GenControl: Generative AI-Driven Autonomous Design of Control Algorithms

Chenggang Cui; Chuanlin Zhang; Jiaming Liu; Peifeng Hui; Pengfeng Lin

REVIEW 5 major objections 5 minor 16 references

GenControl: Generative AI-Driven Autonomous Design of Control Algorithms

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read GenControl's two-level loop of LLM structural edits and PSO parameter tuning autonomously evolves a standard sliding-mode controller into an adaptive one that meets all stated performance specifications.

desk verdict Timely framework, but the validation is missing: the final controller is never shown, so the central claim rests on a narrative. read the letter →

arxiv 2506.12554 v3 pith:RODYSWYW submitted 2025-06-14 eess.SY cs.SY

Chenggang Cui , Jiaming Liu , Peifeng Hui , Pengfeng Lin , Chuanlin Zhang This is my paper

classification eess.SYcs.SY

keywords AutonomousDesignLargeLanguageModel(LLM)ControlSystemsBi-LevelOptimizationParticleSwarm(PSO)IndustrialElectronicsPowerAlgorithmSynthesis

verification ladder T0 review T1 audit T2 compute T3 formal

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 proposes GenControl, a framework that treats control-algorithm design as a joint search over controller structure and parameters, and automates that search with two cooperating levels. An LLM acts as the upper level, reading performance feedback from simulations and proposing structural changes to the controller; particle swarm optimization acts as the lower level, tuning the parameters of whatever structure is proposed. The central claim is that this loop can start from a standard sliding-mode controller template and evolve it into an adaptive sliding-mode controller that meets all stated specifications—overshoot under 5 percent and steady-state error under 2 percent—on a simulated DC-DC boost converter. If correct, the framework offers an end-to-end alternative to the slow, expert-dependent cycle of modeling, tuning, and redesign that dominates power-electronics controller design.

What carries the argument

The load-bearing object is the bi-level optimization loop. At the upper level, the LLM implements a policy $\pi_{\text{LLM}}$ that takes the current controller, the performance specifications, the plant model, historical feedback, and a prompt template, and outputs a structural modification $\Delta S$ or a new structure. At the lower level, PSO solves $\theta_k^* = \arg\min_{\theta \in \Theta_{S_k}} J(\{S_k,\theta\})$ for each proposed structure, so every structure is judged at its best achievable parameters. Performance feedback $J(p_k)$ from closed-loop simulation closes the loop, letting the LLM reason about which part of the controller is causing a shortfall and what structural change should fix it. The framework also provides an implicit formal representation of control laws and a primitive library to make LLM-generated structures executable.

What would settle it

Run the same design loop with the LLM replaced by a random or fixed-rule structural mutator while keeping PSO intact; if the resulting controller meets the same specifications in a comparable number of iterations, then the LLM is not the source of the improvement. Equivalently, test whether PSO alone, restricted to the initial standard SMC structure, already meets all design specifications—if it does, the structural evolution is unnecessary.

Watch

Extended reading notes

Core claim

The paper's central discovery claim is that an LLM, guided only by a prompt template and numerical performance feedback, can act as an intelligent structural optimizer: it diagnoses deficiencies such as chattering and steady-state error, then proposes targeted structural modifications such as adding a boundary layer or an adaptive law. Combined with PSO parameter refinement, this iterative bi-level loop converged to an adaptive sliding-mode controller whose simulated response met the design specifications of less than 5 percent peak overshoot and less than 2 V steady-state error under load steps. The paper states that this outcome validates the framework's capability to autonomously evolve a basic controller into an advanced, high-performance solution for a complex control problem entirely within a simulation environment.

Load-bearing premise

The load-bearing premise is that the LLM's structural modifications, not PSO's parameter tuning, are what turn the basic template into a high-performance controller, yet the paper gives no ablation separating the two.

Editorial extensions

If this is right

Each candidate structure is evaluated only after its parameters are optimized, so comparisons between structures reflect structural merit rather than poor tuning.
The same loop can be pointed at a different plant, specification set, or prompt template without re-engineering the optimization machinery, supporting the paper's claim of end-to-end automation.
Because the LLM can propose non-traditional structures, the framework can escape local optima in the structural space and potentially find designs a human engineer would not try.
Targeted structural fixes—like adding an integral term for steady-state error or a boundary layer for chattering—emerge automatically from performance feedback rather than from a human diagnosis.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural ablation not reported in the paper would replace the LLM with random or fixed-rule structural mutations under the same PSO loop; if the resulting controller meets the same specifications in a comparable number of iterations, the LLM's reasoning is not load-bearing.
The single boost-converter case leaves generalization open; a direct extension would be to run the same framework on a buck converter, motor drive, or grid-tied inverter and compare iteration counts and final performance indices.
The claim of semantic understanding could be probed by feeding misleading performance feedback—for example, attributing a slow transient to steady-state error—and checking whether the LLM's proposed structural change follows the false diagnosis or the true simulation data.
If the framework heads toward safety-critical hardware, the generated structures will need formal verification, which the paper itself lists as future work and therefore does not currently provide.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Chenggang and colleagues have a good instinct here. The idea of using an LLM to propose structural modifications to a control law, with PSO handling parameter fitting underneath, is a natural and timely combination. The paper does a decent job of framing the problem as a bi-level optimization and of describing the LLM's role as a policy that maps performance feedback to structural actions. I also appreciate the explicit primitive library and the feedback loop; the architecture is plausible and the prompt-design discussion is useful.

What the paper does not do is show any evidence that this loop actually works. The only closed-loop plot is for the initial, unoptimized standard SMC. The final adaptive SMC is described but never simulated. The convergence figure is captioned as a "conceptual illustration", which is an honest label but means it is not data. There are no numerical metrics, no overshoot/steady-state error values, no error bars. The abstract and conclusion claim hardware validation, but Section VI describes only simulation; no hardware results appear anywhere. The manuscript also has unresolved cross-references and a leftover note ("assuming Section V is relabeled"). On top of that, the central causal claim — that the LLM's structural changes, not PSO parameter tuning, caused the improvement — has no ablation or baseline to support it. A PSO-tuned standard SMC or a boundary-layer parameter added to the primitive library might meet the specs equally well. As written, the paper is a framework proposal plus a success story that is told, not shown.

That said, I don't want to overstate the problems. The circularity concern (optimizing and evaluating with the same J) is real but not fatal for a design-loop paper; what matters is independent validation, which is absent here. The novelty is modest — the pieces are all from the LLM-agent literature — but the application to power electronics controller structure synthesis is new and worth exploring.

Who is this for? Researchers working on LLM-driven control design. They would get a clear architecture and a set of open questions about validation. It deserves a serious referee in the sense that a rigorous revision could turn it into a legitimate contribution. But the current preprint is not close to acceptable. My recommendation: send it back for major revision, and require the final controller results, a baseline without the LLM (or with a random structure generator), and either hardware or a clear retraction of the hardware claim.

Referee Report

5 major / 5 minor

Summary. This paper proposes GenControl, a bi-level optimization framework for autonomous control algorithm design. The upper level uses an LLM to propose or modify control-law structures, and the lower level uses PSO to tune the parameters of each proposed structure; the process iterates on a performance index computed from simulation. The method is demonstrated on a DC-DC Boost converter case study, where the authors claim the framework evolves a standard sliding-mode controller into an adaptive SMC that meets specifications of <5% overshoot and <2% steady-state error. The abstract and conclusion also claim hardware-experiment validation, although no hardware results appear in the body.

Significance. The problem addressed is relevant and timely: automating control design for power electronics could reduce human effort and enable novel structures. The bi-level formulation and the idea of LLM-driven structural search are potentially useful. However, the manuscript currently does not provide the evidence needed to support its central claim: the final optimized controller is never shown, the only convergence figure is explicitly labeled 'conceptual illustration', and there is no baseline or ablation to separate the contributions of the LLM from PSO. These are foundational gaps, not presentation issues.

major comments (5)

[§VI-C, Fig. 3] The central validation claim rests on the final Adaptive SMC, but no simulation results, performance numbers, or comparison to G_spec are reported for it. Figure 3 only shows the initial standard SMC, and the text asserts 'excellent performance' without quantitative support. The manuscript needs the final controller's output-voltage and control-signal traces, plus numeric overshoot, steady-state error, settling time, and chattering measures.
[§VI-B, Fig. 2] Figure 2 is captioned 'Conceptual illustration of the objective function J(p) convergence during the LLM-driven iterative optimization process.' A conceptual illustration is not data. No measured J(p_k) values, iteration counts, or termination criteria are given, so the claimed decreasing trend of the objective function is not demonstrated.
[§II-C and §VI-C] Validation is circular in the current design: the performance index J(p) in Eq. (3) incorporates the same metrics (overshoot, steady-state error) as the design specifications, and the load-step simulation used for optimization is also the scenario used to claim specification satisfaction. The final controller's compliance with the specs is therefore a fitted outcome. An independent test (e.g., different operating points, parameter uncertainty, or a holdout disturbance profile) is required, and the claimed 'robustness' is neither defined nor tested.
[§IV-C and §VI-C] The distinctive contribution is the LLM's structural evolution, yet no baseline or ablation is provided to show that this evolution matters. The initial standard SMC, with parameters optimized by PSO (and possibly with a boundary-layer term added to the primitive library), might already meet the specifications. Without a parameter-only PSO baseline or random-structure-mutation control, the causal role of the LLM is assumed.
[Abstract and §VII-A vs §VII-B] The abstract and Section VII-A state that the case study was 'validated by both simulation and hardware experiments,' but the manuscript contains no hardware setup, results, or discussion. Section VII-B instead lists hardware-in-the-loop and real-world experiments as future work. This internal inconsistency must be resolved before the paper can be evaluated.

minor comments (5)

[§III, §IV, §IV-C] Multiple cross-references appear as unresolved 'Section ??'; these need to be fixed before review.
[§IV-B] The parenthetical 'assuming Section V is relabeled' indicates an internal planning note that should not appear in a submitted manuscript.
[§V-D] Notation is inconsistent: the prompt template is called T_prompt in Sections II–IV but T_template in Section V-D, and the controller structure is denoted S in the problem formulation but C in Section V-D.
[§VI-A] The plant model equations and the exact simulation environment (tool, solver, switching model) are not provided, which hinders reproducibility of the simulated results.
[§VI] The LLM version/model and the iterative prompt structure are not specified, so the experiments cannot be reproduced or assessed for prompt-sensitivity.

Circularity Check

1 steps flagged · score 6.0 of 10

Partial circularity: the central 'meets all specifications' result is the same performance index J that PSO was optimized against, so the success claim reduces to the optimization objective by construction.

fitted input called prediction [Section VI-B/VI-C, with Eq. (3) (Section II-C) and Eq. (5) (Section IV-C)]
"For each candidate control algorithm structure Sk supplied by the upper-level LLM, the objective of the lower-level optimization is to identify the parameter vector θ∗k ... that optimizes the performance index J(p). ... θ∗k = arg minθ∈ΘSk J({Sk, θ}) (5). ... The simulation results of the final optimized Adaptive SMC demonstrate its excellent performance. It achieved minimal steady-state voltage error and exhibited a fast, stable dynamic response with small overshoot under load step disturbances, fully meeting all predefined design objectives."

The performance index J(p) is defined in Eq. (3) as a weighted sum of individual performance metrics Mi(p), and the design specifications Gspec are stated in Section VI-A as overshoot < 5% and steady-state error < 2%. The lower-level PSO explicitly minimizes J for each structure, so the final controller is selected precisely to make J small on the same simulation scenario. The concluding claim that the final Adaptive SMC 'fully meeting all predefined design objectives' is therefore not an independent prediction or validation; it is a restatement of having optimized against the same objective and the same load-step simulation.

full rationale

The paper's central derivation is a bi-level optimization loop in which the LLM proposes structures and PSO tunes parameters to minimize J(p), defined in Eq. (3) from the same metrics (overshoot, steady-state error) that constitute Gspec. The reported success, that the final Adaptive SMC meets all predefined design objectives, is thus a fitted outcome of minimizing J on the same simulation scenario, not an independent prediction. This is the main circularity. There is no load-bearing self-citation: references such as [10], [11], and [16] are background and do not carry the derivation. However, the manuscript also contains unresolved cross-references ('Section ??', 'assuming Section V is relabeled'), a 'conceptual illustration' convergence plot (Fig. 2) instead of measured iterations, no final controller structure or final simulation curves, and a Section VII claim of 'both simulation and hardware experimental validation' despite no hardware results appearing anywhere. These deficiencies corroborate that the claimed autonomous evolution is a narrative assertion rather than a demonstrated, independently verified result. Because the central validation reduces, at least in part, to the optimization objective itself, the circularity score is 6.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced. The free parameters are the performance weights, the PSO-tuned controller parameters, the unspecified PSO hyperparameters, and the stopping criteria; all are fitted or chosen by hand and none are reported numerically. The axioms are the unverified background assumptions the central claim depends on: simulation fidelity, the causal benefit of LLM structural changes, PSO convergence, the use of the same simulation for optimization and validation, and the sufficiency of LLM control knowledge.

free parameters (4)

Performance index weights w_i in J(p) = Not reported
Eq. (3) defines J as a weighted sum, but the case study never states the weights. These weights determine what 'meets specifications' means and are chosen by the authors, not derived from first principles.
Adaptive SMC controller parameters (boundary layer thickness, adaptive gains) = Not reported
These are tuned by PSO against J and are central to the reported performance, but the final numerical values are never given, so the result cannot be audited or reproduced.
PSO hyperparameters (swarm size, number of iterations, inertia, cognitive and social coefficients) = Not reported
PSO convergence and the claimed improvement depend on these settings. None are stated, so the lower-level optimization is not reproducible.
Convergence criteria and design budget Kmax = Not reported
The stopping rule for the iterative loop controls when the 'optimal' controller is declared final. Without it, the claim that the framework converged to a valid controller is underdetermined.

assumptions (5)

domain assumption The simulation model Mplant faithfully represents the physical boost converter.
All evaluations and the final claim rely on the simulation. No model uncertainty analysis is given, and the hardware experiments mentioned in the conclusion are not presented.
ad hoc to paper The LLM's structural modifications are beneficial and not reducible to random mutations or PSO tuning alone.
No ablation or baseline is provided to show that the LLM's structural choices cause the improvement. This assumption is load-bearing for the central claim of autonomous design.
domain assumption PSO converges to an optimal or near-optimal parameter vector for each structure within the allowed budget.
Eq. (5) assumes the lower-level problem is solved, but PSO has no convergence guarantee and no PSO results or convergence diagnostics are shown.
ad hoc to paper The same simulation and performance index used for optimization is a valid independent test of the design specifications.
Optimization and validation share the same J and the same load-step scenario, so meeting the specification targets is partly by construction. No separate test set or robustness scenario is described.
domain assumption The LLM's internal control-theory knowledge is sufficient to propose valid, stable controller structures.
The upper-level policy pi_LLM relies on pretrained knowledge to generate structures. No stability or realizability verification is provided beyond the evaluation loop, which uses the same model that generated the feedback.

how reviews work

0 comments

Cite this review

Pith. "Pith review of GenControl: Generative AI-Driven Autonomous Design of Control Algorithms." pith.science (2026). https://pith.science/paper/RODYSWYW

@misc{pith2026250612554,
  author       = {Pith},
  title        = {Pith review of: GenControl: Generative AI-Driven Autonomous Design of Control Algorithms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RODYSWYW}},
  note         = {Machine review of arXiv:2506.12554}
}

read the original abstract

Designing controllers for complex industrial electronic systems is challenging due to nonlinearities and parameter uncertainties, and traditional methods are often slow and costly. To address this, we propose a novel autonomous design framework driven by Large Language Models (LLMs). Our approach employs a bi-level optimization strategy: an LLM intelligently explores and iteratively improves the control algorithm's structure, while a Particle Swarm Optimization (PSO) algorithm efficiently refines the parameters for any given structure. This method achieves end-to-end automated design. Validated through a simulation of a DC-DC Boost converter, our framework successfully evolved a basic controller into a high-performance adaptive version that met all stringent design specifications for fast response, low error, and robustness. This work presents a new paradigm for control design that significantly enhances automation and efficiency.

Figures

Figures reproduced from arXiv: 2506.12554 by the authors.

**Figure 2.** Conceptual illustration of the objective function [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Simulation results of the initial standard SMC controller, showing [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 12 canonical work pages

[1]

Harmonic stability in power electronic-based power systems: Concept, modeling, and analysis,

X. Wang and F. Blaabjerg, “Harmonic stability in power electronic-based power systems: Concept, modeling, and analysis,” IEEE Transactions on Smart Grid, vol. 10, no. 3, pp. 2858–2870, 2019

work page 2019
[2]

Buso and P

S. Buso and P. Mattavelli, Digital Control in Power Electronics. Cham, Switzerland: Morgan & Claypool Publishers, 2022

work page 2022
[3]

Overview of power con- verter control in microgrids—Challenges, advances, and future trends,

J. Hu, Y . Shan, K. W. E. Cheng, and S. Islam, “Overview of power con- verter control in microgrids—Challenges, advances, and future trends,” IEEE Transactions on Power Electronics, vol. 37, no. 8, pp. 9907–9922, 2022

work page 2022
[4]

Advanced control of power electronic systems – an overview of methods,

C. Wu and F. Blaabjerg, “Advanced control of power electronic systems – an overview of methods,” in Control of Power Electronic Converters and Systems. San Diego, CA, USA: Academic Press, 2021, pp. 1–33

work page 2021
[5]

Controller design automa- tion for power electronics: A model-free approach,

W. Li, S. Li, H. Yuan, Y . Zhang, and J. Zhu, “Controller design automa- tion for power electronics: A model-free approach,” IEEE Transactions on Power Electronics, vol. 39, no. 2, pp. 2155–2168, 2024

work page 2024
[6]

Model free predictive control of motor drives and power converters: A review,

M. Khalilzadeh, S. Vaez-Zadeh, J. Rodriguez, and R. Heydari, “Model free predictive control of motor drives and power converters: A review,” IEEE Access, vol. 9, pp. 105 733–105 747, 2021

work page 2021
[7]

Power electronics design methods and au- tomation in the digital era: Evolution of design automation tools,

A. J. Marques Cardoso, “Power electronics design methods and au- tomation in the digital era: Evolution of design automation tools,” IEEE Power Electronics Magazine, vol. 7, no. 2, pp. 36–40, 2020

work page 2020
[8]

Artificial intelligence in the hierarchical control of AC, DC, and hybrid AC/DC microgrids: A review,

B. Mohandes, H. E. G. Moussa, S. Kamel, A. T. Elsayed, E. E. Elattar, and E. H. Houssein, “Artificial intelligence in the hierarchical control of AC, DC, and hybrid AC/DC microgrids: A review,” Energies, vol. 17, no. 1, p. 266, 2024

work page 2024

Show all 16 references

[9]

A survey of large language models,

W. X. Zhao, K. Zhou, J. Li, T. Tang, X. Wang, Y . Hou, Y . Min, B. Zhang, J. Zhang, Z. Dong et al. , “A survey of large language models,” arXiv preprint arXiv:2303.18223, 2023

2023 arXiv
[10]

PE-GPT: A New Paradigm for Power Electronics Design,

F. Lin, X. Li, W. Lei, J. J. Rodriguez-Andina, J. M. Guerrero, C. Wen, X. Zhang, and H. Ma, “PE-GPT: A New Paradigm for Power Electronics Design,” IEEE Transactions on Industrial Electronics , 2024

2024
[11]

MetaGPT: Meta programming for a multi-agent collaborative framework,

S. Hong, M. Zhuge, J. Chen, X. Zheng, Y . Cheng, C. Zhang, J. Wang, Z. Wang, S. K. S. Yau, Z. Lin et al. , “MetaGPT: Meta programming for a multi-agent collaborative framework,” in Proceedings of the International Conference on Learning Representations (ICLR) , 2024

2024
[12]

AutoGen: Enabling next-gen LLM applications via multi- agent conversation framework,

Q. Wu, G. Bansal, J. Zhang, Y . Wu, S. Zhang, E. Zhu, B. Wang, L. Li, and C. Wang, “AutoGen: Enabling next-gen LLM applications via multi- agent conversation framework,” arXiv preprint arXiv:2308.08155, 2023

2023 arXiv
[13]

A survey on large language model based autonomous agents,

L. Wang, C. Ma, X. Feng, Z. Zhang, H. Yang, J. Zhang, Z. Chen, J. Tang, X. Chen, Y . Luo et al. , “A survey on large language model based autonomous agents,” arXiv preprint arXiv:2308.11432 , 2023

2023 arXiv
[14]

Wooldridge, An Introduction to MultiAgent Systems, 2nd ed

M. Wooldridge, An Introduction to MultiAgent Systems, 2nd ed. Chich- ester, UK: John Wiley & Sons, 2009

2009
[15]

Multiagent systems,

K. P. Sycara, “Multiagent systems,” AI Magazine , vol. 19, no. 2, pp. 79–92, 1998

1998
[16]

Exploring large language model based intelligent agents: Definitions, methods, and prospects,

Y . Cheng, C. Zhang, Z. Zhang, X. Meng, S. Hong, W. Li, Z. Wang, Z. Wang, F. Yin, J. Zhao et al., “Exploring large language model based intelligent agents: Definitions, methods, and prospects,” arXiv preprint arXiv:2401.03428, 2024. 9

2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Harmonic stability in power electronic-based power systems: Concept, modeling, and analysis,

X. Wang and F. Blaabjerg, “Harmonic stability in power electronic-based power systems: Concept, modeling, and analysis,” IEEE Transactions on Smart Grid, vol. 10, no. 3, pp. 2858–2870, 2019

work page 2019

[2] [2]

Buso and P

S. Buso and P. Mattavelli, Digital Control in Power Electronics. Cham, Switzerland: Morgan & Claypool Publishers, 2022

work page 2022

[3] [3]

Overview of power con- verter control in microgrids—Challenges, advances, and future trends,

J. Hu, Y . Shan, K. W. E. Cheng, and S. Islam, “Overview of power con- verter control in microgrids—Challenges, advances, and future trends,” IEEE Transactions on Power Electronics, vol. 37, no. 8, pp. 9907–9922, 2022

work page 2022

[4] [4]

Advanced control of power electronic systems – an overview of methods,

C. Wu and F. Blaabjerg, “Advanced control of power electronic systems – an overview of methods,” in Control of Power Electronic Converters and Systems. San Diego, CA, USA: Academic Press, 2021, pp. 1–33

work page 2021

[5] [5]

Controller design automa- tion for power electronics: A model-free approach,

W. Li, S. Li, H. Yuan, Y . Zhang, and J. Zhu, “Controller design automa- tion for power electronics: A model-free approach,” IEEE Transactions on Power Electronics, vol. 39, no. 2, pp. 2155–2168, 2024

work page 2024

[6] [6]

Model free predictive control of motor drives and power converters: A review,

M. Khalilzadeh, S. Vaez-Zadeh, J. Rodriguez, and R. Heydari, “Model free predictive control of motor drives and power converters: A review,” IEEE Access, vol. 9, pp. 105 733–105 747, 2021

work page 2021

[7] [7]

Power electronics design methods and au- tomation in the digital era: Evolution of design automation tools,

A. J. Marques Cardoso, “Power electronics design methods and au- tomation in the digital era: Evolution of design automation tools,” IEEE Power Electronics Magazine, vol. 7, no. 2, pp. 36–40, 2020

work page 2020

[8] [8]

Artificial intelligence in the hierarchical control of AC, DC, and hybrid AC/DC microgrids: A review,

B. Mohandes, H. E. G. Moussa, S. Kamel, A. T. Elsayed, E. E. Elattar, and E. H. Houssein, “Artificial intelligence in the hierarchical control of AC, DC, and hybrid AC/DC microgrids: A review,” Energies, vol. 17, no. 1, p. 266, 2024

work page 2024