Pith. sign in

REVIEW 1 major objections 5 minor 25 references

CS-P2L shows that pairing covariance steering with the Pick-to-Learn loop yields a certified violation-probability bound for controllers evaluated on high-fidelity simulators, closing the surrogate–simulator gap.

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 ↗

CS-P2L couples covariance steering with Pick-to-Learn to synthesize and certify controllers against high-fidelity simulators, certifying a 4.9% violation bound on a spacecraft powered-descent problem.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Solid P2L instantiation for covariance steering with a real demonstration, but data-dependent early stopping in Alg. 3 undercuts the stated confidence; fixable and worth refereeing. the 1 major comments →

arxiv 2607.21086 v1 pith:VKJNLAMD submitted 2026-07-23 eess.SY cs.SY

Certified Stochastic Control via Covariance Steering with Pick-to-Learn

Chun-Wei Kong , Zachary Donovan , Morteza Lahijanian , Jay McMahon This is my paper
classification eess.SY cs.SY
keywords covariance steeringpick-to-learncertified controlprobabilistic guaranteessimulator-based synthesiscompression-based boundschance constraintsspacecraft powered descent
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 reading

The paper takes on a basic reliability problem in safety-critical control: policies optimised on a simplified surrogate model are usually validated by Monte Carlo simulation, which gives an empirical violation rate but no guarantee about the true rate. CS-P2L replaces that leap of faith with an iterative loop: it runs the current policy on a fixed set of simulator rollouts, picks the worst violating trajectory, tightens the surrogate's state, control, and terminal-covariance constraints accordingly, and re-solves a covariance-steering problem. The central result is a compression-based certificate: for any confidence delta, the returned policy's true violation probability is at most a bound that depends only on the number of worst-case rollouts selected, not on the total rollout count. In the paper's demonstration on a spacecraft powered-descent problem with uncertain gravity, the certified bound is 4.9% with 600 rollouts, while standalone covariance steering's nominal 5% chance constraint is empirically violated 9.6% of the time. If correct, the method turns any chance-constrained surrogate solver into a simulator-certified controller without convexity assumptions or state-space discretization.

Core claim

The discovery is that the Pick-to-Learn meta-algorithm can be instantiated with covariance steering as the inner solver, and that the resulting CS-P2L algorithm satisfies Theorem 1: with probability at least 1−δ over the random realizations used for synthesis and certification, the true violation probability p(π^(r)) is no larger than ε̄(|T^(r)|,δ,N), where ε̄ is defined by a binomial-tail equation. The proof works by identifying the algorithm as an instance of the P2L meta-algorithm, where the compression set is the set of violating realizations selected to reconfigure the surrogate. The practical significance is that the guarantee holds against the full nonlinear, non-Gaussian simulator wi

What carries the argument

The core mechanism is a closed loop between a high-fidelity simulator and a lightweight convex surrogate. The inner solver, covariance steering via successive convexification, linearizes the simulator around a reference trajectory and solves a chance-constrained problem over state means, covariances, feedforward controls, and feedback gains. The outer Pick-to-Learn loop selects the realization that violates the specification most severely, adds it to a compression set, and calls a configuration algorithm that tightens the surrogate's safe-set bounds, control bound, and terminal covariance. Theorem 1's bound ε̄(i,δ,N), the unique solution of a binomial equation, is what converts the size of t

Load-bearing premise

The load-bearing premise is that stopping the certification phase as soon as the computed bound falls below the target, using the very data that produced that bound, does not compromise the stated confidence 1−δ; without a sequential-testing correction, that early-stopping rule is the step most likely to invalidate the guarantee.

What would settle it

Take a simulator whose true violation probability can be computed analytically (e.g., a linear-Gaussian system with known distribution). Run the full CS-P2L procedure many times with fixed δ and N, and record the empirical frequency of the event p(π^(r)) > ε̄. If that frequency exceeds δ, Theorem 1 is false. For the tightening strategy, run Alg. 3 repeatedly with early stopping and check whether the empirical coverage of the event p(π) ≤ ε̄ at the stopping stage is at least 1−δ; systematic under-coverage would confirm the need for a multiple-comparison correction.

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

If this is right

  • Any policy returned by CS-P2L comes with a computable upper bound on its true violation probability, so a designer can certify safety without an analytical model of the full dynamics.
  • The certificate holds for nonlinear, non-Gaussian simulators with uncertain drift, because the bound is based purely on rollout outcomes.
  • The tightening procedure (Alg. 3) lets the user keep adding independent rollouts until the bound falls below a target, and each stage's guarantee is self-contained.
  • Standalone covariance steering's prescribed chance constraints can be badly optimistic (5% intended, 9.6% observed in the validation set); the P2L loop corrects for that surrogate–simulator mismatch.
  • The approach avoids state-space discretization and does not require convexity of the original problem, unlike scenario methods and neural supermartingale certificates.

Where Pith is reading between the lines

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

  • An inference the authors leave implicit: the same outer certification loop could be wrapped around other uncertainty-aware control solvers (the paper notes this in Remark 2); the only requirements are a configurable surrogate and a rule that tightens it in response to worst-case violations.
  • Because the bound depends on the compression-set size rather than the total number of rollouts, sample efficiency is governed by how many violating trajectories the final policy leaves; policies close to satisfying the specification certify with far fewer rollouts than naive Monte Carlo would need.
  • The reported 1−δ for the stage at which Alg. 3 stops assumes that data-dependent early stopping does not inflate error; a designer wanting a single confidence across the whole sequential algorithm should apply a multiple-comparison correction (δ/S), which the paper flags in a footnote.
  • The gap between the certified bound (0.049) and the validation-set violation rate (0.031) in the experiment suggests the bound is conservative; testing alternative violation measures and tightening schedules could reduce this conservatism.
Share X Bluesky LinkedIn Reddit HN

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

1 major / 5 minor

Summary. The paper proposes CS-P2L, a framework that couples covariance steering (CS) with the Pick-to-Learn (P2L) meta-algorithm to synthesize controllers for high-fidelity stochastic simulators while providing a compression-based probabilistic guarantee on the true violation probability. The authors formalize the simulator as an SDE with uncertain parameters, define a reach-avoid specification, and introduce a surrogate CS problem with configuration parameters. Algorithm 2 implements the P2L loop: it iteratively selects the most violating rollout, adds it to a compression set, updates the surrogate configuration, and re-solves the CS problem until all remaining rollouts satisfy the specification. Theorem 1 claims a 1−δ confidence bound on the violation probability of the returned policy, where the bound depends only on the compression set size, N, and δ. Section III-D and Algorithm 3 propose a two-phase tightening strategy with an optional calibration phase and a certification phase that stops as soon as the bound falls below a target ε. The experiment on a spacecraft powered-descent problem with uncertain gravity reports a certified violation bound of 0.049 with 600 rollouts and δ=10^-3, while the standalone CS controller exhibits an empirical violation rate of 0.096.

Significance. If the guarantee is valid, this is a meaningful contribution: it extends P2L—a general compression-based certification meta-algorithm—to a tractable inner solver (covariance steering) for continuous-time stochastic systems, avoiding the convexity requirements of the scenario approach and the discretization burden of neural certificate methods. The paper correctly separates the calibration data used to tune the configuration algorithm from the certification data, and the proof of Theorem 1 is a faithful identification of Algorithm 2 with P2L. The experimental demonstration on a powered-descent problem is compelling and shows that the CS-P2L controller reduces the empirical violation rate relative to standalone CS while providing a nontrivial certified bound. The main weakness lies in the early-stopping rule of Algorithm 3, which invalidates the stated confidence level without a multiple-comparison correction.

major comments (1)
  1. [Section IV] The experimental certification claim depends directly on the flawed early-stopping rule. The abstract and Section IV state that CS-P2L 'certifies a violation bound of 4.9% with 600 rollouts' at δ=10^-3. Since the stopping time is data-dependent, the reported confidence is not valid. The authors should either revise the algorithm to incorporate a multiple-comparison correction and recompute the reported bound, or explicitly state that the bound holds stagewise and that the returned controller's confidence is only (1−δ) after applying the correction. This is not merely a presentation issue; it affects the validity of the main numerical result.
minor comments (5)
  1. [Equation (15c)] The update rule for P_tf uses 'min' of matrices, but the intended order (Loewner or elementwise) is unspecified. Please clarify, as this affects the conic constraint in the surrogate CS problem.
  2. [Abstract and Section I] The phrase 'certifies a violation bound of 4.9%' should be qualified to indicate that this bound is contingent on the early-stopping rule being corrected (e.g., via a union bound). As written, it overstates the confidence for the stopping-time selection.
  3. [Section III-D, Footnote 2] The footnote correctly identifies the union-bound correction for 'selecting the best bound across S stages,' but does not mention the stopping-rule selection. Adding an explicit statement that early stopping requires the same correction would prevent the misinterpretation.
  4. [Proof of Lemma 4] The assertion that P^{1/2}Σ^{-1}P^{1/2} and Σ^{-1/2}PΣ^{-1/2} have the same nonzero eigenvalues is correct but should be justified in one sentence (e.g., by noting they are both congruent to P under appropriate transformations). This would improve readability.
  5. [Section IV] The notation ε̄ for the bound and ε for the target is standard, but the figure caption uses 'ε̄' while the text uses 'bound' and 'target'. Please ensure consistent notation in the figure and text.

Circularity Check

0 steps flagged

No significant circularity: the certified bound is a compression-set bound using |T|, N, and δ, with calibration data explicitly excluded from certification; the flagged early-stopping issue is a statistical confidence-selection concern, not a circular derivation.

full rationale

The central guarantee (Thm. 1) is an instantiation of compression-based generalization: the bound ε̄ is defined by Ψ_{i,δ}(ε)=1 in Eq. (12), depending only on the compression-set size |T|, the number of rollouts N, and the confidence parameter δ. It is not defined in terms of the fitted policy parameters, the tightening factors, or the empirical violation count. The adaptive construction of π^(r) from the same data is exactly the setting that compression bounds are designed to cover, so this is not a fitted-input-called-prediction pattern. In Sec. III-D the paper explicitly separates calibration from certification: the configuration algorithm L may be tuned on Z^(0), but 'the tuning batch Z^(0) (if used) is excluded from the certification data,' and the frozen L⋆ is fixed before Z^(1), Z^(2), ... are observed. Thus the tightening factors affect the certificate only by determining which rollouts enter T, and that dependence is accounted for by counting |T| in the compression bound. Self-citations appear only in related work (neural certificates, ref. [11]) and are not load-bearing. The one substantive concern is the early-stopping rule in Alg. 3: the paper states 'Each stage's guarantee is self-contained; one may therefore stop as soon as ε̄≤ε,' while the theorem is per fixed stage. The stopping time τ depends on the same data that define the returned policy and bound, and footnote 2 only applies a union-bound correction for 'selecting the best bound across S stages,' not for stopping at the first acceptable stage. Without a δ→δ/S or alpha-spending correction, the 1−δ confidence for the returned controller is not justified. This is a multiple-comparison/selection validity issue, not circularity: Eq. (12) is not equal to its inputs by construction, and no load-bearing claim is reduced to a self-citation or to the empirical fit. Therefore no circular step is exhibited, and the appropriate score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The central claim rests on external compression-based generalization theory (P2L) and standard SDE/convex-optimization machinery. The free parameters are algorithmic hyperparameters that influence tightness and feasibility but not the validity of the certificate itself. No new physical entities are introduced.

free parameters (3)
  • Tightening factors γ_b, γ_u, γ_P and cap γ̄_b = not reported
    Hand-chosen hyperparameters of the configuration algorithm (Eqs. 15a-15c). They affect which trajectories join the compression set and hence the bound's tightness, but not its validity.
  • Feasibility boundary terms b_min_m, u_min_max, P_max_tf = not reported
    Determined 'in practice by running Alg. 1 before Alg. 2' (Sec. III-C) to keep the tightened surrogate feasible. If computed from certification data, the independence requirement for Thm. 1 would be violated.
  • CS solver weights (w_u=1.0, w_vc=100, w_tr=0.1) and tolerances δ_vc=δ_tr=10^-6 = given in App. C
    Fixed hyperparameters of the successive convexification inner solver; they do not enter the probabilistic guarantee directly.
axioms (3)
  • standard math The SDE in Def. 1 admits a unique strong solution on [t0, tf]
    Invoked in Sec. III-A via Lipschitz conditions on f,g over a compact set excluding the singular set; standard SDE theory [16].
  • domain assumption P2L compression-based generalization theorem (ref [13, Thm. 1]) is valid for the synthesis map defined by Alg. 2
    The proof of Thm. 1 is an identification argument; the paper does not independently prove the compression bound. If the theorem's conditions (deterministic synthesis map from compression set, fixed selection rule) fail, the guarantee breaks.
  • domain assumption The discrete-time noise after linearization/discretization is standard Gaussian, justifying Lemmas 2-4
    The surrogate uses a Gaussian noise assumption for chance constraints; this is an approximation for the nonlinear simulator with Brownian noise and zero-order-hold control.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Certified Stochastic Control via Covariance Steering with Pick-to-Learn." pith.science (2026). https://pith.science/paper/VKJNLAMD

@misc{pith2026260721086,
  author       = {Pith},
  title        = {Pith review of: Certified Stochastic Control via Covariance Steering with Pick-to-Learn},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VKJNLAMD}},
  note         = {Machine review of arXiv:2607.21086}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We present CS-P2L, a framework coupling covariance steering (CS) with the Pick-to-Learn (P2L) meta-algorithm for certified controller synthesis over high-fidelity stochastic simulators. The method iteratively evaluates policies on simulator rollouts, tightens surrogate constraints using the worst-case violations, and provides compression-based probabilistic guarantees on the true violation probability given a confidence level. On a spacecraft powered-descent problem with uncertain gravity, CS-P2L certifies a violation bound of 4.9\% with 600 rollouts, whereas standalone covariance steering underestimates the violation rate by roughly a factor of two.

Figures

Figures reproduced from arXiv: 2607.21086 by Chun-Wei Kong, Jay McMahon, Morteza Lahijanian, Zachary Donovan.

Figure 2
Figure 2. Figure 2: Position trajectories of πCS and πCS−P2L on 1000 val￾idation realizations; the violating trajectories are highlighted in the zoomed view; note that no trajectories of πCS−P2L violate the state constraints. 0.0 0.5 1.0 1.5 2.0 2.5 time [-] 0.6 0.8 1.0 u 2 [ - ] CS mean ±3 CS P2L mean ±3 umax [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Control (norm) trajectories of πCS and πCS−P2L on 1000 validation realizations. 1.00 1.05 1.10 1.15 x [-] 0.075 0.050 0.025 0.000 0.025 0.050 0.075 y [-] 0.2 0.0 0.2 0.4 0.6 vx [-] 0.2 0.0 0.2 0.4 0.6 vy [-] CS CS P2L terminal set violating [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Terminal samples of πCS and πCS−P2L on 1000 validation realizations. Σ −1/2 tf Ptf Σ −1/2 tf ⪯ αInx . Moreover, P 1/2 tf Σ −1 tf P 1/2 tf and Σ −1/2 tf Ptf Σ −1/2 tf have the same nonzero eigenvalues; since both matrices are symmetric positive semidefinite, it follows that P 1/2 tf Σ −1 tf P 1/2 tf ⪯ αInx . Substituting this bound into (16) yields (x(tf ) − µtf ) ⊤Σ −1 tf (x(tf ) − µtf ) ≤ α z⊤z. Therefore… 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

25 extracted references · 3 linked inside Pith

  1. [1]

    The scenario approach to robust control design,

    G. C. Calafiore and M. C. Campi, “The scenario approach to robust control design,”IEEE Transactions on automatic control, vol. 51, no. 5, pp. 742–753, 2006

    2006

  2. [2]

    Au- tomatic scenario generation for efficient solution of robust optimal control problems,

    M. Zagorowska, P. Falugi, E. O’Dwyer, and E. C. Kerrigan, “Au- tomatic scenario generation for efficient solution of robust optimal control problems,”International Journal of Robust and Nonlinear Control, vol. 34, no. 2, pp. 1370–1396, 2024

    2024

  3. [3]

    Finite-horizon covariance control of linear time-varying systems,

    M. Goldshtein and P. Tsiotras, “Finite-horizon covariance control of linear time-varying systems,” in2017 IEEE 56th Annual Conference on Decision and Control (CDC). IEEE, 2017, pp. 3606–3611

    2017

  4. [4]

    Optimal stochastic vehicle path planning using covariance steering,

    K. Okamoto and P. Tsiotras, “Optimal stochastic vehicle path planning using covariance steering,”IEEE Robotics and Automation Letters, vol. 4, no. 3, pp. 2276–2281, 2019

    2019

  5. [5]

    Nonlinear uncertainty con- trol with iterative covariance steering,

    J. Ridderhof, K. Okamoto, and P. Tsiotras, “Nonlinear uncertainty con- trol with iterative covariance steering,” in2019 IEEE 58th Conference on Decision and Control (CDC). IEEE, 2019, pp. 3484–3490

    2019

  6. [6]

    Chance-constrained covariance steering in a gaussian random field via successive convex programming,

    J. Ridderhof and P. Tsiotras, “Chance-constrained covariance steering in a gaussian random field via successive convex programming,” Journal of Guidance, Control, and Dynamics, vol. 45, no. 4, pp. 599– 610, 2022

    2022

  7. [7]

    Pac-bayes control: learning policies that provably generalize to novel environments,

    A. Majumdar, A. Farid, and A. Sonar, “Pac-bayes control: learning policies that provably generalize to novel environments,”International Journal of Robotics Research, vol. 40, no. 2–3, pp. 574–593, 2021

    2021

  8. [8]

    Sim-to-lab- to-real: Safe reinforcement learning with shielding and generalization guarantees,

    K.-C. Hsu, V . Rubies-Royo, C. J. Tomlin, and J. F. Fisac, “Sim-to-lab- to-real: Safe reinforcement learning with shielding and generalization guarantees,”Artificial Intelligence, vol. 314, p. 103811, 2023

    2023

  9. [9]

    A pac-bayesian framework for optimal control with stability guarantees,

    M. G. Boroujeni, C. L. Galimberti, A. Krause, and G. Ferrari- Trecate, “A pac-bayesian framework for optimal control with stability guarantees,” in2024 IEEE 63rd Conference on Decision and Control (CDC). IEEE, 2024, pp. 8237–8244

    2024

  10. [10]

    Neural continuous-time supermartingale certificates,

    G. Neustroev, M. Giacobbe, and A. Lukina, “Neural continuous-time supermartingale certificates,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 26, 2025, pp. 27 538–27 546

    2025

  11. [11]

    Training with hard constraints: Learning neural certificates and con- trollers for sdes,

    C.-W. Kong, S. Escobar, I. Gracia, J. McMahon, and M. Lahijanian, “Training with hard constraints: Learning neural certificates and con- trollers for sdes,”arXiv preprint arXiv:2602.23526, 2026

    Pith/arXiv arXiv 2026

  12. [12]

    The pick-to-learn algo- rithm: Empowering compression for tight generalization bounds and improved post-training performance,

    D. Paccagnan, M. Campi, and S. Garatti, “The pick-to-learn algo- rithm: Empowering compression for tight generalization bounds and improved post-training performance,”Advances in Neural Information Processing Systems, vol. 36, pp. 18 165–18 185, 2023

    2023

  13. [13]

    Pick-to-learn for systems and control: Data-driven synthesis with state-of-the-art safety guarantees,

    D. Paccagnan, D. Marks, M. C. Campi, and S. Garatti, “Pick-to-learn for systems and control: Data-driven synthesis with state-of-the-art safety guarantees,”arXiv preprint arXiv:2512.04781, 2025

    arXiv 2025

  14. [14]

    Tutorial on practical prediction theory for classification

    J. Langford and R. Schapire, “Tutorial on practical prediction theory for classification.”Journal of machine learning research, vol. 6, no. 3, 2005

    2005

  15. [15]

    Compression, generalization and learn- ing,

    M. C. Campi and S. Garatti, “Compression, generalization and learn- ing,”Journal of Machine Learning Research, vol. 24, no. 339, pp. 1–74, 2023

    2023

  16. [16]

    Oksendal,Stochastic differential equations (3rd ed.): an introduc- tion with applications

    B. Oksendal,Stochastic differential equations (3rd ed.): an introduc- tion with applications. Berlin, Heidelberg: Springer-Verlag, 1992

    1992

  17. [17]

    Successive convexification: A superlinearly convergent algorithm for non- convex optimal control problems,

    Y . Mao, M. Szmuk, X. Xu, and B. Acikmese, “Successive convexification: A superlinearly convergent algorithm for non- convex optimal control problems,” 2019. [Online]. Available: https://arxiv.org/abs/1804.06539

    Pith/arXiv arXiv 2019

  18. [18]

    Convex optimization for trajectory generation: A tutorial on generating dynamically feasible trajectories reliably and efficiently,

    D. Malyuta, T. P. Reynolds, M. Szmuk, T. Lew, R. Bonalli, M. Pavone, and B. Ac ¸ıkmes ¸e, “Convex optimization for trajectory generation: A tutorial on generating dynamically feasible trajectories reliably and efficiently,”IEEE Control Systems, vol. 42, no. 5, p. 40–113, Oct. 2022

    2022

  19. [19]

    Optimal covariance steering for discrete-time linear stochastic systems,

    F. Liu, G. Rapakoulias, and P. Tsiotras, “Optimal covariance steering for discrete-time linear stochastic systems,”IEEE Transactions on Automatic Control, no. arXiv:2211.00618, Oct. 2024, arXiv:2211.00618 [eess]. [Online]. Available: http://arxiv.org/abs/2211.00618

    Pith/arXiv arXiv 2024

  20. [20]

    Variations and extension of the convex–concave procedure,

    T. Lipp and S. Boyd, “Variations and extension of the convex–concave procedure,”Optimization and Engineering, vol. 17, no. 2, pp. 263–287, Jun. 2016

    2016

  21. [21]

    The concave-convex procedure,

    A. L. Yuille and A. Rangarajan, “The concave-convex procedure,” Neural Computation, vol. 15, no. 4, pp. 915–936, Apr. 2003

    2003

  22. [22]

    Computationally efficient chance con- strained covariance control with output feedback,

    J. Pilipovsky and P. Tsiotras, “Computationally efficient chance con- strained covariance control with output feedback,” in2024 IEEE 63rd Conference on Decision and Control (CDC), 2024, pp. 677–682

    2024

  23. [23]

    Chance-constrained control for safe spacecraft autonomy: Convex programming approach,

    K. Oguri, “Chance-constrained control for safe spacecraft autonomy: Convex programming approach,” in2024 American Control Confer- ence (ACC), 2024, pp. 2318–2324

    2024

  24. [24]

    Robust cislunar low-thrust trajectory optimization under uncertainties via sequential covariance steering,

    N. Kumagai and K. Oguri, “Robust cislunar low-thrust trajectory optimization under uncertainties via sequential covariance steering,” Journal of Guidance, Control, and Dynamics, p. 1–19, Aug. 2025

    2025

  25. [25]

    Successive convexifica- tion for real-time 6-dof powered descent guidance with state-triggered constraints,

    M. Szmuk, T. P. Reynolds, and B. Acikmese, “Successive convexifica- tion for real-time 6-dof powered descent guidance with state-triggered constraints,”Journal of Guidance, Control, and Dynamics, vol. 43, no. 8, p. 1399–1413, Aug. 2020

    2020

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.