neuromorphic computation. This simple architecture enables straightforward fabrication, electronic control, and high scalability. Through experiments and advanced simulations, we show that it exhibits spiking voltage dynamics driven by the interplay of resistive switching and relaxation, with threshold, firing frequency, and refractory time tunable through applied current, temperature, and shunt resistance. We demonstrate neural network operation by synaptically training temporal voltage signals generated by individual neurons. In this approach, trainable temporal weights act directly on the time-dependent wirelet-neuron responses rather than on static neuron outputs alone, allowing the computation to exploit the full temporal structure of the superconducting spikes. As an illustrative example, we apply this framework to handwritten digit recognition and show accurate classification using only three superconducting wirelet neurons. We further discuss on-chip training based on related gated wirelets as tunable synaptic elements, establishing shunted superconducting wirelets as scalable, energy-efficient building blocks for cryogenic artificial intelligence hardware that can be integrated with other emerging superconducting technologies.

Figures

Figures reproduced from arXiv: 2602.14330 by Khalil Harrabi, Leonardo Cadorim, Milorad Milosevic.

Figure 1
Figure 1. Figure 1: Simulation insight into the resistive state of a shunted superconducting filament. a, The calculated current-voltage characteristics of a shunted nanowire (width 10ξ, schematic circuit in inset, with Lk = 500LGL and Rsh = 0.5RGL; see details in Methods), with four main phases labeled. b, The spiking temporal behavior of the output voltage, for increased applied current (shown on each curve, in units of jGL… view at source ↗
Figure 2
Figure 2. Figure 2: Experimental characterization of a spiking superconducting filament. Transport measurement on a 3µm-wide NbTiN filament shunted by 1Ω resistor, at T = 8K. a, The exper￾imental setup. b, The measured spiking behavior of the voltage, for increasing applied current (shown in mA). Curves are vertically displaced for visibility. c, The death of the superconducting neuron, corresponding to Fig. 1c,d. The behavio… view at source ↗
Figure 3
Figure 3. Figure 3: Control of spiking frequency. a, Spiking frequency of the superconducting filament from [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Proof-of-concept pattern recognition with three superconducting neurons. a, Schematic diagram of the circuit, where synaptic multiplexing was software-based. b, Simulated performance of three identical filaments (width 10ξ, and shunt resistances 0.5, 0.75, and 1RGL) in a neural network: (i) exemplified output voltages of each neuron corresponding to the same train of input current pulses for pixels 1-9. (i… view at source ↗
Figure 5
Figure 5. Figure 5: A complex task realized with same three-neuron network. a, Illustrative examples of hand-written number 8 from the MNIST database. b, The temporal voltage output of the three neurons for the input sequence of 420 image pixels of the image a(i). c, Output result after synaptic processing of 1000 test images of digit 8. d, The full input-output matrix on 1000 test images of each digit, demonstrating 92.9% ac… 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

34 extracted references · 1 linked inside Pith

  1. [1]

    & Das, A

    Goswami, S., Hazra, D., Patra, B. & Das, A. Superconductin g optoelectronic neurons for neuromorphic computing. Journal of Applied Physics 128, 050902 (2020)

    2020

  2. [2]

    M., Buckley, S

    Shainline, J. M., Buckley, S. M., Mirin, R. P . & Nam, S. W. Su perconducting optoelectronic circuits for neuromorphic computing. Physical Review Applied 12, 034022 (2019)

    2019

  3. [3]

    & Berggren, K

    Segall, K., Toomey, E., Basaran, A. & Berggren, K. K. Synch ronization and control of super- conducting nanowire neurons. Physical Review E 100, 062408 (2019)

    2019

  4. [4]

    M., Buckley, S

    Shainline, J. M., Buckley, S. M., Mirin, R. P . & Nam, S. W. Su perconducting optoelectronic neurons i: General principles. Journal of Applied Physics 124, 152130 (2018)

    2018

  5. [5]

    Carnevale, G., Suri, B., Gol’tsman, G. N. & Rybin, D. S. Sup erconducting neurons for large- scale neuromorphic networks. Superconductor Science and Technology 34, 094001 (2021)

    2021

  6. [6]

    Toomey, E., Berggren, K. K. & Segall, K. Operation of a supe rconducting nanowire as a spiking neuron. Physical Review Applied 11, 034056 (2019)

    2019

  7. [7]

    Ltspice (2023)

    Analog Devices, Inc. Ltspice (2023). URL https://www.analog.com/ltspice. Lin- ear Circuit Simulation Software

    2023

  8. [8]

    Izhikevich, E. M. Which model to use for cortical spiking n eurons? IEEE transactions on neural networks 15, 1063–1070 (2004). 23

    2004

  9. [9]

    & Schult, D

    Crotty, P ., Segall, K. & Schult, D. Biologically realisti c behaviors from a superconducting neuron model. IEEE Transactions on Applied Superconductivity 33, 1–6 (2023)

    2023

  10. [10]

    J., Beasley, M

    Skocpol, W. J., Beasley, M. R. & Tinkham, M. Self-heating hotspots in superconducting thin-film microbridges. Journal of Applied Physics 45, 4054–4066 (1974)

    1974

  11. [11]

    J., Y ang, J

    Kerman, A. J., Y ang, J. K. W., Molnar, R. J., Dauler, E. A. & Berggren, K. K. Electrothermal feedback in superconducting nanowire single-photon detectors. Physical Review B 79, 100509 (2009)

    2009

  12. [12]

    Zotova, A. N. & V odolazov, D. Y . Hotspot formation and relaxation in current-carrying super- conducting films. Physical Review B 85, 024509 (2012)

    2012

  13. [13]

    & Miloˇ sevi´ c, M

    Harrabi, K., Mekki, A. & Miloˇ sevi´ c, M. V . Characteristic times for gap relaxation and heat es- cape in nanothin nbti superconducting filaments: thickness dependence and effect of substrate. Nanomaterials 14, 1585 (2024)

    2024

  14. [14]

    Langer, J. S. & Ambegaokar, V . Intrinsic resistive trans ition in narrow superconducting chan- nels. Physical Review 164, 498–510 (1967)

    1967

  15. [15]

    McCumber, D. E. & Halperin, B. I. Time scale of intrinsic r esistive fluctuations in thin super- conducting wires. Physical Review B 1, 1054–1070 (1970)

    1970

  16. [16]

    Y ., Golubev, D

    Arutyunov, K. Y ., Golubev, D. S. & Zaikin, A. D. Superconductivity in one dimension. Physics Reports 464, 1–70 (2008). 24

    2008

  17. [17]

    E., Klenov, N

    Schegolev, A. E., Klenov, N. V ., Gubochkin, G. I., Kupriy anov, M. Y . & Soloviev, I. I. Bio- inspired design of superconducting spiking neuron and syna pse. Nanomaterials 13, 2101 (2023)

    2023

  18. [18]

    & Courville, A

    Goodfellow, I., Bengio, Y . & Courville, A. Deep Learning (MIT Press, 2016). URL https://www.deeplearningbook.org

    2016

  19. [19]

    & Haffner, P

    LeCun, Y ., Bottou, L., Bengio, Y . & Haffner, P . Gradient-based learning applied to document recognition. In Proceedings of the IEEE , vol. 86, 2278–2324 (1998)

    1998

  20. [20]

    & Hinton, G

    LeCun, Y ., Bengio, Y . & Hinton, G. Deep learning. Nature 521, 436–444 (2015)

    2015

  21. [21]

    Hestness, J. et al. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409 (2017). URL https://arxiv.org/abs/1712.00409

    Pith/arXiv arXiv 2017

  22. [22]

    Chaudhary, K. et al. Superconducting optoelectronic hardware for neuromorphi c computing. Nature Electronics 6, 686–698 (2023)

    2023

  23. [23]

    & Giazotto, F

    De Simoni, G., Paolucci, F., Solinas, P ., Strambini, E. & Giazotto, F. Metallic supercurrent field-effect transistor. Nature Nanotechnology 15, 801–806 (2020)

    2020

  24. [24]

    Paolucci, F. et al. Field-effect control of metallic superconducting channel s. Physical Review Applied 13, 024061 (2020)

    2020

  25. [25]

    & Bol, D

    Frenkel, C., Lefebvre, M., Legat, J.-D. & Bol, D. A 0.086- mm2 12.7-pj/sop 64k-synapse 256-neuron online-learning digital spiking neuromorphic processor in 28-nm cmos. IEEE Transactions on Biomedical Circuits and Systems 13, 145–158 (2019). 25

    2019

  26. [26]

    Davies, M. et al. Advancing neuromorphic computing with loihi: A survey of re sults and outlook. Proceedings of the IEEE 109, 911–934 (2021)

    2021

  27. [27]

    Pehle, C. et al. The brainscales-2 accelerated neuromorphic system with hy brid plasticity. Frontiers in Neuroscience 16, 795876 (2022)

    2022

  28. [28]

    B., Xiao, X., Proietti, R

    Lee, Y .-J., On, M. B., Xiao, X., Proietti, R. & Y oo, S. J. B. Photonic spiking neural networks with event-driven femtojoule optoelectronic neurons based on izhikevich-inspired model. Op- tics Express 30, 19360–19389 (2022)

    2022

  29. [29]

    Zeng, Y . et al. Photonic spiking neural network based on dml and dfb-sa lase r chip for pattern classification. Optics Express 33, 12045–12058 (2025). Reported energy efficiency of 0.625 pJ/MAC, implying sub-pJ per synaptic event

    2025

  30. [30]

    & Berggren, K

    Zhou, Z., Toomey, E., Segall, K. & Berggren, K. K. Behavio r of superconducting nanowire neurons with different geometries. IEEE Transactions on Applied Superconductivity 29, 1–5 (2019)

    2019

  31. [31]

    K., Toomey, E., Zhao, Q.-Y

    Berggren, K. K., Toomey, E., Zhao, Q.-Y . & Segall, K. Supe rconducting nanowire devices for neuromorphic computing. IEEE Transactions on Applied Superconductivity 28, 1–5 (2018)

    2018

  32. [32]

    Cheng, R., Berggren, K. K. & Zhao, Q.-Y . Flux dynamics in s uperconducting nanowires: Implications for neuromorphic circuits. Applied Physics Letters 111, 122602 (2017)

    2017

  33. [33]

    M., Simmonds, R

    Harrabi, K., Levenson-Falk, E. M., Simmonds, R. W. & Well stood, F. C. Dynamics of super- conducting josephson junctions under pulsed excitation. Applied Physics Letters 114, 052601 (2019). 26

    2019

  34. [34]

    & Geurts, R

    Miloˇ sevi´ c, M. & Geurts, R. The ginzburg–landau theoryin application. Physica C: Supercon- ductivity 470, 791–795 (2010). Acknowledgments K.H. gratefully acknowledges the support of the King Fahd Un iversity of Petroleum and Minerals, Saudi Arabia, under project ISP23212. L.R.C. and M.V .M. acknowledge support from the Research Foundation-Flanders (FWO...

    2010

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.